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 8674ae819SStefano 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); 361575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini PetscInt subset_size; 51862806e4SStefano Zampini 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini 5608122e43SStefano Zampini /* min/max and threshold */ 5708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 58f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 5908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 60f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 61f6f667cfSStefano Zampini if (nmin) { 62f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 63f6f667cfSStefano Zampini } 6408122e43SStefano Zampini 6508122e43SStefano Zampini /* allocate lapack workspace */ 6608122e43SStefano Zampini cum = cum2 = 0; 6708122e43SStefano Zampini maxneigs = 0; 6808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6908122e43SStefano Zampini PetscInt n,subset_size; 70f6f667cfSStefano Zampini 7108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 739162d606SStefano Zampini cum += subset_size; 749162d606SStefano Zampini cum2 += subset_size*n; 7508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7608122e43SStefano Zampini } 7708122e43SStefano Zampini if (mss) { 789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7908122e43SStefano Zampini PetscBLASInt B_itype = 1; 8008122e43SStefano Zampini PetscBLASInt B_N = mss; 814c6709b3SStefano Zampini PetscReal zero = 0.0; 824c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8308122e43SStefano Zampini 8408122e43SStefano Zampini B_lwork = -1; 8508122e43SStefano Zampini S = NULL; 8608122e43SStefano Zampini St = NULL; 87a58a30b4SStefano Zampini eigs = NULL; 88a58a30b4SStefano Zampini eigv = NULL; 89a58a30b4SStefano Zampini B_iwork = NULL; 90a58a30b4SStefano Zampini B_ifail = NULL; 91d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 92d1710679SStefano Zampini rwork = NULL; 93d1710679SStefano Zampini #endif 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 15508122e43SStefano Zampini const PetscInt *idxs; 156f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 157862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15808122e43SStefano Zampini PetscBLASInt B_N; 159aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 16008122e43SStefano Zampini 161862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 162f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 163f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1649ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 165aff50787SStefano Zampini PetscInt j,k; 166aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 167aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 168aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 16908122e43SStefano Zampini } 17008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 171aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 172aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 173aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 174aff50787SStefano Zampini } 17508122e43SStefano Zampini } 17608122e43SStefano Zampini } else { 17708122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17808122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17908122e43SStefano Zampini } 1808bec7fa6SStefano Zampini } else { 181f6f667cfSStefano Zampini S = Sarray + cumarray; 182f6f667cfSStefano Zampini St = Starray + cumarray; 1838bec7fa6SStefano Zampini } 18408122e43SStefano Zampini 185f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 186aff50787SStefano Zampini /* see if we can save some work */ 187aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 188aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 189aff50787SStefano Zampini } 190aff50787SStefano Zampini 191aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 192aff50787SStefano Zampini B_neigs = 0; 193aff50787SStefano Zampini } else { 194aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 195f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 196f6f667cfSStefano Zampini 1979ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 19808122e43SStefano Zampini PetscBLASInt B_itype = 1; 199f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2004c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2019552c7c7SStefano Zampini PetscInt nmin_s; 20208122e43SStefano Zampini 203f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 204fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2058bec7fa6SStefano 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]]); 206fd14bc51SStefano Zampini } 20708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 20808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 209f6f667cfSStefano 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)); 21008122e43SStefano Zampini #else 211f6f667cfSStefano 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)); 21208122e43SStefano Zampini #endif 21308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 21408122e43SStefano Zampini if (B_ierr) { 21508122e43SStefano Zampini if (B_ierr < 0 ) { 21608122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 21708122e43SStefano Zampini } else if (B_ierr <= B_N) { 21808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 21908122e43SStefano Zampini } else { 2209552c7c7SStefano 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); 22108122e43SStefano Zampini } 22208122e43SStefano Zampini } 22308122e43SStefano Zampini 22408122e43SStefano Zampini if (B_neigs > nmax) { 225fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 226fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 227fd14bc51SStefano Zampini } 228f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 22908122e43SStefano Zampini B_neigs = nmax; 23008122e43SStefano Zampini } 23108122e43SStefano Zampini 2329552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2339552c7c7SStefano Zampini if (B_neigs < nmin_s) { 23408122e43SStefano Zampini PetscBLASInt B_neigs2; 23508122e43SStefano Zampini 236f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 237f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 238fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 239fd14bc51SStefano 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); 240fd14bc51SStefano Zampini } 2419ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 24208122e43SStefano Zampini PetscInt j; 24308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24408122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24508122e43SStefano Zampini } 24608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24708122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24808122e43SStefano Zampini } 24908122e43SStefano Zampini } else { 25008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25208122e43SStefano Zampini } 25308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 255f6f667cfSStefano 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)); 25608122e43SStefano Zampini #else 257f6f667cfSStefano 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)); 25808122e43SStefano Zampini #endif 25908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 26008122e43SStefano Zampini B_neigs += B_neigs2; 26108122e43SStefano Zampini } 26208122e43SStefano Zampini if (B_ierr) { 26308122e43SStefano Zampini if (B_ierr < 0 ) { 26408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 26508122e43SStefano Zampini } else if (B_ierr <= B_N) { 26608122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 26708122e43SStefano Zampini } else { 2689552c7c7SStefano 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); 26908122e43SStefano Zampini } 27008122e43SStefano Zampini } 271fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 272ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 27308122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 27408122e43SStefano Zampini if (eigs[j] == 0.0) { 275ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 27608122e43SStefano Zampini } else { 277ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 278fd14bc51SStefano Zampini } 27908122e43SStefano Zampini } 28008122e43SStefano Zampini } 28108122e43SStefano Zampini } else { 28208122e43SStefano Zampini /* TODO */ 28308122e43SStefano Zampini } 284aff50787SStefano Zampini } 2858bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2868bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2879162d606SStefano Zampini if (B_neigs) { 2889162d606SStefano 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); 289fd14bc51SStefano Zampini 290fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2919552c7c7SStefano Zampini PetscInt ii; 2929552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 293ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 2949552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 295ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 296ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 297ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 298ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 299ac47001eSStefano Zampini #else 300ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 301ac47001eSStefano Zampini #endif 3029552c7c7SStefano Zampini } 3039552c7c7SStefano Zampini } 304fd14bc51SStefano Zampini } 30508122e43SStefano Zampini #if 0 3069162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 30708122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 30808122e43SStefano Zampini PetscScalar norm; 30908122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3109162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3119162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 31208122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 31308122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 31408122e43SStefano Zampini } else { 31508122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 31608122e43SStefano Zampini } 3179162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 318b1b3d7a2SStefano Zampini } 319b1b3d7a2SStefano Zampini #endif 3209162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3219162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3229162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3239162d606SStefano Zampini cum++; 32408122e43SStefano Zampini } 32508122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 32608122e43SStefano Zampini /* shift for next computation */ 32708122e43SStefano Zampini cumarray += subset_size*subset_size; 32808122e43SStefano Zampini } 329fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 330fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 331fd14bc51SStefano Zampini } 33208122e43SStefano Zampini 33308122e43SStefano Zampini if (mss) { 33408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 33508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 336f6f667cfSStefano Zampini /* destroy matrices (junk) */ 337f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 338f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 33908122e43SStefano Zampini } 340f6f667cfSStefano Zampini if (allocated_S_St) { 341f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 342f6f667cfSStefano Zampini } 343f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 34408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 34508122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 34608122e43SStefano Zampini #endif 34708122e43SStefano Zampini if (pcbddc->dbg_flag) { 3481b968477SStefano Zampini PetscInt maxneigs_r; 34908122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3509b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 35108122e43SStefano Zampini } 35208122e43SStefano Zampini PetscFunctionReturn(0); 35308122e43SStefano Zampini } 354b1b3d7a2SStefano Zampini 355674ae819SStefano Zampini #undef __FUNCT__ 356c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 357c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 358c8587f34SStefano Zampini { 359c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3608629588bSStefano Zampini PetscScalar *coarse_submat_vals; 361c8587f34SStefano Zampini PetscErrorCode ierr; 362c8587f34SStefano Zampini 363c8587f34SStefano Zampini PetscFunctionBegin; 364f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3655e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 366c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 367c8587f34SStefano Zampini 368684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3690fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 370684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 371c8587f34SStefano Zampini 372c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 373b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 374c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 375c8587f34SStefano Zampini } 376c8587f34SStefano Zampini 3778629588bSStefano Zampini /* 3788629588bSStefano Zampini Setup local correction and local part of coarse basis. 3798629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3808629588bSStefano Zampini */ 38147f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3828629588bSStefano Zampini 3838629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3848629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3858629588bSStefano Zampini 3868629588bSStefano Zampini /* free */ 3878629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 388c8587f34SStefano Zampini PetscFunctionReturn(0); 389c8587f34SStefano Zampini } 390c8587f34SStefano Zampini 391c8587f34SStefano Zampini #undef __FUNCT__ 392674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 393674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 394674ae819SStefano Zampini { 395674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 396674ae819SStefano Zampini PetscErrorCode ierr; 397674ae819SStefano Zampini 398674ae819SStefano Zampini PetscFunctionBegin; 399674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 400674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 401674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 402674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 403785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 404674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 405f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 406f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 407785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 40863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 40963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 410674ae819SStefano Zampini PetscFunctionReturn(0); 411674ae819SStefano Zampini } 412674ae819SStefano Zampini 413674ae819SStefano Zampini #undef __FUNCT__ 414674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 415674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 416674ae819SStefano Zampini { 417674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 418674ae819SStefano Zampini PetscErrorCode ierr; 419674ae819SStefano Zampini 420674ae819SStefano Zampini PetscFunctionBegin; 421b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 422674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 423674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 424674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 425b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 426674ae819SStefano Zampini PetscFunctionReturn(0); 427674ae819SStefano Zampini } 428674ae819SStefano Zampini 429674ae819SStefano Zampini #undef __FUNCT__ 430674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 431674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 432674ae819SStefano Zampini { 433674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 43406656605SStefano Zampini PetscScalar *array; 435674ae819SStefano Zampini PetscErrorCode ierr; 436674ae819SStefano Zampini 437674ae819SStefano Zampini PetscFunctionBegin; 438674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 44006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 44106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 44258da7f69SStefano Zampini } 443674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 444674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 44515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 44615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 447674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 448674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 449674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 45006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 451674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 452674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4538ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 454674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 455674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 456674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 457f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 458f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 459f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 460f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 461727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4620e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 463f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 46470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4656e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 466674ae819SStefano Zampini PetscFunctionReturn(0); 467674ae819SStefano Zampini } 468674ae819SStefano Zampini 469674ae819SStefano Zampini #undef __FUNCT__ 470f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 471f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4726bfb1811SStefano Zampini { 4736bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4746bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4756bfb1811SStefano Zampini VecType impVecType; 476e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4776bfb1811SStefano Zampini PetscErrorCode ierr; 4786bfb1811SStefano Zampini 4796bfb1811SStefano Zampini PetscFunctionBegin; 480f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 481f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 482f4ddd8eeSStefano Zampini } 483e7b262bdSStefano Zampini /* get sizes */ 484b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 485b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4866bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 487e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 488e7b262bdSStefano Zampini /* R nodes */ 489e7b262bdSStefano Zampini old_size = -1; 490e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 491e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 492e7b262bdSStefano Zampini } 493e7b262bdSStefano Zampini if (n_R != old_size) { 494e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 495e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4966bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4976bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4986bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4996bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 500e7b262bdSStefano Zampini } 501e7b262bdSStefano Zampini /* local primal dofs */ 502e7b262bdSStefano Zampini old_size = -1; 503e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 504e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 505e7b262bdSStefano Zampini } 506e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 507e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 50883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 509e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5106bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 511e7b262bdSStefano Zampini } 512e7b262bdSStefano Zampini /* local explicit constraints */ 513e7b262bdSStefano Zampini old_size = -1; 514e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 515e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 516e7b262bdSStefano Zampini } 517e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 518e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 51983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 52083b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 52183b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 52283b7ccabSStefano Zampini } 5236bfb1811SStefano Zampini PetscFunctionReturn(0); 5246bfb1811SStefano Zampini } 5256bfb1811SStefano Zampini 5266bfb1811SStefano Zampini #undef __FUNCT__ 52747f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 52847f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 52988ebb749SStefano Zampini { 53025084f0cSStefano Zampini PetscErrorCode ierr; 53125084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 53288ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 53388ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 534d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 53525084f0cSStefano Zampini /* submatrices of local problem */ 53680677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 53706656605SStefano Zampini /* submatrices of local coarse problem */ 53806656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 53925084f0cSStefano Zampini /* working matrices */ 54006656605SStefano Zampini Mat C_CR; 54125084f0cSStefano Zampini /* additional working stuff */ 54206656605SStefano Zampini PC pc_R; 54306656605SStefano Zampini Mat F; 54406656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 54506656605SStefano Zampini 54625084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 54706656605SStefano Zampini PetscScalar *work; 54806656605SStefano Zampini PetscInt *idx_V_B; 54906656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 55006656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 551b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 55245a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 55388ebb749SStefano Zampini MatType impMatType; 55425084f0cSStefano Zampini /* some shortcuts to scalars */ 55506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 55688ebb749SStefano Zampini 55788ebb749SStefano Zampini PetscFunctionBegin; 558b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 55988ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 56088ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 561b371cd4fSStefano Zampini n_B = pcis->n_B; 562b371cd4fSStefano Zampini n_D = pcis->n - n_B; 56388ebb749SStefano Zampini n_R = pcis->n - n_vertices; 56488ebb749SStefano Zampini 56588ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 56688ebb749SStefano Zampini impMatType = MATSEQDENSE; 56788ebb749SStefano Zampini 56888ebb749SStefano Zampini /* vertices in boundary numbering */ 569785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5700e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 57188ebb749SStefano Zampini if (i != n_vertices) { 57222d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 57388ebb749SStefano Zampini } 57488ebb749SStefano Zampini 57506656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 57606656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 57706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 57806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 57906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 58206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 58306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 58406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 58506656605SStefano Zampini 58606656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 58706656605SStefano Zampini /* allocate workspace */ 58806656605SStefano Zampini n = 0; 58906656605SStefano Zampini if (n_constraints) { 59006656605SStefano Zampini n += n_R*n_constraints; 59106656605SStefano Zampini } 59206656605SStefano Zampini if (n_vertices) { 59306656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 59480677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 59506656605SStefano Zampini } 5963301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 59706656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 59806656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 59906656605SStefano Zampini } 60006656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 60106656605SStefano Zampini 60206656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 60306656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 60406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 60506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 60606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 60706656605SStefano Zampini if (isLU || isILU || isCHOL) { 60806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 609d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 610d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 611d62866d3SStefano Zampini MatFactorType type; 612d62866d3SStefano Zampini 6136816873aSStefano Zampini F = reuse_mumps->F; 6146816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 615d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 61606656605SStefano Zampini } else { 61706656605SStefano Zampini F = NULL; 61806656605SStefano Zampini } 61906656605SStefano Zampini 62088ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 62188ebb749SStefano Zampini if (n_constraints) { 62206656605SStefano Zampini Mat M1,M2,M3; 62380677318SStefano Zampini Mat auxmat; 62406656605SStefano Zampini IS is_aux; 62580677318SStefano Zampini PetscScalar *array,*array2; 62606656605SStefano Zampini 627f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 62880677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 62988ebb749SStefano Zampini 63025084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 63125084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6328ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 63380677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 63488ebb749SStefano Zampini 63580677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 63680677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 63706656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 63888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63906656605SStefano Zampini const PetscScalar *row_cmat_values; 64006656605SStefano Zampini const PetscInt *row_cmat_indices; 64106656605SStefano Zampini PetscInt size_of_constraint,j; 64288ebb749SStefano Zampini 64306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 64506656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64606656605SStefano Zampini } 64706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64806656605SStefano Zampini } 64980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 65006656605SStefano Zampini if (F) { 65106656605SStefano Zampini Mat B; 65206656605SStefano Zampini 65306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 65480677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 65506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 65606656605SStefano Zampini } else { 65780677318SStefano Zampini PetscScalar *marr; 65880677318SStefano Zampini 65980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 66106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 66306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 66406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 66606656605SStefano Zampini } 66780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66806656605SStefano Zampini } 66980677318SStefano Zampini if (!pcbddc->switch_static) { 67080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 67180677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 67280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67380677318SStefano Zampini for (i=0;i<n_constraints;i++) { 67480677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 67580677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 67680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 67980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 68080677318SStefano Zampini } 68180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 68280677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 68380677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68480677318SStefano Zampini } else { 68580677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 68680677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 68725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68880677318SStefano Zampini } 68980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 69080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 69180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 69206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 69306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 69480677318SStefano Zampini if (isCHOL) { 69580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 69680677318SStefano Zampini } else { 69725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 69880677318SStefano Zampini } 69980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 70006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 70125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 70225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 70325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 70480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 70580677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 70680677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 70706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 70806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 709f4ddd8eeSStefano Zampini } 71088ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 71188ebb749SStefano Zampini if (n_vertices) { 71206656605SStefano Zampini IS is_aux; 7133a50541eSStefano Zampini 7146816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7156816873aSStefano Zampini IS tis; 7166816873aSStefano Zampini 7176816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7186816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7196816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7206816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7216816873aSStefano Zampini } else { 7223a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7236816873aSStefano Zampini } 7249577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72604708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 72725084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 72888ebb749SStefano Zampini } 72988ebb749SStefano Zampini 73088ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 731f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 73206656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 73306656605SStefano Zampini if (pcbddc->coarse_phi_D) { 73406656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73506656605SStefano Zampini } 736f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 73706656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 73806656605SStefano Zampini PetscScalar *marray; 73906656605SStefano Zampini 74006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 74106656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 742f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 743f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 744f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 746f4ddd8eeSStefano Zampini } 747f4ddd8eeSStefano Zampini } 74806656605SStefano Zampini 749f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 75006656605SStefano Zampini PetscScalar *marray; 75188ebb749SStefano Zampini 75206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7538eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75406656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75588ebb749SStefano Zampini } 7563301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75706656605SStefano Zampini n *= 2; 75888ebb749SStefano Zampini } 75906656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 76006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 76106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 76406656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76588ebb749SStefano Zampini } 7663301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 76706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7688eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 77188ebb749SStefano Zampini } 77288ebb749SStefano Zampini } else { 773c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 774c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7751b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 776c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 777c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 778c0553b1fSStefano Zampini } 77988ebb749SStefano Zampini } 78006656605SStefano Zampini } 78106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 78206656605SStefano Zampini /* vertices */ 78306656605SStefano Zampini if (n_vertices) { 78416f15bc4SStefano Zampini 78504708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 78604708bb6SStefano Zampini 78716f15bc4SStefano Zampini if (n_R) { 78806656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 78906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 79016f15bc4SStefano Zampini PetscScalar *x,*y; 79104708bb6SStefano Zampini PetscBool isseqaij; 79206656605SStefano Zampini 79321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 79406656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 79504708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7966816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 79706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 79806656605SStefano Zampini } else { 79906656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 80106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 80206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 80306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 80606656605SStefano Zampini } 80706656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 80806656605SStefano Zampini } 80980677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81006656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 81106656605SStefano Zampini if (n_constraints) { 81206656605SStefano Zampini Mat B; 81380677318SStefano Zampini 814b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 81580677318SStefano Zampini for (i=0;i<n_vertices;i++) { 81680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 81780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 81880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82280677318SStefano Zampini } 82380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82480677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 82580677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 83106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83206656605SStefano Zampini } 83304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 83404708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 83504708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 83604708bb6SStefano Zampini } 83706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 83880677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 83906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 84206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 84716f15bc4SStefano Zampini } else { 84816f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84916f15bc4SStefano Zampini } 85021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 85106656605SStefano Zampini /* coarse basis functions */ 85206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85316f15bc4SStefano Zampini PetscScalar *y; 85416f15bc4SStefano Zampini 85506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 86106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86306656605SStefano Zampini 86406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87106656605SStefano Zampini } 87206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87306656605SStefano Zampini } 87404708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 87504708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 87606656605SStefano Zampini } 87706656605SStefano Zampini 87806656605SStefano Zampini if (n_constraints) { 87906656605SStefano Zampini Mat B; 88006656605SStefano Zampini 88106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 88206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88506656605SStefano Zampini if (n_vertices) { 88680677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 88780677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 88880677318SStefano Zampini } else { 88980677318SStefano Zampini Mat S_VCt; 89080677318SStefano Zampini 89180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 89280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89480677318SStefano Zampini } 89506656605SStefano Zampini } 89606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 89706656605SStefano Zampini /* coarse basis functions */ 89806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 89906656605SStefano Zampini PetscScalar *y; 90006656605SStefano Zampini 90106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 90806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 90906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 91106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91506656605SStefano Zampini } 91606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91706656605SStefano Zampini } 91806656605SStefano Zampini } 91980677318SStefano Zampini if (n_constraints) { 92080677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 92180677318SStefano Zampini } 92206656605SStefano Zampini 92306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9243301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 92506656605SStefano Zampini 92606656605SStefano Zampini if (n_constraints) { 92716f15bc4SStefano Zampini Mat S_CCT,B_C; 92806656605SStefano Zampini 92980677318SStefano Zampini /* this is a lazy thing */ 93080677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 93106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 93206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93506656605SStefano Zampini if (n_vertices) { 93616f15bc4SStefano Zampini Mat B_V,S_VCT; 93706656605SStefano Zampini 93806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 93906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 94006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 94216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94306656605SStefano Zampini } 94406656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94580677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94680677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 94706656605SStefano Zampini } 94816f15bc4SStefano Zampini if (n_vertices && n_R) { 94906656605SStefano Zampini Mat A_VRT; 95080677318SStefano Zampini PetscScalar *marray; 95106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 95206656605SStefano Zampini 95380677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95480677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95580677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 95780677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 95880677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 95916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 96006656605SStefano Zampini } 96106656605SStefano Zampini 96206656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96606656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96906656605SStefano Zampini } 97006656605SStefano Zampini } else { 97106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97706656605SStefano Zampini } 97806656605SStefano Zampini } 97906656605SStefano Zampini /* coarse basis functions */ 98006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98106656605SStefano Zampini PetscScalar *y; 98206656605SStefano Zampini 98306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98806656605SStefano Zampini if (i<n_vertices) { 98906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 99006656605SStefano Zampini } 99106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99306656605SStefano Zampini 99406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 99706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100106656605SStefano Zampini } 100206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100306656605SStefano Zampini } 100406656605SStefano Zampini } 1005d62866d3SStefano Zampini /* free memory */ 100688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 100706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 100906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1011d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1012d62866d3SStefano Zampini if (n_vertices) { 1013d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1014d62866d3SStefano Zampini } 1015d62866d3SStefano Zampini if (n_constraints) { 1016d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1017d62866d3SStefano Zampini } 101888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 101988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102125084f0cSStefano Zampini if (pcbddc->dbg_flag) { 102288ebb749SStefano Zampini Mat coarse_sub_mat; 102325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10278bec7fa6SStefano Zampini Mat C_B,CPHI; 10288bec7fa6SStefano Zampini IS is_dummy; 10298bec7fa6SStefano Zampini Vec mones; 103088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 103188ebb749SStefano Zampini PetscReal real_value; 103288ebb749SStefano Zampini 103388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1039c0553b1fSStefano Zampini if (unsymmetric_check) { 104088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 104288ebb749SStefano Zampini } 104388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104488ebb749SStefano Zampini 104525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10463301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 104725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1048c0553b1fSStefano Zampini if (unsymmetric_check) { 104988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini } else { 106288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107088ebb749SStefano Zampini } 107188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10768bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10771575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 107806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini 10808bec7fa6SStefano Zampini /* check constraints */ 10818bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10838bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10848bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10858bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10868bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1088bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1089c0553b1fSStefano Zampini if (unsymmetric_check) { 1090bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1091bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1092bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1093bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1094bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 109588ebb749SStefano Zampini } 10968bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10978bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10988bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10998bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11008bec7fa6SStefano Zampini 110125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 110388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 110888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 110988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 111088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 111188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1112c0553b1fSStefano Zampini if (unsymmetric_check) { 111388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111588ebb749SStefano Zampini } 111688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 111788ebb749SStefano Zampini } 11188629588bSStefano Zampini /* get back data */ 11198629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 112088ebb749SStefano Zampini PetscFunctionReturn(0); 112188ebb749SStefano Zampini } 112288ebb749SStefano Zampini 112388ebb749SStefano Zampini #undef __FUNCT__ 1124d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1125d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1126aa0d41d4SStefano Zampini { 1127d65f70fdSStefano Zampini Mat *work_mat; 1128d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1129d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1130d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1131aa0d41d4SStefano Zampini PetscErrorCode ierr; 1132aa0d41d4SStefano Zampini 1133aa0d41d4SStefano Zampini PetscFunctionBegin; 1134d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1136d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1137d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1138aa0d41d4SStefano Zampini 1139d65f70fdSStefano Zampini if (!rsorted) { 1140906d46d4SStefano Zampini const PetscInt *idxs; 1141906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1142aa0d41d4SStefano Zampini 1143d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1144d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1146d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1147aa0d41d4SStefano Zampini } 1148d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1149d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1150d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1151d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1152aa0d41d4SStefano Zampini } 1153d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1155d65f70fdSStefano Zampini } else { 1156d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini isrow_s = isrow; 1158aa0d41d4SStefano Zampini } 1159906d46d4SStefano Zampini 1160d65f70fdSStefano Zampini if (!csorted) { 1161d65f70fdSStefano Zampini if (isrow == iscol) { 1162d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1163d65f70fdSStefano Zampini iscol_s = isrow_s; 1164d65f70fdSStefano Zampini } else { 1165d65f70fdSStefano Zampini const PetscInt *idxs; 1166d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1167906d46d4SStefano Zampini 1168d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1169d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1171d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1172d65f70fdSStefano Zampini } 1173d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1174d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1175d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1176d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1177d65f70fdSStefano Zampini } 1178d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini } else { 1182d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini iscol_s = iscol; 1184d65f70fdSStefano Zampini } 1185d65f70fdSStefano Zampini 1186d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1187d65f70fdSStefano Zampini 1188d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1189906d46d4SStefano Zampini Mat new_mat; 1190d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1191906d46d4SStefano Zampini 1192d65f70fdSStefano Zampini if (!rsorted) { 1193d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1194d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1196d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1197906d46d4SStefano Zampini } 1198d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini } else { 1201d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1202906d46d4SStefano Zampini } 1203d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini 1205d65f70fdSStefano Zampini if (!csorted) { 1206d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1207d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1209d65f70fdSStefano Zampini } else { 1210d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1211d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1213d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1214d65f70fdSStefano Zampini } 1215d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1217d65f70fdSStefano Zampini } 1218d65f70fdSStefano Zampini } else { 1219d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini } 1221d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini 1223d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1225d65f70fdSStefano Zampini work_mat[0] = new_mat; 1226d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1228d65f70fdSStefano Zampini } 1229d65f70fdSStefano Zampini 1230d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini *B = work_mat[0]; 1232d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1233d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1234d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1235d65f70fdSStefano Zampini PetscFunctionReturn(0); 1236d65f70fdSStefano Zampini } 1237d65f70fdSStefano Zampini 1238d65f70fdSStefano Zampini #undef __FUNCT__ 12395e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12405e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1241aa0d41d4SStefano Zampini { 1242aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12433bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 12445e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1245d65f70fdSStefano Zampini Mat new_mat; 12465e8657edSStefano Zampini IS is_local,is_global; 1247d65f70fdSStefano Zampini PetscInt local_size; 1248d65f70fdSStefano Zampini PetscBool isseqaij; 1249aa0d41d4SStefano Zampini PetscErrorCode ierr; 1250aa0d41d4SStefano Zampini 1251aa0d41d4SStefano Zampini PetscFunctionBegin; 1252aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12535e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12545e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 12553bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1256aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1257d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1258aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1259906d46d4SStefano Zampini 1260906d46d4SStefano Zampini /* check */ 1261906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1262906d46d4SStefano Zampini Vec x,x_change; 1263906d46d4SStefano Zampini PetscReal error; 1264906d46d4SStefano Zampini 12655e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1266906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12675e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1268e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1269e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1270d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1271e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1272e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1273906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1274906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1275906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1276906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1277906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1278906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1279906d46d4SStefano Zampini } 1280906d46d4SStefano Zampini 128122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12829b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 128322d5777bSStefano Zampini if (isseqaij) { 1284d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1285aa0d41d4SStefano Zampini } else { 1286aa0d41d4SStefano Zampini Mat work_mat; 1287aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1288d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1289aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1290aa0d41d4SStefano Zampini } 12913301b35fSStefano Zampini if (matis->A->symmetric_set) { 12923301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1293e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12943301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1295e496cd5dSStefano Zampini #endif 12963301b35fSStefano Zampini } 129745a1bb75SStefano Zampini /* 129845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1299d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 130045a1bb75SStefano Zampini */ 1301d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1302aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1303aa0d41d4SStefano Zampini } 1304aa0d41d4SStefano Zampini 1305aa0d41d4SStefano Zampini #undef __FUNCT__ 1306a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13078ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1308a64d13efSStefano Zampini { 1309a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1310a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1311d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 131253892102SStefano Zampini PetscInt *idx_R_local=NULL; 13133a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13143a50541eSStefano Zampini PetscInt vbs,bs; 13156816873aSStefano Zampini PetscBT bitmask=NULL; 1316a64d13efSStefano Zampini PetscErrorCode ierr; 1317a64d13efSStefano Zampini 1318a64d13efSStefano Zampini PetscFunctionBegin; 1319b23d619eSStefano Zampini /* 1320b23d619eSStefano Zampini No need to setup local scatters if 1321b23d619eSStefano Zampini - primal space is unchanged 1322b23d619eSStefano Zampini AND 1323b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1324b23d619eSStefano Zampini AND 1325b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1326b23d619eSStefano Zampini */ 1327b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1328f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1329f4ddd8eeSStefano Zampini } 1330f4ddd8eeSStefano Zampini /* destroy old objects */ 1331f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1332f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1333f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1334a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1335b371cd4fSStefano Zampini n_B = pcis->n_B; 1336b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1337b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13383a50541eSStefano Zampini 1339a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13406816873aSStefano Zampini 134153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13426816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1343854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1344a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1345a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13460e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1347a64d13efSStefano Zampini } 1348a64d13efSStefano Zampini 1349a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13504641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13516816873aSStefano Zampini idx_R_local[n_R++] = i; 1352a64d13efSStefano Zampini } 1353a64d13efSStefano Zampini } 135453892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13556816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13566816873aSStefano Zampini 135753892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 135853892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13596816873aSStefano Zampini } 13603a50541eSStefano Zampini 13613a50541eSStefano Zampini /* Block code */ 13623a50541eSStefano Zampini vbs = 1; 13633a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13643a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13653a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13663a50541eSStefano Zampini PetscInt *vary; 1367d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1368785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13693a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1370d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1371d3df7717SStefano Zampini /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */ 13720e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1373d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13743a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13753a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13763a50541eSStefano Zampini break; 13773a50541eSStefano Zampini } 13783a50541eSStefano Zampini } 1379d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1380d3df7717SStefano Zampini } else { 1381d3df7717SStefano Zampini /* Verify directly the R set */ 1382d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1383d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1384d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1385d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1386d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1387d3df7717SStefano Zampini break; 1388d3df7717SStefano Zampini } 1389d3df7717SStefano Zampini } 1390d3df7717SStefano Zampini } 1391d3df7717SStefano Zampini } 13923a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13933a50541eSStefano Zampini vbs = bs; 13943a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13953a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13963a50541eSStefano Zampini } 13973a50541eSStefano Zampini } 13983a50541eSStefano Zampini } 13993a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 140053892102SStefano Zampini if (sub_schurs->reuse_mumps) { 140153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 140253892102SStefano Zampini 140353892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140453892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140653892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140753892102SStefano Zampini } else { 14083a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 140953892102SStefano Zampini } 1410a64d13efSStefano Zampini 1411a64d13efSStefano Zampini /* print some info if requested */ 1412a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1413a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1414a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14151575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1416a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1417a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14183a50541eSStefano 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); 1419a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1420a64d13efSStefano Zampini } 1421a64d13efSStefano Zampini 1422a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14236816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14246816873aSStefano Zampini IS is_aux1,is_aux2; 14256816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14266816873aSStefano Zampini 14273a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1428854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1429854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1430a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14314641a718SStefano Zampini for (i=0; i<n_D; i++) { 14324641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14334641a718SStefano Zampini } 1434a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1435a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14364641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14374641a718SStefano Zampini aux_array1[j++] = i; 1438a64d13efSStefano Zampini } 1439a64d13efSStefano Zampini } 1440a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1441a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1442a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14434641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14444641a718SStefano Zampini aux_array2[j++] = i; 1445a64d13efSStefano Zampini } 1446a64d13efSStefano Zampini } 1447a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1448a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1449a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1450a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1451a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1452a64d13efSStefano Zampini 14538eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1454785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1455a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14564641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14574641a718SStefano Zampini aux_array1[j++] = i; 1458a64d13efSStefano Zampini } 1459a64d13efSStefano Zampini } 1460a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1461a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1462a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1463a64d13efSStefano Zampini } 14644641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14653a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1466d62866d3SStefano Zampini } else { 146753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14686816873aSStefano Zampini IS tis; 14696816873aSStefano Zampini PetscInt schur_size; 14706816873aSStefano Zampini 147153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14726816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 147353892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14746816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14756816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14766816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14776816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14786816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1479d62866d3SStefano Zampini } 1480d62866d3SStefano Zampini } 1481a64d13efSStefano Zampini PetscFunctionReturn(0); 1482a64d13efSStefano Zampini } 1483a64d13efSStefano Zampini 1484304d26faSStefano Zampini 1485304d26faSStefano Zampini #undef __FUNCT__ 1486304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1487684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1488304d26faSStefano Zampini { 1489304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1490304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1491304d26faSStefano Zampini PC pc_temp; 1492304d26faSStefano Zampini Mat A_RR; 1493f4ddd8eeSStefano Zampini MatReuse reuse; 1494304d26faSStefano Zampini PetscScalar m_one = -1.0; 1495304d26faSStefano Zampini PetscReal value; 149604708bb6SStefano Zampini PetscInt n_D,n_R; 14979577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1498304d26faSStefano Zampini PetscErrorCode ierr; 1499e604994aSStefano Zampini /* prefixes stuff */ 1500312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1501e604994aSStefano Zampini size_t len; 1502304d26faSStefano Zampini 1503304d26faSStefano Zampini PetscFunctionBegin; 1504304d26faSStefano Zampini 1505e604994aSStefano Zampini /* compute prefixes */ 1506e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1507e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1508e604994aSStefano Zampini if (!pcbddc->current_level) { 1509e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1510e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1511e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1512e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1513e604994aSStefano Zampini } else { 1514e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1515312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1516e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1517e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1518312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1519312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 152034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 152134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1522e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1523e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1524e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1525e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1526e604994aSStefano Zampini } 1527e604994aSStefano Zampini 1528304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1529684f6988SStefano Zampini if (dirichlet) { 1530d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15313301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15323301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1533964fefecSStefano Zampini } 1534ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1535964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1536304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1537304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1538304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1539304d26faSStefano Zampini /* default */ 1540304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1541e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15429577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1543304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15449577ea80SStefano Zampini if (issbaij) { 15459577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15469577ea80SStefano Zampini } else { 1547304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15489577ea80SStefano Zampini } 1549304d26faSStefano Zampini /* Allow user's customization */ 1550304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1551304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1552304d26faSStefano Zampini } 1553d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1554d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1555d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1556d62866d3SStefano Zampini 1557d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1558d5574798SStefano Zampini } 1559304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1560304d26faSStefano Zampini if (!n_D) { 1561304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1562304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1563304d26faSStefano Zampini } 1564304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1565304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1566304d26faSStefano Zampini /* set ksp_D into pcis data */ 1567304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1568304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1569304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1570684f6988SStefano Zampini } 1571304d26faSStefano Zampini 1572304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1573684f6988SStefano Zampini A_RR = 0; 1574684f6988SStefano Zampini if (neumann) { 1575d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 157604708bb6SStefano Zampini PetscInt ibs,mbs; 157704708bb6SStefano Zampini PetscBool issbaij; 157804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1579f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15808ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1581f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1582f4ddd8eeSStefano Zampini PetscInt nn_R; 158381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1585f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1586f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1587f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1588f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1589f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1590f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1591727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1592f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1593f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1594f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1595f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1596f4ddd8eeSStefano Zampini } 1597f4ddd8eeSStefano Zampini } 1598f4ddd8eeSStefano Zampini /* last check */ 1599d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1600f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1601f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1602f4ddd8eeSStefano Zampini } 1603f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1604f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1605f4ddd8eeSStefano Zampini } 1606f4ddd8eeSStefano Zampini /* extract A_RR */ 1607af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1608af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 160904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 161004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 161104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 161204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 161304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1614af732b37SStefano Zampini } else { 161504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16166816873aSStefano Zampini } 161704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 161804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 161904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 162004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162104708bb6SStefano Zampini } else { 162204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162304708bb6SStefano Zampini } 162404708bb6SStefano Zampini } 162504708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1626f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16283301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16296816873aSStefano Zampini } 16306816873aSStefano Zampini } else { 16316816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16326816873aSStefano Zampini 16336816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16346816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16356816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1636af732b37SStefano Zampini } 1637f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1638304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1639304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1640304d26faSStefano Zampini /* default */ 1641304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1642e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1643304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16449577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16459577ea80SStefano Zampini if (issbaij) { 16469577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16479577ea80SStefano Zampini } else { 1648304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16499577ea80SStefano Zampini } 1650304d26faSStefano Zampini /* Allow user's customization */ 1651304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1652304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1653304d26faSStefano Zampini } 1654d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1655304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1656304d26faSStefano Zampini if (!n_R) { 1657304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1658304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1659304d26faSStefano Zampini } 1660d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1661d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1662d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1663d62866d3SStefano Zampini 1664d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1665d62866d3SStefano Zampini } 1666304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1667304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1668684f6988SStefano Zampini } 16696816873aSStefano Zampini /* free Neumann problem's matrix */ 16706816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1671304d26faSStefano Zampini 1672304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16730fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1674684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1675684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16761575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1677684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1678684f6988SStefano Zampini } 1679684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16800fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16810fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16820fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1685304d26faSStefano Zampini /* need to be adapted? */ 1686b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1687b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1688b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1689304d26faSStefano Zampini /* print info */ 1690304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1691e604994aSStefano 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); 1692304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1693304d26faSStefano Zampini } 1694b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1695298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1696304d26faSStefano Zampini } 1697684f6988SStefano Zampini } 1698684f6988SStefano Zampini if (neumann) { /* Neumann */ 16996816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 17000fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 17010fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 17020fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 17030fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17040fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1705304d26faSStefano Zampini /* need to be adapted? */ 1706b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1707b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1708304d26faSStefano Zampini /* print info */ 1709304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1710e604994aSStefano 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); 1711304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1712304d26faSStefano Zampini } 1713b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1714298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1715304d26faSStefano Zampini } 17160fccc4e9SStefano Zampini } 1717684f6988SStefano Zampini } 1718304d26faSStefano Zampini PetscFunctionReturn(0); 1719304d26faSStefano Zampini } 1720304d26faSStefano Zampini 1721304d26faSStefano Zampini #undef __FUNCT__ 1722ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 172380677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1724674ae819SStefano Zampini { 1725674ae819SStefano Zampini PetscErrorCode ierr; 1726674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1727be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1728674ae819SStefano Zampini 1729674ae819SStefano Zampini PetscFunctionBegin; 1730be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 173220c7b377SStefano Zampini } 173380677318SStefano Zampini if (!pcbddc->switch_static) { 173480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 173680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173720c7b377SStefano Zampini } 1738be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174120c7b377SStefano Zampini } else { 1742be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1743be83ff47SStefano Zampini 174453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174620c7b377SStefano Zampini } 1747be83ff47SStefano Zampini } else { 174880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 175380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 175480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1757674ae819SStefano Zampini } 1758674ae819SStefano Zampini } 1759be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176080677318SStefano Zampini if (applytranspose) { 176180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176280677318SStefano Zampini } else { 176380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176480677318SStefano Zampini } 1765be83ff47SStefano Zampini } else { 1766be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1767be83ff47SStefano Zampini 1768be83ff47SStefano Zampini if (applytranspose) { 1769*5a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1770be83ff47SStefano Zampini } else { 1771*5a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1772be83ff47SStefano Zampini } 1773be83ff47SStefano Zampini } 177480677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 177580677318SStefano Zampini if (!pcbddc->switch_static) { 1776be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 177780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1779be83ff47SStefano Zampini } else { 1780be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1781be83ff47SStefano Zampini 178253892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178353892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1784be83ff47SStefano Zampini } 178580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 178880677318SStefano Zampini } 178980677318SStefano Zampini } else { 179080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 179680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 179780677318SStefano Zampini } 179880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1802674ae819SStefano Zampini } 1803674ae819SStefano Zampini PetscFunctionReturn(0); 1804674ae819SStefano Zampini } 1805674ae819SStefano Zampini 1806dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1807674ae819SStefano Zampini #undef __FUNCT__ 1808674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1809dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1810674ae819SStefano Zampini { 1811674ae819SStefano Zampini PetscErrorCode ierr; 1812674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1813674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1814674ae819SStefano Zampini const PetscScalar zero = 0.0; 1815674ae819SStefano Zampini 1816674ae819SStefano Zampini PetscFunctionBegin; 1817dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1818dc359a40SStefano Zampini if (applytranspose) { 1819674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18208eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1821dc359a40SStefano Zampini } else { 1822674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1823674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 182415aaf578SStefano Zampini } 182512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 182612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 182712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182912edc857SStefano Zampini 18309f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 183112edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 183212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1833964fefecSStefano Zampini Vec rhs,sol; 1834964fefecSStefano Zampini 1835964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1836964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 183712edc857SStefano Zampini if (applytranspose) { 1838964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183912edc857SStefano Zampini } else { 1840964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184112edc857SStefano Zampini } 184212edc857SStefano Zampini } 1843674ae819SStefano Zampini 1844674ae819SStefano Zampini /* Local solution on R nodes */ 184580677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 184680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18479f00e9b4SStefano Zampini } 1848674ae819SStefano Zampini 18499f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18509f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 185112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1852674ae819SStefano Zampini 1853674ae819SStefano Zampini /* Sum contributions from two levels */ 1854dc359a40SStefano Zampini if (applytranspose) { 1855dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1856dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1857dc359a40SStefano Zampini } else { 1858674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18598eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1860dc359a40SStefano Zampini } 1861674ae819SStefano Zampini PetscFunctionReturn(0); 1862674ae819SStefano Zampini } 1863674ae819SStefano Zampini 1864674ae819SStefano Zampini #undef __FUNCT__ 1865674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 186612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1867674ae819SStefano Zampini { 1868674ae819SStefano Zampini PetscErrorCode ierr; 1869674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187058da7f69SStefano Zampini PetscScalar *array; 187112edc857SStefano Zampini Vec from,to; 1872674ae819SStefano Zampini 1873674ae819SStefano Zampini PetscFunctionBegin; 187412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187512edc857SStefano Zampini from = pcbddc->coarse_vec; 187612edc857SStefano Zampini to = pcbddc->vec1_P; 187712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 187812edc857SStefano Zampini Vec tvec; 187958da7f69SStefano Zampini 188058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 188212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 188458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 188558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 188612edc857SStefano Zampini } 188712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188812edc857SStefano Zampini from = pcbddc->vec1_P; 188912edc857SStefano Zampini to = pcbddc->coarse_vec; 189012edc857SStefano Zampini } 189112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1892674ae819SStefano Zampini PetscFunctionReturn(0); 1893674ae819SStefano Zampini } 1894674ae819SStefano Zampini 1895674ae819SStefano Zampini #undef __FUNCT__ 1896674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 189712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1898674ae819SStefano Zampini { 1899674ae819SStefano Zampini PetscErrorCode ierr; 1900674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 190158da7f69SStefano Zampini PetscScalar *array; 190212edc857SStefano Zampini Vec from,to; 1903674ae819SStefano Zampini 1904674ae819SStefano Zampini PetscFunctionBegin; 190512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 190612edc857SStefano Zampini from = pcbddc->coarse_vec; 190712edc857SStefano Zampini to = pcbddc->vec1_P; 190812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 190912edc857SStefano Zampini from = pcbddc->vec1_P; 191012edc857SStefano Zampini to = pcbddc->coarse_vec; 191112edc857SStefano Zampini } 191212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 191312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 191412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 191512edc857SStefano Zampini Vec tvec; 191658da7f69SStefano Zampini 191712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 191858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 191958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 192058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 192158da7f69SStefano Zampini } 192258da7f69SStefano Zampini } else { 192358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 192458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 192512edc857SStefano Zampini } 192612edc857SStefano Zampini } 1927674ae819SStefano Zampini PetscFunctionReturn(0); 1928674ae819SStefano Zampini } 1929674ae819SStefano Zampini 1930984c4197SStefano Zampini /* uncomment for testing purposes */ 1931984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1932674ae819SStefano Zampini #undef __FUNCT__ 1933674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1934674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1935674ae819SStefano Zampini { 1936674ae819SStefano Zampini PetscErrorCode ierr; 1937674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1938674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1939674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1940984c4197SStefano Zampini /* one and zero */ 1941984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1942984c4197SStefano Zampini /* space to store constraints and their local indices */ 19439162d606SStefano Zampini PetscScalar *constraints_data; 19449162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19459162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19469162d606SStefano Zampini PetscInt *constraints_n; 1947984c4197SStefano Zampini /* iterators */ 1948b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1949984c4197SStefano Zampini /* BLAS integers */ 1950e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1951e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1952c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1953727cdba6SStefano Zampini /* reuse */ 19540e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19550e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1956984c4197SStefano Zampini /* change of basis */ 1957b3d85658SStefano Zampini PetscBool qr_needed; 19589162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1959984c4197SStefano Zampini /* auxiliary stuff */ 196064efe560SStefano Zampini PetscInt *nnz,*is_indices; 19618a0068c3SStefano Zampini PetscInt ncc; 1962984c4197SStefano Zampini /* some quantities */ 196345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1964a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1965984c4197SStefano Zampini 1966674ae819SStefano Zampini PetscFunctionBegin; 19678e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19688e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19698e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1970088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1971088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19720e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19730e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19740e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19750e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19760e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1977088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1978cf5a6209SStefano Zampini 1979cf5a6209SStefano Zampini /* print some info */ 1980cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1981cf5a6209SStefano Zampini IS vertices; 1982cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1983cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1984cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1985cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 19861575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1987cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1988cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1989fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1990fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1991cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 19921575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1993cf5a6209SStefano Zampini } 1994cf5a6209SStefano Zampini 1995cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19969162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1997cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1998cf5a6209SStefano Zampini const Vec *nearnullvecs; 1999cf5a6209SStefano Zampini Vec *localnearnullsp; 2000cf5a6209SStefano Zampini PetscScalar *array; 2001cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2002cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2003674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2004b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2005674ae819SStefano Zampini PetscScalar *work; 2006674ae819SStefano Zampini PetscReal *singular_vals; 2007674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2008674ae819SStefano Zampini PetscReal *rwork; 2009674ae819SStefano Zampini #endif 2010674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2011674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2012674ae819SStefano Zampini #else 2013964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2014964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2015674ae819SStefano Zampini #endif 2016674ae819SStefano Zampini 2017674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2018d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2019d06fc5fdSStefano Zampini /* free unneeded index sets */ 2020d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2021d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2022674ae819SStefano Zampini } 2023d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2024d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2025d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2026d06fc5fdSStefano Zampini } 2027d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2028d06fc5fdSStefano Zampini n_ISForEdges = 0; 2029d06fc5fdSStefano Zampini } 2030d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2031d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2032d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2033d06fc5fdSStefano Zampini } 2034d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2035d06fc5fdSStefano Zampini n_ISForFaces = 0; 2036d06fc5fdSStefano Zampini } 203770022509SStefano Zampini 203870022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 203970022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 204070022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 204170022509SStefano Zampini if (pcbddc->NullSpace) { 204270022509SStefano Zampini PetscBool tbool[2],gbool[2]; 204370022509SStefano Zampini 204470022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2045b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2046d06fc5fdSStefano Zampini if (!ISForEdges) { 2047d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2048d06fc5fdSStefano Zampini } 2049b8ffe317SStefano Zampini } 2050d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2051d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2052d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2053d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2054d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 205598a51de6SStefano Zampini } 205670022509SStefano Zampini #endif 205708122e43SStefano Zampini 2058674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2059674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2060674ae819SStefano Zampini if (nearnullsp) { 2061674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2062f4ddd8eeSStefano Zampini /* remove any stored info */ 2063f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2064f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2065f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2066f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2067f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2068473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2069f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2070f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2071f4ddd8eeSStefano Zampini } 2072984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2073984c4197SStefano Zampini nnsp_size = 0; 2074674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2075674ae819SStefano Zampini } 2076984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2077984c4197SStefano Zampini max_constraints = nnsp_size; 2078984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2079984c4197SStefano Zampini 2080674ae819SStefano Zampini /* 2081674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20829162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20839162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20849162d606SStefano Zampini There can be multiple constraints per connected component 2085674ae819SStefano Zampini */ 2086674ae819SStefano Zampini n_vertices = 0; 2087674ae819SStefano Zampini if (ISForVertices) { 2088674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2089674ae819SStefano Zampini } 20909162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20919162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20929162d606SStefano Zampini 20939162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20949162d606SStefano Zampini total_counts *= max_constraints; 2095674ae819SStefano Zampini total_counts += n_vertices; 20964641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20979162d606SStefano Zampini 2098674ae819SStefano Zampini total_counts = 0; 2099674ae819SStefano Zampini max_size_of_constraint = 0; 2100674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 21019162d606SStefano Zampini IS used_is; 2102674ae819SStefano Zampini if (i<n_ISForEdges) { 21039162d606SStefano Zampini used_is = ISForEdges[i]; 2104674ae819SStefano Zampini } else { 21059162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2106674ae819SStefano Zampini } 21079162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2108674ae819SStefano Zampini total_counts += j; 2109674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2110674ae819SStefano Zampini } 21119162d606SStefano 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); 21129162d606SStefano Zampini 2113984c4197SStefano Zampini /* get local part of global near null space vectors */ 2114785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2115984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2116984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2117e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2118e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2119984c4197SStefano Zampini } 2120674ae819SStefano Zampini 2121242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2122242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2123a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2124242a89d7SStefano Zampini 2125984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2126a773dcb8SStefano Zampini if (!skip_lapack) { 2127674ae819SStefano Zampini PetscScalar temp_work; 2128911cabfeSStefano Zampini 2129674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2130984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2131785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2132785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2133785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2134674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2135785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2136674ae819SStefano Zampini #endif 2137674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2138c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2139c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2140674ae819SStefano Zampini lwork = -1; 2141674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2142674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2143c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2144674ae819SStefano Zampini #else 2145c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2146674ae819SStefano Zampini #endif 2147674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2148984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2149674ae819SStefano Zampini #else /* on missing GESVD */ 2150674ae819SStefano Zampini /* SVD */ 2151674ae819SStefano Zampini PetscInt max_n,min_n; 2152674ae819SStefano Zampini max_n = max_size_of_constraint; 2153984c4197SStefano Zampini min_n = max_constraints; 2154984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2155674ae819SStefano Zampini min_n = max_size_of_constraint; 2156984c4197SStefano Zampini max_n = max_constraints; 2157674ae819SStefano Zampini } 2158785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2159674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2160785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2161674ae819SStefano Zampini #endif 2162674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2163674ae819SStefano Zampini lwork = -1; 2164e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2165e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2166b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2167674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2168674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21699162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr)); 2170674ae819SStefano Zampini #else 21719162d606SStefano 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)); 2172674ae819SStefano Zampini #endif 2173674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2174984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2175984c4197SStefano Zampini #endif /* on missing GESVD */ 2176674ae819SStefano Zampini /* Allocate optimal workspace */ 2177674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2178854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2179674ae819SStefano Zampini } 2180674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2181674ae819SStefano Zampini total_counts = 0; 21829162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21839162d606SStefano Zampini constraints_data_ptr[0] = 0; 2184674ae819SStefano Zampini /* vertices */ 21859162d606SStefano Zampini if (n_vertices) { 2186674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21879162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21889162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2189674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21909162d606SStefano Zampini constraints_n[total_counts] = 1; 21919162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21929162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21939162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2194674ae819SStefano Zampini total_counts++; 2195674ae819SStefano Zampini } 2196674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2197984c4197SStefano Zampini PetscBool used_vertex; 2198674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2199674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2200674ae819SStefano Zampini k = 0; 2201674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2202984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2203984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22049162d606SStefano Zampini constraints_n[total_counts] = 1; 22059162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22069162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22079162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22089162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2209674ae819SStefano Zampini total_counts++; 2210674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2211674ae819SStefano Zampini } 2212984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2213674ae819SStefano Zampini k++; 2214674ae819SStefano Zampini } 2215674ae819SStefano Zampini } 2216674ae819SStefano Zampini } 2217674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2218674ae819SStefano Zampini n_vertices = total_counts; 2219674ae819SStefano Zampini } 2220984c4197SStefano Zampini 2221674ae819SStefano Zampini /* edges and faces */ 22229162d606SStefano Zampini total_counts_cc = total_counts; 2223911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22249162d606SStefano Zampini IS used_is; 22259162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22269162d606SStefano Zampini 2227911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22289162d606SStefano Zampini used_is = ISForEdges[ncc]; 2229984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2230674ae819SStefano Zampini } else { 22319162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2232984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2233674ae819SStefano Zampini } 2234674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22359162d606SStefano Zampini 22369162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22379162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2238984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2239984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2240674ae819SStefano Zampini if (nnsp_has_cnst) { 22415b08dc53SStefano Zampini PetscScalar quad_value; 22429162d606SStefano Zampini 22439162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22449162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22459162d606SStefano Zampini 2246a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2247674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2248a773dcb8SStefano Zampini } else { 2249a773dcb8SStefano Zampini quad_value = 1.0; 2250a773dcb8SStefano Zampini } 2251674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22529162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2253674ae819SStefano Zampini } 22549162d606SStefano Zampini temp_constraints++; 2255674ae819SStefano Zampini total_counts++; 2256674ae819SStefano Zampini } 2257674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2258984c4197SStefano Zampini PetscReal real_value; 22599162d606SStefano Zampini PetscScalar *ptr_to_data; 22609162d606SStefano Zampini 2261984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22629162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2263674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22649162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2265674ae819SStefano Zampini } 2266984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2267984c4197SStefano Zampini /* check if array is null on the connected component */ 2268e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22699162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22705b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2271674ae819SStefano Zampini temp_constraints++; 2272674ae819SStefano Zampini total_counts++; 22739162d606SStefano Zampini if (!idxs_copied) { 22749162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22759162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2276674ae819SStefano Zampini } 2277674ae819SStefano Zampini } 22789162d606SStefano Zampini } 22799162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 228045a1bb75SStefano Zampini valid_constraints = temp_constraints; 2281eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2282a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22839162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22849162d606SStefano Zampini 22859162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2286a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22879162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2288a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22899162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2290a773dcb8SStefano Zampini } else { /* perform SVD */ 2291984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22929162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2293674ae819SStefano Zampini 2294674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2295984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2296984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2297984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2298984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2299984c4197SStefano Zampini from that computed using LAPACKgesvd 2300984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2301984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2302984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2303674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2304e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2305984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2306674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2307674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23089162d606SStefano 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)); 2309674ae819SStefano Zampini } 2310674ae819SStefano Zampini } 2311e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2312e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2313e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2314674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2315c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2316674ae819SStefano Zampini #else 2317c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2318674ae819SStefano Zampini #endif 2319674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2320984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2321984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2322674ae819SStefano Zampini j = 0; 2323984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2324674ae819SStefano Zampini total_counts = total_counts-j; 232545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2326e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2327c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2328c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2329c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2330c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2331c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2332c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2333674ae819SStefano Zampini if (j<temp_constraints) { 2334984c4197SStefano Zampini PetscInt ii; 2335984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2336674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23379162d606SStefano 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)); 2338674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2339984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2340674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23419162d606SStefano 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]; 2342674ae819SStefano Zampini } 2343674ae819SStefano Zampini } 2344674ae819SStefano Zampini } 2345674ae819SStefano Zampini #else /* on missing GESVD */ 2346e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2347e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2348b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2349674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2350674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23519162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr)); 2352674ae819SStefano Zampini #else 23539162d606SStefano 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)); 2354674ae819SStefano Zampini #endif 2355984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2356674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2357984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2358e310c8b4SStefano Zampini k = temp_constraints; 2359e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2360674ae819SStefano Zampini j = 0; 2361e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 236245a1bb75SStefano Zampini valid_constraints = k-j; 2363911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2364984c4197SStefano Zampini #endif /* on missing GESVD */ 2365674ae819SStefano Zampini } 2366a773dcb8SStefano Zampini } 23679162d606SStefano Zampini /* update pointers information */ 23689162d606SStefano Zampini if (valid_constraints) { 23699162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23709162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23719162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23729162d606SStefano Zampini /* set change_of_basis flag */ 237345a1bb75SStefano Zampini if (boolforchange) { 2374b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23759162d606SStefano Zampini } 2376b3d85658SStefano Zampini total_counts_cc++; 237745a1bb75SStefano Zampini } 237845a1bb75SStefano Zampini } 2379984c4197SStefano Zampini /* free workspace */ 23808f1c130eSStefano Zampini if (!skip_lapack) { 2381984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2382984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2383984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2384984c4197SStefano Zampini #endif 2385984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2386984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2387984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2388984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2389984c4197SStefano Zampini #endif 2390984c4197SStefano Zampini } 2391984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2392984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2393984c4197SStefano Zampini } 2394984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2395cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2396cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2397cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2398cf5a6209SStefano Zampini } 2399cf5a6209SStefano Zampini if (n_ISForFaces) { 2400cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2401cf5a6209SStefano Zampini } 2402cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2403cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2404cf5a6209SStefano Zampini } 2405cf5a6209SStefano Zampini if (n_ISForEdges) { 2406cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2407cf5a6209SStefano Zampini } 2408cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 240908122e43SStefano Zampini } else { 241008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2411984c4197SStefano Zampini 241208122e43SStefano Zampini total_counts = 0; 241308122e43SStefano Zampini n_vertices = 0; 2414d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2415d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 241608122e43SStefano Zampini } 241708122e43SStefano Zampini max_constraints = 0; 24189162d606SStefano Zampini total_counts_cc = 0; 241908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 242008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24219162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 242208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 242308122e43SStefano Zampini } 24249162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24259162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24269162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24279162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 242874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24299162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24309162d606SStefano Zampini total_counts_cc = 0; 24319162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24329162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24339162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 243408122e43SStefano Zampini } 243508122e43SStefano Zampini } 24369162d606SStefano Zampini #if 0 24379162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24389162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24399162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24409162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24419162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24429162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24439162d606SStefano Zampini } 24449162d606SStefano Zampini printf("\n"); 24459162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24469162d606SStefano Zampini } 24471b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24488bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24491b968477SStefano Zampini } 24501b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24518bec7fa6SStefano 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]); 24521b968477SStefano Zampini } 245308122e43SStefano Zampini #endif 245408122e43SStefano Zampini 24558bec7fa6SStefano Zampini max_size_of_constraint = 0; 24569162d606SStefano 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]); 24579162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 245808122e43SStefano Zampini /* Change of basis */ 2459b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 246008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 246108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 246208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2463b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 246408122e43SStefano Zampini } 246508122e43SStefano Zampini } 246608122e43SStefano Zampini } 246708122e43SStefano Zampini } 2468984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24690e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 247008122e43SStefano Zampini 24719162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24729162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24739162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24749162d606SStefano 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); 247508122e43SStefano Zampini } 2476674ae819SStefano Zampini 2477674ae819SStefano Zampini /* Create constraint matrix */ 2478674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 247916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2480984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2481984c4197SStefano Zampini 2482984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2483a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2484a717540cSStefano Zampini qr_needed = PETSC_FALSE; 248574d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2486984c4197SStefano Zampini total_primal_vertices=0; 2487b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24889162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24899162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2490984c4197SStefano Zampini if (size_of_constraint == 1) { 24919162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2492b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 249364efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24949162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24959162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2496a717540cSStefano Zampini } 2497b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 249874d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2499a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2500a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2501a717540cSStefano Zampini } 2502fa434743SStefano Zampini } else { 2503b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2504fa434743SStefano Zampini } 2505a717540cSStefano Zampini } 2506b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2507b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2508674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 250970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2510b3d85658SStefano Zampini 2511b3d85658SStefano 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); 25120e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25130e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2514984c4197SStefano Zampini 2515984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 251674d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2517785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2518984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 251974d5cdf7SStefano Zampini 2520984c4197SStefano Zampini j = total_primal_vertices; 252174d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2522b3d85658SStefano Zampini cum = total_primal_vertices; 25239162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25244641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2525b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2526b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2527b3d85658SStefano Zampini cum++; 25289162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 252974d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 253074d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 253174d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 253274d5cdf7SStefano Zampini } 25339162d606SStefano Zampini j += constraints_n[i]; 2534674ae819SStefano Zampini } 2535674ae819SStefano Zampini } 2536674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2537674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2538088faed8SStefano Zampini 2539674ae819SStefano Zampini /* set values in constraint matrix */ 2540984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25410e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2542674ae819SStefano Zampini } 2543984c4197SStefano Zampini total_counts = total_primal_vertices; 25449162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25454641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25469162d606SStefano Zampini PetscInt *cols; 25479162d606SStefano Zampini 25489162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25499162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25509162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25519162d606SStefano Zampini PetscInt row = total_counts+k; 25529162d606SStefano Zampini PetscScalar *vals; 25539162d606SStefano Zampini 25549162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25559162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25569162d606SStefano Zampini } 25579162d606SStefano Zampini total_counts += constraints_n[i]; 2558674ae819SStefano Zampini } 2559674ae819SStefano Zampini } 2560674ae819SStefano Zampini /* assembling */ 2561674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2562674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2563088faed8SStefano Zampini 2564984c4197SStefano Zampini /* 256545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2566984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2567984c4197SStefano Zampini */ 2568674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2569674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2570026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2571984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2572984c4197SStefano Zampini /* working stuff for GEQRF */ 257381d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2574984c4197SStefano Zampini PetscBLASInt lqr_work; 2575984c4197SStefano Zampini /* working stuff for UNGQR */ 2576984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2577984c4197SStefano Zampini PetscBLASInt lgqr_work; 2578984c4197SStefano Zampini /* working stuff for TRTRS */ 2579984c4197SStefano Zampini PetscScalar *trs_rhs; 25803f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2581984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2582984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2583984c4197SStefano Zampini PetscScalar *start_vals; 2584984c4197SStefano Zampini /* working stuff for values insertion */ 25854641a718SStefano Zampini PetscBT is_primal; 258664efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2587906d46d4SStefano Zampini /* matrix sizes */ 2588906d46d4SStefano Zampini PetscInt global_size,local_size; 2589906d46d4SStefano Zampini /* temporary change of basis */ 2590906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2591cf5a6209SStefano Zampini /* extra space for debugging */ 2592cf5a6209SStefano Zampini PetscScalar *dbg_work; 2593984c4197SStefano Zampini 2594906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2595906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 259616f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2597bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2598906d46d4SStefano Zampini /* nonzeros for local mat */ 2599bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2600bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 26019162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2602a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26039162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2604a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26059162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2606a717540cSStefano Zampini } else { 26079162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26089162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2609a717540cSStefano Zampini } 2610a717540cSStefano Zampini } 2611a717540cSStefano Zampini } 2612906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2613bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2614a717540cSStefano Zampini /* Set initial identity in the matrix */ 2615bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2616906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2617a717540cSStefano Zampini } 2618a717540cSStefano Zampini 2619a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2620a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2621a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2622a717540cSStefano Zampini } 2623a717540cSStefano Zampini 2624a717540cSStefano Zampini 2625a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2626a717540cSStefano Zampini /* 2627a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2628a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2629a717540cSStefano Zampini 2630a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2631a717540cSStefano Zampini 2632a6b551f4SStefano 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) 2633a6b551f4SStefano Zampini 2634a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2635a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2636a717540cSStefano Zampini | ... | 2637a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2638a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2639a717540cSStefano Zampini 2640a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2641a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2642a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2643a6b551f4SStefano Zampini 2644a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2645a717540cSStefano Zampini */ 2646a717540cSStefano Zampini if (qr_needed) { 2647984c4197SStefano Zampini /* space to store Q */ 2648854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2649984c4197SStefano Zampini /* first we issue queries for optimal work */ 26503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26523f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2653984c4197SStefano Zampini lqr_work = -1; 26543f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2655984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2656984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2657785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2658984c4197SStefano Zampini lgqr_work = -1; 26593f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26603f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26613f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26623f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26633f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26643f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2665984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2666984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2667785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2668984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2669785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2670984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2671785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2672a717540cSStefano Zampini /* allocating workspace for check */ 2673a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2674cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2675a717540cSStefano Zampini } 2676a717540cSStefano Zampini } 2677984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26784641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2679473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26800e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 268139e2fb2aSStefano Zampini if (i != total_primal_vertices) { 268239e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26834641a718SStefano Zampini } 268439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 268539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 268639e2fb2aSStefano Zampini } 268739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2688984c4197SStefano Zampini 2689a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26909162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26919162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26924641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2693984c4197SStefano Zampini /* get constraint info */ 26949162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2695984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2696984c4197SStefano Zampini 2697984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26989162d606SStefano 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); 2699674ae819SStefano Zampini } 2700984c4197SStefano Zampini 2701fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2702a717540cSStefano Zampini 2703a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2704a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27059162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2706a717540cSStefano Zampini } 2707984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27089162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2709984c4197SStefano Zampini 2710984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2714674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27153f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2716984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2717674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2718984c4197SStefano Zampini 2719984c4197SStefano Zampini /* explictly compute R^-T */ 2720984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2721984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27223f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2726984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2728984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2729984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2730984c4197SStefano Zampini 2731a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27333f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27343f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2736984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27373f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2738984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2739984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2740984c4197SStefano Zampini 2741984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2742984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2743984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27483f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27493f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2750984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27519162d606SStefano 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)); 2752984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27539162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2754984c4197SStefano Zampini 2755984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27569162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2757984c4197SStefano Zampini /* insert cols for primal dofs */ 2758984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2759984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27609162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2761906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2762984c4197SStefano Zampini } 2763984c4197SStefano Zampini /* insert cols for dual dofs */ 2764984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27659162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2766984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27679162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2768906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2769984c4197SStefano Zampini j++; 2770674ae819SStefano Zampini } 2771674ae819SStefano Zampini } 2772984c4197SStefano Zampini 2773984c4197SStefano Zampini /* check change of basis */ 2774984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2775984c4197SStefano Zampini PetscInt ii,jj; 2776984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2777c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2778c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2779c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2780c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2781c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2782c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2783984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2784cf5a6209SStefano 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)); 2785984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2786984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2787984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2788cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2789cf5a6209SStefano 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; 2790674ae819SStefano Zampini } 2791674ae819SStefano Zampini } 2792984c4197SStefano Zampini if (!valid_qr) { 279322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2794984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2795984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2796cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2797cf5a6209SStefano 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])); 2798674ae819SStefano Zampini } 2799cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2800cf5a6209SStefano 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])); 2801984c4197SStefano Zampini } 2802984c4197SStefano Zampini } 2803984c4197SStefano Zampini } 2804674ae819SStefano Zampini } else { 280522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2806674ae819SStefano Zampini } 2807674ae819SStefano Zampini } 2808a717540cSStefano Zampini } else { /* simple transformation block */ 2809a717540cSStefano Zampini PetscInt row,col; 2810a6b551f4SStefano Zampini PetscScalar val,norm; 2811a6b551f4SStefano Zampini 2812a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28139162d606SStefano 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)); 2814a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28159162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28169162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2817bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28189162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2819906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28209162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2821a717540cSStefano Zampini } else { 2822a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28239162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2824a717540cSStefano Zampini if (row != col) { 28259162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2826a717540cSStefano Zampini } else { 28279162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2828a717540cSStefano Zampini } 2829906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2830a717540cSStefano Zampini } 2831a717540cSStefano Zampini } 2832a717540cSStefano Zampini } 283398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 283422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2835a717540cSStefano Zampini } 2836674ae819SStefano Zampini } 2837984c4197SStefano Zampini } else { 2838984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28399162d606SStefano 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); 2840674ae819SStefano Zampini } 2841674ae819SStefano Zampini } 2842674ae819SStefano Zampini } 2843a717540cSStefano Zampini 2844a717540cSStefano Zampini /* free workspace */ 2845a717540cSStefano Zampini if (qr_needed) { 2846984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2847cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2848984c4197SStefano Zampini } 2849984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2850984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2851984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2852984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2853984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2854674ae819SStefano Zampini } 2855a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2856906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2857906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2858906d46d4SStefano Zampini 2859906d46d4SStefano Zampini /* assembling of global change of variable */ 2860bbb9e6c6SStefano Zampini { 2861bbb9e6c6SStefano Zampini Mat tmat; 286216f15bc4SStefano Zampini PetscInt bs; 286316f15bc4SStefano Zampini 2864906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2865906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2866bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2867bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2868bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 287016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 287116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2873bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2874bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2875bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2878e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2879e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2880bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2881bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2882906d46d4SStefano Zampini } 2883906d46d4SStefano Zampini /* check */ 2884906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2885906d46d4SStefano Zampini PetscReal error; 2886906d46d4SStefano Zampini Vec x,x_change; 2887906d46d4SStefano Zampini 2888906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2889906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2890906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2891906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2892e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2893e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2894bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2895e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2896e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2897906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2898906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2899906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2900906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2901bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2902906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2903906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2904906d46d4SStefano Zampini } 2905b96c3477SStefano Zampini 2906b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2907b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2908b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2909b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2910ac632422SStefano Zampini Mat S_new,tmat; 2911bbb9e6c6SStefano Zampini IS is_all_N; 2912bbb9e6c6SStefano Zampini 2913bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29146816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2915bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2916ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2917b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2918ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2919ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2920ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2921ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2922ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2923b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2924ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2925ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2926ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2927ac632422SStefano Zampini } 2928b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2929b96c3477SStefano Zampini } 2930b96c3477SStefano Zampini } 2931906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2932906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2933b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2934b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2935b9b85e73SStefano Zampini } 2936906d46d4SStefano Zampini 2937906d46d4SStefano Zampini /* set up change of basis context */ 2938906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2939906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2940906d46d4SStefano Zampini 2941906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2942906d46d4SStefano Zampini PetscInt global_size,local_size; 2943906d46d4SStefano Zampini 2944906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2945906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2946906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2951906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2953906d46d4SStefano Zampini } else { 2954906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2955906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2956906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2957906d46d4SStefano Zampini } 2958906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2959906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2960906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2961906d46d4SStefano Zampini } else { 2962906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2963906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2964906d46d4SStefano Zampini } 2965906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2966906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2967906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2968906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2969b9b85e73SStefano Zampini } 2970a717540cSStefano Zampini 2971727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2972727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2973727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2974aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2975c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29760e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2977aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2978727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2979727cdba6SStefano Zampini } 29800e6343abSStefano Zampini } 29810e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2982727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2983727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2984727cdba6SStefano Zampini 2985a717540cSStefano Zampini /* flush dbg viewer */ 2986b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2987b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2988b8ffe317SStefano Zampini } 2989a717540cSStefano Zampini 2990e310c8b4SStefano Zampini /* free workspace */ 2991a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29924641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 299308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29949162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29959162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 299608122e43SStefano Zampini } else { 29979162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29989162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29999162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 300008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 300108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 30029162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30039162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 300408122e43SStefano Zampini } 3005674ae819SStefano Zampini PetscFunctionReturn(0); 3006674ae819SStefano Zampini } 3007674ae819SStefano Zampini 3008674ae819SStefano Zampini #undef __FUNCT__ 3009674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3010674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3011674ae819SStefano Zampini { 3012674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3013674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3014674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30157fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3016674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3017674ae819SStefano Zampini 3018674ae819SStefano Zampini PetscFunctionBegin; 30198e61c736SStefano Zampini /* Reset previously computed graph */ 30208e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3021674ae819SStefano Zampini /* Init local Graph struct */ 30227fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30233bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3024674ae819SStefano Zampini 3025575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3026575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3027575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3028575ad6abSStefano Zampini } 30299577ea80SStefano Zampini 3030674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30314d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30324d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30334d379d7bSStefano Zampini PetscInt nvtxs; 3034e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3035674ae819SStefano Zampini 30364d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30372fffb893SStefano Zampini 30382fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30392fffb893SStefano Zampini if (flg_row) { 30404d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3041b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30422fffb893SStefano Zampini } 30432fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30449b28b941SStefano 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 */ 30454d379d7bSStefano Zampini IS is_dummy; 30464d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30474d379d7bSStefano Zampini PetscInt j,sum; 30484d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30494d379d7bSStefano Zampini const PetscInt *idxs; 30504d379d7bSStefano Zampini PCBDDCGraph graph; 30514d379d7bSStefano Zampini PetscBT is_on_boundary; 30524d379d7bSStefano Zampini 30534d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30544d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30554d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30564d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30577fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30584d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3059e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3060e496cd5dSStefano Zampini if (flg_row) { 30614d379d7bSStefano Zampini graph->xadj = xadj; 30624d379d7bSStefano Zampini graph->adjncy = adjncy; 3063e496cd5dSStefano Zampini } 30644d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30654d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3066e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30674d379d7bSStefano Zampini 30684d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30699b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30704d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30714d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30724d379d7bSStefano Zampini } 30734d379d7bSStefano Zampini } 30744d379d7bSStefano Zampini 3075e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30764d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30774d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30784d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30794d379d7bSStefano Zampini } 30804d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30814d379d7bSStefano Zampini 3082e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30834d379d7bSStefano Zampini sum = 0; 30844d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30854d379d7bSStefano Zampini PetscInt sizecc = 0; 30864d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30874d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30884d379d7bSStefano Zampini sizecc++; 30894d379d7bSStefano Zampini } 30904d379d7bSStefano Zampini } 30914d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30924d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30934d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30944d379d7bSStefano Zampini } 30954d379d7bSStefano Zampini } 30964d379d7bSStefano Zampini sum += sizecc*sizecc; 30974d379d7bSStefano Zampini } 30984d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30994d379d7bSStefano Zampini sum = 0; 3100e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 31014d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 31024d379d7bSStefano Zampini cxadj[i] = sum; 31034d379d7bSStefano Zampini sum += temp; 31044d379d7bSStefano Zampini } 3105e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31064d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31074d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31084d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31094d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31104d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31114d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31124d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31134d379d7bSStefano Zampini sizecc++; 31144d379d7bSStefano Zampini } 31154d379d7bSStefano Zampini } 31164d379d7bSStefano Zampini } 31174d379d7bSStefano Zampini } 31184d379d7bSStefano Zampini } 31199b28b941SStefano Zampini if (sum) { 3120e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31214d379d7bSStefano Zampini } else { 31224d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31234d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31244d379d7bSStefano Zampini } 31254d379d7bSStefano Zampini graph->xadj = 0; 31264d379d7bSStefano Zampini graph->adjncy = 0; 31274d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31284d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31294d379d7bSStefano Zampini } 3130674ae819SStefano Zampini } 31319b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31329b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3133674ae819SStefano Zampini } 3134674ae819SStefano Zampini 313563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3136674ae819SStefano Zampini vertex_size = 1; 313763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 313863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 313995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 314063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3141e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3143674ae819SStefano Zampini } 314463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 314563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 314663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3147674ae819SStefano Zampini } 314863602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3149674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 315063602bcaSStefano Zampini } else { 315163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 315263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3153854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 315463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 315563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 315663602bcaSStefano Zampini } 315763602bcaSStefano Zampini } 3158674ae819SStefano Zampini } 3159674ae819SStefano Zampini 3160674ae819SStefano Zampini /* Setup of Graph */ 3161785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3162e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3163785d1243SStefano Zampini } 3164785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3165e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3166785d1243SStefano Zampini } 3167302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3168674ae819SStefano Zampini 3169674ae819SStefano Zampini /* Graph's connected components analysis */ 3170674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3171674ae819SStefano Zampini 3172674ae819SStefano Zampini /* print some info to stdout */ 3173674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3174302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3175674ae819SStefano Zampini } 3176fb180af4SStefano Zampini 3177fb180af4SStefano Zampini /* mark topography has done */ 3178fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3179674ae819SStefano Zampini PetscFunctionReturn(0); 3180674ae819SStefano Zampini } 3181674ae819SStefano Zampini 3182dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3183674ae819SStefano Zampini #undef __FUNCT__ 3184674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3185dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3186674ae819SStefano Zampini { 3187dc456d91SStefano Zampini PetscSF sf; 3188dc456d91SStefano Zampini PetscLayout map; 3189dc456d91SStefano Zampini const PetscInt *idxs; 3190dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3191dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3192dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3193dc456d91SStefano Zampini PetscMPIInt commsize; 3194674ae819SStefano Zampini PetscBool first_found; 3195674ae819SStefano Zampini PetscErrorCode ierr; 3196674ae819SStefano Zampini 3197674ae819SStefano Zampini PetscFunctionBegin; 3198dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3199dc456d91SStefano Zampini if (subset_mult) { 3200dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3201dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3202dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3203674ae819SStefano Zampini } 3204dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3205dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3206dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3207dc456d91SStefano Zampini for (i=0;i<n;i++) { 3208dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3209dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3210674ae819SStefano Zampini } 3211dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3212dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3213dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3214dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3215dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3216dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3217dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3218dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3219dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3220dc456d91SStefano Zampini 3221dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3222dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3223dc456d91SStefano Zampini if (subset_mult) { 3224dc456d91SStefano Zampini const PetscInt* idxs_mult; 3225dc456d91SStefano Zampini 3226dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3227dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3228dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3229674ae819SStefano Zampini } else { 3230dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3231674ae819SStefano Zampini } 3232dc456d91SStefano Zampini /* local size of new subset */ 3233dc456d91SStefano Zampini n_n = 0; 3234dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3235dc456d91SStefano Zampini 3236dc456d91SStefano Zampini /* global indexes in layout */ 3237dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3238dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3239dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3240dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3241dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3242dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3243dc456d91SStefano Zampini 3244dc456d91SStefano Zampini /* reduce from leaves to roots */ 3245dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 324664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 324764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3248dc456d91SStefano Zampini 3249dc456d91SStefano Zampini /* count indexes in local part of layout */ 3250674ae819SStefano Zampini nlocals = 0; 3251674ae819SStefano Zampini first_index = -1; 3252674ae819SStefano Zampini first_found = PETSC_FALSE; 3253dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3254dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3255674ae819SStefano Zampini first_found = PETSC_TRUE; 3256674ae819SStefano Zampini first_index = i; 3257674ae819SStefano Zampini } 3258dc456d91SStefano Zampini nlocals += root_data[i]; 3259674ae819SStefano Zampini } 3260dc456d91SStefano Zampini 3261dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32625fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3263dc456d91SStefano Zampini start = 0; 326464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32655fa240b1SStefano Zampini #else 326664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32675fa240b1SStefano Zampini start = start-nlocals; 32685fa240b1SStefano Zampini #endif 32695fa240b1SStefano Zampini 3270dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3271dc456d91SStefano Zampini *N_n = start + nlocals; 3272dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3273dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3274674ae819SStefano Zampini } 32755fa240b1SStefano Zampini 32765fa240b1SStefano Zampini /* adapt root data with cumulative */ 3277674ae819SStefano Zampini if (first_found) { 3278dc456d91SStefano Zampini PetscInt old_index; 3279dc456d91SStefano Zampini 3280dc456d91SStefano Zampini root_data[first_index] += start; 3281674ae819SStefano Zampini old_index = first_index; 3282dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3283dc456d91SStefano Zampini if (root_data[i]) { 3284dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3285674ae819SStefano Zampini old_index = i; 3286674ae819SStefano Zampini } 3287674ae819SStefano Zampini } 3288674ae819SStefano Zampini } 3289dc456d91SStefano Zampini 3290dc456d91SStefano Zampini /* from roots to leaves */ 3291dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3292dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3293dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3294dc456d91SStefano Zampini 3295dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3296dc456d91SStefano Zampini if (subset_mult) { 3297dc456d91SStefano Zampini const PetscInt* idxs_mult; 3298dc456d91SStefano Zampini PetscInt cum; 3299dc456d91SStefano Zampini 3300dc456d91SStefano Zampini cum = 0; 3301dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3302dc456d91SStefano Zampini for (i=0;i<n;i++) { 3303dc456d91SStefano Zampini PetscInt j; 3304dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3305674ae819SStefano Zampini } 3306dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3307674ae819SStefano Zampini } else { 3308dc456d91SStefano Zampini for (i=0;i<n;i++) { 3309dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3310674ae819SStefano Zampini } 3311674ae819SStefano Zampini } 3312dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3313dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3314674ae819SStefano Zampini PetscFunctionReturn(0); 3315674ae819SStefano Zampini } 33169a7d3425SStefano Zampini 33179a7d3425SStefano Zampini #undef __FUNCT__ 33189a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33199a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33209a7d3425SStefano Zampini { 33219a7d3425SStefano Zampini PetscInt i,j; 33229a7d3425SStefano Zampini PetscScalar *alphas; 33239a7d3425SStefano Zampini PetscErrorCode ierr; 33249a7d3425SStefano Zampini 33259a7d3425SStefano Zampini PetscFunctionBegin; 33269a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3327785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33289a7d3425SStefano Zampini for (i=0;i<n;i++) { 33299a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33309a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33319a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33329a7d3425SStefano Zampini } 33339a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33349a7d3425SStefano Zampini PetscFunctionReturn(0); 33359a7d3425SStefano Zampini } 33369a7d3425SStefano Zampini 3337e7931f94SStefano Zampini #undef __FUNCT__ 333870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3339b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3340e7931f94SStefano Zampini { 334152e5ac9dSStefano Zampini IS ranks_send_to; 3342e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3343e7931f94SStefano Zampini PetscMPIInt size,rank,color; 334452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 334552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33463837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33472b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3348e7931f94SStefano Zampini PetscSubcomm subcomm; 334952e5ac9dSStefano Zampini PetscErrorCode ierr; 3350a57a6d2fSStefano Zampini 3351e7931f94SStefano Zampini PetscFunctionBegin; 33522b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33532b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33542b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3355e7931f94SStefano Zampini 3356e7931f94SStefano Zampini /* Get info on mapping */ 33573bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3359e7931f94SStefano Zampini 3360e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3361785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3362e7931f94SStefano Zampini xadj[0] = 0; 3363e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3364785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3365785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3366e7931f94SStefano Zampini 33672b510759SStefano Zampini if (threshold) { 3368d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33692b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3370d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3371d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3372d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3373d023bfaeSStefano Zampini xadj_count++; 3374e7931f94SStefano Zampini } 3375e7931f94SStefano Zampini } 3376d023bfaeSStefano Zampini xadj[1] = xadj_count; 3377c8587f34SStefano Zampini } else { 3378e7931f94SStefano Zampini if (xadj[1]) { 3379e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3380e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3381c8587f34SStefano Zampini } 3382e7931f94SStefano Zampini } 33833bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3384e7931f94SStefano Zampini if (use_square) { 3385e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3386e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3387e7931f94SStefano Zampini } 3388e7931f94SStefano Zampini } 3389e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3390e7931f94SStefano Zampini 33913837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3392e7931f94SStefano Zampini 3393e7931f94SStefano Zampini /* 3394e7931f94SStefano Zampini Restrict work on active processes only. 3395e7931f94SStefano Zampini */ 3396e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3397e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3398e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33992b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3400d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3401e7931f94SStefano Zampini if (color) { 3402e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3403e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3404e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3405c8587f34SStefano Zampini } else { 340652e5ac9dSStefano Zampini Mat subdomain_adj; 340752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 340852e5ac9dSStefano Zampini MatPartitioning partitioner; 340952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 341052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3411b0c7d250SStefano Zampini PetscBool aggregate; 3412b0c7d250SStefano Zampini 3413306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3414785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3415e7931f94SStefano Zampini prank = rank; 3416306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34178002ef2cSStefano Zampini /* 3418e7931f94SStefano Zampini for (i=0;i<size;i++) { 3419e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3420c8587f34SStefano Zampini } 34218002ef2cSStefano Zampini */ 3422e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3423e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3424c8587f34SStefano Zampini } 3425e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3426b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3427b0c7d250SStefano Zampini if (aggregate) { 3428b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3429b0c7d250SStefano Zampini PetscMPIInt nrank; 3430b0c7d250SStefano Zampini PetscScalar *vals; 3431b0c7d250SStefano Zampini 3432b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3433b0c7d250SStefano Zampini lrows = 0; 3434b0c7d250SStefano Zampini if (nrank<redprocs) { 3435b0c7d250SStefano Zampini lrows = size/redprocs; 3436b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3437b0c7d250SStefano Zampini } 34385fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3439b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3440b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3441b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3442b0c7d250SStefano Zampini row = nrank; 3443b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3444b0c7d250SStefano Zampini cols = adjncy; 3445b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3446b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3447b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3448b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3449b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 345052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 345152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 345252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3453b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3454b0c7d250SStefano Zampini } else { 3455306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3456b0c7d250SStefano Zampini } 345722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3458e7931f94SStefano Zampini 3459e7931f94SStefano Zampini /* Partition */ 3460306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3461e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3462e7931f94SStefano Zampini if (use_vwgt) { 34633837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3464e7931f94SStefano Zampini v_wgt[0] = local_size; 3465e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3466c8587f34SStefano Zampini } 346728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 346828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3469e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3470e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 347122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3472e7931f94SStefano Zampini 347352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 347452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 347552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 347652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3477b0c7d250SStefano Zampini if (!redprocs) { 3478b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 347928143c3dSStefano Zampini } else { 3480b0c7d250SStefano Zampini PetscInt idxs[1]; 3481b0c7d250SStefano Zampini PetscMPIInt tag; 3482b0c7d250SStefano Zampini MPI_Request *reqs; 3483b0c7d250SStefano Zampini 3484b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3485b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3486b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3487b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 348828143c3dSStefano Zampini } 3489b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3490b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3491b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3492b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3493e7931f94SStefano Zampini } 349452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3495e7931f94SStefano Zampini /* clean up */ 3496e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 349752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3498e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3499e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3500e7931f94SStefano Zampini } 3501e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3502e7931f94SStefano Zampini 3503e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3504e7931f94SStefano Zampini i = 1; 3505e7931f94SStefano Zampini if (color) i=0; 3506e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3507e7931f94SStefano Zampini /* get back IS */ 3508e7931f94SStefano Zampini *is_sends = ranks_send_to; 3509e7931f94SStefano Zampini PetscFunctionReturn(0); 3510e7931f94SStefano Zampini } 3511e7931f94SStefano Zampini 3512e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3513e7931f94SStefano Zampini 3514e7931f94SStefano Zampini #undef __FUNCT__ 3515e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 351653a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3517e7931f94SStefano Zampini { 351870cf5478SStefano Zampini Mat local_mat; 3519e7931f94SStefano Zampini IS is_sends_internal; 35209d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 352128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35229d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3523e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3524e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3525e7931f94SStefano Zampini const PetscInt* is_indices; 3526e7931f94SStefano Zampini MatType new_local_type; 3527e7931f94SStefano Zampini /* buffers */ 3528e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 352928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35309d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3531e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3532e7931f94SStefano Zampini /* MPI */ 353328143c3dSStefano Zampini MPI_Comm comm,comm_n; 353428143c3dSStefano Zampini PetscSubcomm subcomm; 3535e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 353628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 353728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 353828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 353928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 354028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3541e7931f94SStefano Zampini PetscErrorCode ierr; 3542e7931f94SStefano Zampini 3543e7931f94SStefano Zampini PetscFunctionBegin; 354428143c3dSStefano Zampini /* TODO: add missing checks */ 354528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 354628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 354728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 354828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3549e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 355028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3551e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3552e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3553e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3554e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3555e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 355628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 355770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 355870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 355928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 356070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 356170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 356270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 356370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 356470cf5478SStefano Zampini } 3565e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3566e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3567e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3568e7931f94SStefano Zampini if (!is_sends) { 356928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3570b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3571c8587f34SStefano Zampini } else { 3572e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3573e7931f94SStefano Zampini is_sends_internal = is_sends; 3574c8587f34SStefano Zampini } 3575e7931f94SStefano Zampini 3576e7931f94SStefano Zampini /* get comm */ 3577a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3578e7931f94SStefano Zampini 3579e7931f94SStefano Zampini /* compute number of sends */ 3580e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3581e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3582e7931f94SStefano Zampini 3583e7931f94SStefano Zampini /* compute number of receives */ 3584e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3585785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3586e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3587e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3588e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3589e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3590e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3591e7931f94SStefano Zampini 359228143c3dSStefano Zampini /* restrict comm if requested */ 359328143c3dSStefano Zampini subcomm = 0; 359428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 359528143c3dSStefano Zampini if (restrict_comm) { 3596779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3597779c1cceSStefano Zampini 359828143c3dSStefano Zampini color = 0; 359953a05cb3SStefano Zampini if (restrict_full) { 360053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 360153a05cb3SStefano Zampini } else { 360253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 360353a05cb3SStefano Zampini } 360428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 360528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 360628143c3dSStefano Zampini /* check if reuse has been requested */ 360728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 360828143c3dSStefano Zampini if (*mat_n) { 360928143c3dSStefano Zampini PetscMPIInt subcommsize2; 361028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 361128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 361228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 361328143c3dSStefano Zampini } else { 361428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 361528143c3dSStefano Zampini } 361628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3617779c1cceSStefano Zampini PetscMPIInt rank; 3618779c1cceSStefano Zampini 3619779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 362028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 362128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 362228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3623306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 362428143c3dSStefano Zampini } 362528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 362628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 362728143c3dSStefano Zampini } else { 362828143c3dSStefano Zampini comm_n = comm; 362928143c3dSStefano Zampini } 363028143c3dSStefano Zampini 3631e7931f94SStefano Zampini /* prepare send/receive buffers */ 3632785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3633e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3634785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3635e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 363628143c3dSStefano Zampini if (nis) { 3637854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 363828143c3dSStefano Zampini } 3639e7931f94SStefano Zampini 364028143c3dSStefano Zampini /* Get data from local matrices */ 3641e7931f94SStefano Zampini if (!isdense) { 3642a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3643e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3644e7931f94SStefano Zampini /* 3645e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3646e7931f94SStefano Zampini send_buffer_idxs should contain: 3647e7931f94SStefano Zampini - MatType_PRIVATE type 3648e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3649e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3650e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3651e7931f94SStefano Zampini */ 3652e7931f94SStefano Zampini } else { 3653e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36543bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3655854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3656e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3657e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3659e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36603bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3661e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3662e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3663e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3664e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3665c8587f34SStefano Zampini } 3666c8587f34SStefano Zampini } 3667e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 366828143c3dSStefano Zampini /* additional is (if any) */ 366928143c3dSStefano Zampini if (nis) { 367028143c3dSStefano Zampini PetscMPIInt psum; 367128143c3dSStefano Zampini PetscInt j; 367228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 367328143c3dSStefano Zampini PetscInt plen; 367428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 367528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 367628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 367728143c3dSStefano Zampini } 3678854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 367928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 368028143c3dSStefano Zampini PetscInt plen; 368128143c3dSStefano Zampini const PetscInt *is_array_idxs; 368228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 368328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 368428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 368528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 368628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 368728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 368828143c3dSStefano Zampini } 368928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 369028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 369128143c3dSStefano Zampini } 369228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 369328143c3dSStefano Zampini } 369428143c3dSStefano Zampini 3695e7931f94SStefano Zampini buf_size_idxs = 0; 3696e7931f94SStefano Zampini buf_size_vals = 0; 369728143c3dSStefano Zampini buf_size_idxs_is = 0; 3698e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3699e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3700e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 370128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3702e7931f94SStefano Zampini } 3703785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3704785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 370595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3706e7931f94SStefano Zampini 3707e7931f94SStefano Zampini /* get new tags for clean communications */ 3708e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3709e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 371028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3711e7931f94SStefano Zampini 3712e7931f94SStefano Zampini /* allocate for requests */ 3713785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3714785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 371595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3716785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3717785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 371895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3719e7931f94SStefano Zampini 3720e7931f94SStefano Zampini /* communications */ 3721e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3722e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 372328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3724e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3725e7931f94SStefano Zampini source_dest = onodes[i]; 3726e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3727e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3728e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3729e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 373028143c3dSStefano Zampini if (nis) { 373128143c3dSStefano 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); 373228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 373328143c3dSStefano Zampini } 3734e7931f94SStefano Zampini } 3735e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3736e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3737e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3738e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 373928143c3dSStefano Zampini if (nis) { 374028143c3dSStefano 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); 374128143c3dSStefano Zampini } 3742e7931f94SStefano Zampini } 3743e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3744e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3745e7931f94SStefano Zampini 3746e7931f94SStefano Zampini /* assemble new l2g map */ 3747e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3748e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37499d30be91SStefano Zampini new_local_rows = 0; 3750e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37519d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3752e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3753e7931f94SStefano Zampini } 37549d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3755e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37569d30be91SStefano Zampini new_local_rows = 0; 3757e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37589d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37599d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3760e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3761e7931f94SStefano Zampini } 37629d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37639d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3764e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3765e7931f94SStefano Zampini 3766e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3767e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3768e7931f94SStefano 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) */ 3769e7931f94SStefano Zampini if (n_recvs) { 377028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3771e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3772e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3773e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3774e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3775e7931f94SStefano Zampini break; 3776e7931f94SStefano Zampini } 3777e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3778e7931f94SStefano Zampini } 3779e7931f94SStefano Zampini switch (new_local_type_private) { 378028143c3dSStefano Zampini case MATDENSE_PRIVATE: 378128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3782e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3783e7931f94SStefano Zampini bs = 1; 378428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 378528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 378628143c3dSStefano Zampini bs = 1; 378728143c3dSStefano Zampini } 3788e7931f94SStefano Zampini break; 3789e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3790e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3791e7931f94SStefano Zampini bs = 1; 3792e7931f94SStefano Zampini break; 3793e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3794e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3795e7931f94SStefano Zampini break; 3796e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3797e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3798e7931f94SStefano Zampini break; 3799e7931f94SStefano Zampini default: 38009d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3801e7931f94SStefano Zampini break; 3802e7931f94SStefano Zampini } 380328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 380428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 380528143c3dSStefano Zampini bs = 1; 3806e7931f94SStefano Zampini } 3807e7931f94SStefano Zampini 380870cf5478SStefano Zampini /* create MATIS object if needed */ 380970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3810e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3811e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 381270cf5478SStefano Zampini } else { 381370cf5478SStefano Zampini /* it also destroys the local matrices */ 381470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 381570cf5478SStefano Zampini } 381670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3817e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 38189d30be91SStefano Zampini 38199d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38209d30be91SStefano Zampini 38219d30be91SStefano Zampini /* Global to local map of received indices */ 38229d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38239d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38249d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38259d30be91SStefano Zampini 38269d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38279d30be91SStefano Zampini buf_size_idxs = 0; 38289d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38299d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38309d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38319d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38329d30be91SStefano Zampini } 38339d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38349d30be91SStefano Zampini 38359d30be91SStefano Zampini /* set preallocation */ 38369d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38379d30be91SStefano Zampini if (!newisdense) { 38389d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38399d30be91SStefano Zampini 38409d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38419d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38429d30be91SStefano Zampini if (n_recvs) { 38439d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38449d30be91SStefano Zampini } 38459d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38469d30be91SStefano Zampini PetscInt j; 38479d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38489d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38499d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38509d30be91SStefano Zampini } 38519d30be91SStefano Zampini } else { 38529d30be91SStefano Zampini /* TODO */ 38539d30be91SStefano Zampini } 38549d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38559d30be91SStefano Zampini } 38569d30be91SStefano Zampini if (new_local_nnz) { 38579d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38589d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38599d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38609d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38619d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38629d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38639d30be91SStefano Zampini } else { 38649d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38659d30be91SStefano Zampini } 38669d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38679d30be91SStefano Zampini } else { 38689d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38699d30be91SStefano Zampini } 3870e7931f94SStefano Zampini 3871e7931f94SStefano Zampini /* set values */ 3872e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38739d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3874e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3875e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3876e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38779d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3878e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3879e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3880e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 388128143c3dSStefano Zampini } else { 388228143c3dSStefano Zampini /* TODO */ 3883e7931f94SStefano Zampini } 3884e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3885e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3886e7931f94SStefano Zampini } 3887e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3888e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 388970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38919d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38929d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3893e7931f94SStefano Zampini 3894dfd14d43SStefano Zampini #if 0 389528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3896e7931f94SStefano Zampini Vec lvec,rvec; 3897e7931f94SStefano Zampini PetscReal infty_error; 3898e7931f94SStefano Zampini 38992a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3900e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3901e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3902e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 390370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3904e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3906e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3907e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3908e7931f94SStefano Zampini } 390928143c3dSStefano Zampini #endif 3910e7931f94SStefano Zampini 391128143c3dSStefano Zampini /* assemble new additional is (if any) */ 391228143c3dSStefano Zampini if (nis) { 391328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 391428143c3dSStefano Zampini 391528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3916854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 391728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 391828143c3dSStefano Zampini psum = 0; 391928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 392028143c3dSStefano Zampini for (j=0;j<nis;j++) { 392128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 392228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 392328143c3dSStefano Zampini psum += plen; 392428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 392528143c3dSStefano Zampini } 392628143c3dSStefano Zampini } 3927854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3928854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 392928143c3dSStefano Zampini for (i=1;i<nis;i++) { 393028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 393128143c3dSStefano Zampini } 393228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 393328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 393428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 393528143c3dSStefano Zampini for (j=0;j<nis;j++) { 393628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 393728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 393828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 393928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 394028143c3dSStefano Zampini } 394128143c3dSStefano Zampini } 394228143c3dSStefano Zampini for (i=0;i<nis;i++) { 394328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 394428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 394528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 394628143c3dSStefano Zampini } 394728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 394828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 394928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 395028143c3dSStefano Zampini } 3951e7931f94SStefano Zampini /* free workspace */ 395228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3953e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3954e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3955e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3956e7931f94SStefano Zampini if (isdense) { 3957e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3958e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3959e7931f94SStefano Zampini } else { 3960e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3961e7931f94SStefano Zampini } 396228143c3dSStefano Zampini if (nis) { 396328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 396428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 396528143c3dSStefano Zampini } 3966e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3967e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 396828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3969e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3970e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 397128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3972e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3973e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3974e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3975e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3976e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 397728143c3dSStefano Zampini if (nis) { 397828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 397928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 398028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 398128143c3dSStefano Zampini } 398228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 398328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 398428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 398528143c3dSStefano Zampini for (i=0;i<nis;i++) { 398628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 398728143c3dSStefano Zampini } 398853a05cb3SStefano Zampini *mat_n = NULL; 398928143c3dSStefano Zampini } 3990e7931f94SStefano Zampini PetscFunctionReturn(0); 3991e7931f94SStefano Zampini } 3992a57a6d2fSStefano Zampini 399312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3994af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 399512edc857SStefano Zampini 3996c8587f34SStefano Zampini #undef __FUNCT__ 3997c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3998c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3999c8587f34SStefano Zampini { 4000c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4001c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 400220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 40039881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 400420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 40056e683305SStefano Zampini IS coarse_is,*isarray; 40066e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 40076e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4008f9eb5b7dSStefano Zampini PC pc_temp; 4009c8587f34SStefano Zampini PCType coarse_pc_type; 4010c8587f34SStefano Zampini KSPType coarse_ksp_type; 4011f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 40124f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 40136e683305SStefano Zampini Mat t_coarse_mat_is; 40146e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 40156e683305SStefano Zampini PetscMPIInt all_procs; 401674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 401768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 401822bc73bbSStefano Zampini PetscScalar *array; 40199881197aSStefano Zampini PetscErrorCode ierr; 4020fdc09c96SStefano Zampini 4021c8587f34SStefano Zampini PetscFunctionBegin; 4022c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 402368457ee5SStefano 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 */ 4024fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40255a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4026fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4027f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4028f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4029f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4030fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 403151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 403251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4033dc4bcba2SStefano Zampini PC pc; 4034dc4bcba2SStefano Zampini PetscBool isbddc; 4035dc4bcba2SStefano Zampini 4036dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4037dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4038dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4039dc4bcba2SStefano Zampini if (isbddc) { 4040dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4041dc4bcba2SStefano Zampini } 4042727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4043fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4044fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4045fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4046f4ddd8eeSStefano Zampini } 4047fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4048fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4049f4ddd8eeSStefano Zampini } 405070cf5478SStefano Zampini /* reset any subassembling information */ 405170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40526e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40536e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4054fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4055f4ddd8eeSStefano Zampini } 4056c8587f34SStefano Zampini 40576e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40582b510759SStefano Zampini im_active = !!(pcis->n); 40592b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40606e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40616e683305SStefano Zampini void_procs = all_procs-active_procs; 40626e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 406374e2c79eSStefano Zampini redist = PETSC_FALSE; 406422bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40656e683305SStefano Zampini csin_ml = PETSC_TRUE; 40666e683305SStefano Zampini ncoarse_ml = void_procs; 4067779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4068779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40696e683305SStefano Zampini csin_ds = PETSC_TRUE; 407018a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 407118a45a71SStefano Zampini redist = PETSC_TRUE; 407218a45a71SStefano Zampini } else { 40736e683305SStefano Zampini csin_ds = PETSC_TRUE; 4074779c1cceSStefano Zampini ncoarse_ds = active_procs; 4075779c1cceSStefano Zampini redist = PETSC_TRUE; 407618a45a71SStefano Zampini } 40776e683305SStefano Zampini } else { 40786e683305SStefano Zampini csin_ml = PETSC_FALSE; 40796e683305SStefano Zampini ncoarse_ml = all_procs; 40806e683305SStefano Zampini if (void_procs) { 40816e683305SStefano Zampini csin_ds = PETSC_TRUE; 40826e683305SStefano Zampini ncoarse_ds = void_procs; 40836e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40846e683305SStefano Zampini } else { 4085779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 408674e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 408774e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 408874e2c79eSStefano Zampini redist = PETSC_TRUE; 408974e2c79eSStefano Zampini } else { 40906e683305SStefano Zampini csin_ds = PETSC_FALSE; 40916e683305SStefano Zampini ncoarse_ds = all_procs; 40926e683305SStefano Zampini } 40936e683305SStefano Zampini } 409474e2c79eSStefano Zampini } 40956e683305SStefano Zampini 40966e683305SStefano Zampini /* 40976e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40986e683305SStefano Zampini - we have not exceeded the number of levels requested 40996e683305SStefano Zampini - we can actually subassemble the active processes 41006e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 41016e683305SStefano Zampini */ 41026e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 41036e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 41046e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 41056e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 41066e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4107f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 41082b510759SStefano Zampini } else { 4109f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4110c8587f34SStefano Zampini } 4111c8587f34SStefano Zampini } 41126e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 41136e683305SStefano Zampini if (multilevel_allowed) { 41146e683305SStefano Zampini ncoarse = ncoarse_ml; 41156e683305SStefano Zampini csin = csin_ml; 411658da7f69SStefano Zampini redist = PETSC_FALSE; 41176e683305SStefano Zampini } else { 41186e683305SStefano Zampini ncoarse = ncoarse_ds; 41196e683305SStefano Zampini csin = csin_ds; 41206e683305SStefano Zampini } 4121e7931f94SStefano Zampini 4122abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4123abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4124abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4125abbbba34SStefano Zampini 4126abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 412722bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 412822bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 412922bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 413022bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4131b9b85e73SStefano Zampini #if 0 4132b9b85e73SStefano Zampini { 4133b9b85e73SStefano Zampini PetscViewer viewer; 4134b9b85e73SStefano Zampini char filename[256]; 4135b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4136b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4137b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4138b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4139b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4140b9b85e73SStefano Zampini } 4141b9b85e73SStefano Zampini #endif 4142e176bc59SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,NULL,&t_coarse_mat_is);CHKERRQ(ierr); 41436e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41446e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41456e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4146abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4147abbbba34SStefano Zampini 41486e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41496e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41506e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41516e683305SStefano Zampini const PetscInt *idxs; 41526e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41536e683305SStefano Zampini 41546e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41550be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41566e683305SStefano Zampini /* allocate space for temporary storage */ 4157854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4158854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41596e683305SStefano Zampini /* allocate for IS array */ 41606e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41616e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41626e683305SStefano Zampini nis = nisdofs + nisneu; 4163854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41646e683305SStefano Zampini /* dofs splitting */ 41656e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41666e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41676e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41686e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41696e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41706e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41716e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41726e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41736e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41746e683305SStefano Zampini } 41756e683305SStefano Zampini /* neumann boundaries */ 41766e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41776e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41786e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41796e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41806e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41816e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41836e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41846e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41856e683305SStefano Zampini } 41866e683305SStefano Zampini /* free memory */ 41876e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41886e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41896e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41906e683305SStefano Zampini } else { 41916e683305SStefano Zampini nis = 0; 41926e683305SStefano Zampini nisdofs = 0; 41936e683305SStefano Zampini nisneu = 0; 41946e683305SStefano Zampini isarray = NULL; 41956e683305SStefano Zampini } 41966e683305SStefano Zampini /* destroy no longer needed map */ 41976e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41986e683305SStefano Zampini 41996e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 42006e683305SStefano Zampini coarse_mat_is = NULL; 42016e683305SStefano Zampini if (csin) { 42026e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 420374e2c79eSStefano Zampini if (redist) { 420474e2c79eSStefano Zampini PetscMPIInt rank; 4205779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 420674e2c79eSStefano Zampini 420774e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 420858da7f69SStefano Zampini spc = active_procs/ncoarse; 420958da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4210779c1cceSStefano Zampini if (im_active) { 4211779c1cceSStefano Zampini destsize = 1; 421274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 421374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 421474e2c79eSStefano Zampini } else { 421574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 421674e2c79eSStefano Zampini } 421774e2c79eSStefano Zampini } else { 4218779c1cceSStefano Zampini destsize = 0; 42196e683305SStefano Zampini } 4220779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4221779c1cceSStefano Zampini } else if (csin_type_simple) { 42226e683305SStefano Zampini PetscMPIInt rank; 42236e683305SStefano Zampini PetscInt issize,isidx; 4224779c1cceSStefano Zampini 42256e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42266e683305SStefano Zampini if (im_active) { 42276e683305SStefano Zampini issize = 1; 42286e683305SStefano Zampini isidx = (PetscInt)rank; 42296e683305SStefano Zampini } else { 42306e683305SStefano Zampini issize = 0; 42316e683305SStefano Zampini isidx = -1; 42326e683305SStefano Zampini } 42336e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4234779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4235b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42366e683305SStefano Zampini } 4237779c1cceSStefano Zampini 4238779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4239779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4240779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4241779c1cceSStefano Zampini PetscInt *coarse_candidates; 4242779c1cceSStefano Zampini const PetscInt* tisindices; 4243779c1cceSStefano Zampini 4244779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4245779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4246779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4247779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4248779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4249779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4250779c1cceSStefano Zampini } 4251779c1cceSStefano Zampini } 4252779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4253779c1cceSStefano Zampini 4254779c1cceSStefano Zampini 42556e683305SStefano Zampini if (pcbddc->dbg_flag) { 42566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42586e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4260779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42626e683305SStefano Zampini } 42636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42646e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42656e683305SStefano Zampini } 42666e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42676e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42686e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4269854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42706e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42716e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42726e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42736e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42746e683305SStefano Zampini } 42756e683305SStefano Zampini if (pcbddc->dbg_flag) { 42766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42786e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42796e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42806e683305SStefano Zampini } 4281779c1cceSStefano Zampini } 42826e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 428353a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 428453a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 428553a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 428653a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 428753a05cb3SStefano Zampini } 42886e683305SStefano Zampini } else { 42896e683305SStefano Zampini if (pcbddc->dbg_flag) { 42906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42926e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42936e683305SStefano Zampini } 42946e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42956e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42966e683305SStefano Zampini } 42976e683305SStefano Zampini 42986e683305SStefano Zampini /* create local to global scatters for coarse problem */ 429968457ee5SStefano Zampini if (compute_vecs) { 43006e683305SStefano Zampini PetscInt lrows; 43016e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43026e683305SStefano Zampini if (coarse_mat_is) { 43036e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 43046e683305SStefano Zampini } else { 43056e683305SStefano Zampini lrows = 0; 43066e683305SStefano Zampini } 43076e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 43086e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 43096e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 43106e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43116e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43126e683305SStefano Zampini } 43136e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 43146e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4315c8587f34SStefano Zampini 4316f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4317f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4318f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4319f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4320f9eb5b7dSStefano Zampini } else { 4321f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4322f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4323c8587f34SStefano Zampini } 4324c8587f34SStefano Zampini 43256e683305SStefano Zampini /* print some info if requested */ 43266e683305SStefano Zampini if (pcbddc->dbg_flag) { 43276e683305SStefano Zampini if (!multilevel_allowed) { 43286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43296e683305SStefano Zampini if (multilevel_requested) { 43306e683305SStefano 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); 43316e683305SStefano Zampini } else if (pcbddc->max_levels) { 43326e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43336e683305SStefano Zampini } 43346e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43356e683305SStefano Zampini } 43366e683305SStefano Zampini } 43376e683305SStefano Zampini 4338f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43396e683305SStefano Zampini if (coarse_mat_is) { 43406e683305SStefano Zampini MatReuse coarse_mat_reuse; 43416a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43426e683305SStefano Zampini if (pcbddc->dbg_flag) { 43436e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43446e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43456e683305SStefano Zampini } 4346f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4347312be037SStefano Zampini char prefix[256],str_level[16]; 4348e604994aSStefano Zampini size_t len; 43496e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4350422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4351c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4352f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43535f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4354c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43556e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4356c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4357c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4358e604994aSStefano Zampini /* prefix */ 4359e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4360e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4361e604994aSStefano Zampini if (!pcbddc->current_level) { 4362e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4363e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4364c8587f34SStefano Zampini } else { 4365e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4366312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4367312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 436834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4369312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4370e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4371e604994aSStefano Zampini } 4372e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43733e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43743e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43753e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43763e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4377f9eb5b7dSStefano Zampini /* allow user customization */ 4378f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43793e3c6dadSStefano Zampini } 43803e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 438151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43823e3c6dadSStefano Zampini if (nisdofs) { 43833e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43843e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43853e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43863e3c6dadSStefano Zampini } 43873e3c6dadSStefano Zampini } 43883e3c6dadSStefano Zampini if (nisneu) { 43893e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43903e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4391312be037SStefano Zampini } 4392f9eb5b7dSStefano Zampini 4393f9eb5b7dSStefano Zampini /* get some info after set from options */ 4394f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4395f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43964f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43976e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4398f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4399f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4400f9eb5b7dSStefano Zampini } 440139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 44024f3a063dSStefano Zampini if (isredundant) { 44034f3a063dSStefano Zampini KSP inner_ksp; 44044f3a063dSStefano Zampini PC inner_pc; 44054f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 44064f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 44074f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 44084f3a063dSStefano Zampini } 4409f9eb5b7dSStefano Zampini 4410f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4411fa7f1dd8SStefano Zampini if (coarse_reuse) { 441281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4413fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 44146e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4415fa7f1dd8SStefano Zampini } else { 44166e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4417fa7f1dd8SStefano Zampini } 4418c8587f34SStefano Zampini if (isbddc || isnn) { 441922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 442070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4421b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 442222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44256e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44266e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 442722b6e8a2SStefano Zampini } 442870cf5478SStefano Zampini } 442953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 443070cf5478SStefano Zampini } else { 443122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 443222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 443322bc73bbSStefano Zampini } 443422bc73bbSStefano Zampini } else { 44352e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4436c8587f34SStefano Zampini } 4437c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4438c8587f34SStefano Zampini 44393301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44405a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44413301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44423301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44433301b35fSStefano Zampini } 44443301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44453301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44463301b35fSStefano Zampini } 44473301b35fSStefano Zampini if (pc->pmat->spd_set) { 44483301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44493301b35fSStefano Zampini } 44506e683305SStefano Zampini /* set operators */ 44515f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44526e683305SStefano Zampini if (pcbddc->dbg_flag) { 44536e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44546e683305SStefano Zampini } 44556e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44566e683305SStefano Zampini coarse_mat = 0; 44576e683305SStefano Zampini } 44586e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4459b9b85e73SStefano Zampini #if 0 4460b9b85e73SStefano Zampini { 4461b9b85e73SStefano Zampini PetscViewer viewer; 4462b9b85e73SStefano Zampini char filename[256]; 4463b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4464b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4465b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4466b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4467b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4468b9b85e73SStefano Zampini } 4469b9b85e73SStefano Zampini #endif 4470c8587f34SStefano Zampini 4471c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4472298c0119SStefano Zampini #if 0 4473c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4474c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 447598a51de6SStefano Zampini } 4476298c0119SStefano Zampini #endif 447798a51de6SStefano Zampini 447898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 447998a51de6SStefano Zampini Vec crhs,csol; 448098a51de6SStefano Zampini PetscBool ispreonly; 448104708bb6SStefano Zampini 448298a51de6SStefano Zampini if (CoarseNullSpace) { 4483c8587f34SStefano Zampini if (isbddc) { 4484c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4485c8587f34SStefano Zampini } else { 44865fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4487c8587f34SStefano Zampini } 4488c8587f34SStefano Zampini } 4489f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4490f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4491f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4492f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44936e683305SStefano Zampini /* hack */ 4494f347579bSStefano Zampini if (!csol) { 44952a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4496f9eb5b7dSStefano Zampini } 4497f347579bSStefano Zampini if (!crhs) { 44982a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4499f347579bSStefano Zampini } 4500cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4501cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 45026e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4503c8587f34SStefano Zampini KSP check_ksp; 45042b510759SStefano Zampini KSPType check_ksp_type; 4505c8587f34SStefano Zampini PC check_pc; 45066e683305SStefano Zampini Vec check_vec,coarse_vec; 45076a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 45082b510759SStefano Zampini PetscInt its; 45096e683305SStefano Zampini PetscBool compute_eigs; 45106e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 45116e683305SStefano Zampini PetscInt neigs; 45128e185a42SStefano Zampini const char *prefix; 4513c8587f34SStefano Zampini 45142b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 45156e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4516422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 451723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4518f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 45192b510759SStefano Zampini if (ispreonly) { 45202b510759SStefano Zampini check_ksp_type = KSPPREONLY; 45216e683305SStefano Zampini compute_eigs = PETSC_FALSE; 45222b510759SStefano Zampini } else { 4523cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45246e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4525c8587f34SStefano Zampini } 4526c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45276e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45286e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45296e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4530a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4531a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4532a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4533a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4534c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4535c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4536c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4537c8587f34SStefano Zampini /* create random vec */ 45386e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45396e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4540c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4541c8587f34SStefano Zampini if (CoarseNullSpace) { 4542c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4543c8587f34SStefano Zampini } 45446e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4545c8587f34SStefano Zampini /* solve coarse problem */ 45466e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4547c8587f34SStefano Zampini if (CoarseNullSpace) { 45486e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4549c8587f34SStefano Zampini } 4550cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45516e683305SStefano Zampini if (compute_eigs) { 4552854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4553854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45546e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45556e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45566e683305SStefano Zampini lambda_min = eigs_r[0]; 45576e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45586e683305SStefano Zampini if (lambda_max>lambda_min) { 4559cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4560cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4561cbcc2c2aSStefano Zampini } 4562c8587f34SStefano Zampini } 4563c8587f34SStefano Zampini } 4564cbcc2c2aSStefano Zampini 4565c8587f34SStefano Zampini /* check coarse problem residual error */ 45666e683305SStefano Zampini if (pcbddc->dbg_flag) { 45676e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45686e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45696e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4570c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45716e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45726e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4573c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4574779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45756e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45766e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45796e683305SStefano Zampini if (compute_eigs) { 45806e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4581deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4582c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45836e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45846e683305SStefano 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); 45856e683305SStefano Zampini for (i=0;i<neigs;i++) { 45866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4587c8587f34SStefano Zampini } 45886e683305SStefano Zampini } 45896e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45906e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45916e683305SStefano Zampini } 4592c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45936e683305SStefano Zampini if (compute_eigs) { 45946e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45956e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4596c8587f34SStefano Zampini } 45976e683305SStefano Zampini } 45986e683305SStefano Zampini } 4599cbcc2c2aSStefano Zampini /* print additional info */ 4600cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 46016e683305SStefano Zampini /* waits until all processes reaches this point */ 46026e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4603cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4604cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4605cbcc2c2aSStefano Zampini } 4606cbcc2c2aSStefano Zampini 46072b510759SStefano Zampini /* free memory */ 4608c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4609fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4610c8587f34SStefano Zampini PetscFunctionReturn(0); 4611c8587f34SStefano Zampini } 4612674ae819SStefano Zampini 4613f34684f1SStefano Zampini #undef __FUNCT__ 4614f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4615f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4616f34684f1SStefano Zampini { 4617f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4618f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4619f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4620dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4621dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 462273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4623dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4624f34684f1SStefano Zampini PetscErrorCode ierr; 4625f34684f1SStefano Zampini 4626f34684f1SStefano Zampini PetscFunctionBegin; 4627f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46280e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 46290e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4630727cdba6SStefano Zampini } 4631dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4633dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4634dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4635dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4636dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4637dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4638dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46390e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 46400e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 46410e6343abSStefano Zampini } 4642dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4643dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4644dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4645dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4646dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4647f34684f1SStefano Zampini 4648f34684f1SStefano Zampini /* check numbering */ 4649f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4650f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4651dc456d91SStefano Zampini PetscInt i; 4652b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4653f34684f1SStefano Zampini 4654f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4655f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4656f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46571575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4658f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4659f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4660727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4661f34684f1SStefano Zampini } 4662f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4663f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4664f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4665e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4666e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4667e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4668e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4669f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4670f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4671f34684f1SStefano Zampini if (array[i] == 1.0) { 4672b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4673f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4674f34684f1SStefano Zampini } 4675f34684f1SStefano Zampini } 4676b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4677f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4678f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4679f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4680f34684f1SStefano Zampini } 4681f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4682f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4683e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4684e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4685f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4686f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4687b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4688ca8b9ea9SStefano Zampini PetscInt *gidxs; 4689ca8b9ea9SStefano Zampini 4690ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46913bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4692f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4693f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4694f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4695f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46964bc2dc4bSStefano 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); 4697f34684f1SStefano Zampini } 4698f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4699ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4700f34684f1SStefano Zampini } 4701f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 47021575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4703302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4704f34684f1SStefano Zampini } 47058bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4706f34684f1SStefano Zampini /* get back data */ 4707f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4708f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4709674ae819SStefano Zampini PetscFunctionReturn(0); 4710674ae819SStefano Zampini } 4711674ae819SStefano Zampini 4712e456f2a8SStefano Zampini #undef __FUNCT__ 4713e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4714a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4715e456f2a8SStefano Zampini { 4716e456f2a8SStefano Zampini IS localis_t; 4717a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4718e456f2a8SStefano Zampini PetscScalar *vals; 4719e456f2a8SStefano Zampini PetscErrorCode ierr; 4720e456f2a8SStefano Zampini 4721e456f2a8SStefano Zampini PetscFunctionBegin; 4722a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4723e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4724854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4725e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4726e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4727a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4728a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47291035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4730a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47311035eff8SStefano Zampini } 4732a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4733e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4734e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4735a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4736a7dc3881SStefano Zampini /* now compute set in local ordering */ 4737a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4738a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4739a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4740a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4741a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4742ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4743e456f2a8SStefano Zampini lsize++; 4744e456f2a8SStefano Zampini } 4745e456f2a8SStefano Zampini } 4746854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4747a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4748ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4749e456f2a8SStefano Zampini idxs[lsize++] = i; 4750e456f2a8SStefano Zampini } 4751e456f2a8SStefano Zampini } 4752a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4753a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4754e456f2a8SStefano Zampini *localis = localis_t; 4755e456f2a8SStefano Zampini PetscFunctionReturn(0); 4756e456f2a8SStefano Zampini } 4757906d46d4SStefano Zampini 4758906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4759906d46d4SStefano Zampini #undef __FUNCT__ 4760906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4761906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4762906d46d4SStefano Zampini { 4763906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4764906d46d4SStefano Zampini PetscErrorCode ierr; 4765906d46d4SStefano Zampini 4766906d46d4SStefano Zampini PetscFunctionBegin; 4767906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4768906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4769906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4770906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4771906d46d4SStefano Zampini PetscFunctionReturn(0); 4772906d46d4SStefano Zampini } 4773906d46d4SStefano Zampini 4774906d46d4SStefano Zampini #undef __FUNCT__ 4775906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4776906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4777906d46d4SStefano Zampini { 4778906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4779906d46d4SStefano Zampini PetscErrorCode ierr; 4780906d46d4SStefano Zampini 4781906d46d4SStefano Zampini PetscFunctionBegin; 4782906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4783906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4784906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4785906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4786906d46d4SStefano Zampini PetscFunctionReturn(0); 4787906d46d4SStefano Zampini } 4788b96c3477SStefano Zampini 4789b96c3477SStefano Zampini #undef __FUNCT__ 4790b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 479108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4792b96c3477SStefano Zampini { 4793a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4794b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4795b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4796a64f4aa4SStefano Zampini Mat S_j; 4797b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4798b96c3477SStefano Zampini PetscBool free_used_adj; 4799b96c3477SStefano Zampini PetscErrorCode ierr; 4800b96c3477SStefano Zampini 4801b96c3477SStefano Zampini PetscFunctionBegin; 4802b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4803b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 480408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4805b96c3477SStefano Zampini used_xadj = NULL; 4806b96c3477SStefano Zampini used_adjncy = NULL; 4807b96c3477SStefano Zampini } else { 480808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 480908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 481008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 481108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4812b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4813b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4814b96c3477SStefano Zampini } else { 48152fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4816b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4817b96c3477SStefano Zampini PetscInt nvtxs; 4818b96c3477SStefano Zampini 48192fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 48202fffb893SStefano Zampini if (flg_row) { 4821b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4822b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4823b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4824b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48252fffb893SStefano Zampini } else { 48262fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48272fffb893SStefano Zampini used_xadj = NULL; 48282fffb893SStefano Zampini used_adjncy = NULL; 48292fffb893SStefano Zampini } 48302fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4831b96c3477SStefano Zampini } 4832b96c3477SStefano Zampini } 4833d5574798SStefano Zampini 4834d5574798SStefano Zampini /* setup sub_schurs data */ 4835a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4836a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4837a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4838a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4839862806e4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE);CHKERRQ(ierr); 4840a64f4aa4SStefano Zampini } else { 48416816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 484204708bb6SStefano Zampini PetscBool isseqaij; 48435feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48445feab87aSStefano Zampini PetscInt n_vertices; 48455feab87aSStefano Zampini 48465feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48472034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48485feab87aSStefano Zampini } 484904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 485004708bb6SStefano Zampini if (!isseqaij) { 485104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 485204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 485304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 485404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 485504708bb6SStefano Zampini } else { 485604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 485704708bb6SStefano Zampini } 485804708bb6SStefano Zampini } 4859862806e4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers);CHKERRQ(ierr); 4860a64f4aa4SStefano Zampini } 4861a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4862b96c3477SStefano Zampini 4863b96c3477SStefano Zampini /* free adjacency */ 4864b96c3477SStefano Zampini if (free_used_adj) { 4865b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4866b96c3477SStefano Zampini } 4867b96c3477SStefano Zampini PetscFunctionReturn(0); 4868b96c3477SStefano Zampini } 4869b96c3477SStefano Zampini 4870b96c3477SStefano Zampini #undef __FUNCT__ 4871b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 487208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4873b96c3477SStefano Zampini { 4874b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4875b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4876b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4877b96c3477SStefano Zampini PCBDDCGraph graph; 4878b96c3477SStefano Zampini PetscErrorCode ierr; 4879b96c3477SStefano Zampini 4880b96c3477SStefano Zampini PetscFunctionBegin; 4881b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 488208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48833301b35fSStefano Zampini IS verticesIS,verticescomm; 48843301b35fSStefano Zampini PetscInt vsize,*idxs; 4885b96c3477SStefano Zampini 4886b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48873301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48883301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48893301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48903301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48913301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4892b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48937fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48943301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48953301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4896b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4897b96c3477SStefano Zampini /* 4898b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4899b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4900b96c3477SStefano Zampini } 4901b96c3477SStefano Zampini */ 4902b96c3477SStefano Zampini } else { 4903b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4904b96c3477SStefano Zampini } 4905b96c3477SStefano Zampini 4906b96c3477SStefano Zampini /* sub_schurs init */ 4907a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4908a64f4aa4SStefano Zampini 4909b96c3477SStefano Zampini /* free graph struct */ 491008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4911b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4912b96c3477SStefano Zampini } 4913b96c3477SStefano Zampini PetscFunctionReturn(0); 4914b96c3477SStefano Zampini } 4915