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; 111*d62866d3SStefano 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) */ 112*d62866d3SStefano 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; 133*d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13408122e43SStefano Zampini const PetscInt *idxs; 13508122e43SStefano Zampini 136*d62866d3SStefano 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; 145*d62866d3SStefano 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; 518*d62866d3SStefano 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); 593*d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 594*d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 595*d62866d3SStefano Zampini MatFactorType type; 596*d62866d3SStefano Zampini 597*d62866d3SStefano Zampini F = NULL; 598*d62866d3SStefano Zampini ierr = MatGetFactorType(reuse_mumps->F,&type);CHKERRQ(ierr); 599*d62866d3SStefano 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 7013a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7029577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 7039577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 7043a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 7059577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7069577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7079577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7089577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 70980677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7109577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7119577ea80SStefano Zampini } else { 7123a50541eSStefano Zampini /* this is safe */ 7139577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 7143a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7153a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 716b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 717b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7183a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7193a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72006656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7213a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7223a50541eSStefano Zampini } else { 723b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7249577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7269577ea80SStefano Zampini } 72780677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 72880677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7293a50541eSStefano Zampini } 73025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 73188ebb749SStefano Zampini } 73288ebb749SStefano Zampini 73388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 734f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 73506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 73606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 73706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73806656605SStefano Zampini } 739f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 74006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 74106656605SStefano Zampini PetscScalar *marray; 74206656605SStefano Zampini 74306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 74406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 747f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 748f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 749f4ddd8eeSStefano Zampini } 750f4ddd8eeSStefano Zampini } 75106656605SStefano Zampini 752f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 75306656605SStefano Zampini PetscScalar *marray; 75488ebb749SStefano Zampini 75506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7568eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75888ebb749SStefano Zampini } 759b9d89cd5SStefano Zampini if (!pcbddc->issym) { 76006656605SStefano Zampini n *= 2; 761f4ddd8eeSStefano Zampini } 76206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 76306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 76406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 76506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 76706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76888ebb749SStefano Zampini } 76906656605SStefano Zampini if (!pcbddc->issym) { 77006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7711b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 77488ebb749SStefano Zampini } 77588ebb749SStefano Zampini } else { 776c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 777c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7781b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 779c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 780c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 781c0553b1fSStefano Zampini } 78288ebb749SStefano Zampini } 78306656605SStefano Zampini } 78406656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 78506656605SStefano Zampini /* vertices */ 78606656605SStefano Zampini if (n_vertices) { 78716f15bc4SStefano Zampini 78816f15bc4SStefano Zampini if (n_R) { 78906656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 79006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 79116f15bc4SStefano Zampini PetscScalar *x,*y; 79206656605SStefano Zampini 79306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 79421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 79506656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 79606656605SStefano Zampini if (F) { 79706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 79806656605SStefano Zampini } else { 79906656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 80106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 80206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 80306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 80606656605SStefano Zampini } 80706656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 80806656605SStefano Zampini } 80980677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81006656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 81106656605SStefano Zampini if (n_constraints) { 81206656605SStefano Zampini Mat B; 81380677318SStefano Zampini 814b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 81580677318SStefano Zampini for (i=0;i<n_vertices;i++) { 81680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 81780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 81880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82280677318SStefano Zampini } 82380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82480677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 82580677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 83106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83206656605SStefano Zampini } 83306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 83480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 83606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 83706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 83806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 83906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 84316f15bc4SStefano Zampini } else { 84416f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84516f15bc4SStefano Zampini } 84621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 84706656605SStefano Zampini /* coarse basis functions */ 84806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 84916f15bc4SStefano Zampini PetscScalar *y; 85016f15bc4SStefano Zampini 85106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 85706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 85906656605SStefano Zampini 86006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86706656605SStefano Zampini } 86806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 86906656605SStefano Zampini } 87006656605SStefano Zampini } 87106656605SStefano Zampini 87206656605SStefano Zampini if (n_constraints) { 87306656605SStefano Zampini Mat B; 87406656605SStefano Zampini 87506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 87606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 87806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87906656605SStefano Zampini if (n_vertices) { 88080677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 88180677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 88280677318SStefano Zampini } else { 88380677318SStefano Zampini Mat S_VCt; 88480677318SStefano Zampini 88580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 88680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 88780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 88880677318SStefano Zampini } 88906656605SStefano Zampini } 89006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 89106656605SStefano Zampini /* coarse basis functions */ 89206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 89306656605SStefano Zampini PetscScalar *y; 89406656605SStefano Zampini 89506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 89806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 90206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 90306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 90506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90906656605SStefano Zampini } 91006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91106656605SStefano Zampini } 91206656605SStefano Zampini } 91380677318SStefano Zampini if (n_constraints) { 91480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 91580677318SStefano Zampini } 91606656605SStefano Zampini 91706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 91806656605SStefano Zampini if (!pcbddc->issym) { 91906656605SStefano Zampini 92006656605SStefano Zampini if (n_constraints) { 92116f15bc4SStefano Zampini Mat S_CCT,B_C; 92206656605SStefano Zampini 92380677318SStefano Zampini /* this is a lazy thing */ 92480677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 92606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 92706656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 92816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 92906656605SStefano Zampini if (n_vertices) { 93016f15bc4SStefano Zampini Mat B_V,S_VCT; 93106656605SStefano Zampini 93206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 93306656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 93616f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 93706656605SStefano Zampini } 93806656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 93980677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94080677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 94106656605SStefano Zampini } 94216f15bc4SStefano Zampini if (n_vertices && n_R) { 94306656605SStefano Zampini Mat A_VRT; 94480677318SStefano Zampini PetscScalar *marray; 94506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 94606656605SStefano Zampini 94780677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 94880677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 94980677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 95180677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 95280677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 95316f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 95406656605SStefano Zampini } 95506656605SStefano Zampini 95606656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 95706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96306656605SStefano Zampini } 96406656605SStefano Zampini } else { 96506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97106656605SStefano Zampini } 97206656605SStefano Zampini } 97306656605SStefano Zampini /* coarse basis functions */ 97406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97506656605SStefano Zampini PetscScalar *y; 97606656605SStefano Zampini 97706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 97806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98206656605SStefano Zampini if (i<n_vertices) { 98306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 98406656605SStefano Zampini } 98506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 98706656605SStefano Zampini 98806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 98906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99506656605SStefano Zampini } 99606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 99706656605SStefano Zampini } 99806656605SStefano Zampini } 999*d62866d3SStefano Zampini /* free memory */ 100088ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 100406656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1005*d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1006*d62866d3SStefano Zampini if (n_vertices) { 1007*d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1008*d62866d3SStefano Zampini } 1009*d62866d3SStefano Zampini if (n_constraints) { 1010*d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1011*d62866d3SStefano Zampini } 101288ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 101388ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101488ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101525084f0cSStefano Zampini if (pcbddc->dbg_flag) { 101688ebb749SStefano Zampini Mat coarse_sub_mat; 101725084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 101888ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 101988ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102088ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10218bec7fa6SStefano Zampini Mat C_B,CPHI; 10228bec7fa6SStefano Zampini IS is_dummy; 10238bec7fa6SStefano Zampini Vec mones; 102488ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 102588ebb749SStefano Zampini PetscReal real_value; 102688ebb749SStefano Zampini 102788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1033c0553b1fSStefano Zampini if (unsymmetric_check) { 103488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 103688ebb749SStefano Zampini } 103788ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 103888ebb749SStefano Zampini 103925084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 104006656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 104125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1042c0553b1fSStefano Zampini if (unsymmetric_check) { 104388ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini } else { 105688ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_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_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106488ebb749SStefano Zampini } 106588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 106981d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10710fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 107206656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10738bec7fa6SStefano Zampini 10748bec7fa6SStefano Zampini /* check constraints */ 10758bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10768bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10778bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10788bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10808bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10818bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10838bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10848bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10858bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10868bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini 108825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 108988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 109188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 109288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 109388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 109488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 109588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 109688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 109788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 109888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1099c0553b1fSStefano Zampini if (unsymmetric_check) { 110088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 110288ebb749SStefano Zampini } 110388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 110488ebb749SStefano Zampini } 11058629588bSStefano Zampini /* get back data */ 11068629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 110788ebb749SStefano Zampini PetscFunctionReturn(0); 110888ebb749SStefano Zampini } 110988ebb749SStefano Zampini 111088ebb749SStefano Zampini #undef __FUNCT__ 1111d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1112d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1113d65f70fdSStefano Zampini { 1114d65f70fdSStefano Zampini Mat *work_mat; 1115d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1116d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1117d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1118d65f70fdSStefano Zampini PetscErrorCode ierr; 1119d65f70fdSStefano Zampini 1120d65f70fdSStefano Zampini PetscFunctionBegin; 1121d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1122d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1123d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1124d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1125d65f70fdSStefano Zampini 1126d65f70fdSStefano Zampini if (!rsorted) { 1127d65f70fdSStefano Zampini const PetscInt *idxs; 1128d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1129d65f70fdSStefano Zampini 1130d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1133d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1134d65f70fdSStefano Zampini } 1135d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1136d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1137d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1138d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1139d65f70fdSStefano Zampini } 1140d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1142d65f70fdSStefano Zampini } else { 1143d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1144d65f70fdSStefano Zampini isrow_s = isrow; 1145d65f70fdSStefano Zampini } 1146d65f70fdSStefano Zampini 1147d65f70fdSStefano Zampini if (!csorted) { 1148d65f70fdSStefano Zampini if (isrow == iscol) { 1149d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1150d65f70fdSStefano Zampini iscol_s = isrow_s; 1151d65f70fdSStefano Zampini } else { 1152d65f70fdSStefano Zampini const PetscInt *idxs; 1153d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1154d65f70fdSStefano Zampini 1155d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1156d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1158d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1159d65f70fdSStefano Zampini } 1160d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1161d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1162d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1163d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1164d65f70fdSStefano Zampini } 1165d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1167d65f70fdSStefano Zampini } 1168d65f70fdSStefano Zampini } else { 1169d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini iscol_s = iscol; 1171d65f70fdSStefano Zampini } 1172d65f70fdSStefano Zampini 1173d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1174d65f70fdSStefano Zampini 1175d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1176d65f70fdSStefano Zampini Mat new_mat; 1177d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1178d65f70fdSStefano Zampini 1179d65f70fdSStefano Zampini if (!rsorted) { 1180d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1181d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1182d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1183d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1184d65f70fdSStefano Zampini } 1185d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1187d65f70fdSStefano Zampini } else { 1188d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1189d65f70fdSStefano Zampini } 1190d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini 1192d65f70fdSStefano Zampini if (!csorted) { 1193d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1194d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1196d65f70fdSStefano Zampini } else { 1197d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1198d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1200d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1201d65f70fdSStefano Zampini } 1202d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini } 1205d65f70fdSStefano Zampini } else { 1206d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini } 1208d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini 1210d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini work_mat[0] = new_mat; 1213d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1215d65f70fdSStefano Zampini } 1216d65f70fdSStefano Zampini 1217d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini *B = work_mat[0]; 1219d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini PetscFunctionReturn(0); 1223d65f70fdSStefano Zampini } 1224d65f70fdSStefano Zampini 1225d65f70fdSStefano Zampini #undef __FUNCT__ 12265e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12275e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1228aa0d41d4SStefano Zampini { 1229aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12305e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1231d65f70fdSStefano Zampini Mat new_mat; 12325e8657edSStefano Zampini IS is_local,is_global; 1233d65f70fdSStefano Zampini PetscInt local_size; 1234d65f70fdSStefano Zampini PetscBool isseqaij; 1235aa0d41d4SStefano Zampini PetscErrorCode ierr; 1236aa0d41d4SStefano Zampini 1237aa0d41d4SStefano Zampini PetscFunctionBegin; 1238aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12395e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12405e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1241906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1242906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1243d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1244906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1245906d46d4SStefano Zampini 1246906d46d4SStefano Zampini /* check */ 1247906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1248906d46d4SStefano Zampini Vec x,x_change; 1249906d46d4SStefano Zampini PetscReal error; 1250906d46d4SStefano Zampini 12515e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1252906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12535e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12545e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12555e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1256d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12575e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12585e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1259906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1260906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1261906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1262906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1263906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1264906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1265906d46d4SStefano Zampini } 1266906d46d4SStefano Zampini 126722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12689b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 126922d5777bSStefano Zampini if (isseqaij) { 1270d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1271aa0d41d4SStefano Zampini } else { 1272aa0d41d4SStefano Zampini Mat work_mat; 1273aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1274d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1275aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1276aa0d41d4SStefano Zampini } 1277d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1278e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1279e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1280e496cd5dSStefano Zampini #endif 128145a1bb75SStefano Zampini /* 128245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1283d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 128445a1bb75SStefano Zampini */ 1285d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1286aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1287aa0d41d4SStefano Zampini } 1288aa0d41d4SStefano Zampini 1289aa0d41d4SStefano Zampini #undef __FUNCT__ 1290a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12918ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1292a64d13efSStefano Zampini { 1293a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1294a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1295*d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 12968ce42a96SStefano Zampini IS is_aux1,is_aux2; 1297e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 12983a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12993a50541eSStefano Zampini PetscInt vbs,bs; 13004641a718SStefano Zampini PetscBT bitmask; 1301a64d13efSStefano Zampini PetscErrorCode ierr; 1302a64d13efSStefano Zampini 1303a64d13efSStefano Zampini PetscFunctionBegin; 1304b23d619eSStefano Zampini /* 1305b23d619eSStefano Zampini No need to setup local scatters if 1306b23d619eSStefano Zampini - primal space is unchanged 1307b23d619eSStefano Zampini AND 1308b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1309b23d619eSStefano Zampini AND 1310b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1311b23d619eSStefano Zampini */ 1312b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1313f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1314f4ddd8eeSStefano Zampini } 1315f4ddd8eeSStefano Zampini /* destroy old objects */ 1316f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1317f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1318f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1319a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1320b371cd4fSStefano Zampini n_B = pcis->n_B; 1321b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1322b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13234641a718SStefano Zampini /* create auxiliary bitmask */ 13244641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13254641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13260e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13274641a718SStefano Zampini } 13283a50541eSStefano Zampini 1329a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 1330854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1331a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13324641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 1333a64d13efSStefano Zampini idx_R_local[n_R] = i; 1334a64d13efSStefano Zampini n_R++; 1335a64d13efSStefano Zampini } 1336a64d13efSStefano Zampini } 13373a50541eSStefano Zampini 13383a50541eSStefano Zampini /* Block code */ 13393a50541eSStefano Zampini vbs = 1; 13403a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13413a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13423a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13433a50541eSStefano Zampini PetscInt *vary; 13443a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1345785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13463a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13470e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1348964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13493a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13503a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13513a50541eSStefano Zampini break; 13523a50541eSStefano Zampini } 13533a50541eSStefano Zampini } 13543a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13553a50541eSStefano Zampini vbs = bs; 13563a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13573a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13583a50541eSStefano Zampini } 13593a50541eSStefano Zampini } 13603a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13613a50541eSStefano Zampini } 13623a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 13633a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1364a64d13efSStefano Zampini 1365a64d13efSStefano Zampini /* print some info if requested */ 1366a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1367a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1368a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13690fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1370a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1371a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13723a50541eSStefano 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); 1373a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1374a64d13efSStefano Zampini } 1375a64d13efSStefano Zampini 1376a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13773a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1378854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1379854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1380a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 13814641a718SStefano Zampini for (i=0; i<n_D; i++) { 13824641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 13834641a718SStefano Zampini } 1384a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1385a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13864641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 13874641a718SStefano Zampini aux_array1[j++] = i; 1388a64d13efSStefano Zampini } 1389a64d13efSStefano Zampini } 1390a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1391a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1392a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 13934641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 13944641a718SStefano Zampini aux_array2[j++] = i; 1395a64d13efSStefano Zampini } 1396a64d13efSStefano Zampini } 1397a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1398a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1399a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1400a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1401a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1402a64d13efSStefano Zampini 14038eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1404785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1405a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14064641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14074641a718SStefano Zampini aux_array1[j++] = i; 1408a64d13efSStefano Zampini } 1409a64d13efSStefano Zampini } 1410a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1411a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1412a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1413a64d13efSStefano Zampini } 14144641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14153a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1416*d62866d3SStefano Zampini 1417*d62866d3SStefano Zampini /* auxiliary scatters are needed if we are reusing MUMPS Schur solver */ 1418*d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1419*d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1420*d62866d3SStefano Zampini Vec tvec; 1421*d62866d3SStefano Zampini PetscInt schur_size; 1422*d62866d3SStefano Zampini ISLocalToGlobalMapping R_to_N; 1423*d62866d3SStefano Zampini IS list[2],tisB,tisB_N,tisR_R,tisR; 1424*d62866d3SStefano Zampini 1425*d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse_mumps->correction_scatter_B);CHKERRQ(ierr); 1426*d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse_mumps->correction_scatter_R);CHKERRQ(ierr); 1427*d62866d3SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&R_to_N);CHKERRQ(ierr); 1428*d62866d3SStefano Zampini if (sub_schurs->is_dir) { 1429*d62866d3SStefano Zampini list[0] = sub_schurs->is_Ej_all; 1430*d62866d3SStefano Zampini list[1] = sub_schurs->is_dir; 1431*d62866d3SStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&tisB);CHKERRQ(ierr); 1432*d62866d3SStefano Zampini } else { 1433*d62866d3SStefano Zampini ierr = ISDuplicate(sub_schurs->is_Ej_all,&tisB);CHKERRQ(ierr); 1434*d62866d3SStefano Zampini } 1435*d62866d3SStefano Zampini ierr = ISGetLocalSize(tisB,&schur_size);CHKERRQ(ierr); 1436*d62866d3SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&tvec);CHKERRQ(ierr); 1437*d62866d3SStefano Zampini ierr = VecSetSizes(tvec,PETSC_DECIDE,schur_size);CHKERRQ(ierr); 1438*d62866d3SStefano Zampini ierr = VecSetType(tvec,VECSEQ);CHKERRQ(ierr); 1439*d62866d3SStefano Zampini ierr = VecScatterCreate(pcis->vec1_B,tisB,tvec,NULL,&reuse_mumps->correction_scatter_B);CHKERRQ(ierr); 1440*d62866d3SStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1441*d62866d3SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->BtoNmap,tisB,&tisB_N);CHKERRQ(ierr); 1442*d62866d3SStefano Zampini ierr = ISDestroy(&tisB);CHKERRQ(ierr); 1443*d62866d3SStefano Zampini list[0] = pcis->is_I_local; 1444*d62866d3SStefano Zampini list[1] = tisB_N; 1445*d62866d3SStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&tisR);CHKERRQ(ierr); 1446*d62866d3SStefano Zampini ierr = ISDestroy(&tisB_N);CHKERRQ(ierr); 1447*d62866d3SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(R_to_N,IS_GTOLM_DROP,tisR,&tisR_R);CHKERRQ(ierr); 1448*d62866d3SStefano Zampini ierr = ISDestroy(&tisR);CHKERRQ(ierr); 1449*d62866d3SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tisR_R,pcbddc->vec1_R,NULL,&reuse_mumps->correction_scatter_R);CHKERRQ(ierr); 1450*d62866d3SStefano Zampini ierr = ISDestroy(&tisR_R);CHKERRQ(ierr); 1451*d62866d3SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&R_to_N);CHKERRQ(ierr); 1452*d62866d3SStefano Zampini } 1453a64d13efSStefano Zampini PetscFunctionReturn(0); 1454a64d13efSStefano Zampini } 1455a64d13efSStefano Zampini 1456304d26faSStefano Zampini 1457304d26faSStefano Zampini #undef __FUNCT__ 1458304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1459684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1460304d26faSStefano Zampini { 1461304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1462304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1463304d26faSStefano Zampini PC pc_temp; 1464304d26faSStefano Zampini Mat A_RR; 1465f4ddd8eeSStefano Zampini MatReuse reuse; 1466304d26faSStefano Zampini PetscScalar m_one = -1.0; 1467304d26faSStefano Zampini PetscReal value; 1468af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14699577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1470304d26faSStefano Zampini PetscErrorCode ierr; 1471e604994aSStefano Zampini /* prefixes stuff */ 1472312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1473e604994aSStefano Zampini size_t len; 1474304d26faSStefano Zampini 1475304d26faSStefano Zampini PetscFunctionBegin; 1476304d26faSStefano Zampini 1477e604994aSStefano Zampini /* compute prefixes */ 1478e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1479e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1480e604994aSStefano Zampini if (!pcbddc->current_level) { 1481e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1482e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1483e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1484e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1485e604994aSStefano Zampini } else { 1486e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1487312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1488e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1489e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1490312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1491312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 149234d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 149334d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1494e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1495e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1496e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1497e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1498e604994aSStefano Zampini } 1499e604994aSStefano Zampini 1500304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1501684f6988SStefano Zampini if (dirichlet) { 1502d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1503964fefecSStefano Zampini if (pcbddc->issym) { 1504964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1505964fefecSStefano Zampini } 1506ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1507964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1508304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1509304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1510304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1511304d26faSStefano Zampini /* default */ 1512304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1513e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15149577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1515304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15169577ea80SStefano Zampini if (issbaij) { 15179577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15189577ea80SStefano Zampini } else { 1519304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15209577ea80SStefano Zampini } 1521304d26faSStefano Zampini /* Allow user's customization */ 1522304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1523304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1524304d26faSStefano Zampini } 1525d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1526*d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1527*d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1528*d62866d3SStefano Zampini 1529*d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1530d5574798SStefano Zampini } 1531304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1532304d26faSStefano Zampini if (!n_D) { 1533304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1534304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1535304d26faSStefano Zampini } 1536304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1537304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1538304d26faSStefano Zampini /* set ksp_D into pcis data */ 1539304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1540304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1541304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1542684f6988SStefano Zampini } 1543304d26faSStefano Zampini 1544304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1545684f6988SStefano Zampini A_RR = 0; 1546684f6988SStefano Zampini if (neumann) { 1547*d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1548f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15498ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1550f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1551f4ddd8eeSStefano Zampini PetscInt nn_R; 155281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1553f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1554f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1555f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1556f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1557f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1558f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1559f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1560727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1561f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1562f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1563f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1564f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1565f4ddd8eeSStefano Zampini } 1566f4ddd8eeSStefano Zampini } 1567f4ddd8eeSStefano Zampini /* last check */ 1568d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1569f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1570f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1571f4ddd8eeSStefano Zampini } 1572f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1573f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1574f4ddd8eeSStefano Zampini } 1575f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1576304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1577304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1578304d26faSStefano Zampini /* default */ 1579304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1580e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1581304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 15829577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 15839577ea80SStefano Zampini if (issbaij) { 15849577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15859577ea80SStefano Zampini } else { 1586304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15879577ea80SStefano Zampini } 1588304d26faSStefano Zampini /* Allow user's customization */ 1589304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1590304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1591304d26faSStefano Zampini } 1592*d62866d3SStefano Zampini /* extract A_RR */ 1593*d62866d3SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1594*d62866d3SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 1595*d62866d3SStefano Zampini if (ibs != mbs) { 1596*d62866d3SStefano Zampini Mat newmat; 1597*d62866d3SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1598*d62866d3SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1599*d62866d3SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1600*d62866d3SStefano Zampini } else { 1601*d62866d3SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1602*d62866d3SStefano Zampini } 1603*d62866d3SStefano Zampini if (pcbddc->issym) { 1604*d62866d3SStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1605*d62866d3SStefano Zampini } 1606d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1607304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1608304d26faSStefano Zampini if (!n_R) { 1609304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1610304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1611304d26faSStefano Zampini } 1612*d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1613*d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1614*d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1615*d62866d3SStefano Zampini 1616*d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1617*d62866d3SStefano Zampini } 1618304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1619304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1620684f6988SStefano Zampini } 1621304d26faSStefano Zampini 1622304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16230fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1624684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1625684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1626684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1627684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1628684f6988SStefano Zampini } 1629684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16300fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16310fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16320fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16330fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16340fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1635304d26faSStefano Zampini /* need to be adapted? */ 1636b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1637b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1638b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1639304d26faSStefano Zampini /* print info */ 1640304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1641e604994aSStefano 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); 1642304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1643304d26faSStefano Zampini } 1644b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16458ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1646304d26faSStefano Zampini } 1647684f6988SStefano Zampini } 1648684f6988SStefano Zampini if (neumann) { /* Neumann */ 16490fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16500fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16510fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16520fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16530fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1654304d26faSStefano Zampini /* need to be adapted? */ 1655b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1656b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1657304d26faSStefano Zampini /* print info */ 1658304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1659e604994aSStefano 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); 1660304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1661304d26faSStefano Zampini } 1662b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16638ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1664304d26faSStefano Zampini } 16650fccc4e9SStefano Zampini } 1666684f6988SStefano Zampini } 1667304d26faSStefano Zampini /* free Neumann problem's matrix */ 1668304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1669304d26faSStefano Zampini PetscFunctionReturn(0); 1670304d26faSStefano Zampini } 1671304d26faSStefano Zampini 1672304d26faSStefano Zampini #undef __FUNCT__ 1673ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 167480677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1675674ae819SStefano Zampini { 1676674ae819SStefano Zampini PetscErrorCode ierr; 1677674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1678674ae819SStefano Zampini 1679674ae819SStefano Zampini PetscFunctionBegin; 168080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 168180677318SStefano Zampini if (!pcbddc->switch_static) { 168280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 168380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 168480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 168520c7b377SStefano Zampini } 168680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 168780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 168820c7b377SStefano Zampini } else { 168980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 169480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 169580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 169680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 169820c7b377SStefano Zampini } 1699674ae819SStefano Zampini } 170080677318SStefano Zampini if (applytranspose) { 170180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 170280677318SStefano Zampini } else { 170380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 170480677318SStefano Zampini } 170580677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 170680677318SStefano Zampini if (!pcbddc->switch_static) { 170780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 170880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 170980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 171080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 171180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 171280677318SStefano Zampini } 171380677318SStefano Zampini } else { 171480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 171980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 172080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 172180677318SStefano Zampini } 172280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172680677318SStefano Zampini } 1727674ae819SStefano Zampini PetscFunctionReturn(0); 1728674ae819SStefano Zampini } 1729674ae819SStefano Zampini 1730dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1731674ae819SStefano Zampini #undef __FUNCT__ 1732674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1733dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1734674ae819SStefano Zampini { 1735674ae819SStefano Zampini PetscErrorCode ierr; 1736674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1737674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1738674ae819SStefano Zampini const PetscScalar zero = 0.0; 1739674ae819SStefano Zampini 1740674ae819SStefano Zampini PetscFunctionBegin; 1741dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1742dc359a40SStefano Zampini if (applytranspose) { 1743674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 17448eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1745dc359a40SStefano Zampini } else { 1746674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1747674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 174815aaf578SStefano Zampini } 174912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 175012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 175112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 175212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 175312edc857SStefano Zampini 17549f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 175512edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 175612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1757964fefecSStefano Zampini Vec rhs,sol; 1758964fefecSStefano Zampini 1759964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1760964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 176112edc857SStefano Zampini if (applytranspose) { 1762964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 176312edc857SStefano Zampini } else { 1764964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 176512edc857SStefano Zampini } 176612edc857SStefano Zampini } 1767674ae819SStefano Zampini 1768674ae819SStefano Zampini /* Local solution on R nodes */ 176980677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 177080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 17719f00e9b4SStefano Zampini } 1772674ae819SStefano Zampini 17739f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 17749f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 177512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1776674ae819SStefano Zampini 1777674ae819SStefano Zampini /* Sum contributions from two levels */ 1778dc359a40SStefano Zampini if (applytranspose) { 1779dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1780dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1781dc359a40SStefano Zampini } else { 1782674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 17838eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1784dc359a40SStefano Zampini } 1785674ae819SStefano Zampini PetscFunctionReturn(0); 1786674ae819SStefano Zampini } 1787674ae819SStefano Zampini 1788674ae819SStefano Zampini #undef __FUNCT__ 1789674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 179012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1791674ae819SStefano Zampini { 1792674ae819SStefano Zampini PetscErrorCode ierr; 1793674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 179458da7f69SStefano Zampini PetscScalar *array; 179512edc857SStefano Zampini Vec from,to; 1796674ae819SStefano Zampini 1797674ae819SStefano Zampini PetscFunctionBegin; 179812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 179912edc857SStefano Zampini from = pcbddc->coarse_vec; 180012edc857SStefano Zampini to = pcbddc->vec1_P; 180112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 180212edc857SStefano Zampini Vec tvec; 180358da7f69SStefano Zampini 180458da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 180558da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 180612edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 180758da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 180858da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 180958da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 181012edc857SStefano Zampini } 181112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 181212edc857SStefano Zampini from = pcbddc->vec1_P; 181312edc857SStefano Zampini to = pcbddc->coarse_vec; 181412edc857SStefano Zampini } 181512edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1816674ae819SStefano Zampini PetscFunctionReturn(0); 1817674ae819SStefano Zampini } 1818674ae819SStefano Zampini 1819674ae819SStefano Zampini #undef __FUNCT__ 1820674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 182112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1822674ae819SStefano Zampini { 1823674ae819SStefano Zampini PetscErrorCode ierr; 1824674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 182558da7f69SStefano Zampini PetscScalar *array; 182612edc857SStefano Zampini Vec from,to; 1827674ae819SStefano Zampini 1828674ae819SStefano Zampini PetscFunctionBegin; 182912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 183012edc857SStefano Zampini from = pcbddc->coarse_vec; 183112edc857SStefano Zampini to = pcbddc->vec1_P; 183212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 183312edc857SStefano Zampini from = pcbddc->vec1_P; 183412edc857SStefano Zampini to = pcbddc->coarse_vec; 183512edc857SStefano Zampini } 183612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 183712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 183812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 183912edc857SStefano Zampini Vec tvec; 184058da7f69SStefano Zampini 184112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 184258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 184358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 184458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 184558da7f69SStefano Zampini } 184658da7f69SStefano Zampini } else { 184758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 184858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 184912edc857SStefano Zampini } 185012edc857SStefano Zampini } 1851674ae819SStefano Zampini PetscFunctionReturn(0); 1852674ae819SStefano Zampini } 1853674ae819SStefano Zampini 1854984c4197SStefano Zampini /* uncomment for testing purposes */ 1855984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1856674ae819SStefano Zampini #undef __FUNCT__ 1857674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1858674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1859674ae819SStefano Zampini { 1860674ae819SStefano Zampini PetscErrorCode ierr; 1861674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1862674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1863674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1864984c4197SStefano Zampini /* one and zero */ 1865984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1866984c4197SStefano Zampini /* space to store constraints and their local indices */ 18679162d606SStefano Zampini PetscScalar *constraints_data; 18689162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 18699162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 18709162d606SStefano Zampini PetscInt *constraints_n; 1871984c4197SStefano Zampini /* iterators */ 1872b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1873984c4197SStefano Zampini /* BLAS integers */ 1874e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1875e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1876c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1877727cdba6SStefano Zampini /* reuse */ 18780e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 18790e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1880984c4197SStefano Zampini /* change of basis */ 1881b3d85658SStefano Zampini PetscBool qr_needed; 18829162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1883984c4197SStefano Zampini /* auxiliary stuff */ 188464efe560SStefano Zampini PetscInt *nnz,*is_indices; 18858a0068c3SStefano Zampini PetscInt ncc; 1886984c4197SStefano Zampini /* some quantities */ 188745a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1888a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1889984c4197SStefano Zampini 1890674ae819SStefano Zampini PetscFunctionBegin; 18918e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 18928e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 18938e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1894088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1895088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 18960e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 18970e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 18980e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 18990e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19000e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1901088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1902cf5a6209SStefano Zampini 1903cf5a6209SStefano Zampini /* print some info */ 1904cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1905cf5a6209SStefano Zampini IS vertices; 1906cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1907cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1908cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1909cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1910cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1911cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1912cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1913fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1914fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1915cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1916cf5a6209SStefano Zampini } 1917cf5a6209SStefano Zampini 1918cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19199162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1920cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1921cf5a6209SStefano Zampini const Vec *nearnullvecs; 1922cf5a6209SStefano Zampini Vec *localnearnullsp; 1923cf5a6209SStefano Zampini PetscScalar *array; 1924cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1925cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1926cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1927b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1928cf5a6209SStefano Zampini PetscScalar *work; 1929cf5a6209SStefano Zampini PetscReal *singular_vals; 1930cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1931cf5a6209SStefano Zampini PetscReal *rwork; 1932cf5a6209SStefano Zampini #endif 1933cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1934cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1935cf5a6209SStefano Zampini #else 1936964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1937964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1938cf5a6209SStefano Zampini #endif 1939cf5a6209SStefano Zampini 1940674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1941d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1942d06fc5fdSStefano Zampini /* free unneeded index sets */ 1943d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1944d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1945674ae819SStefano Zampini } 1946d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1947d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1948d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1949d06fc5fdSStefano Zampini } 1950d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1951d06fc5fdSStefano Zampini n_ISForEdges = 0; 1952d06fc5fdSStefano Zampini } 1953d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1954d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1955d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1956d06fc5fdSStefano Zampini } 1957d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1958d06fc5fdSStefano Zampini n_ISForFaces = 0; 1959d06fc5fdSStefano Zampini } 196070022509SStefano Zampini 196170022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 196270022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 196370022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 196470022509SStefano Zampini if (pcbddc->NullSpace) { 196570022509SStefano Zampini PetscBool tbool[2],gbool[2]; 196670022509SStefano Zampini 196770022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 1968b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1969d06fc5fdSStefano Zampini if (!ISForEdges) { 1970d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1971d06fc5fdSStefano Zampini } 1972b8ffe317SStefano Zampini } 1973d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 1974d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 1975d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1976d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1977d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 197898a51de6SStefano Zampini } 197970022509SStefano Zampini #endif 198008122e43SStefano Zampini 1981674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1982674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1983674ae819SStefano Zampini if (nearnullsp) { 1984674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1985f4ddd8eeSStefano Zampini /* remove any stored info */ 1986f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1987f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1988f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1989f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1990f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1991473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1992f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1993f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1994f4ddd8eeSStefano Zampini } 1995984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1996984c4197SStefano Zampini nnsp_size = 0; 1997674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1998674ae819SStefano Zampini } 1999984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2000984c4197SStefano Zampini max_constraints = nnsp_size; 2001984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2002984c4197SStefano Zampini 2003674ae819SStefano Zampini /* 2004674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20059162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20069162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20079162d606SStefano Zampini There can be multiple constraints per connected component 2008674ae819SStefano Zampini */ 2009674ae819SStefano Zampini n_vertices = 0; 2010674ae819SStefano Zampini if (ISForVertices) { 2011674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2012674ae819SStefano Zampini } 20139162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20149162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20159162d606SStefano Zampini 20169162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20179162d606SStefano Zampini total_counts *= max_constraints; 2018674ae819SStefano Zampini total_counts += n_vertices; 20194641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20209162d606SStefano Zampini 2021674ae819SStefano Zampini total_counts = 0; 2022674ae819SStefano Zampini max_size_of_constraint = 0; 2023674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20249162d606SStefano Zampini IS used_is; 2025674ae819SStefano Zampini if (i<n_ISForEdges) { 20269162d606SStefano Zampini used_is = ISForEdges[i]; 2027674ae819SStefano Zampini } else { 20289162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2029674ae819SStefano Zampini } 20309162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2031674ae819SStefano Zampini total_counts += j; 2032674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2033674ae819SStefano Zampini } 20349162d606SStefano 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); 20359162d606SStefano Zampini 2036984c4197SStefano Zampini /* get local part of global near null space vectors */ 2037785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2038984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2039984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2040984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2041984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2042984c4197SStefano Zampini } 2043674ae819SStefano Zampini 2044242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2045242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2046a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2047242a89d7SStefano Zampini 2048984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2049a773dcb8SStefano Zampini if (!skip_lapack) { 2050674ae819SStefano Zampini PetscScalar temp_work; 2051911cabfeSStefano Zampini 2052674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2053984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2054785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2055785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2056785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2057674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2058785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2059674ae819SStefano Zampini #endif 2060674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2061c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2062c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2063674ae819SStefano Zampini lwork = -1; 2064674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2065674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2066c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2067674ae819SStefano Zampini #else 2068c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2069674ae819SStefano Zampini #endif 2070674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2071984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2072674ae819SStefano Zampini #else /* on missing GESVD */ 2073674ae819SStefano Zampini /* SVD */ 2074674ae819SStefano Zampini PetscInt max_n,min_n; 2075674ae819SStefano Zampini max_n = max_size_of_constraint; 2076984c4197SStefano Zampini min_n = max_constraints; 2077984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2078674ae819SStefano Zampini min_n = max_size_of_constraint; 2079984c4197SStefano Zampini max_n = max_constraints; 2080674ae819SStefano Zampini } 2081785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2082674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2083785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2084674ae819SStefano Zampini #endif 2085674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2086674ae819SStefano Zampini lwork = -1; 2087e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2088e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2089b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2090674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2091674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20929162d606SStefano 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)); 2093674ae819SStefano Zampini #else 20949162d606SStefano 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)); 2095674ae819SStefano Zampini #endif 2096674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2097984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2098984c4197SStefano Zampini #endif /* on missing GESVD */ 2099674ae819SStefano Zampini /* Allocate optimal workspace */ 2100674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2101854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2102674ae819SStefano Zampini } 2103674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2104674ae819SStefano Zampini total_counts = 0; 21059162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21069162d606SStefano Zampini constraints_data_ptr[0] = 0; 2107674ae819SStefano Zampini /* vertices */ 21089162d606SStefano Zampini if (n_vertices) { 2109674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21109162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21119162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2112674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21139162d606SStefano Zampini constraints_n[total_counts] = 1; 21149162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21159162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21169162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2117674ae819SStefano Zampini total_counts++; 2118674ae819SStefano Zampini } 2119674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2120984c4197SStefano Zampini PetscBool used_vertex; 2121674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2122674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2123674ae819SStefano Zampini k = 0; 2124674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2125984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2126984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21279162d606SStefano Zampini constraints_n[total_counts] = 1; 21289162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21299162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21309162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21319162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2132674ae819SStefano Zampini total_counts++; 2133674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2134674ae819SStefano Zampini } 2135984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2136674ae819SStefano Zampini k++; 2137674ae819SStefano Zampini } 2138674ae819SStefano Zampini } 2139674ae819SStefano Zampini } 2140674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2141674ae819SStefano Zampini n_vertices = total_counts; 2142674ae819SStefano Zampini } 2143984c4197SStefano Zampini 2144674ae819SStefano Zampini /* edges and faces */ 21459162d606SStefano Zampini total_counts_cc = total_counts; 2146911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 21479162d606SStefano Zampini IS used_is; 21489162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 21499162d606SStefano Zampini 2150911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 21519162d606SStefano Zampini used_is = ISForEdges[ncc]; 2152984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2153674ae819SStefano Zampini } else { 21549162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2155984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2156674ae819SStefano Zampini } 2157674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 21589162d606SStefano Zampini 21599162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 21609162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2161984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2162984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2163674ae819SStefano Zampini if (nnsp_has_cnst) { 21645b08dc53SStefano Zampini PetscScalar quad_value; 21659162d606SStefano Zampini 21669162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21679162d606SStefano Zampini idxs_copied = PETSC_TRUE; 21689162d606SStefano Zampini 2169a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2170674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2171a773dcb8SStefano Zampini } else { 2172a773dcb8SStefano Zampini quad_value = 1.0; 2173a773dcb8SStefano Zampini } 2174674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21759162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2176674ae819SStefano Zampini } 21779162d606SStefano Zampini temp_constraints++; 2178674ae819SStefano Zampini total_counts++; 2179674ae819SStefano Zampini } 2180674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2181984c4197SStefano Zampini PetscReal real_value; 21829162d606SStefano Zampini PetscScalar *ptr_to_data; 21839162d606SStefano Zampini 2184984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 21859162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2186674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21879162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2188674ae819SStefano Zampini } 2189984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2190984c4197SStefano Zampini /* check if array is null on the connected component */ 2191e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21929162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 21935b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2194674ae819SStefano Zampini temp_constraints++; 2195674ae819SStefano Zampini total_counts++; 21969162d606SStefano Zampini if (!idxs_copied) { 21979162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21989162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2199674ae819SStefano Zampini } 2200674ae819SStefano Zampini } 22019162d606SStefano Zampini } 22029162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 220345a1bb75SStefano Zampini valid_constraints = temp_constraints; 2204eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2205a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22069162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22079162d606SStefano Zampini 22089162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2209a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22109162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2211a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22129162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2213a773dcb8SStefano Zampini } else { /* perform SVD */ 2214984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22159162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2216674ae819SStefano Zampini 2217674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2218984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2219984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2220984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2221984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2222984c4197SStefano Zampini from that computed using LAPACKgesvd 2223984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2224984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2225984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2226674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2227e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2228984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2229674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2230674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22319162d606SStefano 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)); 2232674ae819SStefano Zampini } 2233674ae819SStefano Zampini } 2234e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2235e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2236e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2237674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2238c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2239674ae819SStefano Zampini #else 2240c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2241674ae819SStefano Zampini #endif 2242674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2243984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2244984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2245674ae819SStefano Zampini j = 0; 2246984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2247674ae819SStefano Zampini total_counts = total_counts-j; 224845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2249e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2250c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2251c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2252c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2253c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2254c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2255c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2256674ae819SStefano Zampini if (j<temp_constraints) { 2257984c4197SStefano Zampini PetscInt ii; 2258984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2259674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 22609162d606SStefano 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)); 2261674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2262984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2263674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 22649162d606SStefano 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]; 2265674ae819SStefano Zampini } 2266674ae819SStefano Zampini } 2267674ae819SStefano Zampini } 2268674ae819SStefano Zampini #else /* on missing GESVD */ 2269e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2270e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2271b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2272674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2273674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 22749162d606SStefano 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)); 2275674ae819SStefano Zampini #else 22769162d606SStefano 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)); 2277674ae819SStefano Zampini #endif 2278984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2279674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2280984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2281e310c8b4SStefano Zampini k = temp_constraints; 2282e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2283674ae819SStefano Zampini j = 0; 2284e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 228545a1bb75SStefano Zampini valid_constraints = k-j; 2286911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2287984c4197SStefano Zampini #endif /* on missing GESVD */ 2288674ae819SStefano Zampini } 2289a773dcb8SStefano Zampini } 22909162d606SStefano Zampini /* update pointers information */ 22919162d606SStefano Zampini if (valid_constraints) { 22929162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 22939162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 22949162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 22959162d606SStefano Zampini /* set change_of_basis flag */ 229645a1bb75SStefano Zampini if (boolforchange) { 2297b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 22989162d606SStefano Zampini } 2299b3d85658SStefano Zampini total_counts_cc++; 230045a1bb75SStefano Zampini } 2301674ae819SStefano Zampini } 2302984c4197SStefano Zampini /* free workspace */ 23038f1c130eSStefano Zampini if (!skip_lapack) { 2304984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2305984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2306984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2307984c4197SStefano Zampini #endif 2308984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2309984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2310984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2311984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2312984c4197SStefano Zampini #endif 2313984c4197SStefano Zampini } 2314984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2315984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2316984c4197SStefano Zampini } 2317984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2318cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2319cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2320cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2321cf5a6209SStefano Zampini } 2322cf5a6209SStefano Zampini if (n_ISForFaces) { 2323cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2324cf5a6209SStefano Zampini } 2325cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2326cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2327cf5a6209SStefano Zampini } 2328cf5a6209SStefano Zampini if (n_ISForEdges) { 2329cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2330cf5a6209SStefano Zampini } 2331cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 233208122e43SStefano Zampini } else { 233308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 233408122e43SStefano Zampini 233508122e43SStefano Zampini total_counts = 0; 233608122e43SStefano Zampini n_vertices = 0; 2337*d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2338*d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 233908122e43SStefano Zampini } 234008122e43SStefano Zampini max_constraints = 0; 23419162d606SStefano Zampini total_counts_cc = 0; 234208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 234308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 23449162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 234508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 234608122e43SStefano Zampini } 23479162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 23489162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 23499162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 23509162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 235174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 23529162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 23539162d606SStefano Zampini total_counts_cc = 0; 23549162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 23559162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 23569162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 235708122e43SStefano Zampini } 235808122e43SStefano Zampini } 23599162d606SStefano Zampini #if 0 23609162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 23619162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23629162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 23639162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 23649162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 23659162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 23669162d606SStefano Zampini } 23679162d606SStefano Zampini printf("\n"); 23689162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 23699162d606SStefano Zampini } 23701b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 23718bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 23721b968477SStefano Zampini } 23731b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 23748bec7fa6SStefano 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]); 23751b968477SStefano Zampini } 237608122e43SStefano Zampini #endif 237708122e43SStefano Zampini 23788bec7fa6SStefano Zampini max_size_of_constraint = 0; 23799162d606SStefano 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]); 23809162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 238108122e43SStefano Zampini /* Change of basis */ 2382b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 238308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 238408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 238508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2386b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 238708122e43SStefano Zampini } 238808122e43SStefano Zampini } 238908122e43SStefano Zampini } 239008122e43SStefano Zampini } 2391b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 23920e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 239308122e43SStefano Zampini 23949162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 23959162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 23969162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 23979162d606SStefano 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); 239808122e43SStefano Zampini } 2399984c4197SStefano Zampini 2400674ae819SStefano Zampini /* Create constraint matrix */ 2401674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 240216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2403984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2404984c4197SStefano Zampini 2405984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 240674d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 240774d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 240874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2409984c4197SStefano Zampini total_primal_vertices=0; 2410b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24119162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24129162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2413984c4197SStefano Zampini if (size_of_constraint == 1) { 24149162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2415b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 241664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24179162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24189162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24199162d606SStefano Zampini } 2420b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 242174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 242274d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 242374d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 242474d5cdf7SStefano Zampini } 2425b3d85658SStefano Zampini } else { 2426b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2427984c4197SStefano Zampini } 2428984c4197SStefano Zampini } 2429b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2430b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 243170022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 243270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2433b3d85658SStefano Zampini 2434b3d85658SStefano 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); 24350e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24360e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 243770022509SStefano Zampini 2438984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 243974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2440785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2441984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 244274d5cdf7SStefano Zampini 2443984c4197SStefano Zampini j = total_primal_vertices; 244474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2445b3d85658SStefano Zampini cum = total_primal_vertices; 24469162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24474641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2448b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2449b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2450b3d85658SStefano Zampini cum++; 24519162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 245274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 245374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 245474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 245574d5cdf7SStefano Zampini } 24569162d606SStefano Zampini j += constraints_n[i]; 2457674ae819SStefano Zampini } 2458674ae819SStefano Zampini } 2459674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2460674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2461088faed8SStefano Zampini 2462674ae819SStefano Zampini /* set values in constraint matrix */ 2463984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 24640e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2465674ae819SStefano Zampini } 2466984c4197SStefano Zampini total_counts = total_primal_vertices; 24679162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24684641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 24699162d606SStefano Zampini PetscInt *cols; 24709162d606SStefano Zampini 24719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 24729162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 24739162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24749162d606SStefano Zampini PetscInt row = total_counts+k; 24759162d606SStefano Zampini PetscScalar *vals; 24769162d606SStefano Zampini 24779162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 24789162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 24799162d606SStefano Zampini } 24809162d606SStefano Zampini total_counts += constraints_n[i]; 2481674ae819SStefano Zampini } 2482674ae819SStefano Zampini } 2483674ae819SStefano Zampini /* assembling */ 2484674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2485674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2486088faed8SStefano Zampini 2487984c4197SStefano Zampini /* 248845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2489984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2490984c4197SStefano Zampini */ 2491674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2492674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2493026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2494984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2495984c4197SStefano Zampini /* working stuff for GEQRF */ 249681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2497984c4197SStefano Zampini PetscBLASInt lqr_work; 2498984c4197SStefano Zampini /* working stuff for UNGQR */ 2499984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2500984c4197SStefano Zampini PetscBLASInt lgqr_work; 2501984c4197SStefano Zampini /* working stuff for TRTRS */ 2502984c4197SStefano Zampini PetscScalar *trs_rhs; 25033f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2504984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2505984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2506984c4197SStefano Zampini PetscScalar *start_vals; 2507984c4197SStefano Zampini /* working stuff for values insertion */ 25084641a718SStefano Zampini PetscBT is_primal; 250964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2510906d46d4SStefano Zampini /* matrix sizes */ 2511906d46d4SStefano Zampini PetscInt global_size,local_size; 2512906d46d4SStefano Zampini /* temporary change of basis */ 2513906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2514cf5a6209SStefano Zampini /* extra space for debugging */ 2515cf5a6209SStefano Zampini PetscScalar *dbg_work; 2516984c4197SStefano Zampini 2517906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2518906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 251916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2520bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2521906d46d4SStefano Zampini /* nonzeros for local mat */ 2522bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2523bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25249162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2525a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25269162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2527a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25289162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2529a717540cSStefano Zampini } else { 25309162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25319162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2532a717540cSStefano Zampini } 2533a717540cSStefano Zampini } 2534a717540cSStefano Zampini } 2535906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2536bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2537a717540cSStefano Zampini /* Set initial identity in the matrix */ 2538bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2539906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2540a717540cSStefano Zampini } 2541a717540cSStefano Zampini 2542a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2543a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2544a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2545a717540cSStefano Zampini } 2546a717540cSStefano Zampini 2547a717540cSStefano Zampini 2548a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2549a717540cSStefano Zampini /* 2550a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2551a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2552a717540cSStefano Zampini 2553a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2554a717540cSStefano Zampini 2555a6b551f4SStefano 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) 2556a6b551f4SStefano Zampini 2557a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2558a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2559a717540cSStefano Zampini | ... | 2560a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2561a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2562a717540cSStefano Zampini 2563a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2564a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2565a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2566a6b551f4SStefano Zampini 2567a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2568a717540cSStefano Zampini */ 2569a717540cSStefano Zampini if (qr_needed) { 2570984c4197SStefano Zampini /* space to store Q */ 2571854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2572984c4197SStefano Zampini /* first we issue queries for optimal work */ 25733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25743f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 25753f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2576984c4197SStefano Zampini lqr_work = -1; 25773f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2578984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2579984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2580785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2581984c4197SStefano Zampini lgqr_work = -1; 25823f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 25843f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 25853f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25863f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 25873f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2588984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2589984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2590785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2591984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2592785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2593984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2594785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2595a717540cSStefano Zampini /* allocating workspace for check */ 2596a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2597cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2598a717540cSStefano Zampini } 2599a717540cSStefano Zampini } 2600984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26014641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2602473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26030e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 260439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 260539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26064641a718SStefano Zampini } 260739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 260839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 260939e2fb2aSStefano Zampini } 261039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2611984c4197SStefano Zampini 2612a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26139162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26149162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26154641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2616984c4197SStefano Zampini /* get constraint info */ 26179162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2618984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2619984c4197SStefano Zampini 2620984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26219162d606SStefano 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); 2622674ae819SStefano Zampini } 2623984c4197SStefano Zampini 2624fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2625a717540cSStefano Zampini 2626a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2627a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26289162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2629a717540cSStefano Zampini } 2630984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26319162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2632984c4197SStefano Zampini 2633984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26353f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26363f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2637674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2639984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2640674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2641984c4197SStefano Zampini 2642984c4197SStefano Zampini /* explictly compute R^-T */ 2643984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2644984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 26453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 26473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26483f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2649984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26503f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2651984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2652984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2653984c4197SStefano Zampini 2654a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 26553f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2659984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26603f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2661984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2662984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2663984c4197SStefano Zampini 2664984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2665984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2666984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 26673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26703f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 26723f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2673984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26749162d606SStefano 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)); 2675984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 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 /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 26799162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2680984c4197SStefano Zampini /* insert cols for primal dofs */ 2681984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2682984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 26839162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2684906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2685984c4197SStefano Zampini } 2686984c4197SStefano Zampini /* insert cols for dual dofs */ 2687984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 26889162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2689984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 26909162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2691906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2692984c4197SStefano Zampini j++; 2693674ae819SStefano Zampini } 2694674ae819SStefano Zampini } 2695984c4197SStefano Zampini 2696984c4197SStefano Zampini /* check change of basis */ 2697984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2698984c4197SStefano Zampini PetscInt ii,jj; 2699984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2700c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2701c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2702c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2703c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2704c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2705c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2706984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2707cf5a6209SStefano 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)); 2708984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2709984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2710984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2711cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2712cf5a6209SStefano 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; 2713674ae819SStefano Zampini } 2714674ae819SStefano Zampini } 2715984c4197SStefano Zampini if (!valid_qr) { 271622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2717984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2718984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2719cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2720cf5a6209SStefano 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])); 2721674ae819SStefano Zampini } 2722cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2723cf5a6209SStefano 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])); 2724984c4197SStefano Zampini } 2725984c4197SStefano Zampini } 2726984c4197SStefano Zampini } 2727674ae819SStefano Zampini } else { 272822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2729674ae819SStefano Zampini } 2730674ae819SStefano Zampini } 2731a717540cSStefano Zampini } else { /* simple transformation block */ 2732a717540cSStefano Zampini PetscInt row,col; 2733a6b551f4SStefano Zampini PetscScalar val,norm; 2734a6b551f4SStefano Zampini 2735a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27369162d606SStefano 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)); 2737a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27389162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27399162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2740bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 27419162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2742906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 27439162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2744a717540cSStefano Zampini } else { 2745a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 27469162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2747a717540cSStefano Zampini if (row != col) { 27489162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2749a717540cSStefano Zampini } else { 27509162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2751a717540cSStefano Zampini } 2752906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2753a717540cSStefano Zampini } 2754a717540cSStefano Zampini } 2755a717540cSStefano Zampini } 275698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 275722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2758a717540cSStefano Zampini } 2759674ae819SStefano Zampini } 2760984c4197SStefano Zampini } else { 2761984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27629162d606SStefano 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); 2763674ae819SStefano Zampini } 2764674ae819SStefano Zampini } 2765674ae819SStefano Zampini } 2766a717540cSStefano Zampini 2767a717540cSStefano Zampini /* free workspace */ 2768a717540cSStefano Zampini if (qr_needed) { 2769984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2770cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2771984c4197SStefano Zampini } 2772984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2773984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2774984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2775984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2776984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2777674ae819SStefano Zampini } 2778a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2779906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2780906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2781906d46d4SStefano Zampini 2782906d46d4SStefano Zampini /* assembling of global change of variable */ 2783bbb9e6c6SStefano Zampini { 2784bbb9e6c6SStefano Zampini Mat tmat; 278516f15bc4SStefano Zampini PetscInt bs; 278616f15bc4SStefano Zampini 2787906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2788906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2789bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2790bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2791bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2792bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 279316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 279416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2795906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2796bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2797bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2798bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2799bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2800bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2801bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2802bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2803bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2804bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2805906d46d4SStefano Zampini } 2806906d46d4SStefano Zampini /* check */ 2807906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2808906d46d4SStefano Zampini PetscReal error; 2809906d46d4SStefano Zampini Vec x,x_change; 2810906d46d4SStefano Zampini 2811906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2812906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2813906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2814906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2815bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2816bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2817bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2818bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2819bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2820906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2821906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2822906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2823906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2824bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2825906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2826906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2827906d46d4SStefano Zampini } 2828b96c3477SStefano Zampini 2829b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2830b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2831b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2832b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2833ac632422SStefano Zampini Mat S_new,tmat; 2834bbb9e6c6SStefano Zampini IS is_all_N; 2835bbb9e6c6SStefano Zampini 2836bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 2837d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr); 2838bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2839ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2840b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2841ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2842ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2843ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2844ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2845ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2846b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2847ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2848ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2849ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2850ac632422SStefano Zampini } 2851b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2852b96c3477SStefano Zampini } 2853b96c3477SStefano Zampini } 2854906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2855906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2856b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2857b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2858b9b85e73SStefano Zampini } 2859906d46d4SStefano Zampini 2860906d46d4SStefano Zampini /* set up change of basis context */ 2861906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2862906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2863906d46d4SStefano Zampini 2864906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2865906d46d4SStefano Zampini PetscInt global_size,local_size; 2866906d46d4SStefano Zampini 2867906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2868906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2869906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2871906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2874906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2875906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2876906d46d4SStefano Zampini } else { 2877906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2878906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2879906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2880906d46d4SStefano Zampini } 2881906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2882906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2883906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2884906d46d4SStefano Zampini } else { 2885906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2886906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2887906d46d4SStefano Zampini } 2888906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2889906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2890906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2891906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2892b9b85e73SStefano Zampini } 2893a717540cSStefano Zampini 2894727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2895727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2896727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 28970e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 28980e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 28990e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 29000e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2901c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2902727cdba6SStefano Zampini } 29030e6343abSStefano Zampini } 29040e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2905727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2906727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2907727cdba6SStefano Zampini 2908a717540cSStefano Zampini /* flush dbg viewer */ 2909b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2910b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2911b8ffe317SStefano Zampini } 2912a717540cSStefano Zampini 2913e310c8b4SStefano Zampini /* free workspace */ 2914a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29154641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 291608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29179162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29189162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 291908122e43SStefano Zampini } else { 29209162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29219162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29229162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 292308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 292408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29259162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29269162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 292708122e43SStefano Zampini } 2928674ae819SStefano Zampini PetscFunctionReturn(0); 2929674ae819SStefano Zampini } 2930674ae819SStefano Zampini 2931674ae819SStefano Zampini #undef __FUNCT__ 2932674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2933674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2934674ae819SStefano Zampini { 2935674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2936674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2937674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29387fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2939674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2940674ae819SStefano Zampini 2941674ae819SStefano Zampini PetscFunctionBegin; 29428e61c736SStefano Zampini /* Reset previously computed graph */ 29438e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2944674ae819SStefano Zampini /* Init local Graph struct */ 29457fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 29467fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 2947674ae819SStefano Zampini 2948575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2949575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2950575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2951575ad6abSStefano Zampini } 29529577ea80SStefano Zampini 2953674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 29544d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 29554d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 29564d379d7bSStefano Zampini PetscInt nvtxs; 2957e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 2958674ae819SStefano Zampini 29594d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 29602fffb893SStefano Zampini 29612fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29622fffb893SStefano Zampini if (flg_row) { 29634d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2964b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 29652fffb893SStefano Zampini } 29662fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29679b28b941SStefano 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 */ 29684d379d7bSStefano Zampini IS is_dummy; 29694d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 29704d379d7bSStefano Zampini PetscInt j,sum; 29714d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 29724d379d7bSStefano Zampini const PetscInt *idxs; 29734d379d7bSStefano Zampini PCBDDCGraph graph; 29744d379d7bSStefano Zampini PetscBT is_on_boundary; 29754d379d7bSStefano Zampini 29764d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 29774d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 29784d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 29794d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 29807fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 29814d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 2982e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2983e496cd5dSStefano Zampini if (flg_row) { 29844d379d7bSStefano Zampini graph->xadj = xadj; 29854d379d7bSStefano Zampini graph->adjncy = adjncy; 2986e496cd5dSStefano Zampini } 29874d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 29884d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2989e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29904d379d7bSStefano Zampini 29914d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 29929b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 29934d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29944d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 29954d379d7bSStefano Zampini } 29964d379d7bSStefano Zampini } 29974d379d7bSStefano Zampini 2998e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 29994d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30004d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30014d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30024d379d7bSStefano Zampini } 30034d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30044d379d7bSStefano Zampini 3005e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30064d379d7bSStefano Zampini sum = 0; 30074d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30084d379d7bSStefano Zampini PetscInt sizecc = 0; 30094d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30104d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30114d379d7bSStefano Zampini sizecc++; 30124d379d7bSStefano Zampini } 30134d379d7bSStefano Zampini } 30144d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30154d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30164d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30174d379d7bSStefano Zampini } 30184d379d7bSStefano Zampini } 30194d379d7bSStefano Zampini sum += sizecc*sizecc; 30204d379d7bSStefano Zampini } 30214d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30224d379d7bSStefano Zampini sum = 0; 3023e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30244d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30254d379d7bSStefano Zampini cxadj[i] = sum; 30264d379d7bSStefano Zampini sum += temp; 30274d379d7bSStefano Zampini } 3028e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30294d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30304d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30314d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30324d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30334d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30344d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30354d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30364d379d7bSStefano Zampini sizecc++; 30374d379d7bSStefano Zampini } 30384d379d7bSStefano Zampini } 30394d379d7bSStefano Zampini } 30404d379d7bSStefano Zampini } 30414d379d7bSStefano Zampini } 30429b28b941SStefano Zampini if (sum) { 3043e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 30444d379d7bSStefano Zampini } else { 30454d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 30464d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 30474d379d7bSStefano Zampini } 30484d379d7bSStefano Zampini graph->xadj = 0; 30494d379d7bSStefano Zampini graph->adjncy = 0; 30504d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 30514d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 30524d379d7bSStefano Zampini } 3053674ae819SStefano Zampini } 30549b28b941SStefano Zampini if (pcbddc->dbg_flag) { 30559b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30569b28b941SStefano Zampini } 3057674ae819SStefano Zampini 305863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3059674ae819SStefano Zampini vertex_size = 1; 306063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 306163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 306295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 306363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3064a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 306563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3066674ae819SStefano Zampini } 306763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 306863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 306963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3070674ae819SStefano Zampini } 307163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3072674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 307363602bcaSStefano Zampini } else { 307463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 307563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3076854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 307763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 307863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 307963602bcaSStefano Zampini } 308063602bcaSStefano Zampini } 3081674ae819SStefano Zampini } 3082674ae819SStefano Zampini 3083674ae819SStefano Zampini /* Setup of Graph */ 3084785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3085a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3086785d1243SStefano Zampini } 3087785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3088a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3089785d1243SStefano Zampini } 3090302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3091674ae819SStefano Zampini 3092674ae819SStefano Zampini /* Graph's connected components analysis */ 3093674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3094674ae819SStefano Zampini 3095674ae819SStefano Zampini /* print some info to stdout */ 3096674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3097302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3098674ae819SStefano Zampini } 3099fb180af4SStefano Zampini 3100fb180af4SStefano Zampini /* mark topography has done */ 3101fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3102674ae819SStefano Zampini PetscFunctionReturn(0); 3103674ae819SStefano Zampini } 3104674ae819SStefano Zampini 3105674ae819SStefano Zampini #undef __FUNCT__ 3106674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3107674ae819SStefano 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[]) 3108674ae819SStefano Zampini { 3109674ae819SStefano Zampini Vec local_vec,global_vec; 3110674ae819SStefano Zampini IS seqis,paris; 3111674ae819SStefano Zampini VecScatter scatter_ctx; 3112674ae819SStefano Zampini PetscScalar *array; 3113674ae819SStefano Zampini PetscInt *temp_global_dofs; 3114674ae819SStefano Zampini PetscScalar globalsum; 3115674ae819SStefano Zampini PetscInt i,j,s; 311617ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 311717ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 311817ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3119674ae819SStefano Zampini PetscBool first_found; 3120674ae819SStefano Zampini PetscErrorCode ierr; 3121674ae819SStefano Zampini 3122674ae819SStefano Zampini PetscFunctionBegin; 3123674ae819SStefano Zampini /* mpi buffers */ 3124b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3125b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3126674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 312717ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3128674ae819SStefano Zampini /* get maximum size of subset */ 3129785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3130674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3131674ae819SStefano Zampini max_local = 0; 3132b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3133674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3134674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3135674ae819SStefano Zampini } 3136674ae819SStefano Zampini } 3137b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3138674ae819SStefano Zampini max_global++; 3139674ae819SStefano Zampini max_local = 0; 3140b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3141674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3142674ae819SStefano Zampini max_local = local_dofs[i]; 3143674ae819SStefano Zampini } 3144674ae819SStefano Zampini } 3145674ae819SStefano Zampini max_local++; 3146674ae819SStefano Zampini /* allocate workspace */ 3147674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3148674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3149674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3150674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3151674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3152674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3153674ae819SStefano Zampini /* create scatter */ 3154674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3155674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3156674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3157674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3158674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3159674ae819SStefano Zampini /* init array */ 3160674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3161674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3162674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3163674ae819SStefano Zampini if (local_dofs_mult) { 3164674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3165674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3166674ae819SStefano Zampini } 3167674ae819SStefano Zampini } else { 3168674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3169674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3170674ae819SStefano Zampini } 3171674ae819SStefano Zampini } 3172674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3173674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3174674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3175674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3176674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 31775b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3178674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3179674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3180674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3181674ae819SStefano Zampini nlocals = 0; 3182674ae819SStefano Zampini first_index = -1; 3183674ae819SStefano Zampini first_found = PETSC_FALSE; 3184674ae819SStefano Zampini for (i=0;i<s;i++) { 3185b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3186674ae819SStefano Zampini first_found = PETSC_TRUE; 3187674ae819SStefano Zampini first_index = i; 3188674ae819SStefano Zampini } 31895b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3190674ae819SStefano Zampini } 3191674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3192674ae819SStefano Zampini if (!rank_prec_comm) { 3193674ae819SStefano Zampini dof_displs[0]=0; 3194674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3195674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3196674ae819SStefano Zampini } 3197674ae819SStefano Zampini } 3198674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3199674ae819SStefano Zampini if (first_found) { 3200674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3201674ae819SStefano Zampini old_index = first_index; 3202674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3203b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3204674ae819SStefano Zampini array[i] += array[old_index]; 3205674ae819SStefano Zampini old_index = i; 3206674ae819SStefano Zampini } 3207674ae819SStefano Zampini } 3208674ae819SStefano Zampini } 3209674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3210674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3211674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3212674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3213674ae819SStefano Zampini /* get global ordering of local dofs */ 32142b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3215674ae819SStefano Zampini if (local_dofs_mult) { 3216674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32175b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3218674ae819SStefano Zampini } 3219674ae819SStefano Zampini } else { 3220674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32215b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3222674ae819SStefano Zampini } 3223674ae819SStefano Zampini } 32242b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3225674ae819SStefano Zampini /* free workspace */ 3226674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3227674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3228674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 322917ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3230674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3231674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3232674ae819SStefano Zampini PetscFunctionReturn(0); 3233674ae819SStefano Zampini } 32349a7d3425SStefano Zampini 32359a7d3425SStefano Zampini #undef __FUNCT__ 32369a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32379a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32389a7d3425SStefano Zampini { 32399a7d3425SStefano Zampini PetscInt i,j; 32409a7d3425SStefano Zampini PetscScalar *alphas; 32419a7d3425SStefano Zampini PetscErrorCode ierr; 32429a7d3425SStefano Zampini 32439a7d3425SStefano Zampini PetscFunctionBegin; 32449a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3245785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 32469a7d3425SStefano Zampini for (i=0;i<n;i++) { 32479a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 32489a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 32499a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 32509a7d3425SStefano Zampini } 32519a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 32529a7d3425SStefano Zampini PetscFunctionReturn(0); 32539a7d3425SStefano Zampini } 32549a7d3425SStefano Zampini 3255e7931f94SStefano Zampini #undef __FUNCT__ 325670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 325728143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3258e7931f94SStefano Zampini { 3259e7931f94SStefano Zampini Mat subdomain_adj; 3260e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3261e7931f94SStefano Zampini MatPartitioning partitioner; 3262e7931f94SStefano Zampini Mat_IS *matis; 3263e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3264e7931f94SStefano Zampini PetscInt prank; 3265e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3266e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3267e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 32683837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3269e7931f94SStefano Zampini PetscErrorCode ierr; 32702b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3271e7931f94SStefano Zampini PetscSubcomm subcomm; 3272a57a6d2fSStefano Zampini 3273e7931f94SStefano Zampini PetscFunctionBegin; 32742b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 32752b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 32762b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3277e7931f94SStefano Zampini 3278e7931f94SStefano Zampini /* Get info on mapping */ 3279e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3280e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3281e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3282e7931f94SStefano Zampini 3283e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3284785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3285e7931f94SStefano Zampini xadj[0] = 0; 3286e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3287785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3288785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3289e7931f94SStefano Zampini 32902b510759SStefano Zampini if (threshold) { 3291d023bfaeSStefano Zampini PetscInt xadj_count = 0; 32922b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3293d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3294d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3295d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3296d023bfaeSStefano Zampini xadj_count++; 3297e7931f94SStefano Zampini } 3298e7931f94SStefano Zampini } 3299d023bfaeSStefano Zampini xadj[1] = xadj_count; 3300c8587f34SStefano Zampini } else { 3301e7931f94SStefano Zampini if (xadj[1]) { 3302e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3303e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3304c8587f34SStefano Zampini } 3305e7931f94SStefano Zampini } 3306e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3307e7931f94SStefano Zampini if (use_square) { 3308e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3309e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3310e7931f94SStefano Zampini } 3311e7931f94SStefano Zampini } 3312e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3313e7931f94SStefano Zampini 33143837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3315e7931f94SStefano Zampini 3316e7931f94SStefano Zampini /* 3317e7931f94SStefano Zampini Restrict work on active processes only. 3318e7931f94SStefano Zampini */ 3319e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3320e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3321e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33222b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3323d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3324e7931f94SStefano Zampini if (color) { 3325e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3326e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3327e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3328c8587f34SStefano Zampini } else { 332928143c3dSStefano Zampini PetscInt coarsening_ratio; 3330306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3331785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3332e7931f94SStefano Zampini prank = rank; 3333306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33348002ef2cSStefano Zampini /* 3335e7931f94SStefano Zampini for (i=0;i<size;i++) { 3336e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3337c8587f34SStefano Zampini } 33388002ef2cSStefano Zampini */ 3339e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3340e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3341c8587f34SStefano Zampini } 3342e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3343306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 334422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3345e7931f94SStefano Zampini 3346e7931f94SStefano Zampini /* Partition */ 3347306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3348e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3349e7931f94SStefano Zampini if (use_vwgt) { 33503837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3351e7931f94SStefano Zampini v_wgt[0] = local_size; 3352e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3353c8587f34SStefano Zampini } 335428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 335528143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 335628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3357e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3358e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 335922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3360e7931f94SStefano Zampini 3361e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 336228143c3dSStefano Zampini if (contiguous) { 336328143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 336428143c3dSStefano Zampini } else { 336584ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 336628143c3dSStefano Zampini } 3367e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3368e7931f94SStefano Zampini /* clean up */ 3369e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3370e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3371e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3372e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3373e7931f94SStefano Zampini } 3374e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3375e7931f94SStefano Zampini 3376e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3377e7931f94SStefano Zampini i = 1; 3378e7931f94SStefano Zampini if (color) i=0; 3379e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3380e7931f94SStefano Zampini 3381e7931f94SStefano Zampini /* get back IS */ 3382e7931f94SStefano Zampini *is_sends = ranks_send_to; 3383e7931f94SStefano Zampini PetscFunctionReturn(0); 3384e7931f94SStefano Zampini } 3385e7931f94SStefano Zampini 3386e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3387e7931f94SStefano Zampini 3388e7931f94SStefano Zampini #undef __FUNCT__ 3389e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 339028143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3391e7931f94SStefano Zampini { 339270cf5478SStefano Zampini Mat local_mat; 3393e7931f94SStefano Zampini Mat_IS *matis; 3394e7931f94SStefano Zampini IS is_sends_internal; 33959d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 339628143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 33979d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3398e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3399e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3400e7931f94SStefano Zampini const PetscInt* is_indices; 3401e7931f94SStefano Zampini MatType new_local_type; 3402e7931f94SStefano Zampini /* buffers */ 3403e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 340428143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34059d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3406e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3407e7931f94SStefano Zampini /* MPI */ 340828143c3dSStefano Zampini MPI_Comm comm,comm_n; 340928143c3dSStefano Zampini PetscSubcomm subcomm; 3410e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 341128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 341228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 341328143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 341428143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 341528143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3416e7931f94SStefano Zampini PetscErrorCode ierr; 3417e7931f94SStefano Zampini 3418e7931f94SStefano Zampini PetscFunctionBegin; 341928143c3dSStefano Zampini /* TODO: add missing checks */ 342028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 342128143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 342228143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 342328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3424e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 342528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3426e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3427e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3428e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3429e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3430e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 343128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 343270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 343370cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 343428143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 343570cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 343670cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 343770cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 343870cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 343970cf5478SStefano Zampini } 3440e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3441e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3442e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3443e7931f94SStefano Zampini if (!is_sends) { 344428143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 344528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 344628143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3447c8587f34SStefano Zampini } else { 3448e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3449e7931f94SStefano Zampini is_sends_internal = is_sends; 3450c8587f34SStefano Zampini } 3451e7931f94SStefano Zampini 3452e7931f94SStefano Zampini /* get pointer of MATIS data */ 3453e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3454e7931f94SStefano Zampini 3455e7931f94SStefano Zampini /* get comm */ 3456a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3457e7931f94SStefano Zampini 3458e7931f94SStefano Zampini /* compute number of sends */ 3459e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3460e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3461e7931f94SStefano Zampini 3462e7931f94SStefano Zampini /* compute number of receives */ 3463e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3464785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3465e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3466e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3467e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3468e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3469e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3470e7931f94SStefano Zampini 347128143c3dSStefano Zampini /* restrict comm if requested */ 347228143c3dSStefano Zampini subcomm = 0; 347328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 347428143c3dSStefano Zampini if (restrict_comm) { 3475779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3476779c1cceSStefano Zampini 347728143c3dSStefano Zampini color = 0; 3478779c1cceSStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */ 347928143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 348028143c3dSStefano Zampini subcommsize = commsize - subcommsize; 348128143c3dSStefano Zampini /* check if reuse has been requested */ 348228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 348328143c3dSStefano Zampini if (*mat_n) { 348428143c3dSStefano Zampini PetscMPIInt subcommsize2; 348528143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 348628143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 348728143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 348828143c3dSStefano Zampini } else { 348928143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 349028143c3dSStefano Zampini } 349128143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3492779c1cceSStefano Zampini PetscMPIInt rank; 3493779c1cceSStefano Zampini 3494779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 349528143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 349628143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 349728143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3498306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 349928143c3dSStefano Zampini } 350028143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 350128143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 350228143c3dSStefano Zampini } else { 350328143c3dSStefano Zampini comm_n = comm; 350428143c3dSStefano Zampini } 350528143c3dSStefano Zampini 3506e7931f94SStefano Zampini /* prepare send/receive buffers */ 3507785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3508e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3509785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3510e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 351128143c3dSStefano Zampini if (nis) { 3512854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 351328143c3dSStefano Zampini } 3514e7931f94SStefano Zampini 351528143c3dSStefano Zampini /* Get data from local matrices */ 3516e7931f94SStefano Zampini if (!isdense) { 3517a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3518e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3519e7931f94SStefano Zampini /* 3520e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3521e7931f94SStefano Zampini send_buffer_idxs should contain: 3522e7931f94SStefano Zampini - MatType_PRIVATE type 3523e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3524e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3525e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3526e7931f94SStefano Zampini */ 3527e7931f94SStefano Zampini } else { 3528e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3529e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3530854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3531e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3532e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3533e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3534e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3535e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3536e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3537e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3538e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3539e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3540c8587f34SStefano Zampini } 3541c8587f34SStefano Zampini } 3542e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 354328143c3dSStefano Zampini /* additional is (if any) */ 354428143c3dSStefano Zampini if (nis) { 354528143c3dSStefano Zampini PetscMPIInt psum; 354628143c3dSStefano Zampini PetscInt j; 354728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 354828143c3dSStefano Zampini PetscInt plen; 354928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 355028143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 355128143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 355228143c3dSStefano Zampini } 3553854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 355428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 355528143c3dSStefano Zampini PetscInt plen; 355628143c3dSStefano Zampini const PetscInt *is_array_idxs; 355728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 355828143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 355928143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 356028143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 356128143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 356228143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 356328143c3dSStefano Zampini } 356428143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 356528143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 356628143c3dSStefano Zampini } 356728143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 356828143c3dSStefano Zampini } 356928143c3dSStefano Zampini 3570e7931f94SStefano Zampini buf_size_idxs = 0; 3571e7931f94SStefano Zampini buf_size_vals = 0; 357228143c3dSStefano Zampini buf_size_idxs_is = 0; 3573e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3574e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3575e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 357628143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3577e7931f94SStefano Zampini } 3578785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3579785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 358095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3581e7931f94SStefano Zampini 3582e7931f94SStefano Zampini /* get new tags for clean communications */ 3583e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3584e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 358528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3586e7931f94SStefano Zampini 3587e7931f94SStefano Zampini /* allocate for requests */ 3588785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3589785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 359095ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3591785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3592785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 359395ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3594e7931f94SStefano Zampini 3595e7931f94SStefano Zampini /* communications */ 3596e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3597e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 359828143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3599e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3600e7931f94SStefano Zampini source_dest = onodes[i]; 3601e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3602e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3603e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3604e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 360528143c3dSStefano Zampini if (nis) { 360628143c3dSStefano 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); 360728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 360828143c3dSStefano Zampini } 3609e7931f94SStefano Zampini } 3610e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3611e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3612e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3613e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 361428143c3dSStefano Zampini if (nis) { 361528143c3dSStefano 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); 361628143c3dSStefano Zampini } 3617e7931f94SStefano Zampini } 3618e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3619e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3620e7931f94SStefano Zampini 3621e7931f94SStefano Zampini /* assemble new l2g map */ 3622e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3623e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36249d30be91SStefano Zampini new_local_rows = 0; 3625e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36269d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3627e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3628e7931f94SStefano Zampini } 36299d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3630e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36319d30be91SStefano Zampini new_local_rows = 0; 3632e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36339d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36349d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3635e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3636e7931f94SStefano Zampini } 36379d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36389d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3639e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3640e7931f94SStefano Zampini 3641e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3642e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3643e7931f94SStefano 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) */ 3644e7931f94SStefano Zampini if (n_recvs) { 364528143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3646e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3647e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3648e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3649e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3650e7931f94SStefano Zampini break; 3651e7931f94SStefano Zampini } 3652e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3653e7931f94SStefano Zampini } 3654e7931f94SStefano Zampini switch (new_local_type_private) { 365528143c3dSStefano Zampini case MATDENSE_PRIVATE: 365628143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3657e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3658e7931f94SStefano Zampini bs = 1; 365928143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 366028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 366128143c3dSStefano Zampini bs = 1; 366228143c3dSStefano Zampini } 3663e7931f94SStefano Zampini break; 3664e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3665e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3666e7931f94SStefano Zampini bs = 1; 3667e7931f94SStefano Zampini break; 3668e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3669e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3670e7931f94SStefano Zampini break; 3671e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3672e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3673e7931f94SStefano Zampini break; 3674e7931f94SStefano Zampini default: 36759d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3676e7931f94SStefano Zampini break; 3677e7931f94SStefano Zampini } 367828143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 367928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 368028143c3dSStefano Zampini bs = 1; 3681e7931f94SStefano Zampini } 3682e7931f94SStefano Zampini 368370cf5478SStefano Zampini /* create MATIS object if needed */ 368470cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3685e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 368628143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 368770cf5478SStefano Zampini } else { 368870cf5478SStefano Zampini /* it also destroys the local matrices */ 368970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 369070cf5478SStefano Zampini } 369170cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3692e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 36939d30be91SStefano Zampini 36949d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 36959d30be91SStefano Zampini 36969d30be91SStefano Zampini /* Global to local map of received indices */ 36979d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 36989d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 36999d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37009d30be91SStefano Zampini 37019d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37029d30be91SStefano Zampini buf_size_idxs = 0; 37039d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37049d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37059d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37069d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37079d30be91SStefano Zampini } 37089d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37099d30be91SStefano Zampini 37109d30be91SStefano Zampini /* set preallocation */ 37119d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37129d30be91SStefano Zampini if (!newisdense) { 37139d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37149d30be91SStefano Zampini 37159d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37169d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37179d30be91SStefano Zampini if (n_recvs) { 37189d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37199d30be91SStefano Zampini } 37209d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37219d30be91SStefano Zampini PetscInt j; 37229d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37239d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37249d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37259d30be91SStefano Zampini } 37269d30be91SStefano Zampini } else { 37279d30be91SStefano Zampini /* TODO */ 37289d30be91SStefano Zampini } 37299d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37309d30be91SStefano Zampini } 37319d30be91SStefano Zampini if (new_local_nnz) { 37329d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37339d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37349d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37359d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37369d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37379d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37389d30be91SStefano Zampini } else { 37399d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37409d30be91SStefano Zampini } 37419d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 37429d30be91SStefano Zampini } else { 37439d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37449d30be91SStefano Zampini } 3745e7931f94SStefano Zampini 3746e7931f94SStefano Zampini /* set values */ 3747e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 37489d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3749e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3750e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3751e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 37529d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3753e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3754e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3755e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 375628143c3dSStefano Zampini } else { 375728143c3dSStefano Zampini /* TODO */ 3758e7931f94SStefano Zampini } 3759e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3760e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3761e7931f94SStefano Zampini } 3762e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3763e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 376470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 376570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 37669d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 37679d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3768e7931f94SStefano Zampini 3769dfd14d43SStefano Zampini #if 0 377028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3771e7931f94SStefano Zampini Vec lvec,rvec; 3772e7931f94SStefano Zampini PetscReal infty_error; 3773e7931f94SStefano Zampini 37742a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3775e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3776e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3777e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 377870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3779e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3780e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3781e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3782e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3783e7931f94SStefano Zampini } 378428143c3dSStefano Zampini #endif 3785e7931f94SStefano Zampini 378628143c3dSStefano Zampini /* assemble new additional is (if any) */ 378728143c3dSStefano Zampini if (nis) { 378828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 378928143c3dSStefano Zampini 379028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3791854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 379228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 379328143c3dSStefano Zampini psum = 0; 379428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 379528143c3dSStefano Zampini for (j=0;j<nis;j++) { 379628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 379728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 379828143c3dSStefano Zampini psum += plen; 379928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 380028143c3dSStefano Zampini } 380128143c3dSStefano Zampini } 3802854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3803854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 380428143c3dSStefano Zampini for (i=1;i<nis;i++) { 380528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 380628143c3dSStefano Zampini } 380728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 380828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 380928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 381028143c3dSStefano Zampini for (j=0;j<nis;j++) { 381128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 381228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 381328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 381428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 381528143c3dSStefano Zampini } 381628143c3dSStefano Zampini } 381728143c3dSStefano Zampini for (i=0;i<nis;i++) { 381828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 381928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 382028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 382128143c3dSStefano Zampini } 382228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 382328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 382428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 382528143c3dSStefano Zampini } 3826e7931f94SStefano Zampini /* free workspace */ 382728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3828e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3829e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3830e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3831e7931f94SStefano Zampini if (isdense) { 3832e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3833e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3834e7931f94SStefano Zampini } else { 3835e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3836e7931f94SStefano Zampini } 383728143c3dSStefano Zampini if (nis) { 383828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 383928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 384028143c3dSStefano Zampini } 3841e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3842e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 384328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3844e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3845e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 384628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3847e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3848e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3849e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3850e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3851e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 385228143c3dSStefano Zampini if (nis) { 385328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 385428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 385528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 385628143c3dSStefano Zampini } 385728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 385828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 385928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 386028143c3dSStefano Zampini for (i=0;i<nis;i++) { 386128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 386228143c3dSStefano Zampini } 386328143c3dSStefano Zampini } 3864e7931f94SStefano Zampini PetscFunctionReturn(0); 3865e7931f94SStefano Zampini } 3866a57a6d2fSStefano Zampini 386712edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3868af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 386912edc857SStefano Zampini 3870c8587f34SStefano Zampini #undef __FUNCT__ 3871c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3872c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3873c8587f34SStefano Zampini { 3874c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3875c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 387620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 38779881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 387820a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 38796e683305SStefano Zampini IS coarse_is,*isarray; 38806e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 38816e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3882f9eb5b7dSStefano Zampini PC pc_temp; 3883c8587f34SStefano Zampini PCType coarse_pc_type; 3884c8587f34SStefano Zampini KSPType coarse_ksp_type; 3885f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 38864f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 38876e683305SStefano Zampini Mat t_coarse_mat_is; 38886e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 38896e683305SStefano Zampini PetscMPIInt all_procs; 389074e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 389168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 389222bc73bbSStefano Zampini PetscScalar *array; 38939881197aSStefano Zampini PetscErrorCode ierr; 3894fdc09c96SStefano Zampini 3895c8587f34SStefano Zampini PetscFunctionBegin; 3896c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 389768457ee5SStefano 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 */ 3898fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 38995a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3900fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3901f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3902f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3903f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3904fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3905fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3906727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3907fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3908fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3909fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3910f4ddd8eeSStefano Zampini } 3911fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3912fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3913f4ddd8eeSStefano Zampini } 391470cf5478SStefano Zampini /* reset any subassembling information */ 391570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39166e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39176e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3918fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3919f4ddd8eeSStefano Zampini } 3920c8587f34SStefano Zampini 39216e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39222b510759SStefano Zampini im_active = !!(pcis->n); 39232b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39246e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39256e683305SStefano Zampini void_procs = all_procs-active_procs; 39266e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 392774e2c79eSStefano Zampini redist = PETSC_FALSE; 392822bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39296e683305SStefano Zampini csin_ml = PETSC_TRUE; 39306e683305SStefano Zampini ncoarse_ml = void_procs; 3931779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3932779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 393318a45a71SStefano Zampini csin_ds = PETSC_TRUE; 393418a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 393518a45a71SStefano Zampini redist = PETSC_TRUE; 393618a45a71SStefano Zampini } else { 39376e683305SStefano Zampini csin_ds = PETSC_TRUE; 3938779c1cceSStefano Zampini ncoarse_ds = active_procs; 3939779c1cceSStefano Zampini redist = PETSC_TRUE; 394018a45a71SStefano Zampini } 39416e683305SStefano Zampini } else { 39426e683305SStefano Zampini csin_ml = PETSC_FALSE; 39436e683305SStefano Zampini ncoarse_ml = all_procs; 39446e683305SStefano Zampini if (void_procs) { 39456e683305SStefano Zampini csin_ds = PETSC_TRUE; 39466e683305SStefano Zampini ncoarse_ds = void_procs; 39476e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 39486e683305SStefano Zampini } else { 3949779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 395074e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 395174e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 395274e2c79eSStefano Zampini redist = PETSC_TRUE; 395374e2c79eSStefano Zampini } else { 39546e683305SStefano Zampini csin_ds = PETSC_FALSE; 39556e683305SStefano Zampini ncoarse_ds = all_procs; 39566e683305SStefano Zampini } 39576e683305SStefano Zampini } 395874e2c79eSStefano Zampini } 39596e683305SStefano Zampini 39606e683305SStefano Zampini /* 39616e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 39626e683305SStefano Zampini - we have not exceeded the number of levels requested 39636e683305SStefano Zampini - we can actually subassemble the active processes 39646e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 39656e683305SStefano Zampini */ 39666e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 39676e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 39686e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 39696e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 39706e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3971f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 39722b510759SStefano Zampini } else { 3973f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3974c8587f34SStefano Zampini } 3975c8587f34SStefano Zampini } 39766e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 39776e683305SStefano Zampini if (multilevel_allowed) { 39786e683305SStefano Zampini ncoarse = ncoarse_ml; 39796e683305SStefano Zampini csin = csin_ml; 398058da7f69SStefano Zampini redist = PETSC_FALSE; 39816e683305SStefano Zampini } else { 39826e683305SStefano Zampini ncoarse = ncoarse_ds; 39836e683305SStefano Zampini csin = csin_ds; 39846e683305SStefano Zampini } 3985e7931f94SStefano Zampini 3986abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3987abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3988abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3989abbbba34SStefano Zampini 3990abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 399122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 399222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 399322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 399422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3995b9b85e73SStefano Zampini #if 0 3996b9b85e73SStefano Zampini { 3997b9b85e73SStefano Zampini PetscViewer viewer; 3998b9b85e73SStefano Zampini char filename[256]; 3999b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4000b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4001b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4002b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4003b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4004b9b85e73SStefano Zampini } 4005b9b85e73SStefano Zampini #endif 40066e683305SStefano 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); 40076e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40086e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40096e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4010abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4011abbbba34SStefano Zampini 40126e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40136e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40146e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40156e683305SStefano Zampini const PetscInt *idxs; 40166e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40176e683305SStefano Zampini 40186e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40190be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40206e683305SStefano Zampini /* allocate space for temporary storage */ 4021854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4022854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40236e683305SStefano Zampini /* allocate for IS array */ 40246e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40256e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40266e683305SStefano Zampini nis = nisdofs + nisneu; 4027854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40286e683305SStefano Zampini /* dofs splitting */ 40296e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40306e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40316e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40326e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40336e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40346e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40356e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40366e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40376e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40386e683305SStefano Zampini } 40396e683305SStefano Zampini /* neumann boundaries */ 40406e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 40416e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 40426e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 40436e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40446e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40456e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40466e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40476e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 40486e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 40496e683305SStefano Zampini } 40506e683305SStefano Zampini /* free memory */ 40516e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 40526e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 40536e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 40546e683305SStefano Zampini } else { 40556e683305SStefano Zampini nis = 0; 40566e683305SStefano Zampini nisdofs = 0; 40576e683305SStefano Zampini nisneu = 0; 40586e683305SStefano Zampini isarray = NULL; 40596e683305SStefano Zampini } 40606e683305SStefano Zampini /* destroy no longer needed map */ 40616e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 40626e683305SStefano Zampini 40636e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 40646e683305SStefano Zampini coarse_mat_is = NULL; 40656e683305SStefano Zampini if (csin) { 40666e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 406774e2c79eSStefano Zampini if (redist) { 406874e2c79eSStefano Zampini PetscMPIInt rank; 4069779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 407074e2c79eSStefano Zampini 407174e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 407258da7f69SStefano Zampini spc = active_procs/ncoarse; 407358da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4074779c1cceSStefano Zampini if (im_active) { 4075779c1cceSStefano Zampini destsize = 1; 407674e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 407774e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 407874e2c79eSStefano Zampini } else { 407974e2c79eSStefano Zampini dest[0] = rank/(spc+1); 408074e2c79eSStefano Zampini } 408174e2c79eSStefano Zampini } else { 4082779c1cceSStefano Zampini destsize = 0; 40836e683305SStefano Zampini } 4084779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4085779c1cceSStefano Zampini } else if (csin_type_simple) { 40866e683305SStefano Zampini PetscMPIInt rank; 40876e683305SStefano Zampini PetscInt issize,isidx; 4088779c1cceSStefano Zampini 40896e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 40906e683305SStefano Zampini if (im_active) { 40916e683305SStefano Zampini issize = 1; 40926e683305SStefano Zampini isidx = (PetscInt)rank; 40936e683305SStefano Zampini } else { 40946e683305SStefano Zampini issize = 0; 40956e683305SStefano Zampini isidx = -1; 40966e683305SStefano Zampini } 40976e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4098779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 40996e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41006e683305SStefano Zampini } 4101779c1cceSStefano Zampini 4102779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4103779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4104779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4105779c1cceSStefano Zampini PetscInt *coarse_candidates; 4106779c1cceSStefano Zampini const PetscInt* tisindices; 4107779c1cceSStefano Zampini 4108779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4109779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4110779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4111779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4112779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4113779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4114779c1cceSStefano Zampini } 4115779c1cceSStefano Zampini } 4116779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4117779c1cceSStefano Zampini 4118779c1cceSStefano Zampini 41196e683305SStefano Zampini if (pcbddc->dbg_flag) { 41206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41226e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4124779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41266e683305SStefano Zampini } 41276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41286e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41296e683305SStefano Zampini } 41306e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41316e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41326e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4133854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41346e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41356e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41366e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41376e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41386e683305SStefano Zampini } 41396e683305SStefano Zampini if (pcbddc->dbg_flag) { 41406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 41426e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41436e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41446e683305SStefano Zampini } 4145779c1cceSStefano Zampini } 41466e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 41476e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 41486e683305SStefano Zampini } else { 41496e683305SStefano Zampini if (pcbddc->dbg_flag) { 41506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 41526e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41536e683305SStefano Zampini } 41546e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 41556e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 41566e683305SStefano Zampini } 41576e683305SStefano Zampini 41586e683305SStefano Zampini /* create local to global scatters for coarse problem */ 415968457ee5SStefano Zampini if (compute_vecs) { 41606e683305SStefano Zampini PetscInt lrows; 41616e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41626e683305SStefano Zampini if (coarse_mat_is) { 41636e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 41646e683305SStefano Zampini } else { 41656e683305SStefano Zampini lrows = 0; 41666e683305SStefano Zampini } 41676e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 41686e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 41696e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 41706e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41716e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41726e683305SStefano Zampini } 41736e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 41746e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4175c8587f34SStefano Zampini 4176f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4177f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4178f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4179f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4180f9eb5b7dSStefano Zampini } else { 4181f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4182f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4183c8587f34SStefano Zampini } 4184c8587f34SStefano Zampini 41856e683305SStefano Zampini /* print some info if requested */ 41866e683305SStefano Zampini if (pcbddc->dbg_flag) { 41876e683305SStefano Zampini if (!multilevel_allowed) { 41886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41896e683305SStefano Zampini if (multilevel_requested) { 41906e683305SStefano 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); 41916e683305SStefano Zampini } else if (pcbddc->max_levels) { 41926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 41936e683305SStefano Zampini } 41946e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41956e683305SStefano Zampini } 41966e683305SStefano Zampini } 41976e683305SStefano Zampini 4198f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 41996e683305SStefano Zampini if (coarse_mat_is) { 42006e683305SStefano Zampini MatReuse coarse_mat_reuse; 42016a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42026e683305SStefano Zampini if (pcbddc->dbg_flag) { 42036e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42046e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42056e683305SStefano Zampini } 4206f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4207312be037SStefano Zampini char prefix[256],str_level[16]; 4208e604994aSStefano Zampini size_t len; 42096e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4210c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4211f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42125f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4213c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42146e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4215c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4216c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42173e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4218e604994aSStefano Zampini /* prefix */ 4219e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4220e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4221e604994aSStefano Zampini if (!pcbddc->current_level) { 4222e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4223e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4224c8587f34SStefano Zampini } else { 4225e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4226312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4227312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 422834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4229312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4230e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4231e604994aSStefano Zampini } 4232e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 42333e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42343e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 42353e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 42363e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4237f9eb5b7dSStefano Zampini /* allow user customization */ 4238f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 42393e3c6dadSStefano Zampini } 42403e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42413e3c6dadSStefano Zampini if (nisdofs) { 42423e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 42433e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 42443e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 42453e3c6dadSStefano Zampini } 42463e3c6dadSStefano Zampini } 42473e3c6dadSStefano Zampini if (nisneu) { 42483e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 42493e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4250312be037SStefano Zampini } 4251f9eb5b7dSStefano Zampini 4252f9eb5b7dSStefano Zampini /* get some info after set from options */ 4253f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4254f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4255f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 42564f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 42576e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4258f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4259f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4260f9eb5b7dSStefano Zampini } 42614f3a063dSStefano Zampini if (isredundant) { 42624f3a063dSStefano Zampini KSP inner_ksp; 42634f3a063dSStefano Zampini PC inner_pc; 42644f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 42654f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 42664f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 42674f3a063dSStefano Zampini } 4268f9eb5b7dSStefano Zampini 4269f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4270fa7f1dd8SStefano Zampini if (coarse_reuse) { 427181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4272fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 42736e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4274fa7f1dd8SStefano Zampini } else { 42756e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4276fa7f1dd8SStefano Zampini } 4277c8587f34SStefano Zampini if (isbddc || isnn) { 427822bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 427970cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 428028143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 428122b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 42826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 42846e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 42856e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 428622b6e8a2SStefano Zampini } 428770cf5478SStefano Zampini } 42886e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 428970cf5478SStefano Zampini } else { 429022bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 429122bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 429222bc73bbSStefano Zampini } 429322bc73bbSStefano Zampini } else { 42942e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4295c8587f34SStefano Zampini } 4296c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4297c8587f34SStefano Zampini 4298c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4299b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 43005a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4301c8587f34SStefano Zampini 43026e683305SStefano Zampini /* set operators */ 43035f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43046e683305SStefano Zampini if (pcbddc->dbg_flag) { 43056e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43066e683305SStefano Zampini } 43076e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43086e683305SStefano Zampini coarse_mat = 0; 43096e683305SStefano Zampini } 43106e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4311b9b85e73SStefano Zampini #if 0 4312b9b85e73SStefano Zampini { 4313b9b85e73SStefano Zampini PetscViewer viewer; 4314b9b85e73SStefano Zampini char filename[256]; 4315b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4316b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4317b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4318b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4319b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4320b9b85e73SStefano Zampini } 4321b9b85e73SStefano Zampini #endif 4322c8587f34SStefano Zampini 4323c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4324c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4325c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 432698a51de6SStefano Zampini } 432798a51de6SStefano Zampini 432898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 432998a51de6SStefano Zampini Vec crhs,csol; 433098a51de6SStefano Zampini PetscBool ispreonly; 433198a51de6SStefano Zampini if (CoarseNullSpace) { 4332c8587f34SStefano Zampini if (isbddc) { 4333c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4334c8587f34SStefano Zampini } else { 4335c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4336c8587f34SStefano Zampini } 4337c8587f34SStefano Zampini } 4338f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4339f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4340f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4341f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 43426e683305SStefano Zampini /* hack */ 4343f347579bSStefano Zampini if (!csol) { 43442a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4345f9eb5b7dSStefano Zampini } 4346f347579bSStefano Zampini if (!crhs) { 43472a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4348f347579bSStefano Zampini } 4349cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4350cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 43516e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4352c8587f34SStefano Zampini KSP check_ksp; 43532b510759SStefano Zampini KSPType check_ksp_type; 4354c8587f34SStefano Zampini PC check_pc; 43556e683305SStefano Zampini Vec check_vec,coarse_vec; 43566a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 43572b510759SStefano Zampini PetscInt its; 43586e683305SStefano Zampini PetscBool compute_eigs; 43596e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 43606e683305SStefano Zampini PetscInt neigs; 43618e185a42SStefano Zampini const char *prefix; 4362c8587f34SStefano Zampini 43632b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 43646e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 436523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4366f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 43672b510759SStefano Zampini if (ispreonly) { 43682b510759SStefano Zampini check_ksp_type = KSPPREONLY; 43696e683305SStefano Zampini compute_eigs = PETSC_FALSE; 43702b510759SStefano Zampini } else { 4371cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 43726e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4373c8587f34SStefano Zampini } 4374c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 43756e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 43766e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 43776e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4378a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4379a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4380a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4381a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4382c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4383c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4384c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4385c8587f34SStefano Zampini /* create random vec */ 43866e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 43876e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4388c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4389c8587f34SStefano Zampini if (CoarseNullSpace) { 4390c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4391c8587f34SStefano Zampini } 43926e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4393c8587f34SStefano Zampini /* solve coarse problem */ 43946e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4395c8587f34SStefano Zampini if (CoarseNullSpace) { 43966e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4397c8587f34SStefano Zampini } 4398cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 43996e683305SStefano Zampini if (compute_eigs) { 4400854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4401854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44026e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44036e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44046e683305SStefano Zampini lambda_min = eigs_r[0]; 44056e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44066e683305SStefano Zampini if (lambda_max>lambda_min) { 4407cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4408cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4409cbcc2c2aSStefano Zampini } 4410c8587f34SStefano Zampini } 4411c8587f34SStefano Zampini } 4412cbcc2c2aSStefano Zampini 4413c8587f34SStefano Zampini /* check coarse problem residual error */ 44146e683305SStefano Zampini if (pcbddc->dbg_flag) { 44156e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44166e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44176e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4418c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44196e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44206e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4421c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4422779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44236e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44246e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44276e683305SStefano Zampini if (compute_eigs) { 44286e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4429deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4430c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44316e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 44326e683305SStefano 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); 44336e683305SStefano Zampini for (i=0;i<neigs;i++) { 44346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4435c8587f34SStefano Zampini } 44366e683305SStefano Zampini } 44376e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 44386e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44396e683305SStefano Zampini } 4440c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 44416e683305SStefano Zampini if (compute_eigs) { 44426e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 44436e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4444c8587f34SStefano Zampini } 44456e683305SStefano Zampini } 44466e683305SStefano Zampini } 4447cbcc2c2aSStefano Zampini /* print additional info */ 4448cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 44496e683305SStefano Zampini /* waits until all processes reaches this point */ 44506e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4451cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4452cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4453cbcc2c2aSStefano Zampini } 4454cbcc2c2aSStefano Zampini 44552b510759SStefano Zampini /* free memory */ 4456c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4457fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4458c8587f34SStefano Zampini PetscFunctionReturn(0); 4459c8587f34SStefano Zampini } 4460674ae819SStefano Zampini 4461f34684f1SStefano Zampini #undef __FUNCT__ 4462f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4463f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4464f34684f1SStefano Zampini { 4465f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4466f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4467f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 44680e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 446973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 44700e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4471f34684f1SStefano Zampini PetscErrorCode ierr; 4472f34684f1SStefano Zampini 4473f34684f1SStefano Zampini PetscFunctionBegin; 4474f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 44750e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 44760e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4477727cdba6SStefano Zampini } 44780e6343abSStefano 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); 44790e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 44800e6343abSStefano Zampini local_size = 0; 44810e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 44820e6343abSStefano Zampini PetscInt j; 44830e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 44840e6343abSStefano Zampini } 44850e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 44860e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 44870e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 44880e6343abSStefano Zampini } 4489f34684f1SStefano Zampini 4490f34684f1SStefano Zampini /* check numbering */ 4491f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4492f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4493b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4494f34684f1SStefano Zampini 4495f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4496f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4497f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 44980fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4499f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4500f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4501727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4502f34684f1SStefano Zampini } 4503f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4504f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4505f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4506f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4507f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4508f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4509f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4510f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4511f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4512f34684f1SStefano Zampini if (array[i] == 1.0) { 4513b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4514f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4515f34684f1SStefano Zampini } 4516f34684f1SStefano Zampini } 4517b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4518f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4519f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4520f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4521f34684f1SStefano Zampini } 4522f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4523f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4524f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4525f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4526f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4527f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4528b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4529ca8b9ea9SStefano Zampini PetscInt *gidxs; 4530ca8b9ea9SStefano Zampini 4531ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4532ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4533f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4534f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4535f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4536f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 45374bc2dc4bSStefano 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); 4538f34684f1SStefano Zampini } 4539f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4540ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4541f34684f1SStefano Zampini } 4542f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4543302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4544f34684f1SStefano Zampini } 45458bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4546f34684f1SStefano Zampini /* get back data */ 4547f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4548f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4549674ae819SStefano Zampini PetscFunctionReturn(0); 4550674ae819SStefano Zampini } 4551674ae819SStefano Zampini 4552e456f2a8SStefano Zampini #undef __FUNCT__ 4553e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4554a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4555e456f2a8SStefano Zampini { 4556e456f2a8SStefano Zampini IS localis_t; 4557a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4558e456f2a8SStefano Zampini PetscScalar *vals; 4559e456f2a8SStefano Zampini PetscErrorCode ierr; 4560e456f2a8SStefano Zampini 4561e456f2a8SStefano Zampini PetscFunctionBegin; 4562a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4563e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4564854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4565e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4566e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4567a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4568a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 45691035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4570a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 45711035eff8SStefano Zampini } 4572a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4573e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4574e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4575a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4576a7dc3881SStefano Zampini /* now compute set in local ordering */ 4577a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4578a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4579a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4580a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4581a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4582ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4583e456f2a8SStefano Zampini lsize++; 4584e456f2a8SStefano Zampini } 4585e456f2a8SStefano Zampini } 4586854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4587a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4588ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4589e456f2a8SStefano Zampini idxs[lsize++] = i; 4590e456f2a8SStefano Zampini } 4591e456f2a8SStefano Zampini } 4592a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4593a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4594e456f2a8SStefano Zampini *localis = localis_t; 4595e456f2a8SStefano Zampini PetscFunctionReturn(0); 4596e456f2a8SStefano Zampini } 4597906d46d4SStefano Zampini 4598906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4599906d46d4SStefano Zampini #undef __FUNCT__ 4600906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4601906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4602906d46d4SStefano Zampini { 4603906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4604906d46d4SStefano Zampini PetscErrorCode ierr; 4605906d46d4SStefano Zampini 4606906d46d4SStefano Zampini PetscFunctionBegin; 4607906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4608906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4609906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4610906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4611906d46d4SStefano Zampini PetscFunctionReturn(0); 4612906d46d4SStefano Zampini } 4613906d46d4SStefano Zampini 4614906d46d4SStefano Zampini #undef __FUNCT__ 4615906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4616906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4617906d46d4SStefano Zampini { 4618906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4619906d46d4SStefano Zampini PetscErrorCode ierr; 4620906d46d4SStefano Zampini 4621906d46d4SStefano Zampini PetscFunctionBegin; 4622906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4623906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4624906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4625906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4626906d46d4SStefano Zampini PetscFunctionReturn(0); 4627906d46d4SStefano Zampini } 4628b96c3477SStefano Zampini 4629b96c3477SStefano Zampini #undef __FUNCT__ 4630b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 463108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4632b96c3477SStefano Zampini { 4633a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4634b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4635b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4636a64f4aa4SStefano Zampini Mat S_j; 4637b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4638b96c3477SStefano Zampini PetscBool free_used_adj; 4639b96c3477SStefano Zampini PetscErrorCode ierr; 4640b96c3477SStefano Zampini 4641b96c3477SStefano Zampini PetscFunctionBegin; 4642b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4643b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 464408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4645b96c3477SStefano Zampini used_xadj = NULL; 4646b96c3477SStefano Zampini used_adjncy = NULL; 4647b96c3477SStefano Zampini } else { 464808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 464908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 465008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 465108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4652b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4653b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4654b96c3477SStefano Zampini } else { 46552fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4656b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4657b96c3477SStefano Zampini PetscInt nvtxs; 4658b96c3477SStefano Zampini 46592fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 46602fffb893SStefano Zampini if (flg_row) { 4661b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4662b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4663b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4664b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 46652fffb893SStefano Zampini } else { 46662fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 46672fffb893SStefano Zampini used_xadj = NULL; 46682fffb893SStefano Zampini used_adjncy = NULL; 46692fffb893SStefano Zampini } 46702fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4671b96c3477SStefano Zampini } 4672b96c3477SStefano Zampini } 4673d5574798SStefano Zampini 4674d5574798SStefano Zampini /* setup sub_schurs data */ 4675a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4676a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4677a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4678a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4679a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4680a64f4aa4SStefano Zampini } else { 4681a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4682a64f4aa4SStefano Zampini } 4683a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4684b96c3477SStefano Zampini 4685b96c3477SStefano Zampini /* free adjacency */ 4686b96c3477SStefano Zampini if (free_used_adj) { 4687b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4688b96c3477SStefano Zampini } 4689b96c3477SStefano Zampini PetscFunctionReturn(0); 4690b96c3477SStefano Zampini } 4691b96c3477SStefano Zampini 4692b96c3477SStefano Zampini #undef __FUNCT__ 4693b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 469408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4695b96c3477SStefano Zampini { 4696b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4697b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4698b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4699b96c3477SStefano Zampini PCBDDCGraph graph; 4700b96c3477SStefano Zampini PetscErrorCode ierr; 4701b96c3477SStefano Zampini 4702b96c3477SStefano Zampini PetscFunctionBegin; 4703b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 470408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4705b96c3477SStefano Zampini IS verticesIS; 4706b96c3477SStefano Zampini 4707b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4708b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47097fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4710b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4711b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4712b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4713b96c3477SStefano Zampini /* 4714b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4715b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4716b96c3477SStefano Zampini } 4717b96c3477SStefano Zampini */ 4718b96c3477SStefano Zampini } else { 4719b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4720b96c3477SStefano Zampini } 4721b96c3477SStefano Zampini 4722b96c3477SStefano Zampini /* sub_schurs init */ 4723a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4724a64f4aa4SStefano Zampini 4725b96c3477SStefano Zampini /* free graph struct */ 472608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4727b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4728b96c3477SStefano Zampini } 4729b96c3477SStefano Zampini PetscFunctionReturn(0); 4730b96c3477SStefano Zampini } 4731