1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8b1b3d7a2SStefano Zampini #undef __FUNCT__ 9b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 1008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 11b1b3d7a2SStefano Zampini { 12b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 1508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 1608122e43SStefano Zampini PetscScalar *work,lwork; 1708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 1808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 1908122e43SStefano Zampini PetscReal *eigs,thresh; 201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 21f6f667cfSStefano Zampini PetscBool allocated_S_St; 2208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2308122e43SStefano Zampini PetscReal *rwork; 2408122e43SStefano Zampini #endif 25b1b3d7a2SStefano Zampini PetscErrorCode ierr; 26b1b3d7a2SStefano Zampini 27b1b3d7a2SStefano Zampini PetscFunctionBegin; 2808122e43SStefano Zampini if (!sub_schurs->use_mumps) { 2908122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3008122e43SStefano Zampini } 3108122e43SStefano Zampini 32fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 33fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 34fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 35fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 36fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 5108122e43SStefano Zampini PetscInt subset_size; 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini } 5608122e43SStefano Zampini 5708122e43SStefano Zampini /* min/max and threshold */ 5808122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 59f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 6008122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 61f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 62f6f667cfSStefano Zampini if (nmin) { 63f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 64f6f667cfSStefano Zampini } 6508122e43SStefano Zampini 6608122e43SStefano Zampini /* allocate lapack workspace */ 6708122e43SStefano Zampini cum = cum2 = 0; 6808122e43SStefano Zampini maxneigs = 0; 6908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 7108122e43SStefano Zampini PetscInt n,subset_size; 72f6f667cfSStefano Zampini 7308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 759162d606SStefano Zampini cum += subset_size; 769162d606SStefano Zampini cum2 += subset_size*n; 7708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7808122e43SStefano Zampini } 7908122e43SStefano Zampini } 8008122e43SStefano Zampini if (mss) { 819ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 8208122e43SStefano Zampini PetscBLASInt B_itype = 1; 8308122e43SStefano Zampini PetscBLASInt B_N = mss; 844c6709b3SStefano Zampini PetscReal zero = 0.0; 854c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8608122e43SStefano Zampini 8708122e43SStefano Zampini B_lwork = -1; 8808122e43SStefano Zampini S = NULL; 8908122e43SStefano Zampini St = NULL; 90a58a30b4SStefano Zampini eigs = NULL; 91a58a30b4SStefano Zampini eigv = NULL; 92a58a30b4SStefano Zampini B_iwork = NULL; 93a58a30b4SStefano Zampini B_ifail = NULL; 948bec7fa6SStefano Zampini thresh = 1.0; 9508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9708122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 9808122e43SStefano Zampini #else 9908122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 10008122e43SStefano Zampini #endif 10108122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 10208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10308122e43SStefano Zampini } else { 10408122e43SStefano Zampini /* TODO */ 10508122e43SStefano Zampini } 10608122e43SStefano Zampini } else { 10708122e43SStefano Zampini lwork = 0; 10808122e43SStefano Zampini } 10908122e43SStefano Zampini 11008122e43SStefano Zampini nv = 0; 111d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 112d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 11308122e43SStefano Zampini } 1144c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 115f6f667cfSStefano Zampini if (allocated_S_St) { 116f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 117f6f667cfSStefano Zampini } 118f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12008122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 12108122e43SStefano Zampini #endif 1229162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1239162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1249162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1269162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12708122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12808122e43SStefano Zampini 12908122e43SStefano Zampini maxneigs = 0; 13008122e43SStefano Zampini cum = cum2 = cumarray = 0; 1319162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1329162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 133d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13408122e43SStefano Zampini const PetscInt *idxs; 13508122e43SStefano Zampini 136d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13708122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13808122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 14008122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1419162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1429162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 14308122e43SStefano Zampini } 14408122e43SStefano Zampini cum2 = cum; 145d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14608122e43SStefano Zampini } 14708122e43SStefano Zampini 14808122e43SStefano Zampini if (mss) { /* multilevel */ 14908122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 15008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 15108122e43SStefano Zampini } 15208122e43SStefano Zampini 15308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15408122e43SStefano Zampini PetscInt j,subset_size; 15508122e43SStefano Zampini 15608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 15708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 15808122e43SStefano Zampini const PetscInt *idxs; 159f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 160f6f667cfSStefano Zampini PetscInt eigs_start = 0; 16108122e43SStefano Zampini PetscBLASInt B_N; 16208122e43SStefano Zampini 163f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 164f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1659ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 16608122e43SStefano Zampini PetscInt j; 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 16808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 16908122e43SStefano Zampini } 17008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 17108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 17208122e43SStefano Zampini } 17308122e43SStefano Zampini } else { 17408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17608122e43SStefano Zampini } 1778bec7fa6SStefano Zampini } else { 178f6f667cfSStefano Zampini S = Sarray + cumarray; 179f6f667cfSStefano Zampini St = Starray + cumarray; 1808bec7fa6SStefano Zampini } 18108122e43SStefano Zampini 182f6f667cfSStefano Zampini /* Threshold: this is an heuristic for edges */ 183f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 184f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 185f6f667cfSStefano Zampini 1869ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 18708122e43SStefano Zampini PetscBLASInt B_itype = 1; 188f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1894c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1909552c7c7SStefano Zampini PetscInt nmin_s; 19108122e43SStefano Zampini 192f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 193fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1948bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 195fd14bc51SStefano Zampini } 19608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 19708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 198f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 19908122e43SStefano Zampini #else 200f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 20108122e43SStefano Zampini #endif 20208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 20308122e43SStefano Zampini if (B_ierr) { 20408122e43SStefano Zampini if (B_ierr < 0 ) { 20508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 20608122e43SStefano Zampini } else if (B_ierr <= B_N) { 20708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 20808122e43SStefano Zampini } else { 2099552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 21008122e43SStefano Zampini } 21108122e43SStefano Zampini } 21208122e43SStefano Zampini 21308122e43SStefano Zampini if (B_neigs > nmax) { 214fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 215fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 216fd14bc51SStefano Zampini } 217f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 21808122e43SStefano Zampini B_neigs = nmax; 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 2219552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2229552c7c7SStefano Zampini if (B_neigs < nmin_s) { 22308122e43SStefano Zampini PetscBLASInt B_neigs2; 22408122e43SStefano Zampini 225f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 226f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 227fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 228fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 229fd14bc51SStefano Zampini } 2309ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23108122e43SStefano Zampini PetscInt j; 23208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23408122e43SStefano Zampini } 23508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23708122e43SStefano Zampini } 23808122e43SStefano Zampini } else { 23908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24108122e43SStefano Zampini } 24208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 244f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 24508122e43SStefano Zampini #else 246f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 24708122e43SStefano Zampini #endif 24808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 24908122e43SStefano Zampini B_neigs += B_neigs2; 25008122e43SStefano Zampini } 25108122e43SStefano Zampini if (B_ierr) { 25208122e43SStefano Zampini if (B_ierr < 0 ) { 25308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 25408122e43SStefano Zampini } else if (B_ierr <= B_N) { 25508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25608122e43SStefano Zampini } else { 2579552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 25808122e43SStefano Zampini } 25908122e43SStefano Zampini } 260fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 261fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 26208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 26308122e43SStefano Zampini if (eigs[j] == 0.0) { 264fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 26508122e43SStefano Zampini } else { 266f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 267fd14bc51SStefano Zampini } 26808122e43SStefano Zampini } 26908122e43SStefano Zampini } 27008122e43SStefano Zampini } else { 27108122e43SStefano Zampini /* TODO */ 27208122e43SStefano Zampini } 27308122e43SStefano Zampini 2748bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2758bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2769162d606SStefano Zampini if (B_neigs) { 2779162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 278fd14bc51SStefano Zampini 279fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2809552c7c7SStefano Zampini PetscInt ii; 2819552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 282fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2839552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2849162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2859552c7c7SStefano Zampini } 2869552c7c7SStefano Zampini } 287fd14bc51SStefano Zampini } 28808122e43SStefano Zampini #if 0 2899162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29008122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29108122e43SStefano Zampini PetscScalar norm; 29208122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 2939162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 2949162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 29508122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 29608122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 29708122e43SStefano Zampini } else { 29808122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 29908122e43SStefano Zampini } 3009162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 301b1b3d7a2SStefano Zampini } 302b1b3d7a2SStefano Zampini #endif 3039162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3049162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3059162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3069162d606SStefano Zampini cum++; 30708122e43SStefano Zampini } 30808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 30908122e43SStefano Zampini } 31008122e43SStefano Zampini /* shift for next computation */ 31108122e43SStefano Zampini cumarray += subset_size*subset_size; 31208122e43SStefano Zampini } 313fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 314fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 315fd14bc51SStefano Zampini } 31608122e43SStefano Zampini 31708122e43SStefano Zampini if (mss) { 31808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 31908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 320f6f667cfSStefano Zampini /* destroy matrices (junk) */ 321f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 322f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 32308122e43SStefano Zampini } 324f6f667cfSStefano Zampini if (allocated_S_St) { 325f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 326f6f667cfSStefano Zampini } 327f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 32808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33008122e43SStefano Zampini #endif 33108122e43SStefano Zampini if (pcbddc->dbg_flag) { 3321b968477SStefano Zampini PetscInt maxneigs_r; 33308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3349b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33508122e43SStefano Zampini } 33608122e43SStefano Zampini PetscFunctionReturn(0); 33708122e43SStefano Zampini } 338b1b3d7a2SStefano Zampini 339674ae819SStefano Zampini #undef __FUNCT__ 340c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 341c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 342c8587f34SStefano Zampini { 343c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3448629588bSStefano Zampini PetscScalar *coarse_submat_vals; 345c8587f34SStefano Zampini PetscErrorCode ierr; 346c8587f34SStefano Zampini 347c8587f34SStefano Zampini PetscFunctionBegin; 348f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3495e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 350c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 351c8587f34SStefano Zampini 352684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3530fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 354684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 355c8587f34SStefano Zampini 356c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 357b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 358c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 359c8587f34SStefano Zampini } 360c8587f34SStefano Zampini 3618629588bSStefano Zampini /* 3628629588bSStefano Zampini Setup local correction and local part of coarse basis. 3638629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3648629588bSStefano Zampini */ 36547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3668629588bSStefano Zampini 3678629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3688629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3698629588bSStefano Zampini 3708629588bSStefano Zampini /* free */ 3718629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 372c8587f34SStefano Zampini PetscFunctionReturn(0); 373c8587f34SStefano Zampini } 374c8587f34SStefano Zampini 375c8587f34SStefano Zampini #undef __FUNCT__ 376674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 377674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 378674ae819SStefano Zampini { 379674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 380674ae819SStefano Zampini PetscErrorCode ierr; 381674ae819SStefano Zampini 382674ae819SStefano Zampini PetscFunctionBegin; 383674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 384674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 385674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 386674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 387785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 388674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 389f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 390f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 391785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 39363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 394674ae819SStefano Zampini PetscFunctionReturn(0); 395674ae819SStefano Zampini } 396674ae819SStefano Zampini 397674ae819SStefano Zampini #undef __FUNCT__ 398674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 399674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 400674ae819SStefano Zampini { 401674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 402674ae819SStefano Zampini PetscErrorCode ierr; 403674ae819SStefano Zampini 404674ae819SStefano Zampini PetscFunctionBegin; 405b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 406674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 407674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 408674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 409b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 410674ae819SStefano Zampini PetscFunctionReturn(0); 411674ae819SStefano Zampini } 412674ae819SStefano Zampini 413674ae819SStefano Zampini #undef __FUNCT__ 414674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 415674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 416674ae819SStefano Zampini { 417674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 41806656605SStefano Zampini PetscScalar *array; 419674ae819SStefano Zampini PetscErrorCode ierr; 420674ae819SStefano Zampini 421674ae819SStefano Zampini PetscFunctionBegin; 422674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42658da7f69SStefano Zampini } 427674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 428674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 42915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 431674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 432674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 433674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4378ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 441f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 442f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 443f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 444f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 445727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4460e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 44870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4496e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 450674ae819SStefano Zampini PetscFunctionReturn(0); 451674ae819SStefano Zampini } 452674ae819SStefano Zampini 453674ae819SStefano Zampini #undef __FUNCT__ 454f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 455f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4566bfb1811SStefano Zampini { 4576bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4586bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4596bfb1811SStefano Zampini VecType impVecType; 460e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4616bfb1811SStefano Zampini PetscErrorCode ierr; 4626bfb1811SStefano Zampini 4636bfb1811SStefano Zampini PetscFunctionBegin; 464f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 465f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 466f4ddd8eeSStefano Zampini } 467e7b262bdSStefano Zampini /* get sizes */ 468b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 469b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4706bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 471e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 472e7b262bdSStefano Zampini /* R nodes */ 473e7b262bdSStefano Zampini old_size = -1; 474e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 475e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 476e7b262bdSStefano Zampini } 477e7b262bdSStefano Zampini if (n_R != old_size) { 478e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 479e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4806bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4816bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 484e7b262bdSStefano Zampini } 485e7b262bdSStefano Zampini /* local primal dofs */ 486e7b262bdSStefano Zampini old_size = -1; 487e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 493e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4946bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 495e7b262bdSStefano Zampini } 496e7b262bdSStefano Zampini /* local explicit constraints */ 497e7b262bdSStefano Zampini old_size = -1; 498e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 499e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 500e7b262bdSStefano Zampini } 501e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 502e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50683b7ccabSStefano Zampini } 5076bfb1811SStefano Zampini PetscFunctionReturn(0); 5086bfb1811SStefano Zampini } 5096bfb1811SStefano Zampini 5106bfb1811SStefano Zampini #undef __FUNCT__ 51147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51388ebb749SStefano Zampini { 51425084f0cSStefano Zampini PetscErrorCode ierr; 51525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 518d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 51925084f0cSStefano Zampini /* submatrices of local problem */ 52080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52106656605SStefano Zampini /* submatrices of local coarse problem */ 52206656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52325084f0cSStefano Zampini /* working matrices */ 52406656605SStefano Zampini Mat C_CR; 52525084f0cSStefano Zampini /* additional working stuff */ 52606656605SStefano Zampini PC pc_R; 52706656605SStefano Zampini Mat F; 52806656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 52906656605SStefano Zampini 53025084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53106656605SStefano Zampini PetscScalar *work; 53206656605SStefano Zampini PetscInt *idx_V_B; 53306656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53406656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 535b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53645a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53788ebb749SStefano Zampini MatType impMatType; 53825084f0cSStefano Zampini /* some shortcuts to scalars */ 53906656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54088ebb749SStefano Zampini 54188ebb749SStefano Zampini PetscFunctionBegin; 542b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54388ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54488ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 545b371cd4fSStefano Zampini n_B = pcis->n_B; 546b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54788ebb749SStefano Zampini n_R = pcis->n - n_vertices; 54888ebb749SStefano Zampini 54988ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55088ebb749SStefano Zampini impMatType = MATSEQDENSE; 55188ebb749SStefano Zampini 55288ebb749SStefano Zampini /* vertices in boundary numbering */ 553785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5540e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 55588ebb749SStefano Zampini if (i != n_vertices) { 55622d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 55788ebb749SStefano Zampini } 55888ebb749SStefano Zampini 55906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56006656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56906656605SStefano Zampini 57006656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57106656605SStefano Zampini /* allocate workspace */ 57206656605SStefano Zampini n = 0; 57306656605SStefano Zampini if (n_constraints) { 57406656605SStefano Zampini n += n_R*n_constraints; 57506656605SStefano Zampini } 57606656605SStefano Zampini if (n_vertices) { 57706656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57880677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57906656605SStefano Zampini } 58006656605SStefano Zampini if (!pcbddc->issym) { 58106656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58206656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58306656605SStefano Zampini } 58406656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58506656605SStefano Zampini 58606656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58706656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59106656605SStefano Zampini if (isLU || isILU || isCHOL) { 59206656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 593d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 594d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 595d62866d3SStefano Zampini MatFactorType type; 596d62866d3SStefano Zampini 5976816873aSStefano Zampini F = reuse_mumps->F; 5986816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 599d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 60006656605SStefano Zampini } else { 60106656605SStefano Zampini F = NULL; 60206656605SStefano Zampini } 60306656605SStefano Zampini 60488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 60588ebb749SStefano Zampini if (n_constraints) { 60606656605SStefano Zampini Mat M1,M2,M3; 60780677318SStefano Zampini Mat auxmat; 60806656605SStefano Zampini IS is_aux; 60980677318SStefano Zampini PetscScalar *array,*array2; 61006656605SStefano Zampini 611f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 61388ebb749SStefano Zampini 61425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 61525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6168ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 61780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61888ebb749SStefano Zampini 61980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 62080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62106656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 62306656605SStefano Zampini const PetscScalar *row_cmat_values; 62406656605SStefano Zampini const PetscInt *row_cmat_indices; 62506656605SStefano Zampini PetscInt size_of_constraint,j; 62688ebb749SStefano Zampini 62706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62906656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63006656605SStefano Zampini } 63106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63206656605SStefano Zampini } 63380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 63406656605SStefano Zampini if (F) { 63506656605SStefano Zampini Mat B; 63606656605SStefano Zampini 63706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 63880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64006656605SStefano Zampini } else { 64180677318SStefano Zampini PetscScalar *marr; 64280677318SStefano Zampini 64380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 64506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 64680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 64706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65006656605SStefano Zampini } 65180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65206656605SStefano Zampini } 65380677318SStefano Zampini if (!pcbddc->switch_static) { 65480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 66080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 66380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66480677318SStefano Zampini } 66580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66780677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66880677318SStefano Zampini } else { 66980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 67080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67125084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67280677318SStefano Zampini } 67380677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 67480677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 67580677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 67706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 67880677318SStefano Zampini if (isCHOL) { 67980677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 68080677318SStefano Zampini } else { 68125084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68280677318SStefano Zampini } 68380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 68406656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 68525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 68625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 68725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 68880677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68980677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 69080677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 69388ebb749SStefano Zampini } 69488ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 69588ebb749SStefano Zampini if (n_vertices) { 69606656605SStefano Zampini Mat newmat; 69706656605SStefano Zampini IS is_aux; 6983a50541eSStefano Zampini PetscInt ibs,mbs; 6993a50541eSStefano Zampini PetscBool issbaij; 7003a50541eSStefano Zampini 7016816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7026816873aSStefano Zampini IS tis; 7036816873aSStefano Zampini 7046816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7056816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7066816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7076816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7086816873aSStefano Zampini } else { 7093a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7106816873aSStefano Zampini } 7119577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 7129577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 7133a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 7149577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7159577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7169577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7179577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 71880677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7199577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7209577ea80SStefano Zampini } else { 7213a50541eSStefano Zampini /* this is safe */ 7229577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 7233a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7243a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 725b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 726b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7273a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7283a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72906656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7303a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7313a50541eSStefano Zampini } else { 732b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7339577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7349577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7359577ea80SStefano Zampini } 73680677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 73780677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7383a50541eSStefano Zampini } 73925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 74088ebb749SStefano Zampini } 74188ebb749SStefano Zampini 74288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 743f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 74406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 74506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 74606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 74706656605SStefano Zampini } 748f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 74906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 75006656605SStefano Zampini PetscScalar *marray; 75106656605SStefano Zampini 75206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 75306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 754f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 755f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 756f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 757f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 758f4ddd8eeSStefano Zampini } 759f4ddd8eeSStefano Zampini } 76006656605SStefano Zampini 761f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 76206656605SStefano Zampini PetscScalar *marray; 76388ebb749SStefano Zampini 76406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7658eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76788ebb749SStefano Zampini } 768b9d89cd5SStefano Zampini if (!pcbddc->issym) { 76906656605SStefano Zampini n *= 2; 770f4ddd8eeSStefano Zampini } 77106656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 77206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 77306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 77606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 77788ebb749SStefano Zampini } 77806656605SStefano Zampini if (!pcbddc->issym) { 77906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7801b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 78106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 78206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 78388ebb749SStefano Zampini } 78488ebb749SStefano Zampini } else { 785c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 786c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7871b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 788c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 789c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 790c0553b1fSStefano Zampini } 79188ebb749SStefano Zampini } 79206656605SStefano Zampini } 79306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 79406656605SStefano Zampini /* vertices */ 79506656605SStefano Zampini if (n_vertices) { 79616f15bc4SStefano Zampini 79716f15bc4SStefano Zampini if (n_R) { 79806656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 79906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 80016f15bc4SStefano Zampini PetscScalar *x,*y; 80106656605SStefano Zampini 80206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 80321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 8056816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 80606656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 80706656605SStefano Zampini } else { 80806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 81006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 81106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 81206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 81306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 81406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 81506656605SStefano Zampini } 81606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 81706656605SStefano Zampini } 81880677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81906656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 82006656605SStefano Zampini if (n_constraints) { 82106656605SStefano Zampini Mat B; 82280677318SStefano Zampini 823b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 82480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 82580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 82680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 83080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 83180677318SStefano Zampini } 83280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 83380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 83480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 83680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 83706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 83806656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 84006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 84106656605SStefano Zampini } 84206656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 84380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 84706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 85216f15bc4SStefano Zampini } else { 85316f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85416f15bc4SStefano Zampini } 85521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 85606656605SStefano Zampini /* coarse basis functions */ 85706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85816f15bc4SStefano Zampini PetscScalar *y; 85916f15bc4SStefano Zampini 86006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 86106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 86606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86806656605SStefano Zampini 86906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 87006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 87206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87606656605SStefano Zampini } 87706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87806656605SStefano Zampini } 87906656605SStefano Zampini } 88006656605SStefano Zampini 88106656605SStefano Zampini if (n_constraints) { 88206656605SStefano Zampini Mat B; 88306656605SStefano Zampini 88406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88806656605SStefano Zampini if (n_vertices) { 88980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 89080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 89180677318SStefano Zampini } else { 89280677318SStefano Zampini Mat S_VCt; 89380677318SStefano Zampini 89480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 89580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89780677318SStefano Zampini } 89806656605SStefano Zampini } 89906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 90006656605SStefano Zampini /* coarse basis functions */ 90106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 90206656605SStefano Zampini PetscScalar *y; 90306656605SStefano Zampini 90406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 91406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91806656605SStefano Zampini } 91906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini } 92280677318SStefano Zampini if (n_constraints) { 92380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 92480677318SStefano Zampini } 92506656605SStefano Zampini 92606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 92706656605SStefano Zampini if (!pcbddc->issym) { 92806656605SStefano Zampini 92906656605SStefano Zampini if (n_constraints) { 93016f15bc4SStefano Zampini Mat S_CCT,B_C; 93106656605SStefano Zampini 93280677318SStefano Zampini /* this is a lazy thing */ 93380677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93806656605SStefano Zampini if (n_vertices) { 93916f15bc4SStefano Zampini Mat B_V,S_VCT; 94006656605SStefano Zampini 94106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 94516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94606656605SStefano Zampini } 94706656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94880677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94980677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 95006656605SStefano Zampini } 95116f15bc4SStefano Zampini if (n_vertices && n_R) { 95206656605SStefano Zampini Mat A_VRT; 95380677318SStefano Zampini PetscScalar *marray; 95406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 95506656605SStefano Zampini 95680677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95880677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 96080677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 96180677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 96216f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 96306656605SStefano Zampini } 96406656605SStefano Zampini 96506656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97206656605SStefano Zampini } 97306656605SStefano Zampini } else { 97406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 98006656605SStefano Zampini } 98106656605SStefano Zampini } 98206656605SStefano Zampini /* coarse basis functions */ 98306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98406656605SStefano Zampini PetscScalar *y; 98506656605SStefano Zampini 98606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99106656605SStefano Zampini if (i<n_vertices) { 99206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 99306656605SStefano Zampini } 99406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99606656605SStefano Zampini 99706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100406656605SStefano Zampini } 100506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100606656605SStefano Zampini } 100706656605SStefano Zampini } 1008d62866d3SStefano Zampini /* free memory */ 100988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 101106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1014d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1015d62866d3SStefano Zampini if (n_vertices) { 1016d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1017d62866d3SStefano Zampini } 1018d62866d3SStefano Zampini if (n_constraints) { 1019d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1020d62866d3SStefano Zampini } 102188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 102288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102425084f0cSStefano Zampini if (pcbddc->dbg_flag) { 102588ebb749SStefano Zampini Mat coarse_sub_mat; 102625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10308bec7fa6SStefano Zampini Mat C_B,CPHI; 10318bec7fa6SStefano Zampini IS is_dummy; 10328bec7fa6SStefano Zampini Vec mones; 103388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 103488ebb749SStefano Zampini PetscReal real_value; 103588ebb749SStefano Zampini 103688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1042c0553b1fSStefano Zampini if (unsymmetric_check) { 104388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 104588ebb749SStefano Zampini } 104688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104788ebb749SStefano Zampini 104825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 104906656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 105025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1051c0553b1fSStefano Zampini if (unsymmetric_check) { 105288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106488ebb749SStefano Zampini } else { 106588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107388ebb749SStefano Zampini } 107488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10800fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini 10838bec7fa6SStefano Zampini /* check constraints */ 10848bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10858bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10868bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10888bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10908bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10918bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10928bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10938bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10948bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10958bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10968bec7fa6SStefano Zampini 109725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 109888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 109988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 110388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1108c0553b1fSStefano Zampini if (unsymmetric_check) { 110988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111188ebb749SStefano Zampini } 111288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 111388ebb749SStefano Zampini } 11148629588bSStefano Zampini /* get back data */ 11158629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 111688ebb749SStefano Zampini PetscFunctionReturn(0); 111788ebb749SStefano Zampini } 111888ebb749SStefano Zampini 111988ebb749SStefano Zampini #undef __FUNCT__ 1120d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1121d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1122d65f70fdSStefano Zampini { 1123d65f70fdSStefano Zampini Mat *work_mat; 1124d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1125d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1126d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1127d65f70fdSStefano Zampini PetscErrorCode ierr; 1128d65f70fdSStefano Zampini 1129d65f70fdSStefano Zampini PetscFunctionBegin; 1130d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1134d65f70fdSStefano Zampini 1135d65f70fdSStefano Zampini if (!rsorted) { 1136d65f70fdSStefano Zampini const PetscInt *idxs; 1137d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1138d65f70fdSStefano Zampini 1139d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1140d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1142d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1143d65f70fdSStefano Zampini } 1144d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1147d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1148d65f70fdSStefano Zampini } 1149d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1150d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini } else { 1152d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini isrow_s = isrow; 1154d65f70fdSStefano Zampini } 1155d65f70fdSStefano Zampini 1156d65f70fdSStefano Zampini if (!csorted) { 1157d65f70fdSStefano Zampini if (isrow == iscol) { 1158d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini iscol_s = isrow_s; 1160d65f70fdSStefano Zampini } else { 1161d65f70fdSStefano Zampini const PetscInt *idxs; 1162d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1163d65f70fdSStefano Zampini 1164d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1167d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1172d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1173d65f70fdSStefano Zampini } 1174d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1175d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini } 1177d65f70fdSStefano Zampini } else { 1178d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini iscol_s = iscol; 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini 1182d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini 1184d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1185d65f70fdSStefano Zampini Mat new_mat; 1186d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1187d65f70fdSStefano Zampini 1188d65f70fdSStefano Zampini if (!rsorted) { 1189d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1190d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1192d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini } else { 1197d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini } 1199d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini 1201d65f70fdSStefano Zampini if (!csorted) { 1202d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1203d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1205d65f70fdSStefano Zampini } else { 1206d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1207d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1209d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1210d65f70fdSStefano Zampini } 1211d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini } 1214d65f70fdSStefano Zampini } else { 1215d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini } 1217d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini 1219d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini work_mat[0] = new_mat; 1222d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1223d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini } 1225d65f70fdSStefano Zampini 1226d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini *B = work_mat[0]; 1228d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1229d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini PetscFunctionReturn(0); 1232d65f70fdSStefano Zampini } 1233d65f70fdSStefano Zampini 1234d65f70fdSStefano Zampini #undef __FUNCT__ 12355e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12365e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1237aa0d41d4SStefano Zampini { 1238aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12395e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1240d65f70fdSStefano Zampini Mat new_mat; 12415e8657edSStefano Zampini IS is_local,is_global; 1242d65f70fdSStefano Zampini PetscInt local_size; 1243d65f70fdSStefano Zampini PetscBool isseqaij; 1244aa0d41d4SStefano Zampini PetscErrorCode ierr; 1245aa0d41d4SStefano Zampini 1246aa0d41d4SStefano Zampini PetscFunctionBegin; 1247aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12485e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1250906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1252d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1253906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1254906d46d4SStefano Zampini 1255906d46d4SStefano Zampini /* check */ 1256906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1257906d46d4SStefano Zampini Vec x,x_change; 1258906d46d4SStefano Zampini PetscReal error; 1259906d46d4SStefano Zampini 12605e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1261906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12625e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12635e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12645e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1265d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12665e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12675e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1268906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1269906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1270906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1271906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1272906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1273906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1274906d46d4SStefano Zampini } 1275906d46d4SStefano Zampini 127622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12779b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 127822d5777bSStefano Zampini if (isseqaij) { 1279d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1280aa0d41d4SStefano Zampini } else { 1281aa0d41d4SStefano Zampini Mat work_mat; 1282aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1283d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1284aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1285aa0d41d4SStefano Zampini } 1286d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1287e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1288e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1289e496cd5dSStefano Zampini #endif 129045a1bb75SStefano Zampini /* 129145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1292d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 129345a1bb75SStefano Zampini */ 1294d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1295aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1296aa0d41d4SStefano Zampini } 1297aa0d41d4SStefano Zampini 1298aa0d41d4SStefano Zampini #undef __FUNCT__ 1299a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13008ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1301a64d13efSStefano Zampini { 1302a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1303a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1304d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 13056816873aSStefano Zampini IS is_B_reuse_mumps=NULL; 13066816873aSStefano Zampini PetscInt *idx_R_local; 13073a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13083a50541eSStefano Zampini PetscInt vbs,bs; 13096816873aSStefano Zampini PetscBT bitmask=NULL; 1310a64d13efSStefano Zampini PetscErrorCode ierr; 1311a64d13efSStefano Zampini 1312a64d13efSStefano Zampini PetscFunctionBegin; 1313b23d619eSStefano Zampini /* 1314b23d619eSStefano Zampini No need to setup local scatters if 1315b23d619eSStefano Zampini - primal space is unchanged 1316b23d619eSStefano Zampini AND 1317b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1318b23d619eSStefano Zampini AND 1319b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1320b23d619eSStefano Zampini */ 1321b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1322f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1323f4ddd8eeSStefano Zampini } 1324f4ddd8eeSStefano Zampini /* destroy old objects */ 1325f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1326f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1327f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1328a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1329b371cd4fSStefano Zampini n_B = pcis->n_B; 1330b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1331b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13326816873aSStefano Zampini 13336816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13346816873aSStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13356816873aSStefano Zampini 13364641a718SStefano Zampini /* create auxiliary bitmask */ 13376816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 13384641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13394641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13400e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13414641a718SStefano Zampini } 13423a50541eSStefano Zampini 1343a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13444641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13456816873aSStefano Zampini idx_R_local[n_R++] = i; 1346a64d13efSStefano Zampini } 1347a64d13efSStefano Zampini } 13486816873aSStefano Zampini } else { /* A different ordering and auxiliary scatters are needed we are reusing MUMPS Schur solver */ 13496816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13506816873aSStefano Zampini const PetscInt *idxs; 13516816873aSStefano Zampini PetscInt schur_size; 13526816873aSStefano Zampini IS list[2],tisB_N,tisR; 13536816873aSStefano Zampini 13546816873aSStefano Zampini if (sub_schurs->is_dir) { 13556816873aSStefano Zampini list[0] = sub_schurs->is_Ej_all; 13566816873aSStefano Zampini list[1] = sub_schurs->is_dir; 13576816873aSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(pcis->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_dir,&list[1]);CHKERRQ(ierr); 13586816873aSStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&is_B_reuse_mumps);CHKERRQ(ierr); 13596816873aSStefano Zampini ierr = ISDestroy(&list[1]);CHKERRQ(ierr); 13606816873aSStefano Zampini } else { 13616816873aSStefano Zampini ierr = ISDuplicate(sub_schurs->is_Ej_all,&is_B_reuse_mumps);CHKERRQ(ierr); 13626816873aSStefano Zampini } 13636816873aSStefano Zampini ierr = ISGetLocalSize(is_B_reuse_mumps,&schur_size);CHKERRQ(ierr); 13646816873aSStefano Zampini ierr = VecScatterDestroy(&reuse_mumps->correction_scatter_B);CHKERRQ(ierr); 13656816873aSStefano Zampini ierr = VecScatterCreate(pcis->vec1_B,is_B_reuse_mumps,reuse_mumps->solB,NULL,&reuse_mumps->correction_scatter_B);CHKERRQ(ierr); 13666816873aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->BtoNmap,is_B_reuse_mumps,&tisB_N);CHKERRQ(ierr); 13676816873aSStefano Zampini list[0] = pcis->is_I_local; 13686816873aSStefano Zampini list[1] = tisB_N; 13696816873aSStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&tisR);CHKERRQ(ierr); 13706816873aSStefano Zampini ierr = ISDestroy(&tisB_N);CHKERRQ(ierr); 13716816873aSStefano Zampini ierr = ISGetLocalSize(tisR,&n_R);CHKERRQ(ierr); 13726816873aSStefano Zampini ierr = ISGetIndices(tisR,&idxs);CHKERRQ(ierr); 13736816873aSStefano Zampini ierr = PetscMemcpy(idx_R_local,idxs,n_R*sizeof(PetscInt));CHKERRQ(ierr); 13746816873aSStefano Zampini ierr = ISRestoreIndices(tisR,&idxs);CHKERRQ(ierr); 13756816873aSStefano Zampini ierr = ISDestroy(&tisR);CHKERRQ(ierr); 13766816873aSStefano Zampini } 13773a50541eSStefano Zampini 13783a50541eSStefano Zampini /* Block code */ 13793a50541eSStefano Zampini vbs = 1; 13803a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13813a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13823a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13833a50541eSStefano Zampini PetscInt *vary; 13843a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1385785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13863a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13870e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1388964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13893a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13903a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13913a50541eSStefano Zampini break; 13923a50541eSStefano Zampini } 13933a50541eSStefano Zampini } 13943a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13953a50541eSStefano Zampini vbs = bs; 13963a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13973a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13983a50541eSStefano Zampini } 13993a50541eSStefano Zampini } 14003a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 14013a50541eSStefano Zampini } 14023a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 14033a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1404a64d13efSStefano Zampini 1405a64d13efSStefano Zampini /* print some info if requested */ 1406a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1407a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1408a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14090fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1410a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1411a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14123a50541eSStefano 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); 1413a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1414a64d13efSStefano Zampini } 1415a64d13efSStefano Zampini 1416a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14176816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14186816873aSStefano Zampini IS is_aux1,is_aux2; 14196816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14206816873aSStefano Zampini 14213a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1422854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1423854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1424a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14254641a718SStefano Zampini for (i=0; i<n_D; i++) { 14264641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14274641a718SStefano Zampini } 1428a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1429a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14304641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14314641a718SStefano Zampini aux_array1[j++] = i; 1432a64d13efSStefano Zampini } 1433a64d13efSStefano Zampini } 1434a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1435a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1436a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14374641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14384641a718SStefano Zampini aux_array2[j++] = i; 1439a64d13efSStefano Zampini } 1440a64d13efSStefano Zampini } 1441a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1442a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1443a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1444a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1445a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1446a64d13efSStefano Zampini 14478eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1448785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1449a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14504641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14514641a718SStefano Zampini aux_array1[j++] = i; 1452a64d13efSStefano Zampini } 1453a64d13efSStefano Zampini } 1454a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1455a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1456a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1457a64d13efSStefano Zampini } 14584641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14593a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1460d62866d3SStefano Zampini } else { 14616816873aSStefano Zampini IS tis; 14626816873aSStefano Zampini PetscInt schur_size; 14636816873aSStefano Zampini 14646816873aSStefano Zampini ierr = ISGetLocalSize(is_B_reuse_mumps,&schur_size);CHKERRQ(ierr); 14656816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 14666816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,is_B_reuse_mumps,&pcbddc->R_to_B);CHKERRQ(ierr); 14676816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14686816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14696816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14706816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14716816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1472d62866d3SStefano Zampini } 1473d62866d3SStefano Zampini } 14746816873aSStefano Zampini ierr = ISDestroy(&is_B_reuse_mumps);CHKERRQ(ierr); 1475a64d13efSStefano Zampini PetscFunctionReturn(0); 1476a64d13efSStefano Zampini } 1477a64d13efSStefano Zampini 1478304d26faSStefano Zampini 1479304d26faSStefano Zampini #undef __FUNCT__ 1480304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1481684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1482304d26faSStefano Zampini { 1483304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1484304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1485304d26faSStefano Zampini PC pc_temp; 1486304d26faSStefano Zampini Mat A_RR; 1487f4ddd8eeSStefano Zampini MatReuse reuse; 1488304d26faSStefano Zampini PetscScalar m_one = -1.0; 1489304d26faSStefano Zampini PetscReal value; 1490af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14919577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1492304d26faSStefano Zampini PetscErrorCode ierr; 1493e604994aSStefano Zampini /* prefixes stuff */ 1494312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1495e604994aSStefano Zampini size_t len; 1496304d26faSStefano Zampini 1497304d26faSStefano Zampini PetscFunctionBegin; 1498304d26faSStefano Zampini 1499e604994aSStefano Zampini /* compute prefixes */ 1500e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1501e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1502e604994aSStefano Zampini if (!pcbddc->current_level) { 1503e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1504e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1505e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1506e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1507e604994aSStefano Zampini } else { 1508e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1509312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1510e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1511e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1512312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1513312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 151434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 151534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1516e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1517e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1518e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1520e604994aSStefano Zampini } 1521e604994aSStefano Zampini 1522304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1523684f6988SStefano Zampini if (dirichlet) { 1524d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1525964fefecSStefano Zampini if (pcbddc->issym) { 1526964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1527964fefecSStefano Zampini } 1528ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1529964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1530304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1531304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1532304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1533304d26faSStefano Zampini /* default */ 1534304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1535e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15369577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1537304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15389577ea80SStefano Zampini if (issbaij) { 15399577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15409577ea80SStefano Zampini } else { 1541304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15429577ea80SStefano Zampini } 1543304d26faSStefano Zampini /* Allow user's customization */ 1544304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1545304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1546304d26faSStefano Zampini } 1547d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1548d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1549d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1550d62866d3SStefano Zampini 1551d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1552d5574798SStefano Zampini } 1553304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1554304d26faSStefano Zampini if (!n_D) { 1555304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1556304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1557304d26faSStefano Zampini } 1558304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1559304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1560304d26faSStefano Zampini /* set ksp_D into pcis data */ 1561304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1562304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1563304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1564684f6988SStefano Zampini } 1565304d26faSStefano Zampini 1566304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1567684f6988SStefano Zampini A_RR = 0; 1568684f6988SStefano Zampini if (neumann) { 1569d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1570f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15718ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1572f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1573f4ddd8eeSStefano Zampini PetscInt nn_R; 157481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1575f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1576f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1578f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1579f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1580f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1581f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1582727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1583f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1585f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1586f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1587f4ddd8eeSStefano Zampini } 1588f4ddd8eeSStefano Zampini } 1589f4ddd8eeSStefano Zampini /* last check */ 1590d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1591f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1592f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1593f4ddd8eeSStefano Zampini } 1594f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1595f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1596f4ddd8eeSStefano Zampini } 15976816873aSStefano Zampini /* extract A_RR */ 15986816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 15996816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16006816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 16016816873aSStefano Zampini if (ibs != mbs) { 16026816873aSStefano Zampini Mat newmat; 16036816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 16046816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16056816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 16066816873aSStefano Zampini } else { 16076816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16086816873aSStefano Zampini } 16096816873aSStefano Zampini if (pcbddc->issym) { 16106816873aSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 16116816873aSStefano Zampini } 16126816873aSStefano Zampini } else { 16136816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16146816873aSStefano Zampini 16156816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16166816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16176816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16186816873aSStefano Zampini } 1619f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1620304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1621304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1622304d26faSStefano Zampini /* default */ 1623304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1624e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1625304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16269577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16279577ea80SStefano Zampini if (issbaij) { 16289577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16299577ea80SStefano Zampini } else { 1630304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16319577ea80SStefano Zampini } 1632304d26faSStefano Zampini /* Allow user's customization */ 1633304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1634304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1635304d26faSStefano Zampini } 1636d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1637304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1638304d26faSStefano Zampini if (!n_R) { 1639304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1640304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1641304d26faSStefano Zampini } 1642d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1643d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1644d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1645d62866d3SStefano Zampini 1646d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1647d62866d3SStefano Zampini } 1648304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1649304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1650684f6988SStefano Zampini } 16516816873aSStefano Zampini /* free Neumann problem's matrix */ 16526816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1653304d26faSStefano Zampini 1654304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16550fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1656684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1657684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1658684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1659684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1660684f6988SStefano Zampini } 1661684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16620fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16630fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16640fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16650fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16660fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1667304d26faSStefano Zampini /* need to be adapted? */ 1668b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1669b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1670b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1671304d26faSStefano Zampini /* print info */ 1672304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1673e604994aSStefano 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); 1674304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1675304d26faSStefano Zampini } 1676b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16778ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1678304d26faSStefano Zampini } 1679684f6988SStefano Zampini } 1680684f6988SStefano Zampini if (neumann) { /* Neumann */ 16816816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16820fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1687304d26faSStefano Zampini /* need to be adapted? */ 1688b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1689b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1690304d26faSStefano Zampini /* print info */ 1691304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1692e604994aSStefano 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); 1693304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1694304d26faSStefano Zampini } 1695b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16968ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1697304d26faSStefano Zampini } 16980fccc4e9SStefano Zampini } 1699684f6988SStefano Zampini } 1700304d26faSStefano Zampini PetscFunctionReturn(0); 1701304d26faSStefano Zampini } 1702304d26faSStefano Zampini 1703304d26faSStefano Zampini #undef __FUNCT__ 1704ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 170580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1706674ae819SStefano Zampini { 1707674ae819SStefano Zampini PetscErrorCode ierr; 1708674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1709*be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1710674ae819SStefano Zampini 1711674ae819SStefano Zampini PetscFunctionBegin; 1712*be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1714*be83ff47SStefano Zampini } 171580677318SStefano Zampini if (!pcbddc->switch_static) { 171680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 171780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 171880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 171920c7b377SStefano Zampini } 1720*be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172320c7b377SStefano Zampini } else { 1724*be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1725*be83ff47SStefano Zampini 1726*be83ff47SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhsB,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1727*be83ff47SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhsB,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1728*be83ff47SStefano Zampini } 1729*be83ff47SStefano Zampini } else { 173080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 173680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173920c7b377SStefano Zampini } 1740674ae819SStefano Zampini } 1741*be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174280677318SStefano Zampini if (applytranspose) { 174380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174480677318SStefano Zampini } else { 174580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174680677318SStefano Zampini } 1747*be83ff47SStefano Zampini } else { 1748*be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1749*be83ff47SStefano Zampini 1750*be83ff47SStefano Zampini if (applytranspose) { 1751*be83ff47SStefano Zampini /* MatMumpsSolveSchurComplementTranspose */ 1752*be83ff47SStefano Zampini } else { 1753*be83ff47SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhsB,reuse_mumps->solB);CHKERRQ(ierr); 1754*be83ff47SStefano Zampini } 1755*be83ff47SStefano Zampini } 175680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 175780677318SStefano Zampini if (!pcbddc->switch_static) { 1758*be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1761*be83ff47SStefano Zampini } else { 1762*be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1763*be83ff47SStefano Zampini 1764*be83ff47SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->solB,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1765*be83ff47SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->solB,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1766*be83ff47SStefano Zampini } 176780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 176880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 176980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 177080677318SStefano Zampini } 177180677318SStefano Zampini } else { 177280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 177980677318SStefano Zampini } 178080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178480677318SStefano Zampini } 1785674ae819SStefano Zampini PetscFunctionReturn(0); 1786674ae819SStefano Zampini } 1787674ae819SStefano Zampini 1788dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1789674ae819SStefano Zampini #undef __FUNCT__ 1790674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1791dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1792674ae819SStefano Zampini { 1793674ae819SStefano Zampini PetscErrorCode ierr; 1794674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1795674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1796674ae819SStefano Zampini const PetscScalar zero = 0.0; 1797674ae819SStefano Zampini 1798674ae819SStefano Zampini PetscFunctionBegin; 1799dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1800dc359a40SStefano Zampini if (applytranspose) { 1801674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18028eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1803dc359a40SStefano Zampini } else { 1804674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1805674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 180615aaf578SStefano Zampini } 180712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 180812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 180912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181112edc857SStefano Zampini 18129f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 181312edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 181412edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1815964fefecSStefano Zampini Vec rhs,sol; 1816964fefecSStefano Zampini 1817964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1818964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 181912edc857SStefano Zampini if (applytranspose) { 1820964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182112edc857SStefano Zampini } else { 1822964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182312edc857SStefano Zampini } 182412edc857SStefano Zampini } 1825674ae819SStefano Zampini 1826674ae819SStefano Zampini /* Local solution on R nodes */ 182780677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 182880677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18299f00e9b4SStefano Zampini } 1830674ae819SStefano Zampini 18319f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18329f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 183312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1834674ae819SStefano Zampini 1835674ae819SStefano Zampini /* Sum contributions from two levels */ 1836dc359a40SStefano Zampini if (applytranspose) { 1837dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1838dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1839dc359a40SStefano Zampini } else { 1840674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18418eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1842dc359a40SStefano Zampini } 1843674ae819SStefano Zampini PetscFunctionReturn(0); 1844674ae819SStefano Zampini } 1845674ae819SStefano Zampini 1846674ae819SStefano Zampini #undef __FUNCT__ 1847674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 184812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1849674ae819SStefano Zampini { 1850674ae819SStefano Zampini PetscErrorCode ierr; 1851674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 185258da7f69SStefano Zampini PetscScalar *array; 185312edc857SStefano Zampini Vec from,to; 1854674ae819SStefano Zampini 1855674ae819SStefano Zampini PetscFunctionBegin; 185612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 185712edc857SStefano Zampini from = pcbddc->coarse_vec; 185812edc857SStefano Zampini to = pcbddc->vec1_P; 185912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 186012edc857SStefano Zampini Vec tvec; 186158da7f69SStefano Zampini 186258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 186412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 186658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 186758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 186812edc857SStefano Zampini } 186912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187012edc857SStefano Zampini from = pcbddc->vec1_P; 187112edc857SStefano Zampini to = pcbddc->coarse_vec; 187212edc857SStefano Zampini } 187312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1874674ae819SStefano Zampini PetscFunctionReturn(0); 1875674ae819SStefano Zampini } 1876674ae819SStefano Zampini 1877674ae819SStefano Zampini #undef __FUNCT__ 1878674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 187912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1880674ae819SStefano Zampini { 1881674ae819SStefano Zampini PetscErrorCode ierr; 1882674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 188358da7f69SStefano Zampini PetscScalar *array; 188412edc857SStefano Zampini Vec from,to; 1885674ae819SStefano Zampini 1886674ae819SStefano Zampini PetscFunctionBegin; 188712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 188812edc857SStefano Zampini from = pcbddc->coarse_vec; 188912edc857SStefano Zampini to = pcbddc->vec1_P; 189012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189112edc857SStefano Zampini from = pcbddc->vec1_P; 189212edc857SStefano Zampini to = pcbddc->coarse_vec; 189312edc857SStefano Zampini } 189412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 189512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 189612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 189712edc857SStefano Zampini Vec tvec; 189858da7f69SStefano Zampini 189912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 190058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 190158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 190258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 190358da7f69SStefano Zampini } 190458da7f69SStefano Zampini } else { 190558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 190658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 190712edc857SStefano Zampini } 190812edc857SStefano Zampini } 1909674ae819SStefano Zampini PetscFunctionReturn(0); 1910674ae819SStefano Zampini } 1911674ae819SStefano Zampini 1912984c4197SStefano Zampini /* uncomment for testing purposes */ 1913984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1914674ae819SStefano Zampini #undef __FUNCT__ 1915674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1916674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1917674ae819SStefano Zampini { 1918674ae819SStefano Zampini PetscErrorCode ierr; 1919674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1920674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1921674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1922984c4197SStefano Zampini /* one and zero */ 1923984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1924984c4197SStefano Zampini /* space to store constraints and their local indices */ 19259162d606SStefano Zampini PetscScalar *constraints_data; 19269162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19279162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19289162d606SStefano Zampini PetscInt *constraints_n; 1929984c4197SStefano Zampini /* iterators */ 1930b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1931984c4197SStefano Zampini /* BLAS integers */ 1932e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1933e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1934c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1935727cdba6SStefano Zampini /* reuse */ 19360e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19370e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1938984c4197SStefano Zampini /* change of basis */ 1939b3d85658SStefano Zampini PetscBool qr_needed; 19409162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1941984c4197SStefano Zampini /* auxiliary stuff */ 194264efe560SStefano Zampini PetscInt *nnz,*is_indices; 19438a0068c3SStefano Zampini PetscInt ncc; 1944984c4197SStefano Zampini /* some quantities */ 194545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1946a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1947984c4197SStefano Zampini 1948674ae819SStefano Zampini PetscFunctionBegin; 19498e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19508e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19518e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1952088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1953088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19540e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19550e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19560e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19570e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19580e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1959088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1960cf5a6209SStefano Zampini 1961cf5a6209SStefano Zampini /* print some info */ 1962cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1963cf5a6209SStefano Zampini IS vertices; 1964cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1965cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1966cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1967cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1968cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1969cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1970cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1971fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1972fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1973cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1974cf5a6209SStefano Zampini } 1975cf5a6209SStefano Zampini 1976cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19779162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1978cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1979cf5a6209SStefano Zampini const Vec *nearnullvecs; 1980cf5a6209SStefano Zampini Vec *localnearnullsp; 1981cf5a6209SStefano Zampini PetscScalar *array; 1982cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1983cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1984cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1985b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1986cf5a6209SStefano Zampini PetscScalar *work; 1987cf5a6209SStefano Zampini PetscReal *singular_vals; 1988cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1989cf5a6209SStefano Zampini PetscReal *rwork; 1990cf5a6209SStefano Zampini #endif 1991cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1992cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1993cf5a6209SStefano Zampini #else 1994964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1995964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1996cf5a6209SStefano Zampini #endif 1997cf5a6209SStefano Zampini 1998674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1999d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2000d06fc5fdSStefano Zampini /* free unneeded index sets */ 2001d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2002d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2003674ae819SStefano Zampini } 2004d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2005d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2006d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2007d06fc5fdSStefano Zampini } 2008d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2009d06fc5fdSStefano Zampini n_ISForEdges = 0; 2010d06fc5fdSStefano Zampini } 2011d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2012d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2013d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2014d06fc5fdSStefano Zampini } 2015d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2016d06fc5fdSStefano Zampini n_ISForFaces = 0; 2017d06fc5fdSStefano Zampini } 201870022509SStefano Zampini 201970022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 202070022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 202170022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 202270022509SStefano Zampini if (pcbddc->NullSpace) { 202370022509SStefano Zampini PetscBool tbool[2],gbool[2]; 202470022509SStefano Zampini 202570022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2026b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2027d06fc5fdSStefano Zampini if (!ISForEdges) { 2028d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2029d06fc5fdSStefano Zampini } 2030b8ffe317SStefano Zampini } 2031d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2032d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2033d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2034d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2035d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 203698a51de6SStefano Zampini } 203770022509SStefano Zampini #endif 203808122e43SStefano Zampini 2039674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2040674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2041674ae819SStefano Zampini if (nearnullsp) { 2042674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2043f4ddd8eeSStefano Zampini /* remove any stored info */ 2044f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2045f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2046f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2047f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2049473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2050f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2051f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2052f4ddd8eeSStefano Zampini } 2053984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2054984c4197SStefano Zampini nnsp_size = 0; 2055674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2056674ae819SStefano Zampini } 2057984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2058984c4197SStefano Zampini max_constraints = nnsp_size; 2059984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2060984c4197SStefano Zampini 2061674ae819SStefano Zampini /* 2062674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20639162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20649162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20659162d606SStefano Zampini There can be multiple constraints per connected component 2066674ae819SStefano Zampini */ 2067674ae819SStefano Zampini n_vertices = 0; 2068674ae819SStefano Zampini if (ISForVertices) { 2069674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2070674ae819SStefano Zampini } 20719162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20729162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20739162d606SStefano Zampini 20749162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20759162d606SStefano Zampini total_counts *= max_constraints; 2076674ae819SStefano Zampini total_counts += n_vertices; 20774641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20789162d606SStefano Zampini 2079674ae819SStefano Zampini total_counts = 0; 2080674ae819SStefano Zampini max_size_of_constraint = 0; 2081674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20829162d606SStefano Zampini IS used_is; 2083674ae819SStefano Zampini if (i<n_ISForEdges) { 20849162d606SStefano Zampini used_is = ISForEdges[i]; 2085674ae819SStefano Zampini } else { 20869162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2087674ae819SStefano Zampini } 20889162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2089674ae819SStefano Zampini total_counts += j; 2090674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2091674ae819SStefano Zampini } 20929162d606SStefano 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); 20939162d606SStefano Zampini 2094984c4197SStefano Zampini /* get local part of global near null space vectors */ 2095785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2096984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2097984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2098984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2099984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2100984c4197SStefano Zampini } 2101674ae819SStefano Zampini 2102242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2103242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2104a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2105242a89d7SStefano Zampini 2106984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2107a773dcb8SStefano Zampini if (!skip_lapack) { 2108674ae819SStefano Zampini PetscScalar temp_work; 2109911cabfeSStefano Zampini 2110674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2111984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2112785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2113785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2114785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2115674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2116785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2117674ae819SStefano Zampini #endif 2118674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2119c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2120c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2121674ae819SStefano Zampini lwork = -1; 2122674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2123674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2124c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2125674ae819SStefano Zampini #else 2126c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2127674ae819SStefano Zampini #endif 2128674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2129984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2130674ae819SStefano Zampini #else /* on missing GESVD */ 2131674ae819SStefano Zampini /* SVD */ 2132674ae819SStefano Zampini PetscInt max_n,min_n; 2133674ae819SStefano Zampini max_n = max_size_of_constraint; 2134984c4197SStefano Zampini min_n = max_constraints; 2135984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2136674ae819SStefano Zampini min_n = max_size_of_constraint; 2137984c4197SStefano Zampini max_n = max_constraints; 2138674ae819SStefano Zampini } 2139785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2140674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2141785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2142674ae819SStefano Zampini #endif 2143674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2144674ae819SStefano Zampini lwork = -1; 2145e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2146e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2147b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2148674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2149674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21509162d606SStefano 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)); 2151674ae819SStefano Zampini #else 21529162d606SStefano 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)); 2153674ae819SStefano Zampini #endif 2154674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2155984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2156984c4197SStefano Zampini #endif /* on missing GESVD */ 2157674ae819SStefano Zampini /* Allocate optimal workspace */ 2158674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2159854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2160674ae819SStefano Zampini } 2161674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2162674ae819SStefano Zampini total_counts = 0; 21639162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21649162d606SStefano Zampini constraints_data_ptr[0] = 0; 2165674ae819SStefano Zampini /* vertices */ 21669162d606SStefano Zampini if (n_vertices) { 2167674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21689162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21699162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2170674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21719162d606SStefano Zampini constraints_n[total_counts] = 1; 21729162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21739162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21749162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2175674ae819SStefano Zampini total_counts++; 2176674ae819SStefano Zampini } 2177674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2178984c4197SStefano Zampini PetscBool used_vertex; 2179674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2180674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2181674ae819SStefano Zampini k = 0; 2182674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2183984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2184984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21859162d606SStefano Zampini constraints_n[total_counts] = 1; 21869162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21879162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21889162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21899162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2190674ae819SStefano Zampini total_counts++; 2191674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2192674ae819SStefano Zampini } 2193984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2194674ae819SStefano Zampini k++; 2195674ae819SStefano Zampini } 2196674ae819SStefano Zampini } 2197674ae819SStefano Zampini } 2198674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2199674ae819SStefano Zampini n_vertices = total_counts; 2200674ae819SStefano Zampini } 2201984c4197SStefano Zampini 2202674ae819SStefano Zampini /* edges and faces */ 22039162d606SStefano Zampini total_counts_cc = total_counts; 2204911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22059162d606SStefano Zampini IS used_is; 22069162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22079162d606SStefano Zampini 2208911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22099162d606SStefano Zampini used_is = ISForEdges[ncc]; 2210984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2211674ae819SStefano Zampini } else { 22129162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2213984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2214674ae819SStefano Zampini } 2215674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22169162d606SStefano Zampini 22179162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22189162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2219984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2220984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2221674ae819SStefano Zampini if (nnsp_has_cnst) { 22225b08dc53SStefano Zampini PetscScalar quad_value; 22239162d606SStefano Zampini 22249162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22269162d606SStefano Zampini 2227a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2228674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2229a773dcb8SStefano Zampini } else { 2230a773dcb8SStefano Zampini quad_value = 1.0; 2231a773dcb8SStefano Zampini } 2232674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22339162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2234674ae819SStefano Zampini } 22359162d606SStefano Zampini temp_constraints++; 2236674ae819SStefano Zampini total_counts++; 2237674ae819SStefano Zampini } 2238674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2239984c4197SStefano Zampini PetscReal real_value; 22409162d606SStefano Zampini PetscScalar *ptr_to_data; 22419162d606SStefano Zampini 2242984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22439162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2244674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22459162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2246674ae819SStefano Zampini } 2247984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2248984c4197SStefano Zampini /* check if array is null on the connected component */ 2249e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22509162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22515b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2252674ae819SStefano Zampini temp_constraints++; 2253674ae819SStefano Zampini total_counts++; 22549162d606SStefano Zampini if (!idxs_copied) { 22559162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22569162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2257674ae819SStefano Zampini } 2258674ae819SStefano Zampini } 22599162d606SStefano Zampini } 22609162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 226145a1bb75SStefano Zampini valid_constraints = temp_constraints; 2262eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2263a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22649162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22659162d606SStefano Zampini 22669162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2267a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22689162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2269a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22709162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2271a773dcb8SStefano Zampini } else { /* perform SVD */ 2272984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22739162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2274674ae819SStefano Zampini 2275674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2276984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2277984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2278984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2279984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2280984c4197SStefano Zampini from that computed using LAPACKgesvd 2281984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2282984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2283984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2284674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2285e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2286984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2287674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2288674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22899162d606SStefano 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)); 2290674ae819SStefano Zampini } 2291674ae819SStefano Zampini } 2292e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2293e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2294e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2295674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2296c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2297674ae819SStefano Zampini #else 2298c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2299674ae819SStefano Zampini #endif 2300674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2301984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2302984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2303674ae819SStefano Zampini j = 0; 2304984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2305674ae819SStefano Zampini total_counts = total_counts-j; 230645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2307e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2308c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2309c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2310c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2311c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2312c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2313c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2314674ae819SStefano Zampini if (j<temp_constraints) { 2315984c4197SStefano Zampini PetscInt ii; 2316984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2317674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23189162d606SStefano 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)); 2319674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2320984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2321674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23229162d606SStefano 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]; 2323674ae819SStefano Zampini } 2324674ae819SStefano Zampini } 2325674ae819SStefano Zampini } 2326674ae819SStefano Zampini #else /* on missing GESVD */ 2327e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2328e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2329b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2330674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2331674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23329162d606SStefano 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)); 2333674ae819SStefano Zampini #else 23349162d606SStefano 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)); 2335674ae819SStefano Zampini #endif 2336984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2337674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2338984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2339e310c8b4SStefano Zampini k = temp_constraints; 2340e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2341674ae819SStefano Zampini j = 0; 2342e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 234345a1bb75SStefano Zampini valid_constraints = k-j; 2344911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2345984c4197SStefano Zampini #endif /* on missing GESVD */ 2346674ae819SStefano Zampini } 2347a773dcb8SStefano Zampini } 23489162d606SStefano Zampini /* update pointers information */ 23499162d606SStefano Zampini if (valid_constraints) { 23509162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23519162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23529162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23539162d606SStefano Zampini /* set change_of_basis flag */ 235445a1bb75SStefano Zampini if (boolforchange) { 2355b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23569162d606SStefano Zampini } 2357b3d85658SStefano Zampini total_counts_cc++; 235845a1bb75SStefano Zampini } 2359674ae819SStefano Zampini } 2360984c4197SStefano Zampini /* free workspace */ 23618f1c130eSStefano Zampini if (!skip_lapack) { 2362984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2363984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2364984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2365984c4197SStefano Zampini #endif 2366984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2367984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2368984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2369984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2370984c4197SStefano Zampini #endif 2371984c4197SStefano Zampini } 2372984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2373984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2374984c4197SStefano Zampini } 2375984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2376cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2377cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2378cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2379cf5a6209SStefano Zampini } 2380cf5a6209SStefano Zampini if (n_ISForFaces) { 2381cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2382cf5a6209SStefano Zampini } 2383cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2384cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2385cf5a6209SStefano Zampini } 2386cf5a6209SStefano Zampini if (n_ISForEdges) { 2387cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2388cf5a6209SStefano Zampini } 2389cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 239008122e43SStefano Zampini } else { 239108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 239208122e43SStefano Zampini 239308122e43SStefano Zampini total_counts = 0; 239408122e43SStefano Zampini n_vertices = 0; 2395d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2396d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 239708122e43SStefano Zampini } 239808122e43SStefano Zampini max_constraints = 0; 23999162d606SStefano Zampini total_counts_cc = 0; 240008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 240108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24029162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 240308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 240408122e43SStefano Zampini } 24059162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24069162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24079162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24089162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 240974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24109162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24119162d606SStefano Zampini total_counts_cc = 0; 24129162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24139162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24149162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 241508122e43SStefano Zampini } 241608122e43SStefano Zampini } 24179162d606SStefano Zampini #if 0 24189162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24199162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24209162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24219162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24229162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24239162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24249162d606SStefano Zampini } 24259162d606SStefano Zampini printf("\n"); 24269162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24279162d606SStefano Zampini } 24281b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24298bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24301b968477SStefano Zampini } 24311b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24328bec7fa6SStefano 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]); 24331b968477SStefano Zampini } 243408122e43SStefano Zampini #endif 243508122e43SStefano Zampini 24368bec7fa6SStefano Zampini max_size_of_constraint = 0; 24379162d606SStefano 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]); 24389162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 243908122e43SStefano Zampini /* Change of basis */ 2440b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 244108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 244208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 244308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2444b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 244508122e43SStefano Zampini } 244608122e43SStefano Zampini } 244708122e43SStefano Zampini } 244808122e43SStefano Zampini } 2449b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24500e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 245108122e43SStefano Zampini 24529162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24539162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24549162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24559162d606SStefano 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); 245608122e43SStefano Zampini } 2457984c4197SStefano Zampini 2458674ae819SStefano Zampini /* Create constraint matrix */ 2459674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 246016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2461984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2462984c4197SStefano Zampini 2463984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 246474d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 246574d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 246674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2467984c4197SStefano Zampini total_primal_vertices=0; 2468b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24699162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24709162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2471984c4197SStefano Zampini if (size_of_constraint == 1) { 24729162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2473b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 247464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24759162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24769162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24779162d606SStefano Zampini } 2478b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 247974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 248074d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 248174d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 248274d5cdf7SStefano Zampini } 2483b3d85658SStefano Zampini } else { 2484b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2485984c4197SStefano Zampini } 2486984c4197SStefano Zampini } 2487b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2488b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 248970022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 249070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2491b3d85658SStefano Zampini 2492b3d85658SStefano 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); 24930e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24940e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 249570022509SStefano Zampini 2496984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 249774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2498785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2499984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 250074d5cdf7SStefano Zampini 2501984c4197SStefano Zampini j = total_primal_vertices; 250274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2503b3d85658SStefano Zampini cum = total_primal_vertices; 25049162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25054641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2506b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2507b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2508b3d85658SStefano Zampini cum++; 25099162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 251074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 251174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 251274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 251374d5cdf7SStefano Zampini } 25149162d606SStefano Zampini j += constraints_n[i]; 2515674ae819SStefano Zampini } 2516674ae819SStefano Zampini } 2517674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2518674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2519088faed8SStefano Zampini 2520674ae819SStefano Zampini /* set values in constraint matrix */ 2521984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25220e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2523674ae819SStefano Zampini } 2524984c4197SStefano Zampini total_counts = total_primal_vertices; 25259162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25264641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25279162d606SStefano Zampini PetscInt *cols; 25289162d606SStefano Zampini 25299162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25309162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25319162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25329162d606SStefano Zampini PetscInt row = total_counts+k; 25339162d606SStefano Zampini PetscScalar *vals; 25349162d606SStefano Zampini 25359162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25369162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25379162d606SStefano Zampini } 25389162d606SStefano Zampini total_counts += constraints_n[i]; 2539674ae819SStefano Zampini } 2540674ae819SStefano Zampini } 2541674ae819SStefano Zampini /* assembling */ 2542674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2543674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2544088faed8SStefano Zampini 2545984c4197SStefano Zampini /* 254645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2547984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2548984c4197SStefano Zampini */ 2549674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2550674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2551026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2552984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2553984c4197SStefano Zampini /* working stuff for GEQRF */ 255481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2555984c4197SStefano Zampini PetscBLASInt lqr_work; 2556984c4197SStefano Zampini /* working stuff for UNGQR */ 2557984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2558984c4197SStefano Zampini PetscBLASInt lgqr_work; 2559984c4197SStefano Zampini /* working stuff for TRTRS */ 2560984c4197SStefano Zampini PetscScalar *trs_rhs; 25613f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2562984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2563984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2564984c4197SStefano Zampini PetscScalar *start_vals; 2565984c4197SStefano Zampini /* working stuff for values insertion */ 25664641a718SStefano Zampini PetscBT is_primal; 256764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2568906d46d4SStefano Zampini /* matrix sizes */ 2569906d46d4SStefano Zampini PetscInt global_size,local_size; 2570906d46d4SStefano Zampini /* temporary change of basis */ 2571906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2572cf5a6209SStefano Zampini /* extra space for debugging */ 2573cf5a6209SStefano Zampini PetscScalar *dbg_work; 2574984c4197SStefano Zampini 2575906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2576906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 257716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2578bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2579906d46d4SStefano Zampini /* nonzeros for local mat */ 2580bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2581bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25829162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2583a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2585a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25869162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2587a717540cSStefano Zampini } else { 25889162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25899162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2590a717540cSStefano Zampini } 2591a717540cSStefano Zampini } 2592a717540cSStefano Zampini } 2593906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2594bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2595a717540cSStefano Zampini /* Set initial identity in the matrix */ 2596bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2597906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2598a717540cSStefano Zampini } 2599a717540cSStefano Zampini 2600a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2601a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2602a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2603a717540cSStefano Zampini } 2604a717540cSStefano Zampini 2605a717540cSStefano Zampini 2606a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2607a717540cSStefano Zampini /* 2608a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2609a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2610a717540cSStefano Zampini 2611a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2612a717540cSStefano Zampini 2613a6b551f4SStefano 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) 2614a6b551f4SStefano Zampini 2615a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2616a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2617a717540cSStefano Zampini | ... | 2618a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2619a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2620a717540cSStefano Zampini 2621a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2622a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2623a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2624a6b551f4SStefano Zampini 2625a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2626a717540cSStefano Zampini */ 2627a717540cSStefano Zampini if (qr_needed) { 2628984c4197SStefano Zampini /* space to store Q */ 2629854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2630984c4197SStefano Zampini /* first we issue queries for optimal work */ 26313f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26323f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26333f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2634984c4197SStefano Zampini lqr_work = -1; 26353f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2636984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2637984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2638785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2639984c4197SStefano Zampini lgqr_work = -1; 26403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26443f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26453f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2646984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2647984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2648785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2649984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2650785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2651984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2652785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2653a717540cSStefano Zampini /* allocating workspace for check */ 2654a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2655cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2656a717540cSStefano Zampini } 2657a717540cSStefano Zampini } 2658984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26594641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2660473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26610e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 266239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 266339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26644641a718SStefano Zampini } 266539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 266639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 266739e2fb2aSStefano Zampini } 266839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2669984c4197SStefano Zampini 2670a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26719162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26734641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2674984c4197SStefano Zampini /* get constraint info */ 26759162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2676984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2677984c4197SStefano Zampini 2678984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26799162d606SStefano 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); 2680674ae819SStefano Zampini } 2681984c4197SStefano Zampini 2682fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2683a717540cSStefano Zampini 2684a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2685a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26869162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2687a717540cSStefano Zampini } 2688984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26899162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2690984c4197SStefano Zampini 2691984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2695674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26963f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2697984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2698674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2699984c4197SStefano Zampini 2700984c4197SStefano Zampini /* explictly compute R^-T */ 2701984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2702984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27053f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27063f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2707984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27083f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2709984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2710984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2711984c4197SStefano Zampini 2712a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2717984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27183f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2719984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2720984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2721984c4197SStefano Zampini 2722984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2723984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2724984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2731984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27329162d606SStefano 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)); 2733984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27349162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2735984c4197SStefano Zampini 2736984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27379162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2738984c4197SStefano Zampini /* insert cols for primal dofs */ 2739984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2740984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27419162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2742906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2743984c4197SStefano Zampini } 2744984c4197SStefano Zampini /* insert cols for dual dofs */ 2745984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27469162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2747984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27489162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2749906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2750984c4197SStefano Zampini j++; 2751674ae819SStefano Zampini } 2752674ae819SStefano Zampini } 2753984c4197SStefano Zampini 2754984c4197SStefano Zampini /* check change of basis */ 2755984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2756984c4197SStefano Zampini PetscInt ii,jj; 2757984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2758c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2759c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2760c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2761c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2762c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2763c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2764984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2765cf5a6209SStefano 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)); 2766984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2767984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2768984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2769cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2770cf5a6209SStefano 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; 2771674ae819SStefano Zampini } 2772674ae819SStefano Zampini } 2773984c4197SStefano Zampini if (!valid_qr) { 277422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2775984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2776984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2777cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2778cf5a6209SStefano 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])); 2779674ae819SStefano Zampini } 2780cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2781cf5a6209SStefano 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])); 2782984c4197SStefano Zampini } 2783984c4197SStefano Zampini } 2784984c4197SStefano Zampini } 2785674ae819SStefano Zampini } else { 278622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2787674ae819SStefano Zampini } 2788674ae819SStefano Zampini } 2789a717540cSStefano Zampini } else { /* simple transformation block */ 2790a717540cSStefano Zampini PetscInt row,col; 2791a6b551f4SStefano Zampini PetscScalar val,norm; 2792a6b551f4SStefano Zampini 2793a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27949162d606SStefano 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)); 2795a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27969162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27979162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2798bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 27999162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2800906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28019162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2802a717540cSStefano Zampini } else { 2803a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28049162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2805a717540cSStefano Zampini if (row != col) { 28069162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2807a717540cSStefano Zampini } else { 28089162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2809a717540cSStefano Zampini } 2810906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2811a717540cSStefano Zampini } 2812a717540cSStefano Zampini } 2813a717540cSStefano Zampini } 281498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 281522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2816a717540cSStefano Zampini } 2817674ae819SStefano Zampini } 2818984c4197SStefano Zampini } else { 2819984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28209162d606SStefano 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); 2821674ae819SStefano Zampini } 2822674ae819SStefano Zampini } 2823674ae819SStefano Zampini } 2824a717540cSStefano Zampini 2825a717540cSStefano Zampini /* free workspace */ 2826a717540cSStefano Zampini if (qr_needed) { 2827984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2828cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2829984c4197SStefano Zampini } 2830984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2831984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2832984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2833984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2834984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2835674ae819SStefano Zampini } 2836a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2837906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2838906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2839906d46d4SStefano Zampini 2840906d46d4SStefano Zampini /* assembling of global change of variable */ 2841bbb9e6c6SStefano Zampini { 2842bbb9e6c6SStefano Zampini Mat tmat; 284316f15bc4SStefano Zampini PetscInt bs; 284416f15bc4SStefano Zampini 2845906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2846906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2847bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2848bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2850bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 285116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 285216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2853906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2854bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2855bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2860bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2861bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2862bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2863906d46d4SStefano Zampini } 2864906d46d4SStefano Zampini /* check */ 2865906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2866906d46d4SStefano Zampini PetscReal error; 2867906d46d4SStefano Zampini Vec x,x_change; 2868906d46d4SStefano Zampini 2869906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2871906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2873bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2874bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2875bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2878906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2879906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2880906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2882bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2883906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2884906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2885906d46d4SStefano Zampini } 2886b96c3477SStefano Zampini 2887b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2888b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2889b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2890b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2891ac632422SStefano Zampini Mat S_new,tmat; 2892bbb9e6c6SStefano Zampini IS is_all_N; 2893bbb9e6c6SStefano Zampini 2894bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28956816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2896bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2897ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2898b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2899ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2900ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2901ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2902ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2903ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2904b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2905ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2906ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2907ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2908ac632422SStefano Zampini } 2909b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2910b96c3477SStefano Zampini } 2911b96c3477SStefano Zampini } 2912906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2913906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2914b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2915b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2916b9b85e73SStefano Zampini } 2917906d46d4SStefano Zampini 2918906d46d4SStefano Zampini /* set up change of basis context */ 2919906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2920906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2921906d46d4SStefano Zampini 2922906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2923906d46d4SStefano Zampini PetscInt global_size,local_size; 2924906d46d4SStefano Zampini 2925906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2926906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2927906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2928906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2929906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2930906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2931906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2932906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2933906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2934906d46d4SStefano Zampini } else { 2935906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2936906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2937906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2938906d46d4SStefano Zampini } 2939906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2940906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2941906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2942906d46d4SStefano Zampini } else { 2943906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2944906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2945906d46d4SStefano Zampini } 2946906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2950b9b85e73SStefano Zampini } 2951a717540cSStefano Zampini 2952727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2953727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2954727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 29550e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 29560e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29570e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 29580e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2959c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2960727cdba6SStefano Zampini } 29610e6343abSStefano Zampini } 29620e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2963727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2964727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2965727cdba6SStefano Zampini 2966a717540cSStefano Zampini /* flush dbg viewer */ 2967b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2968b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2969b8ffe317SStefano Zampini } 2970a717540cSStefano Zampini 2971e310c8b4SStefano Zampini /* free workspace */ 2972a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29734641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 297408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29759162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29769162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 297708122e43SStefano Zampini } else { 29789162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29799162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29809162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 298108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 298208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29839162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29849162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 298508122e43SStefano Zampini } 2986674ae819SStefano Zampini PetscFunctionReturn(0); 2987674ae819SStefano Zampini } 2988674ae819SStefano Zampini 2989674ae819SStefano Zampini #undef __FUNCT__ 2990674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2991674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2992674ae819SStefano Zampini { 2993674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2994674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2995674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29967fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2997674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2998674ae819SStefano Zampini 2999674ae819SStefano Zampini PetscFunctionBegin; 30008e61c736SStefano Zampini /* Reset previously computed graph */ 30018e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3002674ae819SStefano Zampini /* Init local Graph struct */ 30037fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30047fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3005674ae819SStefano Zampini 3006575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3007575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3008575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3009575ad6abSStefano Zampini } 30109577ea80SStefano Zampini 3011674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30124d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30134d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30144d379d7bSStefano Zampini PetscInt nvtxs; 3015e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3016674ae819SStefano Zampini 30174d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30182fffb893SStefano Zampini 30192fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30202fffb893SStefano Zampini if (flg_row) { 30214d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3022b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30232fffb893SStefano Zampini } 30242fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30259b28b941SStefano 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 */ 30264d379d7bSStefano Zampini IS is_dummy; 30274d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30284d379d7bSStefano Zampini PetscInt j,sum; 30294d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30304d379d7bSStefano Zampini const PetscInt *idxs; 30314d379d7bSStefano Zampini PCBDDCGraph graph; 30324d379d7bSStefano Zampini PetscBT is_on_boundary; 30334d379d7bSStefano Zampini 30344d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30354d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30364d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30374d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30387fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30394d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3040e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3041e496cd5dSStefano Zampini if (flg_row) { 30424d379d7bSStefano Zampini graph->xadj = xadj; 30434d379d7bSStefano Zampini graph->adjncy = adjncy; 3044e496cd5dSStefano Zampini } 30454d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30464d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3047e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30484d379d7bSStefano Zampini 30494d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30509b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30514d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30524d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30534d379d7bSStefano Zampini } 30544d379d7bSStefano Zampini } 30554d379d7bSStefano Zampini 3056e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30574d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30584d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30594d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30604d379d7bSStefano Zampini } 30614d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30624d379d7bSStefano Zampini 3063e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30644d379d7bSStefano Zampini sum = 0; 30654d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30664d379d7bSStefano Zampini PetscInt sizecc = 0; 30674d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30684d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30694d379d7bSStefano Zampini sizecc++; 30704d379d7bSStefano Zampini } 30714d379d7bSStefano Zampini } 30724d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30734d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30744d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30754d379d7bSStefano Zampini } 30764d379d7bSStefano Zampini } 30774d379d7bSStefano Zampini sum += sizecc*sizecc; 30784d379d7bSStefano Zampini } 30794d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30804d379d7bSStefano Zampini sum = 0; 3081e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30824d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30834d379d7bSStefano Zampini cxadj[i] = sum; 30844d379d7bSStefano Zampini sum += temp; 30854d379d7bSStefano Zampini } 3086e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30874d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30884d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30894d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30904d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30914d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30924d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30934d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30944d379d7bSStefano Zampini sizecc++; 30954d379d7bSStefano Zampini } 30964d379d7bSStefano Zampini } 30974d379d7bSStefano Zampini } 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini } 31009b28b941SStefano Zampini if (sum) { 3101e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31024d379d7bSStefano Zampini } else { 31034d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31044d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31054d379d7bSStefano Zampini } 31064d379d7bSStefano Zampini graph->xadj = 0; 31074d379d7bSStefano Zampini graph->adjncy = 0; 31084d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31094d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31104d379d7bSStefano Zampini } 3111674ae819SStefano Zampini } 31129b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31139b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31149b28b941SStefano Zampini } 3115674ae819SStefano Zampini 311663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3117674ae819SStefano Zampini vertex_size = 1; 311863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 311963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 312095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3122a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 312363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3124674ae819SStefano Zampini } 312563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 312663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 312763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3128674ae819SStefano Zampini } 312963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3130674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 313163602bcaSStefano Zampini } else { 313263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 313363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3134854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 313663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313763602bcaSStefano Zampini } 313863602bcaSStefano Zampini } 3139674ae819SStefano Zampini } 3140674ae819SStefano Zampini 3141674ae819SStefano Zampini /* Setup of Graph */ 3142785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3143a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3144785d1243SStefano Zampini } 3145785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3146a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3147785d1243SStefano Zampini } 3148302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3149674ae819SStefano Zampini 3150674ae819SStefano Zampini /* Graph's connected components analysis */ 3151674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3152674ae819SStefano Zampini 3153674ae819SStefano Zampini /* print some info to stdout */ 3154674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3155302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3156674ae819SStefano Zampini } 3157fb180af4SStefano Zampini 3158fb180af4SStefano Zampini /* mark topography has done */ 3159fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3160674ae819SStefano Zampini PetscFunctionReturn(0); 3161674ae819SStefano Zampini } 3162674ae819SStefano Zampini 3163674ae819SStefano Zampini #undef __FUNCT__ 3164674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3165674ae819SStefano 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[]) 3166674ae819SStefano Zampini { 3167674ae819SStefano Zampini Vec local_vec,global_vec; 3168674ae819SStefano Zampini IS seqis,paris; 3169674ae819SStefano Zampini VecScatter scatter_ctx; 3170674ae819SStefano Zampini PetscScalar *array; 3171674ae819SStefano Zampini PetscInt *temp_global_dofs; 3172674ae819SStefano Zampini PetscScalar globalsum; 3173674ae819SStefano Zampini PetscInt i,j,s; 317417ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 317517ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 317617ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3177674ae819SStefano Zampini PetscBool first_found; 3178674ae819SStefano Zampini PetscErrorCode ierr; 3179674ae819SStefano Zampini 3180674ae819SStefano Zampini PetscFunctionBegin; 3181674ae819SStefano Zampini /* mpi buffers */ 3182b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3183b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3184674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 318517ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3186674ae819SStefano Zampini /* get maximum size of subset */ 3187785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3188674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3189674ae819SStefano Zampini max_local = 0; 3190b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3191674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3192674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3193674ae819SStefano Zampini } 3194674ae819SStefano Zampini } 3195b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3196674ae819SStefano Zampini max_global++; 3197674ae819SStefano Zampini max_local = 0; 3198b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3199674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3200674ae819SStefano Zampini max_local = local_dofs[i]; 3201674ae819SStefano Zampini } 3202674ae819SStefano Zampini } 3203674ae819SStefano Zampini max_local++; 3204674ae819SStefano Zampini /* allocate workspace */ 3205674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3206674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3207674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3208674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3209674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3210674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3211674ae819SStefano Zampini /* create scatter */ 3212674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3213674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3214674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3215674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3216674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3217674ae819SStefano Zampini /* init array */ 3218674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3219674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3220674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3221674ae819SStefano Zampini if (local_dofs_mult) { 3222674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3223674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3224674ae819SStefano Zampini } 3225674ae819SStefano Zampini } else { 3226674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3227674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3228674ae819SStefano Zampini } 3229674ae819SStefano Zampini } 3230674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3231674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3232674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3233674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3234674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 32355b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3236674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3237674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3238674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3239674ae819SStefano Zampini nlocals = 0; 3240674ae819SStefano Zampini first_index = -1; 3241674ae819SStefano Zampini first_found = PETSC_FALSE; 3242674ae819SStefano Zampini for (i=0;i<s;i++) { 3243b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3244674ae819SStefano Zampini first_found = PETSC_TRUE; 3245674ae819SStefano Zampini first_index = i; 3246674ae819SStefano Zampini } 32475b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3248674ae819SStefano Zampini } 3249674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3250674ae819SStefano Zampini if (!rank_prec_comm) { 3251674ae819SStefano Zampini dof_displs[0]=0; 3252674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3253674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3254674ae819SStefano Zampini } 3255674ae819SStefano Zampini } 3256674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3257674ae819SStefano Zampini if (first_found) { 3258674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3259674ae819SStefano Zampini old_index = first_index; 3260674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3261b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3262674ae819SStefano Zampini array[i] += array[old_index]; 3263674ae819SStefano Zampini old_index = i; 3264674ae819SStefano Zampini } 3265674ae819SStefano Zampini } 3266674ae819SStefano Zampini } 3267674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3268674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3269674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3270674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3271674ae819SStefano Zampini /* get global ordering of local dofs */ 32722b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3273674ae819SStefano Zampini if (local_dofs_mult) { 3274674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32755b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3276674ae819SStefano Zampini } 3277674ae819SStefano Zampini } else { 3278674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32795b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3280674ae819SStefano Zampini } 3281674ae819SStefano Zampini } 32822b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3283674ae819SStefano Zampini /* free workspace */ 3284674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3285674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3286674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 328717ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3288674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3289674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3290674ae819SStefano Zampini PetscFunctionReturn(0); 3291674ae819SStefano Zampini } 32929a7d3425SStefano Zampini 32939a7d3425SStefano Zampini #undef __FUNCT__ 32949a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32959a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32969a7d3425SStefano Zampini { 32979a7d3425SStefano Zampini PetscInt i,j; 32989a7d3425SStefano Zampini PetscScalar *alphas; 32999a7d3425SStefano Zampini PetscErrorCode ierr; 33009a7d3425SStefano Zampini 33019a7d3425SStefano Zampini PetscFunctionBegin; 33029a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3303785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33049a7d3425SStefano Zampini for (i=0;i<n;i++) { 33059a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33069a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33079a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33089a7d3425SStefano Zampini } 33099a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33109a7d3425SStefano Zampini PetscFunctionReturn(0); 33119a7d3425SStefano Zampini } 33129a7d3425SStefano Zampini 3313e7931f94SStefano Zampini #undef __FUNCT__ 331470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 331528143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3316e7931f94SStefano Zampini { 3317e7931f94SStefano Zampini Mat subdomain_adj; 3318e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3319e7931f94SStefano Zampini MatPartitioning partitioner; 3320e7931f94SStefano Zampini Mat_IS *matis; 3321e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3322e7931f94SStefano Zampini PetscInt prank; 3323e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3324e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3325e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33263837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3327e7931f94SStefano Zampini PetscErrorCode ierr; 33282b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3329e7931f94SStefano Zampini PetscSubcomm subcomm; 3330a57a6d2fSStefano Zampini 3331e7931f94SStefano Zampini PetscFunctionBegin; 33322b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33332b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33342b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3335e7931f94SStefano Zampini 3336e7931f94SStefano Zampini /* Get info on mapping */ 3337e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3338e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3339e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3340e7931f94SStefano Zampini 3341e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3342785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3343e7931f94SStefano Zampini xadj[0] = 0; 3344e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3345785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3346785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3347e7931f94SStefano Zampini 33482b510759SStefano Zampini if (threshold) { 3349d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33502b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3351d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3352d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3353d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3354d023bfaeSStefano Zampini xadj_count++; 3355e7931f94SStefano Zampini } 3356e7931f94SStefano Zampini } 3357d023bfaeSStefano Zampini xadj[1] = xadj_count; 3358c8587f34SStefano Zampini } else { 3359e7931f94SStefano Zampini if (xadj[1]) { 3360e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3361e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3362c8587f34SStefano Zampini } 3363e7931f94SStefano Zampini } 3364e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3365e7931f94SStefano Zampini if (use_square) { 3366e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3367e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3368e7931f94SStefano Zampini } 3369e7931f94SStefano Zampini } 3370e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3371e7931f94SStefano Zampini 33723837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3373e7931f94SStefano Zampini 3374e7931f94SStefano Zampini /* 3375e7931f94SStefano Zampini Restrict work on active processes only. 3376e7931f94SStefano Zampini */ 3377e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3378e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3379e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33802b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3381d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3382e7931f94SStefano Zampini if (color) { 3383e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3384e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3385e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3386c8587f34SStefano Zampini } else { 338728143c3dSStefano Zampini PetscInt coarsening_ratio; 3388306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3389785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3390e7931f94SStefano Zampini prank = rank; 3391306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33928002ef2cSStefano Zampini /* 3393e7931f94SStefano Zampini for (i=0;i<size;i++) { 3394e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3395c8587f34SStefano Zampini } 33968002ef2cSStefano Zampini */ 3397e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3398e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3399c8587f34SStefano Zampini } 3400e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3401306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 340222b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3403e7931f94SStefano Zampini 3404e7931f94SStefano Zampini /* Partition */ 3405306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3406e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3407e7931f94SStefano Zampini if (use_vwgt) { 34083837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3409e7931f94SStefano Zampini v_wgt[0] = local_size; 3410e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3411c8587f34SStefano Zampini } 341228143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 341328143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 341428143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3415e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3416e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 341722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3418e7931f94SStefano Zampini 3419e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 342028143c3dSStefano Zampini if (contiguous) { 342128143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 342228143c3dSStefano Zampini } else { 342384ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 342428143c3dSStefano Zampini } 3425e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3426e7931f94SStefano Zampini /* clean up */ 3427e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3428e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3429e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3431e7931f94SStefano Zampini } 3432e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3433e7931f94SStefano Zampini 3434e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3435e7931f94SStefano Zampini i = 1; 3436e7931f94SStefano Zampini if (color) i=0; 3437e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3438e7931f94SStefano Zampini 3439e7931f94SStefano Zampini /* get back IS */ 3440e7931f94SStefano Zampini *is_sends = ranks_send_to; 3441e7931f94SStefano Zampini PetscFunctionReturn(0); 3442e7931f94SStefano Zampini } 3443e7931f94SStefano Zampini 3444e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3445e7931f94SStefano Zampini 3446e7931f94SStefano Zampini #undef __FUNCT__ 3447e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 344828143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3449e7931f94SStefano Zampini { 345070cf5478SStefano Zampini Mat local_mat; 3451e7931f94SStefano Zampini Mat_IS *matis; 3452e7931f94SStefano Zampini IS is_sends_internal; 34539d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 345428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34559d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3456e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3457e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3458e7931f94SStefano Zampini const PetscInt* is_indices; 3459e7931f94SStefano Zampini MatType new_local_type; 3460e7931f94SStefano Zampini /* buffers */ 3461e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 346228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34639d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3464e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3465e7931f94SStefano Zampini /* MPI */ 346628143c3dSStefano Zampini MPI_Comm comm,comm_n; 346728143c3dSStefano Zampini PetscSubcomm subcomm; 3468e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 346928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 347028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 347128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 347228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 347328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3474e7931f94SStefano Zampini PetscErrorCode ierr; 3475e7931f94SStefano Zampini 3476e7931f94SStefano Zampini PetscFunctionBegin; 347728143c3dSStefano Zampini /* TODO: add missing checks */ 347828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 347928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 348028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 348128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3482e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 348328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3484e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3485e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3486e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3487e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3488e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 348928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 349070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 349170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 349228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 349370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 349470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 349570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 349670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 349770cf5478SStefano Zampini } 3498e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3499e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3500e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3501e7931f94SStefano Zampini if (!is_sends) { 350228143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 350328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 350428143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3505c8587f34SStefano Zampini } else { 3506e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3507e7931f94SStefano Zampini is_sends_internal = is_sends; 3508c8587f34SStefano Zampini } 3509e7931f94SStefano Zampini 3510e7931f94SStefano Zampini /* get pointer of MATIS data */ 3511e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3512e7931f94SStefano Zampini 3513e7931f94SStefano Zampini /* get comm */ 3514a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3515e7931f94SStefano Zampini 3516e7931f94SStefano Zampini /* compute number of sends */ 3517e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3518e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3519e7931f94SStefano Zampini 3520e7931f94SStefano Zampini /* compute number of receives */ 3521e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3522785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3523e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3524e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3525e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3526e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3527e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3528e7931f94SStefano Zampini 352928143c3dSStefano Zampini /* restrict comm if requested */ 353028143c3dSStefano Zampini subcomm = 0; 353128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 353228143c3dSStefano Zampini if (restrict_comm) { 3533779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3534779c1cceSStefano Zampini 353528143c3dSStefano Zampini color = 0; 35366366ad86SStefano Zampini if (!n_recvs && n_sends) color = 1; /* processes sending and not receiving anything will not partecipate in new comm */ 353728143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 353828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 353928143c3dSStefano Zampini /* check if reuse has been requested */ 354028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 354128143c3dSStefano Zampini if (*mat_n) { 354228143c3dSStefano Zampini PetscMPIInt subcommsize2; 354328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 354428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 354528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 354628143c3dSStefano Zampini } else { 354728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 354828143c3dSStefano Zampini } 354928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3550779c1cceSStefano Zampini PetscMPIInt rank; 3551779c1cceSStefano Zampini 3552779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 355328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 355428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 355528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3556306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 355728143c3dSStefano Zampini } 355828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 355928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 356028143c3dSStefano Zampini } else { 356128143c3dSStefano Zampini comm_n = comm; 356228143c3dSStefano Zampini } 356328143c3dSStefano Zampini 3564e7931f94SStefano Zampini /* prepare send/receive buffers */ 3565785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3566e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3567785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3568e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 356928143c3dSStefano Zampini if (nis) { 3570854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 357128143c3dSStefano Zampini } 3572e7931f94SStefano Zampini 357328143c3dSStefano Zampini /* Get data from local matrices */ 3574e7931f94SStefano Zampini if (!isdense) { 3575a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3576e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3577e7931f94SStefano Zampini /* 3578e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3579e7931f94SStefano Zampini send_buffer_idxs should contain: 3580e7931f94SStefano Zampini - MatType_PRIVATE type 3581e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3582e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3583e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3584e7931f94SStefano Zampini */ 3585e7931f94SStefano Zampini } else { 3586e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3587e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3588854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3589e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3590e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3591e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3592e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3593e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3594e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3595e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3596e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3597e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3598c8587f34SStefano Zampini } 3599c8587f34SStefano Zampini } 3600e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 360128143c3dSStefano Zampini /* additional is (if any) */ 360228143c3dSStefano Zampini if (nis) { 360328143c3dSStefano Zampini PetscMPIInt psum; 360428143c3dSStefano Zampini PetscInt j; 360528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 360628143c3dSStefano Zampini PetscInt plen; 360728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 360828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 360928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 361028143c3dSStefano Zampini } 3611854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 361228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 361328143c3dSStefano Zampini PetscInt plen; 361428143c3dSStefano Zampini const PetscInt *is_array_idxs; 361528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 361728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 361828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 361928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 362128143c3dSStefano Zampini } 362228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 362328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 362428143c3dSStefano Zampini } 362528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 362628143c3dSStefano Zampini } 362728143c3dSStefano Zampini 3628e7931f94SStefano Zampini buf_size_idxs = 0; 3629e7931f94SStefano Zampini buf_size_vals = 0; 363028143c3dSStefano Zampini buf_size_idxs_is = 0; 3631e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3632e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3633e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 363428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3635e7931f94SStefano Zampini } 3636785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3637785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 363895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3639e7931f94SStefano Zampini 3640e7931f94SStefano Zampini /* get new tags for clean communications */ 3641e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3642e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 364328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3644e7931f94SStefano Zampini 3645e7931f94SStefano Zampini /* allocate for requests */ 3646785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3647785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 364895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3649785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3650785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 365195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3652e7931f94SStefano Zampini 3653e7931f94SStefano Zampini /* communications */ 3654e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3655e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 365628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3657e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3658e7931f94SStefano Zampini source_dest = onodes[i]; 3659e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3660e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3661e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3662e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 366328143c3dSStefano Zampini if (nis) { 366428143c3dSStefano 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); 366528143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 366628143c3dSStefano Zampini } 3667e7931f94SStefano Zampini } 3668e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3669e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3670e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3671e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 367228143c3dSStefano Zampini if (nis) { 367328143c3dSStefano 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); 367428143c3dSStefano Zampini } 3675e7931f94SStefano Zampini } 3676e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3677e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3678e7931f94SStefano Zampini 3679e7931f94SStefano Zampini /* assemble new l2g map */ 3680e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3681e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36829d30be91SStefano Zampini new_local_rows = 0; 3683e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36849d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3685e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3686e7931f94SStefano Zampini } 36879d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3688e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36899d30be91SStefano Zampini new_local_rows = 0; 3690e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36919d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36929d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3693e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3694e7931f94SStefano Zampini } 36959d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36969d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3697e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3698e7931f94SStefano Zampini 3699e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3700e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3701e7931f94SStefano 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) */ 3702e7931f94SStefano Zampini if (n_recvs) { 370328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3704e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3705e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3706e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3707e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3708e7931f94SStefano Zampini break; 3709e7931f94SStefano Zampini } 3710e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3711e7931f94SStefano Zampini } 3712e7931f94SStefano Zampini switch (new_local_type_private) { 371328143c3dSStefano Zampini case MATDENSE_PRIVATE: 371428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3715e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3716e7931f94SStefano Zampini bs = 1; 371728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 371828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 371928143c3dSStefano Zampini bs = 1; 372028143c3dSStefano Zampini } 3721e7931f94SStefano Zampini break; 3722e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3723e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3724e7931f94SStefano Zampini bs = 1; 3725e7931f94SStefano Zampini break; 3726e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3727e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3728e7931f94SStefano Zampini break; 3729e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3730e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3731e7931f94SStefano Zampini break; 3732e7931f94SStefano Zampini default: 37339d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3734e7931f94SStefano Zampini break; 3735e7931f94SStefano Zampini } 373628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 373728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 373828143c3dSStefano Zampini bs = 1; 3739e7931f94SStefano Zampini } 3740e7931f94SStefano Zampini 374170cf5478SStefano Zampini /* create MATIS object if needed */ 374270cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3743e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 374428143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 374570cf5478SStefano Zampini } else { 374670cf5478SStefano Zampini /* it also destroys the local matrices */ 374770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 374870cf5478SStefano Zampini } 374970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3750e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37519d30be91SStefano Zampini 37529d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37539d30be91SStefano Zampini 37549d30be91SStefano Zampini /* Global to local map of received indices */ 37559d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37569d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37579d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37589d30be91SStefano Zampini 37599d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37609d30be91SStefano Zampini buf_size_idxs = 0; 37619d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37629d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37639d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37649d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37659d30be91SStefano Zampini } 37669d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37679d30be91SStefano Zampini 37689d30be91SStefano Zampini /* set preallocation */ 37699d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37709d30be91SStefano Zampini if (!newisdense) { 37719d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37729d30be91SStefano Zampini 37739d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37749d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37759d30be91SStefano Zampini if (n_recvs) { 37769d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37779d30be91SStefano Zampini } 37789d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37799d30be91SStefano Zampini PetscInt j; 37809d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37819d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37829d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37839d30be91SStefano Zampini } 37849d30be91SStefano Zampini } else { 37859d30be91SStefano Zampini /* TODO */ 37869d30be91SStefano Zampini } 37879d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37889d30be91SStefano Zampini } 37899d30be91SStefano Zampini if (new_local_nnz) { 37909d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37919d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37929d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37939d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37949d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37959d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37969d30be91SStefano Zampini } else { 37979d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37989d30be91SStefano Zampini } 37999d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38009d30be91SStefano Zampini } else { 38019d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38029d30be91SStefano Zampini } 3803e7931f94SStefano Zampini 3804e7931f94SStefano Zampini /* set values */ 3805e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38069d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3807e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3808e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3809e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38109d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3811e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3812e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3813e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 381428143c3dSStefano Zampini } else { 381528143c3dSStefano Zampini /* TODO */ 3816e7931f94SStefano Zampini } 3817e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3818e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3819e7931f94SStefano Zampini } 3820e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3821e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38249d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38259d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3826e7931f94SStefano Zampini 3827dfd14d43SStefano Zampini #if 0 382828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3829e7931f94SStefano Zampini Vec lvec,rvec; 3830e7931f94SStefano Zampini PetscReal infty_error; 3831e7931f94SStefano Zampini 38322a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3833e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3834e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3835e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 383670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3837e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3838e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3839e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3840e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3841e7931f94SStefano Zampini } 384228143c3dSStefano Zampini #endif 3843e7931f94SStefano Zampini 384428143c3dSStefano Zampini /* assemble new additional is (if any) */ 384528143c3dSStefano Zampini if (nis) { 384628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 384728143c3dSStefano Zampini 384828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3849854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 385028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 385128143c3dSStefano Zampini psum = 0; 385228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 385328143c3dSStefano Zampini for (j=0;j<nis;j++) { 385428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 385528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 385628143c3dSStefano Zampini psum += plen; 385728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 385828143c3dSStefano Zampini } 385928143c3dSStefano Zampini } 3860854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3861854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 386228143c3dSStefano Zampini for (i=1;i<nis;i++) { 386328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 386428143c3dSStefano Zampini } 386528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 386628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 386728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 386828143c3dSStefano Zampini for (j=0;j<nis;j++) { 386928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 387128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 387228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387328143c3dSStefano Zampini } 387428143c3dSStefano Zampini } 387528143c3dSStefano Zampini for (i=0;i<nis;i++) { 387628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 387728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 387828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 387928143c3dSStefano Zampini } 388028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 388128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 388228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 388328143c3dSStefano Zampini } 3884e7931f94SStefano Zampini /* free workspace */ 388528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3886e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3887e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3888e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3889e7931f94SStefano Zampini if (isdense) { 3890e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3891e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3892e7931f94SStefano Zampini } else { 3893e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3894e7931f94SStefano Zampini } 389528143c3dSStefano Zampini if (nis) { 389628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 389728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 389828143c3dSStefano Zampini } 3899e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3900e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 390128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3902e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3903e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 390428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3906e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3907e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 391028143c3dSStefano Zampini if (nis) { 391128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 391228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 391328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 391428143c3dSStefano Zampini } 391528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 391628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 391728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 391828143c3dSStefano Zampini for (i=0;i<nis;i++) { 391928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 392028143c3dSStefano Zampini } 392128143c3dSStefano Zampini } 3922e7931f94SStefano Zampini PetscFunctionReturn(0); 3923e7931f94SStefano Zampini } 3924a57a6d2fSStefano Zampini 392512edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3926af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 392712edc857SStefano Zampini 3928c8587f34SStefano Zampini #undef __FUNCT__ 3929c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3930c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3931c8587f34SStefano Zampini { 3932c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3933c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 393420a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39359881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 393620a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39376e683305SStefano Zampini IS coarse_is,*isarray; 39386e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39396e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3940f9eb5b7dSStefano Zampini PC pc_temp; 3941c8587f34SStefano Zampini PCType coarse_pc_type; 3942c8587f34SStefano Zampini KSPType coarse_ksp_type; 3943f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39444f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39456e683305SStefano Zampini Mat t_coarse_mat_is; 39466e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39476e683305SStefano Zampini PetscMPIInt all_procs; 394874e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 394968457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 395022bc73bbSStefano Zampini PetscScalar *array; 39519881197aSStefano Zampini PetscErrorCode ierr; 3952fdc09c96SStefano Zampini 3953c8587f34SStefano Zampini PetscFunctionBegin; 3954c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 395568457ee5SStefano 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 */ 3956fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39575a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3958fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3959f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3960f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3961f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3962fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3963fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3964727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3965fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3966fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3967fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3968f4ddd8eeSStefano Zampini } 3969fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3970fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3971f4ddd8eeSStefano Zampini } 397270cf5478SStefano Zampini /* reset any subassembling information */ 397370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39746e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39756e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3976fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3977f4ddd8eeSStefano Zampini } 3978c8587f34SStefano Zampini 39796e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39802b510759SStefano Zampini im_active = !!(pcis->n); 39812b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39826e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39836e683305SStefano Zampini void_procs = all_procs-active_procs; 39846e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 398574e2c79eSStefano Zampini redist = PETSC_FALSE; 398622bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39876e683305SStefano Zampini csin_ml = PETSC_TRUE; 39886e683305SStefano Zampini ncoarse_ml = void_procs; 3989779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3990779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 399118a45a71SStefano Zampini csin_ds = PETSC_TRUE; 399218a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 399318a45a71SStefano Zampini redist = PETSC_TRUE; 399418a45a71SStefano Zampini } else { 39956e683305SStefano Zampini csin_ds = PETSC_TRUE; 3996779c1cceSStefano Zampini ncoarse_ds = active_procs; 3997779c1cceSStefano Zampini redist = PETSC_TRUE; 399818a45a71SStefano Zampini } 39996e683305SStefano Zampini } else { 40006e683305SStefano Zampini csin_ml = PETSC_FALSE; 40016e683305SStefano Zampini ncoarse_ml = all_procs; 40026e683305SStefano Zampini if (void_procs) { 40036e683305SStefano Zampini csin_ds = PETSC_TRUE; 40046e683305SStefano Zampini ncoarse_ds = void_procs; 40056e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40066e683305SStefano Zampini } else { 4007779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 400874e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 400974e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401074e2c79eSStefano Zampini redist = PETSC_TRUE; 401174e2c79eSStefano Zampini } else { 40126e683305SStefano Zampini csin_ds = PETSC_FALSE; 40136e683305SStefano Zampini ncoarse_ds = all_procs; 40146e683305SStefano Zampini } 40156e683305SStefano Zampini } 401674e2c79eSStefano Zampini } 40176e683305SStefano Zampini 40186e683305SStefano Zampini /* 40196e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40206e683305SStefano Zampini - we have not exceeded the number of levels requested 40216e683305SStefano Zampini - we can actually subassemble the active processes 40226e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40236e683305SStefano Zampini */ 40246e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40256e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40266e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40276e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40286e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4029f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40302b510759SStefano Zampini } else { 4031f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4032c8587f34SStefano Zampini } 4033c8587f34SStefano Zampini } 40346e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40356e683305SStefano Zampini if (multilevel_allowed) { 40366e683305SStefano Zampini ncoarse = ncoarse_ml; 40376e683305SStefano Zampini csin = csin_ml; 403858da7f69SStefano Zampini redist = PETSC_FALSE; 40396e683305SStefano Zampini } else { 40406e683305SStefano Zampini ncoarse = ncoarse_ds; 40416e683305SStefano Zampini csin = csin_ds; 40426e683305SStefano Zampini } 4043e7931f94SStefano Zampini 4044abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4045abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4046abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4047abbbba34SStefano Zampini 4048abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 404922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 405022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 405122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 405222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4053b9b85e73SStefano Zampini #if 0 4054b9b85e73SStefano Zampini { 4055b9b85e73SStefano Zampini PetscViewer viewer; 4056b9b85e73SStefano Zampini char filename[256]; 4057b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4058b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4059b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4060b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4061b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4062b9b85e73SStefano Zampini } 4063b9b85e73SStefano Zampini #endif 40646e683305SStefano 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); 40656e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40666e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40676e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4068abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4069abbbba34SStefano Zampini 40706e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40716e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40726e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40736e683305SStefano Zampini const PetscInt *idxs; 40746e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40756e683305SStefano Zampini 40766e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40770be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40786e683305SStefano Zampini /* allocate space for temporary storage */ 4079854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4080854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40816e683305SStefano Zampini /* allocate for IS array */ 40826e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40836e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40846e683305SStefano Zampini nis = nisdofs + nisneu; 4085854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40866e683305SStefano Zampini /* dofs splitting */ 40876e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40886e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40896e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40906e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40916e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40926e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40936e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40946e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40956e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40966e683305SStefano Zampini } 40976e683305SStefano Zampini /* neumann boundaries */ 40986e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 40996e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41006e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41016e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41026e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41036e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41046e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41056e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41066e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41076e683305SStefano Zampini } 41086e683305SStefano Zampini /* free memory */ 41096e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41106e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41116e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41126e683305SStefano Zampini } else { 41136e683305SStefano Zampini nis = 0; 41146e683305SStefano Zampini nisdofs = 0; 41156e683305SStefano Zampini nisneu = 0; 41166e683305SStefano Zampini isarray = NULL; 41176e683305SStefano Zampini } 41186e683305SStefano Zampini /* destroy no longer needed map */ 41196e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41206e683305SStefano Zampini 41216e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41226e683305SStefano Zampini coarse_mat_is = NULL; 41236e683305SStefano Zampini if (csin) { 41246e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 412574e2c79eSStefano Zampini if (redist) { 412674e2c79eSStefano Zampini PetscMPIInt rank; 4127779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 412874e2c79eSStefano Zampini 412974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 413058da7f69SStefano Zampini spc = active_procs/ncoarse; 413158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4132779c1cceSStefano Zampini if (im_active) { 4133779c1cceSStefano Zampini destsize = 1; 413474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 413574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 413674e2c79eSStefano Zampini } else { 413774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 413874e2c79eSStefano Zampini } 413974e2c79eSStefano Zampini } else { 4140779c1cceSStefano Zampini destsize = 0; 41416e683305SStefano Zampini } 4142779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4143779c1cceSStefano Zampini } else if (csin_type_simple) { 41446e683305SStefano Zampini PetscMPIInt rank; 41456e683305SStefano Zampini PetscInt issize,isidx; 4146779c1cceSStefano Zampini 41476e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41486e683305SStefano Zampini if (im_active) { 41496e683305SStefano Zampini issize = 1; 41506e683305SStefano Zampini isidx = (PetscInt)rank; 41516e683305SStefano Zampini } else { 41526e683305SStefano Zampini issize = 0; 41536e683305SStefano Zampini isidx = -1; 41546e683305SStefano Zampini } 41556e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4156779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41576e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41586e683305SStefano Zampini } 4159779c1cceSStefano Zampini 4160779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4161779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4162779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4163779c1cceSStefano Zampini PetscInt *coarse_candidates; 4164779c1cceSStefano Zampini const PetscInt* tisindices; 4165779c1cceSStefano Zampini 4166779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4167779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4168779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4169779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4170779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4171779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4172779c1cceSStefano Zampini } 4173779c1cceSStefano Zampini } 4174779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4175779c1cceSStefano Zampini 4176779c1cceSStefano Zampini 41776e683305SStefano Zampini if (pcbddc->dbg_flag) { 41786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41806e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4182779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41846e683305SStefano Zampini } 41856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41866e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41876e683305SStefano Zampini } 41886e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41896e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41906e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4191854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41926e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41936e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41946e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41956e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41966e683305SStefano Zampini } 41976e683305SStefano Zampini if (pcbddc->dbg_flag) { 41986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42006e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42016e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42026e683305SStefano Zampini } 4203779c1cceSStefano Zampini } 42046e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 42056e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 42066e683305SStefano Zampini } else { 42076e683305SStefano Zampini if (pcbddc->dbg_flag) { 42086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42106e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42116e683305SStefano Zampini } 42126e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42136e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42146e683305SStefano Zampini } 42156e683305SStefano Zampini 42166e683305SStefano Zampini /* create local to global scatters for coarse problem */ 421768457ee5SStefano Zampini if (compute_vecs) { 42186e683305SStefano Zampini PetscInt lrows; 42196e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42206e683305SStefano Zampini if (coarse_mat_is) { 42216e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42226e683305SStefano Zampini } else { 42236e683305SStefano Zampini lrows = 0; 42246e683305SStefano Zampini } 42256e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42266e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42276e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42286e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42296e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42306e683305SStefano Zampini } 42316e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42326e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4233c8587f34SStefano Zampini 4234f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4235f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4236f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4237f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4238f9eb5b7dSStefano Zampini } else { 4239f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4240f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4241c8587f34SStefano Zampini } 4242c8587f34SStefano Zampini 42436e683305SStefano Zampini /* print some info if requested */ 42446e683305SStefano Zampini if (pcbddc->dbg_flag) { 42456e683305SStefano Zampini if (!multilevel_allowed) { 42466e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42476e683305SStefano Zampini if (multilevel_requested) { 42486e683305SStefano 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); 42496e683305SStefano Zampini } else if (pcbddc->max_levels) { 42506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42516e683305SStefano Zampini } 42526e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42536e683305SStefano Zampini } 42546e683305SStefano Zampini } 42556e683305SStefano Zampini 4256f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42576e683305SStefano Zampini if (coarse_mat_is) { 42586e683305SStefano Zampini MatReuse coarse_mat_reuse; 42596a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42606e683305SStefano Zampini if (pcbddc->dbg_flag) { 42616e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42626e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42636e683305SStefano Zampini } 4264f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4265312be037SStefano Zampini char prefix[256],str_level[16]; 4266e604994aSStefano Zampini size_t len; 42676e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4268c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4269f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42705f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4271c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42726e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4273c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4274c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42753e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4276e604994aSStefano Zampini /* prefix */ 4277e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4278e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4279e604994aSStefano Zampini if (!pcbddc->current_level) { 4280e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4281e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4282c8587f34SStefano Zampini } else { 4283e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4284312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4285312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 428634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4287312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4288e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4289e604994aSStefano Zampini } 4290e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 42913e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42923e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 42933e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 42943e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4295f9eb5b7dSStefano Zampini /* allow user customization */ 4296f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 42973e3c6dadSStefano Zampini } 42983e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42993e3c6dadSStefano Zampini if (nisdofs) { 43003e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43013e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43023e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43033e3c6dadSStefano Zampini } 43043e3c6dadSStefano Zampini } 43053e3c6dadSStefano Zampini if (nisneu) { 43063e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43073e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4308312be037SStefano Zampini } 4309f9eb5b7dSStefano Zampini 4310f9eb5b7dSStefano Zampini /* get some info after set from options */ 4311f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4312f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4313f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43144f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43156e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4316f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4317f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4318f9eb5b7dSStefano Zampini } 43194f3a063dSStefano Zampini if (isredundant) { 43204f3a063dSStefano Zampini KSP inner_ksp; 43214f3a063dSStefano Zampini PC inner_pc; 43224f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43234f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43244f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43254f3a063dSStefano Zampini } 4326f9eb5b7dSStefano Zampini 4327f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4328fa7f1dd8SStefano Zampini if (coarse_reuse) { 432981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4330fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43316e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4332fa7f1dd8SStefano Zampini } else { 43336e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4334fa7f1dd8SStefano Zampini } 4335c8587f34SStefano Zampini if (isbddc || isnn) { 433622bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 433770cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 433828143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 433922b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43426e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43436e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 434422b6e8a2SStefano Zampini } 434570cf5478SStefano Zampini } 43466e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 434770cf5478SStefano Zampini } else { 434822bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 434922bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 435022bc73bbSStefano Zampini } 435122bc73bbSStefano Zampini } else { 43522e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4353c8587f34SStefano Zampini } 4354c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4355c8587f34SStefano Zampini 4356c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4357b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 43585a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4359c8587f34SStefano Zampini 43606e683305SStefano Zampini /* set operators */ 43615f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43626e683305SStefano Zampini if (pcbddc->dbg_flag) { 43636e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43646e683305SStefano Zampini } 43656e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43666e683305SStefano Zampini coarse_mat = 0; 43676e683305SStefano Zampini } 43686e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4369b9b85e73SStefano Zampini #if 0 4370b9b85e73SStefano Zampini { 4371b9b85e73SStefano Zampini PetscViewer viewer; 4372b9b85e73SStefano Zampini char filename[256]; 4373b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4374b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4375b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4376b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4377b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4378b9b85e73SStefano Zampini } 4379b9b85e73SStefano Zampini #endif 4380c8587f34SStefano Zampini 4381c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4382c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4383c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 438498a51de6SStefano Zampini } 438598a51de6SStefano Zampini 438698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 438798a51de6SStefano Zampini Vec crhs,csol; 438898a51de6SStefano Zampini PetscBool ispreonly; 438998a51de6SStefano Zampini if (CoarseNullSpace) { 4390c8587f34SStefano Zampini if (isbddc) { 4391c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4392c8587f34SStefano Zampini } else { 4393c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4394c8587f34SStefano Zampini } 4395c8587f34SStefano Zampini } 4396f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4397f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4398f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4399f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44006e683305SStefano Zampini /* hack */ 4401f347579bSStefano Zampini if (!csol) { 44022a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4403f9eb5b7dSStefano Zampini } 4404f347579bSStefano Zampini if (!crhs) { 44052a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4406f347579bSStefano Zampini } 4407cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4408cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44096e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4410c8587f34SStefano Zampini KSP check_ksp; 44112b510759SStefano Zampini KSPType check_ksp_type; 4412c8587f34SStefano Zampini PC check_pc; 44136e683305SStefano Zampini Vec check_vec,coarse_vec; 44146a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44152b510759SStefano Zampini PetscInt its; 44166e683305SStefano Zampini PetscBool compute_eigs; 44176e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44186e683305SStefano Zampini PetscInt neigs; 44198e185a42SStefano Zampini const char *prefix; 4420c8587f34SStefano Zampini 44212b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44226e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 442323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4424f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44252b510759SStefano Zampini if (ispreonly) { 44262b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44276e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44282b510759SStefano Zampini } else { 4429cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44306e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4431c8587f34SStefano Zampini } 4432c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44336e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44346e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44356e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4436a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4437a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4438a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4439a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4440c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4441c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4442c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4443c8587f34SStefano Zampini /* create random vec */ 44446e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44456e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4446c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4447c8587f34SStefano Zampini if (CoarseNullSpace) { 4448c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4449c8587f34SStefano Zampini } 44506e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4451c8587f34SStefano Zampini /* solve coarse problem */ 44526e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4453c8587f34SStefano Zampini if (CoarseNullSpace) { 44546e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4455c8587f34SStefano Zampini } 4456cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44576e683305SStefano Zampini if (compute_eigs) { 4458854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4459854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44606e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44616e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44626e683305SStefano Zampini lambda_min = eigs_r[0]; 44636e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44646e683305SStefano Zampini if (lambda_max>lambda_min) { 4465cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4466cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4467cbcc2c2aSStefano Zampini } 4468c8587f34SStefano Zampini } 4469c8587f34SStefano Zampini } 4470cbcc2c2aSStefano Zampini 4471c8587f34SStefano Zampini /* check coarse problem residual error */ 44726e683305SStefano Zampini if (pcbddc->dbg_flag) { 44736e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44746e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44756e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4476c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44776e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44786e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4479c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4480779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44816e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44826e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44856e683305SStefano Zampini if (compute_eigs) { 44866e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4487deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4488c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44896e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 44906e683305SStefano 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); 44916e683305SStefano Zampini for (i=0;i<neigs;i++) { 44926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4493c8587f34SStefano Zampini } 44946e683305SStefano Zampini } 44956e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 44966e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44976e683305SStefano Zampini } 4498c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 44996e683305SStefano Zampini if (compute_eigs) { 45006e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45016e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4502c8587f34SStefano Zampini } 45036e683305SStefano Zampini } 45046e683305SStefano Zampini } 4505cbcc2c2aSStefano Zampini /* print additional info */ 4506cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45076e683305SStefano Zampini /* waits until all processes reaches this point */ 45086e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4509cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4510cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4511cbcc2c2aSStefano Zampini } 4512cbcc2c2aSStefano Zampini 45132b510759SStefano Zampini /* free memory */ 4514c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4515fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4516c8587f34SStefano Zampini PetscFunctionReturn(0); 4517c8587f34SStefano Zampini } 4518674ae819SStefano Zampini 4519f34684f1SStefano Zampini #undef __FUNCT__ 4520f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4521f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4522f34684f1SStefano Zampini { 4523f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4524f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4525f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 45260e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 452773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 45280e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4529f34684f1SStefano Zampini PetscErrorCode ierr; 4530f34684f1SStefano Zampini 4531f34684f1SStefano Zampini PetscFunctionBegin; 4532f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45330e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45340e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4535727cdba6SStefano Zampini } 45360e6343abSStefano 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); 45370e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 45380e6343abSStefano Zampini local_size = 0; 45390e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 45400e6343abSStefano Zampini PetscInt j; 45410e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 45420e6343abSStefano Zampini } 45430e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 45440e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45450e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45460e6343abSStefano Zampini } 4547f34684f1SStefano Zampini 4548f34684f1SStefano Zampini /* check numbering */ 4549f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4550f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4551b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4552f34684f1SStefano Zampini 4553f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4554f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4555f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45560fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4557f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4558f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4559727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4560f34684f1SStefano Zampini } 4561f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4562f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4563f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4564f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4565f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4566f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4567f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4568f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4569f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4570f34684f1SStefano Zampini if (array[i] == 1.0) { 4571b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4572f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4573f34684f1SStefano Zampini } 4574f34684f1SStefano Zampini } 4575b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4576f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4577f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4578f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4579f34684f1SStefano Zampini } 4580f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4581f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4582f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4583f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4584f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4585f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4586b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4587ca8b9ea9SStefano Zampini PetscInt *gidxs; 4588ca8b9ea9SStefano Zampini 4589ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4590ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4591f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4592f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4593f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4594f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 45954bc2dc4bSStefano 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); 4596f34684f1SStefano Zampini } 4597f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4598ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4599f34684f1SStefano Zampini } 4600f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4601302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4602f34684f1SStefano Zampini } 46038bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4604f34684f1SStefano Zampini /* get back data */ 4605f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4606f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4607674ae819SStefano Zampini PetscFunctionReturn(0); 4608674ae819SStefano Zampini } 4609674ae819SStefano Zampini 4610e456f2a8SStefano Zampini #undef __FUNCT__ 4611e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4612a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4613e456f2a8SStefano Zampini { 4614e456f2a8SStefano Zampini IS localis_t; 4615a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4616e456f2a8SStefano Zampini PetscScalar *vals; 4617e456f2a8SStefano Zampini PetscErrorCode ierr; 4618e456f2a8SStefano Zampini 4619e456f2a8SStefano Zampini PetscFunctionBegin; 4620a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4621e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4622854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4623e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4624e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4625a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4626a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46271035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4628a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46291035eff8SStefano Zampini } 4630a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4631e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4632e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4633a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4634a7dc3881SStefano Zampini /* now compute set in local ordering */ 4635a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4636a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4637a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4638a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4639a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4640ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4641e456f2a8SStefano Zampini lsize++; 4642e456f2a8SStefano Zampini } 4643e456f2a8SStefano Zampini } 4644854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4645a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4646ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4647e456f2a8SStefano Zampini idxs[lsize++] = i; 4648e456f2a8SStefano Zampini } 4649e456f2a8SStefano Zampini } 4650a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4651a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4652e456f2a8SStefano Zampini *localis = localis_t; 4653e456f2a8SStefano Zampini PetscFunctionReturn(0); 4654e456f2a8SStefano Zampini } 4655906d46d4SStefano Zampini 4656906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4657906d46d4SStefano Zampini #undef __FUNCT__ 4658906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4659906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4660906d46d4SStefano Zampini { 4661906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4662906d46d4SStefano Zampini PetscErrorCode ierr; 4663906d46d4SStefano Zampini 4664906d46d4SStefano Zampini PetscFunctionBegin; 4665906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4666906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4667906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4668906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4669906d46d4SStefano Zampini PetscFunctionReturn(0); 4670906d46d4SStefano Zampini } 4671906d46d4SStefano Zampini 4672906d46d4SStefano Zampini #undef __FUNCT__ 4673906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4674906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4675906d46d4SStefano Zampini { 4676906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4677906d46d4SStefano Zampini PetscErrorCode ierr; 4678906d46d4SStefano Zampini 4679906d46d4SStefano Zampini PetscFunctionBegin; 4680906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4681906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4682906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4683906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4684906d46d4SStefano Zampini PetscFunctionReturn(0); 4685906d46d4SStefano Zampini } 4686b96c3477SStefano Zampini 4687b96c3477SStefano Zampini #undef __FUNCT__ 4688b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 468908122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4690b96c3477SStefano Zampini { 4691a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4692b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4693b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4694a64f4aa4SStefano Zampini Mat S_j; 4695b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4696b96c3477SStefano Zampini PetscBool free_used_adj; 4697b96c3477SStefano Zampini PetscErrorCode ierr; 4698b96c3477SStefano Zampini 4699b96c3477SStefano Zampini PetscFunctionBegin; 4700b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4701b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 470208122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4703b96c3477SStefano Zampini used_xadj = NULL; 4704b96c3477SStefano Zampini used_adjncy = NULL; 4705b96c3477SStefano Zampini } else { 470608122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 470708122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 470808122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 470908122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4710b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4711b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4712b96c3477SStefano Zampini } else { 47132fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4714b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4715b96c3477SStefano Zampini PetscInt nvtxs; 4716b96c3477SStefano Zampini 47172fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47182fffb893SStefano Zampini if (flg_row) { 4719b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4720b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4721b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4722b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47232fffb893SStefano Zampini } else { 47242fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47252fffb893SStefano Zampini used_xadj = NULL; 47262fffb893SStefano Zampini used_adjncy = NULL; 47272fffb893SStefano Zampini } 47282fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4729b96c3477SStefano Zampini } 4730b96c3477SStefano Zampini } 4731d5574798SStefano Zampini 4732d5574798SStefano Zampini /* setup sub_schurs data */ 4733a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4734a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4735a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4736a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 47376816873aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4738a64f4aa4SStefano Zampini } else { 47396816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 47406816873aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4741a64f4aa4SStefano Zampini } 4742a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4743b96c3477SStefano Zampini 4744b96c3477SStefano Zampini /* free adjacency */ 4745b96c3477SStefano Zampini if (free_used_adj) { 4746b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4747b96c3477SStefano Zampini } 4748b96c3477SStefano Zampini PetscFunctionReturn(0); 4749b96c3477SStefano Zampini } 4750b96c3477SStefano Zampini 4751b96c3477SStefano Zampini #undef __FUNCT__ 4752b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 475308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4754b96c3477SStefano Zampini { 4755b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4756b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4757b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4758b96c3477SStefano Zampini PCBDDCGraph graph; 4759b96c3477SStefano Zampini PetscErrorCode ierr; 4760b96c3477SStefano Zampini 4761b96c3477SStefano Zampini PetscFunctionBegin; 4762b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 476308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4764b96c3477SStefano Zampini IS verticesIS; 4765b96c3477SStefano Zampini 4766b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4767b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47687fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4769b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4770b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4771b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4772b96c3477SStefano Zampini /* 4773b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4774b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4775b96c3477SStefano Zampini } 4776b96c3477SStefano Zampini */ 4777b96c3477SStefano Zampini } else { 4778b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4779b96c3477SStefano Zampini } 4780b96c3477SStefano Zampini 4781b96c3477SStefano Zampini /* sub_schurs init */ 4782a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4783a64f4aa4SStefano Zampini 4784b96c3477SStefano Zampini /* free graph struct */ 478508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4786b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4787b96c3477SStefano Zampini } 4788b96c3477SStefano Zampini PetscFunctionReturn(0); 4789b96c3477SStefano Zampini } 4790