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); 36fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini PetscInt subset_size; 51862806e4SStefano Zampini 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini 5608122e43SStefano Zampini /* min/max and threshold */ 5708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 58f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 5908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 60f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 61f6f667cfSStefano Zampini if (nmin) { 62f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 63f6f667cfSStefano Zampini } 6408122e43SStefano Zampini 6508122e43SStefano Zampini /* allocate lapack workspace */ 6608122e43SStefano Zampini cum = cum2 = 0; 6708122e43SStefano Zampini maxneigs = 0; 6808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6908122e43SStefano Zampini PetscInt n,subset_size; 70f6f667cfSStefano Zampini 7108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 739162d606SStefano Zampini cum += subset_size; 749162d606SStefano Zampini cum2 += subset_size*n; 7508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7608122e43SStefano Zampini } 7708122e43SStefano Zampini if (mss) { 789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7908122e43SStefano Zampini PetscBLASInt B_itype = 1; 8008122e43SStefano Zampini PetscBLASInt B_N = mss; 814c6709b3SStefano Zampini PetscReal zero = 0.0; 824c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8308122e43SStefano Zampini 8408122e43SStefano Zampini B_lwork = -1; 8508122e43SStefano Zampini S = NULL; 8608122e43SStefano Zampini St = NULL; 87a58a30b4SStefano Zampini eigs = NULL; 88a58a30b4SStefano Zampini eigv = NULL; 89a58a30b4SStefano Zampini B_iwork = NULL; 90a58a30b4SStefano Zampini B_ifail = NULL; 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); 466*f5fd1fbfSStefano Zampini ierr = ISDestroy(&pcbddc->zerodiag);CHKERRQ(ierr); 467*f5fd1fbfSStefano Zampini ierr = VecDestroy(&pcbddc->saddle_sums);CHKERRQ(ierr); 468674ae819SStefano Zampini PetscFunctionReturn(0); 469674ae819SStefano Zampini } 470674ae819SStefano Zampini 471674ae819SStefano Zampini #undef __FUNCT__ 472f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 473f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4746bfb1811SStefano Zampini { 4756bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4766bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4776bfb1811SStefano Zampini VecType impVecType; 478e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4796bfb1811SStefano Zampini PetscErrorCode ierr; 4806bfb1811SStefano Zampini 4816bfb1811SStefano Zampini PetscFunctionBegin; 482f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 483f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 484f4ddd8eeSStefano Zampini } 485e7b262bdSStefano Zampini /* get sizes */ 486b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 487b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4886bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 489e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 490e7b262bdSStefano Zampini /* R nodes */ 491e7b262bdSStefano Zampini old_size = -1; 492e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 493e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 494e7b262bdSStefano Zampini } 495e7b262bdSStefano Zampini if (n_R != old_size) { 496e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 497e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4986bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4996bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5006bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5016bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e7b262bdSStefano Zampini /* local primal dofs */ 504e7b262bdSStefano Zampini old_size = -1; 505e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 506e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 507e7b262bdSStefano Zampini } 508e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 509e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 51083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 511e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5126bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 513e7b262bdSStefano Zampini } 514e7b262bdSStefano Zampini /* local explicit constraints */ 515e7b262bdSStefano Zampini old_size = -1; 516e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 517e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 518e7b262bdSStefano Zampini } 519e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 520e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 52183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 52283b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 52383b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 52483b7ccabSStefano Zampini } 5256bfb1811SStefano Zampini PetscFunctionReturn(0); 5266bfb1811SStefano Zampini } 5276bfb1811SStefano Zampini 5286bfb1811SStefano Zampini #undef __FUNCT__ 52947f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 53047f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 53188ebb749SStefano Zampini { 53225084f0cSStefano Zampini PetscErrorCode ierr; 53325084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 53488ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 53588ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 536d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 53725084f0cSStefano Zampini /* submatrices of local problem */ 53880677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 53906656605SStefano Zampini /* submatrices of local coarse problem */ 54006656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 54125084f0cSStefano Zampini /* working matrices */ 54206656605SStefano Zampini Mat C_CR; 54325084f0cSStefano Zampini /* additional working stuff */ 54406656605SStefano Zampini PC pc_R; 54506656605SStefano Zampini Mat F; 54606656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 54706656605SStefano Zampini 54825084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 54906656605SStefano Zampini PetscScalar *work; 55006656605SStefano Zampini PetscInt *idx_V_B; 55106656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 55206656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 553b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 55445a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 55588ebb749SStefano Zampini MatType impMatType; 55625084f0cSStefano Zampini /* some shortcuts to scalars */ 55706656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 55888ebb749SStefano Zampini 55988ebb749SStefano Zampini PetscFunctionBegin; 560b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 56188ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 56288ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 563b371cd4fSStefano Zampini n_B = pcis->n_B; 564b371cd4fSStefano Zampini n_D = pcis->n - n_B; 56588ebb749SStefano Zampini n_R = pcis->n - n_vertices; 56688ebb749SStefano Zampini 56788ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 56888ebb749SStefano Zampini impMatType = MATSEQDENSE; 56988ebb749SStefano Zampini 57088ebb749SStefano Zampini /* vertices in boundary numbering */ 571785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5720e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 57388ebb749SStefano Zampini if (i != n_vertices) { 57422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 57588ebb749SStefano Zampini } 57688ebb749SStefano Zampini 57706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 57806656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 57906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 58206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 58306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 58406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 58506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 58606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 58706656605SStefano Zampini 58806656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 58906656605SStefano Zampini /* allocate workspace */ 59006656605SStefano Zampini n = 0; 59106656605SStefano Zampini if (n_constraints) { 59206656605SStefano Zampini n += n_R*n_constraints; 59306656605SStefano Zampini } 59406656605SStefano Zampini if (n_vertices) { 59506656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 59680677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 59706656605SStefano Zampini } 5983301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 59906656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 60006656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 60106656605SStefano Zampini } 60206656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 60306656605SStefano Zampini 60406656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 60506656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 60606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 60706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 60806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 60906656605SStefano Zampini if (isLU || isILU || isCHOL) { 61006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 611d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 612d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 613d62866d3SStefano Zampini MatFactorType type; 614d62866d3SStefano Zampini 6156816873aSStefano Zampini F = reuse_mumps->F; 6166816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 617d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 61806656605SStefano Zampini } else { 61906656605SStefano Zampini F = NULL; 62006656605SStefano Zampini } 62106656605SStefano Zampini 62288ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 62388ebb749SStefano Zampini if (n_constraints) { 62406656605SStefano Zampini Mat M1,M2,M3; 62580677318SStefano Zampini Mat auxmat; 62606656605SStefano Zampini IS is_aux; 62780677318SStefano Zampini PetscScalar *array,*array2; 62806656605SStefano Zampini 629f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 63080677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 63188ebb749SStefano Zampini 63225084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 63325084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6348ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 63580677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 63688ebb749SStefano Zampini 63780677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 63880677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 63906656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 64088ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 64106656605SStefano Zampini const PetscScalar *row_cmat_values; 64206656605SStefano Zampini const PetscInt *row_cmat_indices; 64306656605SStefano Zampini PetscInt size_of_constraint,j; 64488ebb749SStefano Zampini 64506656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64606656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 64706656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64806656605SStefano Zampini } 64906656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 65006656605SStefano Zampini } 65180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 65206656605SStefano Zampini if (F) { 65306656605SStefano Zampini Mat B; 65406656605SStefano Zampini 65506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 65680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 65706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 65806656605SStefano Zampini } else { 65980677318SStefano Zampini PetscScalar *marr; 66080677318SStefano Zampini 66180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 66306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 66480677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 66606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 66806656605SStefano Zampini } 66980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 67006656605SStefano Zampini } 67180677318SStefano Zampini if (!pcbddc->switch_static) { 67280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 67380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 67680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 67780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 67880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 68080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 68180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 68280677318SStefano Zampini } 68380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 68480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 68580677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68680677318SStefano Zampini } else { 68780677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 68880677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 68925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 69080677318SStefano Zampini } 69180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 69280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 69380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 69406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 69506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 69680677318SStefano Zampini if (isCHOL) { 69780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 69880677318SStefano Zampini } else { 69925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 70080677318SStefano Zampini } 70180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 70206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 70325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 70425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 70525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 70680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 70780677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 70880677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 70906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 71006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 711f4ddd8eeSStefano Zampini } 71288ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 71388ebb749SStefano Zampini if (n_vertices) { 71406656605SStefano Zampini IS is_aux; 7153a50541eSStefano Zampini 7166816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7176816873aSStefano Zampini IS tis; 7186816873aSStefano Zampini 7196816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7206816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7216816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7226816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7236816873aSStefano Zampini } else { 7243a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7256816873aSStefano Zampini } 7269577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7279577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72804708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 72925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 73088ebb749SStefano Zampini } 73188ebb749SStefano Zampini 73288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 733f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 73406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 73506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 73606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73706656605SStefano Zampini } 738f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 73906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 74006656605SStefano Zampini PetscScalar *marray; 74106656605SStefano Zampini 74206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 74306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 744f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 747f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 748f4ddd8eeSStefano Zampini } 749f4ddd8eeSStefano Zampini } 75006656605SStefano Zampini 751f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 75206656605SStefano Zampini PetscScalar *marray; 75388ebb749SStefano Zampini 75406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7558eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75788ebb749SStefano Zampini } 7583301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75906656605SStefano Zampini n *= 2; 76088ebb749SStefano Zampini } 76106656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 76206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 76306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7648eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 76606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76788ebb749SStefano Zampini } 7683301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 76906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7708eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 77388ebb749SStefano Zampini } 77488ebb749SStefano Zampini } else { 775c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 776c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7771b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 778c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 779c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 780c0553b1fSStefano Zampini } 78188ebb749SStefano Zampini } 78206656605SStefano Zampini } 78306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 78406656605SStefano Zampini /* vertices */ 78506656605SStefano Zampini if (n_vertices) { 78616f15bc4SStefano Zampini 78704708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 78804708bb6SStefano Zampini 78916f15bc4SStefano Zampini if (n_R) { 79006656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 79106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 79216f15bc4SStefano Zampini PetscScalar *x,*y; 79304708bb6SStefano Zampini PetscBool isseqaij; 79406656605SStefano Zampini 79521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 79606656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 79704708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7986816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 79906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 80006656605SStefano Zampini } else { 80106656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 80306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 80606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 80806656605SStefano Zampini } 80906656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 81006656605SStefano Zampini } 81180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 81306656605SStefano Zampini if (n_constraints) { 81406656605SStefano Zampini Mat B; 81580677318SStefano Zampini 816b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 81780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 81880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 81980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 82080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82480677318SStefano Zampini } 82580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 83006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 83306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83406656605SStefano Zampini } 83504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 83604708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 83704708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 83804708bb6SStefano Zampini } 83906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 84080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 84406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 84706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84806656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 84916f15bc4SStefano Zampini } else { 85016f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85116f15bc4SStefano Zampini } 85221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 85306656605SStefano Zampini /* coarse basis functions */ 85406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85516f15bc4SStefano Zampini PetscScalar *y; 85616f15bc4SStefano Zampini 85706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 86306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86506656605SStefano Zampini 86606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87306656605SStefano Zampini } 87406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87506656605SStefano Zampini } 87604708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 87704708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 87806656605SStefano Zampini } 87906656605SStefano Zampini 88006656605SStefano Zampini if (n_constraints) { 88106656605SStefano Zampini Mat B; 88206656605SStefano Zampini 88306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 88406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88706656605SStefano Zampini if (n_vertices) { 88880677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 88980677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 89080677318SStefano Zampini } else { 89180677318SStefano Zampini Mat S_VCt; 89280677318SStefano Zampini 89380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 89480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89680677318SStefano Zampini } 89706656605SStefano Zampini } 89806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 89906656605SStefano Zampini /* coarse basis functions */ 90006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 90106656605SStefano Zampini PetscScalar *y; 90206656605SStefano Zampini 90306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91706656605SStefano Zampini } 91806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91906656605SStefano Zampini } 92006656605SStefano Zampini } 92180677318SStefano Zampini if (n_constraints) { 92280677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 92380677318SStefano Zampini } 92406656605SStefano Zampini 92506656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9263301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 92706656605SStefano Zampini 92806656605SStefano Zampini if (n_constraints) { 92916f15bc4SStefano Zampini Mat S_CCT,B_C; 93006656605SStefano Zampini 93180677318SStefano Zampini /* this is a lazy thing */ 93280677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 93306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93616f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93706656605SStefano Zampini if (n_vertices) { 93816f15bc4SStefano Zampini Mat B_V,S_VCT; 93906656605SStefano Zampini 94006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 94416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94506656605SStefano Zampini } 94606656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94780677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94880677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 94906656605SStefano Zampini } 95016f15bc4SStefano Zampini if (n_vertices && n_R) { 95106656605SStefano Zampini Mat A_VRT; 95280677318SStefano Zampini PetscScalar *marray; 95306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 95406656605SStefano Zampini 95580677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95680677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 95980677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 96080677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 96116f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 96206656605SStefano Zampini } 96306656605SStefano Zampini 96406656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97106656605SStefano Zampini } 97206656605SStefano Zampini } else { 97306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97906656605SStefano Zampini } 98006656605SStefano Zampini } 98106656605SStefano Zampini /* coarse basis functions */ 98206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98306656605SStefano Zampini PetscScalar *y; 98406656605SStefano Zampini 98506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99006656605SStefano Zampini if (i<n_vertices) { 99106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 99206656605SStefano Zampini } 99306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99506656605SStefano Zampini 99606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100306656605SStefano Zampini } 100406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100506656605SStefano Zampini } 100606656605SStefano Zampini } 1007d62866d3SStefano Zampini /* free memory */ 100888ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 100906656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 101106656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1013d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1014d62866d3SStefano Zampini if (n_vertices) { 1015d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1016d62866d3SStefano Zampini } 1017d62866d3SStefano Zampini if (n_constraints) { 1018d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1019d62866d3SStefano Zampini } 102088ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 102188ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102288ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102325084f0cSStefano Zampini if (pcbddc->dbg_flag) { 102488ebb749SStefano Zampini Mat coarse_sub_mat; 102525084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102688ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102788ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102888ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10298bec7fa6SStefano Zampini Mat C_B,CPHI; 10308bec7fa6SStefano Zampini IS is_dummy; 10318bec7fa6SStefano Zampini Vec mones; 103288ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 103388ebb749SStefano Zampini PetscReal real_value; 103488ebb749SStefano Zampini 103588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1041c0553b1fSStefano Zampini if (unsymmetric_check) { 104288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 104488ebb749SStefano Zampini } 104588ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104688ebb749SStefano Zampini 104725084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10483301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 104925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1050c0553b1fSStefano Zampini if (unsymmetric_check) { 105188ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106388ebb749SStefano Zampini } else { 106488ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106588ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107288ebb749SStefano Zampini } 107388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107781d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10788bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10790fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 108006656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10818bec7fa6SStefano Zampini 10828bec7fa6SStefano Zampini /* check constraints */ 10838bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10848bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10858bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10868bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10888bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1090bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1091c0553b1fSStefano Zampini if (unsymmetric_check) { 1092bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1093bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1094bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1095bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1096bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 109788ebb749SStefano Zampini } 10988bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10998bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11008bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11018bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11028bec7fa6SStefano Zampini 110325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 110988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 111088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 111188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 111288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 111388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1114c0553b1fSStefano Zampini if (unsymmetric_check) { 111588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111788ebb749SStefano Zampini } 111888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 111988ebb749SStefano Zampini } 11208629588bSStefano Zampini /* get back data */ 11218629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 112288ebb749SStefano Zampini PetscFunctionReturn(0); 112388ebb749SStefano Zampini } 112488ebb749SStefano Zampini 112588ebb749SStefano Zampini #undef __FUNCT__ 1126d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1127d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1128aa0d41d4SStefano Zampini { 1129d65f70fdSStefano Zampini Mat *work_mat; 1130d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1131d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1132d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1133aa0d41d4SStefano Zampini PetscErrorCode ierr; 1134aa0d41d4SStefano Zampini 1135aa0d41d4SStefano Zampini PetscFunctionBegin; 1136d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1137d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1138d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1140aa0d41d4SStefano Zampini 1141d65f70fdSStefano Zampini if (!rsorted) { 1142906d46d4SStefano Zampini const PetscInt *idxs; 1143906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1144aa0d41d4SStefano Zampini 1145d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1148d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1149aa0d41d4SStefano Zampini } 1150d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1153d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1154aa0d41d4SStefano Zampini } 1155d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1156d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini } else { 1158d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini isrow_s = isrow; 1160aa0d41d4SStefano Zampini } 1161906d46d4SStefano Zampini 1162d65f70fdSStefano Zampini if (!csorted) { 1163d65f70fdSStefano Zampini if (isrow == iscol) { 1164d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini iscol_s = isrow_s; 1166d65f70fdSStefano Zampini } else { 1167d65f70fdSStefano Zampini const PetscInt *idxs; 1168d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1169906d46d4SStefano Zampini 1170d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1172d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1173d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1174d65f70fdSStefano Zampini } 1175d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1177d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1178d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1179d65f70fdSStefano Zampini } 1180d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1181d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1182d65f70fdSStefano Zampini } 1183d65f70fdSStefano Zampini } else { 1184d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1185d65f70fdSStefano Zampini iscol_s = iscol; 1186d65f70fdSStefano Zampini } 1187d65f70fdSStefano Zampini 1188d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1189d65f70fdSStefano Zampini 1190d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1191906d46d4SStefano Zampini Mat new_mat; 1192d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1193906d46d4SStefano Zampini 1194d65f70fdSStefano Zampini if (!rsorted) { 1195d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1196d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1197d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1198d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1199906d46d4SStefano Zampini } 1200d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1201d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1202d65f70fdSStefano Zampini } else { 1203d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1204906d46d4SStefano Zampini } 1205d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1206d65f70fdSStefano Zampini 1207d65f70fdSStefano Zampini if (!csorted) { 1208d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1209d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1210d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1211d65f70fdSStefano Zampini } else { 1212d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1213d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1215d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1216d65f70fdSStefano Zampini } 1217d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini } 1220d65f70fdSStefano Zampini } else { 1221d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini } 1223d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini 1225d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1226d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini work_mat[0] = new_mat; 1228d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1229d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini } 1231d65f70fdSStefano Zampini 1232d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1233d65f70fdSStefano Zampini *B = work_mat[0]; 1234d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1235d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1236d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1237d65f70fdSStefano Zampini PetscFunctionReturn(0); 1238d65f70fdSStefano Zampini } 1239d65f70fdSStefano Zampini 1240d65f70fdSStefano Zampini #undef __FUNCT__ 12415e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12425e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1243aa0d41d4SStefano Zampini { 1244aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12453bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 12465e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1247d65f70fdSStefano Zampini Mat new_mat; 12485e8657edSStefano Zampini IS is_local,is_global; 1249d65f70fdSStefano Zampini PetscInt local_size; 1250d65f70fdSStefano Zampini PetscBool isseqaij; 1251aa0d41d4SStefano Zampini PetscErrorCode ierr; 1252aa0d41d4SStefano Zampini 1253aa0d41d4SStefano Zampini PetscFunctionBegin; 1254aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12555e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12565e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 12573bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1258aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1259d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1260aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1261906d46d4SStefano Zampini 1262906d46d4SStefano Zampini /* check */ 1263906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1264906d46d4SStefano Zampini Vec x,x_change; 1265906d46d4SStefano Zampini PetscReal error; 1266906d46d4SStefano Zampini 12675e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1268906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12695e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1270e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1271e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1272d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1273e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1274e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1275906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1276906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1277906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1278906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1279906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1280906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1281906d46d4SStefano Zampini } 1282906d46d4SStefano Zampini 128322d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12849b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 128522d5777bSStefano Zampini if (isseqaij) { 1286d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1287aa0d41d4SStefano Zampini } else { 1288aa0d41d4SStefano Zampini Mat work_mat; 1289aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1290d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1291aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1292aa0d41d4SStefano Zampini } 12933301b35fSStefano Zampini if (matis->A->symmetric_set) { 12943301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1295e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12963301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1297e496cd5dSStefano Zampini #endif 12983301b35fSStefano Zampini } 129945a1bb75SStefano Zampini /* 130045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1301d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 130245a1bb75SStefano Zampini */ 1303d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1304aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1305aa0d41d4SStefano Zampini } 1306aa0d41d4SStefano Zampini 1307aa0d41d4SStefano Zampini #undef __FUNCT__ 1308a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13098ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1310a64d13efSStefano Zampini { 1311a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1312a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1313d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 131453892102SStefano Zampini PetscInt *idx_R_local=NULL; 13153a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13163a50541eSStefano Zampini PetscInt vbs,bs; 13176816873aSStefano Zampini PetscBT bitmask=NULL; 1318a64d13efSStefano Zampini PetscErrorCode ierr; 1319a64d13efSStefano Zampini 1320a64d13efSStefano Zampini PetscFunctionBegin; 1321b23d619eSStefano Zampini /* 1322b23d619eSStefano Zampini No need to setup local scatters if 1323b23d619eSStefano Zampini - primal space is unchanged 1324b23d619eSStefano Zampini AND 1325b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1326b23d619eSStefano Zampini AND 1327b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1328b23d619eSStefano Zampini */ 1329b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1330f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1331f4ddd8eeSStefano Zampini } 1332f4ddd8eeSStefano Zampini /* destroy old objects */ 1333f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1334f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1335f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1336a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1337b371cd4fSStefano Zampini n_B = pcis->n_B; 1338b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1339b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13403a50541eSStefano Zampini 1341a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13426816873aSStefano Zampini 134353892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13446816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1345854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1346a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1347a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13480e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1349a64d13efSStefano Zampini } 1350a64d13efSStefano Zampini 1351a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13524641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13536816873aSStefano Zampini idx_R_local[n_R++] = i; 1354a64d13efSStefano Zampini } 1355a64d13efSStefano Zampini } 135653892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13576816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13586816873aSStefano Zampini 135953892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 136053892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13616816873aSStefano Zampini } 13623a50541eSStefano Zampini 13633a50541eSStefano Zampini /* Block code */ 13643a50541eSStefano Zampini vbs = 1; 13653a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13663a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13673a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13683a50541eSStefano Zampini PetscInt *vary; 1369d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1370785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13713a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1372d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1373d3df7717SStefano 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 */ 13740e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1375d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13763a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13773a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13783a50541eSStefano Zampini break; 13793a50541eSStefano Zampini } 13803a50541eSStefano Zampini } 1381d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1382d3df7717SStefano Zampini } else { 1383d3df7717SStefano Zampini /* Verify directly the R set */ 1384d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1385d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1386d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1387d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1388d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1389d3df7717SStefano Zampini break; 1390d3df7717SStefano Zampini } 1391d3df7717SStefano Zampini } 1392d3df7717SStefano Zampini } 1393d3df7717SStefano Zampini } 13943a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13953a50541eSStefano Zampini vbs = bs; 13963a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13973a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13983a50541eSStefano Zampini } 13993a50541eSStefano Zampini } 14003a50541eSStefano Zampini } 14013a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 140253892102SStefano Zampini if (sub_schurs->reuse_mumps) { 140353892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 140453892102SStefano Zampini 140553892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140653892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140753892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140853892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140953892102SStefano Zampini } else { 14103a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 141153892102SStefano Zampini } 1412a64d13efSStefano Zampini 1413a64d13efSStefano Zampini /* print some info if requested */ 1414a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1415a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1416a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14170fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1418a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1419a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14203a50541eSStefano 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); 1421a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1422a64d13efSStefano Zampini } 1423a64d13efSStefano Zampini 1424a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14256816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14266816873aSStefano Zampini IS is_aux1,is_aux2; 14276816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14286816873aSStefano Zampini 14293a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1430854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1431854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1432a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14334641a718SStefano Zampini for (i=0; i<n_D; i++) { 14344641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14354641a718SStefano Zampini } 1436a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1437a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14384641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14394641a718SStefano Zampini aux_array1[j++] = i; 1440a64d13efSStefano Zampini } 1441a64d13efSStefano Zampini } 1442a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1443a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1444a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14454641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14464641a718SStefano Zampini aux_array2[j++] = i; 1447a64d13efSStefano Zampini } 1448a64d13efSStefano Zampini } 1449a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1450a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1451a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1452a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1453a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1454a64d13efSStefano Zampini 14558eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1456785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1457a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14584641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14594641a718SStefano Zampini aux_array1[j++] = i; 1460a64d13efSStefano Zampini } 1461a64d13efSStefano Zampini } 1462a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1463a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1464a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1465a64d13efSStefano Zampini } 14664641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14673a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1468d62866d3SStefano Zampini } else { 146953892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14706816873aSStefano Zampini IS tis; 14716816873aSStefano Zampini PetscInt schur_size; 14726816873aSStefano Zampini 147353892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14746816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 147553892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14766816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14776816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14786816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14796816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14806816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1481d62866d3SStefano Zampini } 1482d62866d3SStefano Zampini } 1483a64d13efSStefano Zampini PetscFunctionReturn(0); 1484a64d13efSStefano Zampini } 1485a64d13efSStefano Zampini 1486304d26faSStefano Zampini 1487304d26faSStefano Zampini #undef __FUNCT__ 1488304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1489684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1490304d26faSStefano Zampini { 1491304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1492304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1493304d26faSStefano Zampini PC pc_temp; 1494304d26faSStefano Zampini Mat A_RR; 1495f4ddd8eeSStefano Zampini MatReuse reuse; 1496304d26faSStefano Zampini PetscScalar m_one = -1.0; 1497304d26faSStefano Zampini PetscReal value; 149804708bb6SStefano Zampini PetscInt n_D,n_R; 14999577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1500304d26faSStefano Zampini PetscErrorCode ierr; 1501e604994aSStefano Zampini /* prefixes stuff */ 1502312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1503e604994aSStefano Zampini size_t len; 1504304d26faSStefano Zampini 1505304d26faSStefano Zampini PetscFunctionBegin; 1506304d26faSStefano Zampini 1507e604994aSStefano Zampini /* compute prefixes */ 1508e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1509e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1510e604994aSStefano Zampini if (!pcbddc->current_level) { 1511e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1512e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1513e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1514e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1515e604994aSStefano Zampini } else { 1516e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1517312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1518e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1519e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1520312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1521312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 152234d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 152334d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1524e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1525e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1526e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1527e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1528e604994aSStefano Zampini } 1529e604994aSStefano Zampini 1530304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1531684f6988SStefano Zampini if (dirichlet) { 1532d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15333301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15343301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1535964fefecSStefano Zampini } 1536ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1537964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1538304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1539304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1540304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1541304d26faSStefano Zampini /* default */ 1542304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1543e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15449577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1545304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15469577ea80SStefano Zampini if (issbaij) { 15479577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15489577ea80SStefano Zampini } else { 1549304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15509577ea80SStefano Zampini } 1551304d26faSStefano Zampini /* Allow user's customization */ 1552304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1553304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1554304d26faSStefano Zampini } 1555d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1556d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1557d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1558d62866d3SStefano Zampini 1559d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1560d5574798SStefano Zampini } 1561304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1562304d26faSStefano Zampini if (!n_D) { 1563304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1564304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1565304d26faSStefano Zampini } 1566304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1567304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1568304d26faSStefano Zampini /* set ksp_D into pcis data */ 1569304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1570304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1571304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1572684f6988SStefano Zampini } 1573304d26faSStefano Zampini 1574304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1575684f6988SStefano Zampini A_RR = 0; 1576684f6988SStefano Zampini if (neumann) { 1577d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 157804708bb6SStefano Zampini PetscInt ibs,mbs; 157904708bb6SStefano Zampini PetscBool issbaij; 158004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1581f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15828ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1583f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1584f4ddd8eeSStefano Zampini PetscInt nn_R; 158581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1586f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1587f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1588f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1589f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1590f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1591f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1592f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1593727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1594f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1595f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1596f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1597f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1598f4ddd8eeSStefano Zampini } 1599f4ddd8eeSStefano Zampini } 1600f4ddd8eeSStefano Zampini /* last check */ 1601d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1602f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1603f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1604f4ddd8eeSStefano Zampini } 1605f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1606f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1607f4ddd8eeSStefano Zampini } 1608f4ddd8eeSStefano Zampini /* extract A_RR */ 1609af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1610af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 161104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 161204708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 161304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 161404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 161504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1616af732b37SStefano Zampini } else { 161704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16186816873aSStefano Zampini } 161904708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 162004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 162104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 162204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162304708bb6SStefano Zampini } else { 162404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162504708bb6SStefano Zampini } 162604708bb6SStefano Zampini } 162704708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1628f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16293301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16303301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16316816873aSStefano Zampini } 16326816873aSStefano Zampini } else { 16336816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16346816873aSStefano Zampini 16356816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16366816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16376816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1638af732b37SStefano Zampini } 1639f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1640304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1641304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1642304d26faSStefano Zampini /* default */ 1643304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1644e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1645304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16469577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16479577ea80SStefano Zampini if (issbaij) { 16489577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16499577ea80SStefano Zampini } else { 1650304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16519577ea80SStefano Zampini } 1652304d26faSStefano Zampini /* Allow user's customization */ 1653304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1654304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1655304d26faSStefano Zampini } 1656d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1657304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1658304d26faSStefano Zampini if (!n_R) { 1659304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1660304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1661304d26faSStefano Zampini } 1662d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1663d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1664d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1665d62866d3SStefano Zampini 1666d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1667d62866d3SStefano Zampini } 1668304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1669304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1670684f6988SStefano Zampini } 16716816873aSStefano Zampini /* free Neumann problem's matrix */ 16726816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1673304d26faSStefano Zampini 1674304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16750fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1676684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1677684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1678684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1679684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1680684f6988SStefano Zampini } 1681684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16820fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1687304d26faSStefano Zampini /* need to be adapted? */ 1688b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1689b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1690b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1691304d26faSStefano Zampini /* print info */ 1692304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1693e604994aSStefano 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); 1694304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1695304d26faSStefano Zampini } 1696b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1697298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1698304d26faSStefano Zampini } 1699684f6988SStefano Zampini } 1700684f6988SStefano Zampini if (neumann) { /* Neumann */ 17016816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 17020fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 17030fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 17040fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 17050fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17060fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1707304d26faSStefano Zampini /* need to be adapted? */ 1708b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1709b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1710304d26faSStefano Zampini /* print info */ 1711304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1712e604994aSStefano 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); 1713304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1714304d26faSStefano Zampini } 1715b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1716298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1717304d26faSStefano Zampini } 17180fccc4e9SStefano Zampini } 1719684f6988SStefano Zampini } 1720304d26faSStefano Zampini PetscFunctionReturn(0); 1721304d26faSStefano Zampini } 1722304d26faSStefano Zampini 1723304d26faSStefano Zampini #undef __FUNCT__ 1724ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 172580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1726674ae819SStefano Zampini { 1727674ae819SStefano Zampini PetscErrorCode ierr; 1728674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1729be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1730674ae819SStefano Zampini 1731674ae819SStefano Zampini PetscFunctionBegin; 1732be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 173420c7b377SStefano Zampini } 173580677318SStefano Zampini if (!pcbddc->switch_static) { 173680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 173880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173920c7b377SStefano Zampini } 1740be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174320c7b377SStefano Zampini } else { 1744be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1745be83ff47SStefano Zampini 174653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174820c7b377SStefano Zampini } 1749be83ff47SStefano Zampini } else { 175080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 175580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 175680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1759674ae819SStefano Zampini } 1760674ae819SStefano Zampini } 1761be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176280677318SStefano Zampini if (applytranspose) { 176380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176480677318SStefano Zampini } else { 176580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176680677318SStefano Zampini } 176753892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1768be83ff47SStefano Zampini } else { 1769be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1770be83ff47SStefano Zampini 1771be83ff47SStefano Zampini if (applytranspose) { 177253892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1773be83ff47SStefano Zampini } else { 177453892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1775be83ff47SStefano Zampini } 177653892102SStefano Zampini #endif 1777be83ff47SStefano Zampini } 177880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 177980677318SStefano Zampini if (!pcbddc->switch_static) { 1780be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 178180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1783be83ff47SStefano Zampini } else { 1784be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1785be83ff47SStefano Zampini 178653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1788be83ff47SStefano Zampini } 178980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 179280677318SStefano Zampini } 179380677318SStefano Zampini } else { 179480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 180080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 180180677318SStefano Zampini } 180280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1806674ae819SStefano Zampini } 1807674ae819SStefano Zampini PetscFunctionReturn(0); 1808674ae819SStefano Zampini } 1809674ae819SStefano Zampini 1810dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1811674ae819SStefano Zampini #undef __FUNCT__ 1812674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1813dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1814674ae819SStefano Zampini { 1815674ae819SStefano Zampini PetscErrorCode ierr; 1816674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1817674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1818674ae819SStefano Zampini const PetscScalar zero = 0.0; 1819674ae819SStefano Zampini 1820674ae819SStefano Zampini PetscFunctionBegin; 1821dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1822dc359a40SStefano Zampini if (applytranspose) { 1823674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18248eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1825dc359a40SStefano Zampini } else { 1826674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1827674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 182815aaf578SStefano Zampini } 182912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 183012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 183112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183312edc857SStefano Zampini 18349f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 183512edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 183612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1837964fefecSStefano Zampini Vec rhs,sol; 1838964fefecSStefano Zampini 1839964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1840964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 184112edc857SStefano Zampini if (applytranspose) { 1842964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184312edc857SStefano Zampini } else { 1844964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184512edc857SStefano Zampini } 184612edc857SStefano Zampini } 1847674ae819SStefano Zampini 1848674ae819SStefano Zampini /* Local solution on R nodes */ 184980677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 185080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18519f00e9b4SStefano Zampini } 1852674ae819SStefano Zampini 18539f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18549f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 185512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1856674ae819SStefano Zampini 1857674ae819SStefano Zampini /* Sum contributions from two levels */ 1858dc359a40SStefano Zampini if (applytranspose) { 1859dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1860dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1861dc359a40SStefano Zampini } else { 1862674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18638eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1864dc359a40SStefano Zampini } 1865674ae819SStefano Zampini PetscFunctionReturn(0); 1866674ae819SStefano Zampini } 1867674ae819SStefano Zampini 1868674ae819SStefano Zampini #undef __FUNCT__ 1869674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 187012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1871674ae819SStefano Zampini { 1872674ae819SStefano Zampini PetscErrorCode ierr; 1873674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187458da7f69SStefano Zampini PetscScalar *array; 187512edc857SStefano Zampini Vec from,to; 1876674ae819SStefano Zampini 1877674ae819SStefano Zampini PetscFunctionBegin; 187812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187912edc857SStefano Zampini from = pcbddc->coarse_vec; 188012edc857SStefano Zampini to = pcbddc->vec1_P; 188112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 188212edc857SStefano Zampini Vec tvec; 188358da7f69SStefano Zampini 188458da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188558da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 188612edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188758da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 188858da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 188958da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 189012edc857SStefano Zampini } 189112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189212edc857SStefano Zampini from = pcbddc->vec1_P; 189312edc857SStefano Zampini to = pcbddc->coarse_vec; 189412edc857SStefano Zampini } 189512edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1896674ae819SStefano Zampini PetscFunctionReturn(0); 1897674ae819SStefano Zampini } 1898674ae819SStefano Zampini 1899674ae819SStefano Zampini #undef __FUNCT__ 1900674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 190112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1902674ae819SStefano Zampini { 1903674ae819SStefano Zampini PetscErrorCode ierr; 1904674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 190558da7f69SStefano Zampini PetscScalar *array; 190612edc857SStefano Zampini Vec from,to; 1907674ae819SStefano Zampini 1908674ae819SStefano Zampini PetscFunctionBegin; 190912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 191012edc857SStefano Zampini from = pcbddc->coarse_vec; 191112edc857SStefano Zampini to = pcbddc->vec1_P; 191212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 191312edc857SStefano Zampini from = pcbddc->vec1_P; 191412edc857SStefano Zampini to = pcbddc->coarse_vec; 191512edc857SStefano Zampini } 191612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 191712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 191812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 191912edc857SStefano Zampini Vec tvec; 192058da7f69SStefano Zampini 192112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 192258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 192358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 192458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 192558da7f69SStefano Zampini } 192658da7f69SStefano Zampini } else { 192758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 192858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 192912edc857SStefano Zampini } 193012edc857SStefano Zampini } 1931674ae819SStefano Zampini PetscFunctionReturn(0); 1932674ae819SStefano Zampini } 1933674ae819SStefano Zampini 1934984c4197SStefano Zampini /* uncomment for testing purposes */ 1935984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1936674ae819SStefano Zampini #undef __FUNCT__ 1937674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1938674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1939674ae819SStefano Zampini { 1940674ae819SStefano Zampini PetscErrorCode ierr; 1941674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1942674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1943674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1944984c4197SStefano Zampini /* one and zero */ 1945984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1946984c4197SStefano Zampini /* space to store constraints and their local indices */ 19479162d606SStefano Zampini PetscScalar *constraints_data; 19489162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19499162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19509162d606SStefano Zampini PetscInt *constraints_n; 1951984c4197SStefano Zampini /* iterators */ 1952b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1953984c4197SStefano Zampini /* BLAS integers */ 1954e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1955e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1956c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1957727cdba6SStefano Zampini /* reuse */ 19580e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19590e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1960984c4197SStefano Zampini /* change of basis */ 1961b3d85658SStefano Zampini PetscBool qr_needed; 19629162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1963984c4197SStefano Zampini /* auxiliary stuff */ 196464efe560SStefano Zampini PetscInt *nnz,*is_indices; 19658a0068c3SStefano Zampini PetscInt ncc; 1966984c4197SStefano Zampini /* some quantities */ 196745a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1968a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1969984c4197SStefano Zampini 1970674ae819SStefano Zampini PetscFunctionBegin; 19718e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19728e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19738e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1974088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1975088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19760e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19770e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19780e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19790e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19800e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1981088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1982cf5a6209SStefano Zampini 1983cf5a6209SStefano Zampini /* print some info */ 1984cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1985cf5a6209SStefano Zampini IS vertices; 1986cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1987cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1988cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1989cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1990cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1991cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1992cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1993fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1994fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1995cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1996cf5a6209SStefano Zampini } 1997cf5a6209SStefano Zampini 1998cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19999162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2000cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2001cf5a6209SStefano Zampini const Vec *nearnullvecs; 2002cf5a6209SStefano Zampini Vec *localnearnullsp; 2003cf5a6209SStefano Zampini PetscScalar *array; 2004cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2005cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2006674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2007b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2008674ae819SStefano Zampini PetscScalar *work; 2009674ae819SStefano Zampini PetscReal *singular_vals; 2010674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2011674ae819SStefano Zampini PetscReal *rwork; 2012674ae819SStefano Zampini #endif 2013674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2014674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2015674ae819SStefano Zampini #else 2016964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2017964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2018674ae819SStefano Zampini #endif 2019674ae819SStefano Zampini 2020674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2021d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2022d06fc5fdSStefano Zampini /* free unneeded index sets */ 2023d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2024d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2025674ae819SStefano Zampini } 2026d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2027d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2028d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2029d06fc5fdSStefano Zampini } 2030d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2031d06fc5fdSStefano Zampini n_ISForEdges = 0; 2032d06fc5fdSStefano Zampini } 2033d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2034d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2035d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2036d06fc5fdSStefano Zampini } 2037d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2038d06fc5fdSStefano Zampini n_ISForFaces = 0; 2039d06fc5fdSStefano Zampini } 204070022509SStefano Zampini 204170022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 204270022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 204370022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 204470022509SStefano Zampini if (pcbddc->NullSpace) { 204570022509SStefano Zampini PetscBool tbool[2],gbool[2]; 204670022509SStefano Zampini 204770022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2048b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2049d06fc5fdSStefano Zampini if (!ISForEdges) { 2050d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2051d06fc5fdSStefano Zampini } 2052b8ffe317SStefano Zampini } 2053d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2054d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2055d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2056d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2057d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 205898a51de6SStefano Zampini } 205970022509SStefano Zampini #endif 206008122e43SStefano Zampini 2061674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2062674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2063674ae819SStefano Zampini if (nearnullsp) { 2064674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2065f4ddd8eeSStefano Zampini /* remove any stored info */ 2066f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2067f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2068f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2069f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2070f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2071473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2072f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2073f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2074f4ddd8eeSStefano Zampini } 2075984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2076984c4197SStefano Zampini nnsp_size = 0; 2077674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2078674ae819SStefano Zampini } 2079984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2080984c4197SStefano Zampini max_constraints = nnsp_size; 2081984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2082984c4197SStefano Zampini 2083674ae819SStefano Zampini /* 2084674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20859162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20869162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20879162d606SStefano Zampini There can be multiple constraints per connected component 2088674ae819SStefano Zampini */ 2089674ae819SStefano Zampini n_vertices = 0; 2090674ae819SStefano Zampini if (ISForVertices) { 2091674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2092674ae819SStefano Zampini } 20939162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20949162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20959162d606SStefano Zampini 20969162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20979162d606SStefano Zampini total_counts *= max_constraints; 2098674ae819SStefano Zampini total_counts += n_vertices; 20994641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 21009162d606SStefano Zampini 2101674ae819SStefano Zampini total_counts = 0; 2102674ae819SStefano Zampini max_size_of_constraint = 0; 2103674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 21049162d606SStefano Zampini IS used_is; 2105674ae819SStefano Zampini if (i<n_ISForEdges) { 21069162d606SStefano Zampini used_is = ISForEdges[i]; 2107674ae819SStefano Zampini } else { 21089162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2109674ae819SStefano Zampini } 21109162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2111674ae819SStefano Zampini total_counts += j; 2112674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2113674ae819SStefano Zampini } 21149162d606SStefano 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); 21159162d606SStefano Zampini 2116984c4197SStefano Zampini /* get local part of global near null space vectors */ 2117785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2118984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2119984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2120e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2121e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2122984c4197SStefano Zampini } 2123674ae819SStefano Zampini 2124242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2125242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2126a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2127242a89d7SStefano Zampini 2128984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2129a773dcb8SStefano Zampini if (!skip_lapack) { 2130674ae819SStefano Zampini PetscScalar temp_work; 2131911cabfeSStefano Zampini 2132674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2133984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2134785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2135785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2136785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2137674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2138785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2139674ae819SStefano Zampini #endif 2140674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2141c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2142c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2143674ae819SStefano Zampini lwork = -1; 2144674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2145674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2146c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2147674ae819SStefano Zampini #else 2148c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2149674ae819SStefano Zampini #endif 2150674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2151984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2152674ae819SStefano Zampini #else /* on missing GESVD */ 2153674ae819SStefano Zampini /* SVD */ 2154674ae819SStefano Zampini PetscInt max_n,min_n; 2155674ae819SStefano Zampini max_n = max_size_of_constraint; 2156984c4197SStefano Zampini min_n = max_constraints; 2157984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2158674ae819SStefano Zampini min_n = max_size_of_constraint; 2159984c4197SStefano Zampini max_n = max_constraints; 2160674ae819SStefano Zampini } 2161785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2162674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2163785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2164674ae819SStefano Zampini #endif 2165674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2166674ae819SStefano Zampini lwork = -1; 2167e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2168e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2169b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2170674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2171674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21729162d606SStefano 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)); 2173674ae819SStefano Zampini #else 21749162d606SStefano 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)); 2175674ae819SStefano Zampini #endif 2176674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2177984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2178984c4197SStefano Zampini #endif /* on missing GESVD */ 2179674ae819SStefano Zampini /* Allocate optimal workspace */ 2180674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2181854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2182674ae819SStefano Zampini } 2183674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2184674ae819SStefano Zampini total_counts = 0; 21859162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21869162d606SStefano Zampini constraints_data_ptr[0] = 0; 2187674ae819SStefano Zampini /* vertices */ 21889162d606SStefano Zampini if (n_vertices) { 2189674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21909162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21919162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2192674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21939162d606SStefano Zampini constraints_n[total_counts] = 1; 21949162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21959162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21969162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2197674ae819SStefano Zampini total_counts++; 2198674ae819SStefano Zampini } 2199674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2200984c4197SStefano Zampini PetscBool used_vertex; 2201674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2202674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2203674ae819SStefano Zampini k = 0; 2204674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2205984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2206984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22079162d606SStefano Zampini constraints_n[total_counts] = 1; 22089162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22099162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22109162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22119162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2212674ae819SStefano Zampini total_counts++; 2213674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2214674ae819SStefano Zampini } 2215984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2216674ae819SStefano Zampini k++; 2217674ae819SStefano Zampini } 2218674ae819SStefano Zampini } 2219674ae819SStefano Zampini } 2220674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2221674ae819SStefano Zampini n_vertices = total_counts; 2222674ae819SStefano Zampini } 2223984c4197SStefano Zampini 2224674ae819SStefano Zampini /* edges and faces */ 22259162d606SStefano Zampini total_counts_cc = total_counts; 2226911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22279162d606SStefano Zampini IS used_is; 22289162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22299162d606SStefano Zampini 2230911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22319162d606SStefano Zampini used_is = ISForEdges[ncc]; 2232984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2233674ae819SStefano Zampini } else { 22349162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2235984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2236674ae819SStefano Zampini } 2237674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22389162d606SStefano Zampini 22399162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22409162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2241984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2242984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2243674ae819SStefano Zampini if (nnsp_has_cnst) { 22445b08dc53SStefano Zampini PetscScalar quad_value; 22459162d606SStefano Zampini 22469162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22479162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22489162d606SStefano Zampini 2249a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2250674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2251a773dcb8SStefano Zampini } else { 2252a773dcb8SStefano Zampini quad_value = 1.0; 2253a773dcb8SStefano Zampini } 2254674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22559162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2256674ae819SStefano Zampini } 22579162d606SStefano Zampini temp_constraints++; 2258674ae819SStefano Zampini total_counts++; 2259674ae819SStefano Zampini } 2260674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2261984c4197SStefano Zampini PetscReal real_value; 22629162d606SStefano Zampini PetscScalar *ptr_to_data; 22639162d606SStefano Zampini 2264984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22659162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2266674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22679162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2268674ae819SStefano Zampini } 2269984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2270984c4197SStefano Zampini /* check if array is null on the connected component */ 2271e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22729162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22735b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2274674ae819SStefano Zampini temp_constraints++; 2275674ae819SStefano Zampini total_counts++; 22769162d606SStefano Zampini if (!idxs_copied) { 22779162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22789162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2279674ae819SStefano Zampini } 2280674ae819SStefano Zampini } 22819162d606SStefano Zampini } 22829162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 228345a1bb75SStefano Zampini valid_constraints = temp_constraints; 2284eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2285a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22869162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22879162d606SStefano Zampini 22889162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2289a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22909162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2291a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22929162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2293a773dcb8SStefano Zampini } else { /* perform SVD */ 2294984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22959162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2296674ae819SStefano Zampini 2297674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2298984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2299984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2300984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2301984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2302984c4197SStefano Zampini from that computed using LAPACKgesvd 2303984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2304984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2305984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2306674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2307e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2308984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2309674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2310674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23119162d606SStefano 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)); 2312674ae819SStefano Zampini } 2313674ae819SStefano Zampini } 2314e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2315e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2316e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2317674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2318c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2319674ae819SStefano Zampini #else 2320c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2321674ae819SStefano Zampini #endif 2322674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2323984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2324984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2325674ae819SStefano Zampini j = 0; 2326984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2327674ae819SStefano Zampini total_counts = total_counts-j; 232845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2329e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2330c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2331c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2332c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2333c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2334c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2335c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2336674ae819SStefano Zampini if (j<temp_constraints) { 2337984c4197SStefano Zampini PetscInt ii; 2338984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2339674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23409162d606SStefano 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)); 2341674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2342984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2343674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23449162d606SStefano 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]; 2345674ae819SStefano Zampini } 2346674ae819SStefano Zampini } 2347674ae819SStefano Zampini } 2348674ae819SStefano Zampini #else /* on missing GESVD */ 2349e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2350e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2351b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2352674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2353674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23549162d606SStefano 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)); 2355674ae819SStefano Zampini #else 23569162d606SStefano 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)); 2357674ae819SStefano Zampini #endif 2358984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2359674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2360984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2361e310c8b4SStefano Zampini k = temp_constraints; 2362e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2363674ae819SStefano Zampini j = 0; 2364e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 236545a1bb75SStefano Zampini valid_constraints = k-j; 2366911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2367984c4197SStefano Zampini #endif /* on missing GESVD */ 2368674ae819SStefano Zampini } 2369a773dcb8SStefano Zampini } 23709162d606SStefano Zampini /* update pointers information */ 23719162d606SStefano Zampini if (valid_constraints) { 23729162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23739162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23749162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23759162d606SStefano Zampini /* set change_of_basis flag */ 237645a1bb75SStefano Zampini if (boolforchange) { 2377b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23789162d606SStefano Zampini } 2379b3d85658SStefano Zampini total_counts_cc++; 238045a1bb75SStefano Zampini } 238145a1bb75SStefano Zampini } 2382984c4197SStefano Zampini /* free workspace */ 23838f1c130eSStefano Zampini if (!skip_lapack) { 2384984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2385984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2386984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2387984c4197SStefano Zampini #endif 2388984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2389984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2390984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2391984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2392984c4197SStefano Zampini #endif 2393984c4197SStefano Zampini } 2394984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2395984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2396984c4197SStefano Zampini } 2397984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2398cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2399cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2400cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2401cf5a6209SStefano Zampini } 2402cf5a6209SStefano Zampini if (n_ISForFaces) { 2403cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2404cf5a6209SStefano Zampini } 2405cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2406cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2407cf5a6209SStefano Zampini } 2408cf5a6209SStefano Zampini if (n_ISForEdges) { 2409cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2410cf5a6209SStefano Zampini } 2411cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 241208122e43SStefano Zampini } else { 241308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2414984c4197SStefano Zampini 241508122e43SStefano Zampini total_counts = 0; 241608122e43SStefano Zampini n_vertices = 0; 2417d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2418d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 241908122e43SStefano Zampini } 242008122e43SStefano Zampini max_constraints = 0; 24219162d606SStefano Zampini total_counts_cc = 0; 242208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 242308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24249162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 242508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 242608122e43SStefano Zampini } 24279162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24289162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24299162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24309162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 243174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24329162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24339162d606SStefano Zampini total_counts_cc = 0; 24349162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24359162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24369162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 243708122e43SStefano Zampini } 243808122e43SStefano Zampini } 24399162d606SStefano Zampini #if 0 24409162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24419162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24429162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24439162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24449162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24459162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24469162d606SStefano Zampini } 24479162d606SStefano Zampini printf("\n"); 24489162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24499162d606SStefano Zampini } 24501b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24518bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24521b968477SStefano Zampini } 24531b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24548bec7fa6SStefano 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]); 24551b968477SStefano Zampini } 245608122e43SStefano Zampini #endif 245708122e43SStefano Zampini 24588bec7fa6SStefano Zampini max_size_of_constraint = 0; 24599162d606SStefano 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]); 24609162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 246108122e43SStefano Zampini /* Change of basis */ 2462b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 246308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 246408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 246508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2466b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 246708122e43SStefano Zampini } 246808122e43SStefano Zampini } 246908122e43SStefano Zampini } 247008122e43SStefano Zampini } 2471984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24720e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 247308122e43SStefano Zampini 24749162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24759162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24769162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24779162d606SStefano 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); 247808122e43SStefano Zampini } 2479674ae819SStefano Zampini 2480674ae819SStefano Zampini /* Create constraint matrix */ 2481674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 248216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2483984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2484984c4197SStefano Zampini 2485984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2486a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2487a717540cSStefano Zampini qr_needed = PETSC_FALSE; 248874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2489984c4197SStefano Zampini total_primal_vertices=0; 2490b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24919162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24929162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2493984c4197SStefano Zampini if (size_of_constraint == 1) { 24949162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2495b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 249664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24979162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24989162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2499a717540cSStefano Zampini } 2500b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 250174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2502a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2503a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2504a717540cSStefano Zampini } 2505fa434743SStefano Zampini } else { 2506b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2507fa434743SStefano Zampini } 2508a717540cSStefano Zampini } 2509b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2510b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2511674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 251270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2513b3d85658SStefano Zampini 2514b3d85658SStefano 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); 25150e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25160e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2517984c4197SStefano Zampini 2518984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 251974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2520785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2521984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 252274d5cdf7SStefano Zampini 2523984c4197SStefano Zampini j = total_primal_vertices; 252474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2525b3d85658SStefano Zampini cum = total_primal_vertices; 25269162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25274641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2528b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2529b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2530b3d85658SStefano Zampini cum++; 25319162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 253274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 253374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 253474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 253574d5cdf7SStefano Zampini } 25369162d606SStefano Zampini j += constraints_n[i]; 2537674ae819SStefano Zampini } 2538674ae819SStefano Zampini } 2539674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2540674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2541088faed8SStefano Zampini 2542674ae819SStefano Zampini /* set values in constraint matrix */ 2543984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25440e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2545674ae819SStefano Zampini } 2546984c4197SStefano Zampini total_counts = total_primal_vertices; 25479162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25484641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25499162d606SStefano Zampini PetscInt *cols; 25509162d606SStefano Zampini 25519162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25529162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25539162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25549162d606SStefano Zampini PetscInt row = total_counts+k; 25559162d606SStefano Zampini PetscScalar *vals; 25569162d606SStefano Zampini 25579162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25589162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25599162d606SStefano Zampini } 25609162d606SStefano Zampini total_counts += constraints_n[i]; 2561674ae819SStefano Zampini } 2562674ae819SStefano Zampini } 2563674ae819SStefano Zampini /* assembling */ 2564674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2565674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2566088faed8SStefano Zampini 2567984c4197SStefano Zampini /* 256845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2569984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2570984c4197SStefano Zampini */ 2571674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2572674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2573026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2574984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2575984c4197SStefano Zampini /* working stuff for GEQRF */ 257681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2577984c4197SStefano Zampini PetscBLASInt lqr_work; 2578984c4197SStefano Zampini /* working stuff for UNGQR */ 2579984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2580984c4197SStefano Zampini PetscBLASInt lgqr_work; 2581984c4197SStefano Zampini /* working stuff for TRTRS */ 2582984c4197SStefano Zampini PetscScalar *trs_rhs; 25833f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2584984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2585984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2586984c4197SStefano Zampini PetscScalar *start_vals; 2587984c4197SStefano Zampini /* working stuff for values insertion */ 25884641a718SStefano Zampini PetscBT is_primal; 258964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2590906d46d4SStefano Zampini /* matrix sizes */ 2591906d46d4SStefano Zampini PetscInt global_size,local_size; 2592906d46d4SStefano Zampini /* temporary change of basis */ 2593906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2594cf5a6209SStefano Zampini /* extra space for debugging */ 2595cf5a6209SStefano Zampini PetscScalar *dbg_work; 2596984c4197SStefano Zampini 2597906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2598906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 259916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2600bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2601906d46d4SStefano Zampini /* nonzeros for local mat */ 2602bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2603bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 26049162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2605a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26069162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2607a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26089162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2609a717540cSStefano Zampini } else { 26109162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26119162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2612a717540cSStefano Zampini } 2613a717540cSStefano Zampini } 2614a717540cSStefano Zampini } 2615906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2616bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2617a717540cSStefano Zampini /* Set initial identity in the matrix */ 2618bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2619906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2620a717540cSStefano Zampini } 2621a717540cSStefano Zampini 2622a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2623a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2624a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2625a717540cSStefano Zampini } 2626a717540cSStefano Zampini 2627a717540cSStefano Zampini 2628a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2629a717540cSStefano Zampini /* 2630a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2631a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2632a717540cSStefano Zampini 2633a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2634a717540cSStefano Zampini 2635a6b551f4SStefano 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) 2636a6b551f4SStefano Zampini 2637a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2638a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2639a717540cSStefano Zampini | ... | 2640a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2641a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2642a717540cSStefano Zampini 2643a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2644a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2645a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2646a6b551f4SStefano Zampini 2647a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2648a717540cSStefano Zampini */ 2649a717540cSStefano Zampini if (qr_needed) { 2650984c4197SStefano Zampini /* space to store Q */ 2651854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2652984c4197SStefano Zampini /* first we issue queries for optimal work */ 26533f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26543f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26553f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2656984c4197SStefano Zampini lqr_work = -1; 26573f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2658984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2659984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2660785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2661984c4197SStefano Zampini lgqr_work = -1; 26623f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26643f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26653f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26663f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26673f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2668984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2669984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2670785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2671984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2672785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2673984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2674785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2675a717540cSStefano Zampini /* allocating workspace for check */ 2676a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2677cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2678a717540cSStefano Zampini } 2679a717540cSStefano Zampini } 2680984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26814641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2682473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26830e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 268439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 268539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26864641a718SStefano Zampini } 268739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 268839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 268939e2fb2aSStefano Zampini } 269039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2691984c4197SStefano Zampini 2692a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26939162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26949162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26954641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2696984c4197SStefano Zampini /* get constraint info */ 26979162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2698984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2699984c4197SStefano Zampini 2700984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27019162d606SStefano 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); 2702674ae819SStefano Zampini } 2703984c4197SStefano Zampini 2704fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2705a717540cSStefano Zampini 2706a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2707a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27089162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2709a717540cSStefano Zampini } 2710984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27119162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2712984c4197SStefano Zampini 2713984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2717674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27183f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2719984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2720674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2721984c4197SStefano Zampini 2722984c4197SStefano Zampini /* explictly compute R^-T */ 2723984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2724984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27273f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2729984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27303f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2731984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2732984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2733984c4197SStefano Zampini 2734a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27363f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27373f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2739984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27403f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2741984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2742984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2743984c4197SStefano Zampini 2744984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2745984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2746984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27483f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27503f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27513f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27523f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2753984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27549162d606SStefano 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)); 2755984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27569162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2757984c4197SStefano Zampini 2758984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27599162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2760984c4197SStefano Zampini /* insert cols for primal dofs */ 2761984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2762984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27639162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2764906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2765984c4197SStefano Zampini } 2766984c4197SStefano Zampini /* insert cols for dual dofs */ 2767984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27689162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2769984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27709162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2771906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2772984c4197SStefano Zampini j++; 2773674ae819SStefano Zampini } 2774674ae819SStefano Zampini } 2775984c4197SStefano Zampini 2776984c4197SStefano Zampini /* check change of basis */ 2777984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2778984c4197SStefano Zampini PetscInt ii,jj; 2779984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2780c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2781c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2782c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2783c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2784c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2785c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2786984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2787cf5a6209SStefano 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)); 2788984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2789984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2790984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2791cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2792cf5a6209SStefano 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; 2793674ae819SStefano Zampini } 2794674ae819SStefano Zampini } 2795984c4197SStefano Zampini if (!valid_qr) { 279622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2797984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2798984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2799cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2800cf5a6209SStefano 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])); 2801674ae819SStefano Zampini } 2802cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2803cf5a6209SStefano 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])); 2804984c4197SStefano Zampini } 2805984c4197SStefano Zampini } 2806984c4197SStefano Zampini } 2807674ae819SStefano Zampini } else { 280822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2809674ae819SStefano Zampini } 2810674ae819SStefano Zampini } 2811a717540cSStefano Zampini } else { /* simple transformation block */ 2812a717540cSStefano Zampini PetscInt row,col; 2813a6b551f4SStefano Zampini PetscScalar val,norm; 2814a6b551f4SStefano Zampini 2815a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28169162d606SStefano 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)); 2817a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28189162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28199162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2820bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28219162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2822906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28239162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2824a717540cSStefano Zampini } else { 2825a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28269162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2827a717540cSStefano Zampini if (row != col) { 28289162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2829a717540cSStefano Zampini } else { 28309162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2831a717540cSStefano Zampini } 2832906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2833a717540cSStefano Zampini } 2834a717540cSStefano Zampini } 2835a717540cSStefano Zampini } 283698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 283722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2838a717540cSStefano Zampini } 2839674ae819SStefano Zampini } 2840984c4197SStefano Zampini } else { 2841984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28429162d606SStefano 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); 2843674ae819SStefano Zampini } 2844674ae819SStefano Zampini } 2845674ae819SStefano Zampini } 2846a717540cSStefano Zampini 2847a717540cSStefano Zampini /* free workspace */ 2848a717540cSStefano Zampini if (qr_needed) { 2849984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2850cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2851984c4197SStefano Zampini } 2852984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2853984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2854984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2855984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2856984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2857674ae819SStefano Zampini } 2858a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2859906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2860906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2861906d46d4SStefano Zampini 2862906d46d4SStefano Zampini /* assembling of global change of variable */ 2863bbb9e6c6SStefano Zampini { 2864bbb9e6c6SStefano Zampini Mat tmat; 286516f15bc4SStefano Zampini PetscInt bs; 286616f15bc4SStefano Zampini 2867906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2868906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2870bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2872bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 287316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 287416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2875906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2879bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2880bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2881e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2882e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2883bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2884bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2885906d46d4SStefano Zampini } 2886906d46d4SStefano Zampini /* check */ 2887906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2888906d46d4SStefano Zampini PetscReal error; 2889906d46d4SStefano Zampini Vec x,x_change; 2890906d46d4SStefano Zampini 2891906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2892906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2893906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2894906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2895e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2896e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2897bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2898e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2899e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2900906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2901906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2902906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2903906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2904bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2905906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2906906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2907906d46d4SStefano Zampini } 2908b96c3477SStefano Zampini 2909b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2910b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2911b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2912b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2913ac632422SStefano Zampini Mat S_new,tmat; 2914bbb9e6c6SStefano Zampini IS is_all_N; 2915bbb9e6c6SStefano Zampini 2916bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29176816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2918bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2919ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2920b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2921ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2922ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2923ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2924ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2925ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2926b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2927ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2928ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2929ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2930ac632422SStefano Zampini } 2931b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2932b96c3477SStefano Zampini } 2933b96c3477SStefano Zampini } 2934906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2935906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2936b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2937b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2938b9b85e73SStefano Zampini } 2939906d46d4SStefano Zampini 2940906d46d4SStefano Zampini /* set up change of basis context */ 2941906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2942906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2943906d46d4SStefano Zampini 2944906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2945906d46d4SStefano Zampini PetscInt global_size,local_size; 2946906d46d4SStefano Zampini 2947906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2951906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2953906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2954906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2955906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2956906d46d4SStefano Zampini } else { 2957906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2958906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2959906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2960906d46d4SStefano Zampini } 2961906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2962906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2963906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2964906d46d4SStefano Zampini } else { 2965906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2966906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2967906d46d4SStefano Zampini } 2968906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2969906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2970906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2971906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2972b9b85e73SStefano Zampini } 2973a717540cSStefano Zampini 2974727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2975727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2976727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2977aff50787SStefano 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); 2978c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29790e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2980aff50787SStefano 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); 2981727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2982727cdba6SStefano Zampini } 29830e6343abSStefano Zampini } 29840e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2985727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2986727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2987727cdba6SStefano Zampini 2988a717540cSStefano Zampini /* flush dbg viewer */ 2989b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2990b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2991b8ffe317SStefano Zampini } 2992a717540cSStefano Zampini 2993e310c8b4SStefano Zampini /* free workspace */ 2994a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29954641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 299608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29979162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29989162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 299908122e43SStefano Zampini } else { 30009162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 30019162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 30029162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 300308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 300408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 30059162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30069162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 300708122e43SStefano Zampini } 3008674ae819SStefano Zampini PetscFunctionReturn(0); 3009674ae819SStefano Zampini } 3010674ae819SStefano Zampini 3011674ae819SStefano Zampini #undef __FUNCT__ 3012674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3013674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3014674ae819SStefano Zampini { 3015674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3016674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3017674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30187fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3019674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3020674ae819SStefano Zampini 3021674ae819SStefano Zampini PetscFunctionBegin; 30228e61c736SStefano Zampini /* Reset previously computed graph */ 30238e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3024674ae819SStefano Zampini /* Init local Graph struct */ 30257fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30263bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3027674ae819SStefano Zampini 3028575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3029575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3030575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3031575ad6abSStefano Zampini } 30329577ea80SStefano Zampini 3033674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30344d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30354d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30364d379d7bSStefano Zampini PetscInt nvtxs; 3037e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3038674ae819SStefano Zampini 30394d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30402fffb893SStefano Zampini 30412fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30422fffb893SStefano Zampini if (flg_row) { 30434d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3044b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30452fffb893SStefano Zampini } 30462fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30479b28b941SStefano 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 */ 30484d379d7bSStefano Zampini IS is_dummy; 30494d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30504d379d7bSStefano Zampini PetscInt j,sum; 30514d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30524d379d7bSStefano Zampini const PetscInt *idxs; 30534d379d7bSStefano Zampini PCBDDCGraph graph; 30544d379d7bSStefano Zampini PetscBT is_on_boundary; 30554d379d7bSStefano Zampini 30564d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30574d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30584d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30594d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30607fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30614d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3062e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3063e496cd5dSStefano Zampini if (flg_row) { 30644d379d7bSStefano Zampini graph->xadj = xadj; 30654d379d7bSStefano Zampini graph->adjncy = adjncy; 3066e496cd5dSStefano Zampini } 30674d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30684d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3069e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30704d379d7bSStefano Zampini 30714d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30729b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30734d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30744d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30754d379d7bSStefano Zampini } 30764d379d7bSStefano Zampini } 30774d379d7bSStefano Zampini 3078e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30794d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30804d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30814d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30824d379d7bSStefano Zampini } 30834d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30844d379d7bSStefano Zampini 3085e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30864d379d7bSStefano Zampini sum = 0; 30874d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30884d379d7bSStefano Zampini PetscInt sizecc = 0; 30894d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30904d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30914d379d7bSStefano Zampini sizecc++; 30924d379d7bSStefano Zampini } 30934d379d7bSStefano Zampini } 30944d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30954d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30964d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30974d379d7bSStefano Zampini } 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini sum += sizecc*sizecc; 31004d379d7bSStefano Zampini } 31014d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 31024d379d7bSStefano Zampini sum = 0; 3103e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 31044d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 31054d379d7bSStefano Zampini cxadj[i] = sum; 31064d379d7bSStefano Zampini sum += temp; 31074d379d7bSStefano Zampini } 3108e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31094d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31104d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31114d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31124d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31134d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31144d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31154d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31164d379d7bSStefano Zampini sizecc++; 31174d379d7bSStefano Zampini } 31184d379d7bSStefano Zampini } 31194d379d7bSStefano Zampini } 31204d379d7bSStefano Zampini } 31214d379d7bSStefano Zampini } 31229b28b941SStefano Zampini if (sum) { 3123e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31244d379d7bSStefano Zampini } else { 31254d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31264d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31274d379d7bSStefano Zampini } 31284d379d7bSStefano Zampini graph->xadj = 0; 31294d379d7bSStefano Zampini graph->adjncy = 0; 31304d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31314d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31324d379d7bSStefano Zampini } 3133674ae819SStefano Zampini } 31349b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31359b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3136674ae819SStefano Zampini } 3137674ae819SStefano Zampini 313863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3139674ae819SStefano Zampini vertex_size = 1; 314063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 314163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 314295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 314363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3144e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3146674ae819SStefano Zampini } 314763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 314863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 314963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3150674ae819SStefano Zampini } 315163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3152674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 315363602bcaSStefano Zampini } else { 315463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 315563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3156854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 315763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 315863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 315963602bcaSStefano Zampini } 316063602bcaSStefano Zampini } 3161674ae819SStefano Zampini } 3162674ae819SStefano Zampini 3163674ae819SStefano Zampini /* Setup of Graph */ 3164785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3165e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3166785d1243SStefano Zampini } 3167785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3168e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3169785d1243SStefano Zampini } 3170302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3171674ae819SStefano Zampini 3172674ae819SStefano Zampini /* Graph's connected components analysis */ 3173674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3174674ae819SStefano Zampini 3175674ae819SStefano Zampini /* print some info to stdout */ 3176674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3177302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3178674ae819SStefano Zampini } 3179fb180af4SStefano Zampini 3180fb180af4SStefano Zampini /* mark topography has done */ 3181fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3182674ae819SStefano Zampini PetscFunctionReturn(0); 3183674ae819SStefano Zampini } 3184674ae819SStefano Zampini 3185dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3186674ae819SStefano Zampini #undef __FUNCT__ 3187674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3188dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3189674ae819SStefano Zampini { 3190dc456d91SStefano Zampini PetscSF sf; 3191dc456d91SStefano Zampini PetscLayout map; 3192dc456d91SStefano Zampini const PetscInt *idxs; 3193dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3194dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3195dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3196dc456d91SStefano Zampini PetscMPIInt commsize; 3197674ae819SStefano Zampini PetscBool first_found; 3198674ae819SStefano Zampini PetscErrorCode ierr; 3199674ae819SStefano Zampini 3200674ae819SStefano Zampini PetscFunctionBegin; 3201dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3202dc456d91SStefano Zampini if (subset_mult) { 3203dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3204dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3205dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3206674ae819SStefano Zampini } 3207dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3208dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3209dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3210dc456d91SStefano Zampini for (i=0;i<n;i++) { 3211dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3212dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3213674ae819SStefano Zampini } 3214dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3215dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3216dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3217dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3218dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3219dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3220dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3221dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3222dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3223dc456d91SStefano Zampini 3224dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3225dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3226dc456d91SStefano Zampini if (subset_mult) { 3227dc456d91SStefano Zampini const PetscInt* idxs_mult; 3228dc456d91SStefano Zampini 3229dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3230dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3231dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3232674ae819SStefano Zampini } else { 3233dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3234674ae819SStefano Zampini } 3235dc456d91SStefano Zampini /* local size of new subset */ 3236dc456d91SStefano Zampini n_n = 0; 3237dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3238dc456d91SStefano Zampini 3239dc456d91SStefano Zampini /* global indexes in layout */ 3240dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3241dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3242dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3243dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3244dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3245dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3246dc456d91SStefano Zampini 3247dc456d91SStefano Zampini /* reduce from leaves to roots */ 3248dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 324964a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 325064a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3251dc456d91SStefano Zampini 3252dc456d91SStefano Zampini /* count indexes in local part of layout */ 3253674ae819SStefano Zampini nlocals = 0; 3254674ae819SStefano Zampini first_index = -1; 3255674ae819SStefano Zampini first_found = PETSC_FALSE; 3256dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3257dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3258674ae819SStefano Zampini first_found = PETSC_TRUE; 3259674ae819SStefano Zampini first_index = i; 3260674ae819SStefano Zampini } 3261dc456d91SStefano Zampini nlocals += root_data[i]; 3262674ae819SStefano Zampini } 3263dc456d91SStefano Zampini 3264dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32655fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3266dc456d91SStefano Zampini start = 0; 326764a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32685fa240b1SStefano Zampini #else 326964a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32705fa240b1SStefano Zampini start = start-nlocals; 32715fa240b1SStefano Zampini #endif 32725fa240b1SStefano Zampini 3273dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3274dc456d91SStefano Zampini *N_n = start + nlocals; 3275dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3276dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3277674ae819SStefano Zampini } 32785fa240b1SStefano Zampini 32795fa240b1SStefano Zampini /* adapt root data with cumulative */ 3280674ae819SStefano Zampini if (first_found) { 3281dc456d91SStefano Zampini PetscInt old_index; 3282dc456d91SStefano Zampini 3283dc456d91SStefano Zampini root_data[first_index] += start; 3284674ae819SStefano Zampini old_index = first_index; 3285dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3286dc456d91SStefano Zampini if (root_data[i]) { 3287dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3288674ae819SStefano Zampini old_index = i; 3289674ae819SStefano Zampini } 3290674ae819SStefano Zampini } 3291674ae819SStefano Zampini } 3292dc456d91SStefano Zampini 3293dc456d91SStefano Zampini /* from roots to leaves */ 3294dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3295dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3296dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3297dc456d91SStefano Zampini 3298dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3299dc456d91SStefano Zampini if (subset_mult) { 3300dc456d91SStefano Zampini const PetscInt* idxs_mult; 3301dc456d91SStefano Zampini PetscInt cum; 3302dc456d91SStefano Zampini 3303dc456d91SStefano Zampini cum = 0; 3304dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3305dc456d91SStefano Zampini for (i=0;i<n;i++) { 3306dc456d91SStefano Zampini PetscInt j; 3307dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3308674ae819SStefano Zampini } 3309dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3310674ae819SStefano Zampini } else { 3311dc456d91SStefano Zampini for (i=0;i<n;i++) { 3312dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3313674ae819SStefano Zampini } 3314674ae819SStefano Zampini } 3315dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3316dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3317674ae819SStefano Zampini PetscFunctionReturn(0); 3318674ae819SStefano Zampini } 33199a7d3425SStefano Zampini 33209a7d3425SStefano Zampini #undef __FUNCT__ 33219a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33229a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33239a7d3425SStefano Zampini { 33249a7d3425SStefano Zampini PetscInt i,j; 33259a7d3425SStefano Zampini PetscScalar *alphas; 33269a7d3425SStefano Zampini PetscErrorCode ierr; 33279a7d3425SStefano Zampini 33289a7d3425SStefano Zampini PetscFunctionBegin; 33299a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3330785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33319a7d3425SStefano Zampini for (i=0;i<n;i++) { 33329a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33339a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33349a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33359a7d3425SStefano Zampini } 33369a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33379a7d3425SStefano Zampini PetscFunctionReturn(0); 33389a7d3425SStefano Zampini } 33399a7d3425SStefano Zampini 3340e7931f94SStefano Zampini #undef __FUNCT__ 334170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3342b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3343e7931f94SStefano Zampini { 334452e5ac9dSStefano Zampini IS ranks_send_to; 3345e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3346e7931f94SStefano Zampini PetscMPIInt size,rank,color; 334752e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 334852e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33493837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33502b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3351e7931f94SStefano Zampini PetscSubcomm subcomm; 335252e5ac9dSStefano Zampini PetscErrorCode ierr; 3353a57a6d2fSStefano Zampini 3354e7931f94SStefano Zampini PetscFunctionBegin; 33552b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33562b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33572b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3358e7931f94SStefano Zampini 3359e7931f94SStefano Zampini /* Get info on mapping */ 33603bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3362e7931f94SStefano Zampini 3363e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3364785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3365e7931f94SStefano Zampini xadj[0] = 0; 3366e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3367785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3368785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3369e7931f94SStefano Zampini 33702b510759SStefano Zampini if (threshold) { 3371d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33722b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3373d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3374d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3375d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3376d023bfaeSStefano Zampini xadj_count++; 3377e7931f94SStefano Zampini } 3378e7931f94SStefano Zampini } 3379d023bfaeSStefano Zampini xadj[1] = xadj_count; 3380c8587f34SStefano Zampini } else { 3381e7931f94SStefano Zampini if (xadj[1]) { 3382e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3383e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3384c8587f34SStefano Zampini } 3385e7931f94SStefano Zampini } 33863bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3387e7931f94SStefano Zampini if (use_square) { 3388e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3389e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3390e7931f94SStefano Zampini } 3391e7931f94SStefano Zampini } 3392e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3393e7931f94SStefano Zampini 33943837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3395e7931f94SStefano Zampini 3396e7931f94SStefano Zampini /* 3397e7931f94SStefano Zampini Restrict work on active processes only. 3398e7931f94SStefano Zampini */ 3399e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3400e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3401e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 34022b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3403d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3404e7931f94SStefano Zampini if (color) { 3405e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3406e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3407e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3408c8587f34SStefano Zampini } else { 340952e5ac9dSStefano Zampini Mat subdomain_adj; 341052e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 341152e5ac9dSStefano Zampini MatPartitioning partitioner; 341252e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 341352e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3414b0c7d250SStefano Zampini PetscBool aggregate; 3415b0c7d250SStefano Zampini 3416306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3417785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3418e7931f94SStefano Zampini prank = rank; 3419306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34208002ef2cSStefano Zampini /* 3421e7931f94SStefano Zampini for (i=0;i<size;i++) { 3422e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3423c8587f34SStefano Zampini } 34248002ef2cSStefano Zampini */ 3425e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3426e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3427c8587f34SStefano Zampini } 3428e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3429b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3430b0c7d250SStefano Zampini if (aggregate) { 3431b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3432b0c7d250SStefano Zampini PetscMPIInt nrank; 3433b0c7d250SStefano Zampini PetscScalar *vals; 3434b0c7d250SStefano Zampini 3435b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3436b0c7d250SStefano Zampini lrows = 0; 3437b0c7d250SStefano Zampini if (nrank<redprocs) { 3438b0c7d250SStefano Zampini lrows = size/redprocs; 3439b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3440b0c7d250SStefano Zampini } 34415fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3442b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3443b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3444b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3445b0c7d250SStefano Zampini row = nrank; 3446b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3447b0c7d250SStefano Zampini cols = adjncy; 3448b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3449b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3450b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3451b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3452b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 345352e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 345452e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 345552e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3456b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3457b0c7d250SStefano Zampini } else { 3458306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3459b0c7d250SStefano Zampini } 346022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3461e7931f94SStefano Zampini 3462e7931f94SStefano Zampini /* Partition */ 3463306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3464e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3465e7931f94SStefano Zampini if (use_vwgt) { 34663837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3467e7931f94SStefano Zampini v_wgt[0] = local_size; 3468e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3469c8587f34SStefano Zampini } 347028143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 347128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3472e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3473e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 347422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3475e7931f94SStefano Zampini 347652e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 347752e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 347852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 347952e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3480b0c7d250SStefano Zampini if (!redprocs) { 3481b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 348228143c3dSStefano Zampini } else { 3483b0c7d250SStefano Zampini PetscInt idxs[1]; 3484b0c7d250SStefano Zampini PetscMPIInt tag; 3485b0c7d250SStefano Zampini MPI_Request *reqs; 3486b0c7d250SStefano Zampini 3487b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3488b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3489b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3490b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 349128143c3dSStefano Zampini } 3492b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3493b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3494b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3495b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3496e7931f94SStefano Zampini } 349752e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3498e7931f94SStefano Zampini /* clean up */ 3499e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 350052e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3501e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3502e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3503e7931f94SStefano Zampini } 3504e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3505e7931f94SStefano Zampini 3506e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3507e7931f94SStefano Zampini i = 1; 3508e7931f94SStefano Zampini if (color) i=0; 3509e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3510e7931f94SStefano Zampini /* get back IS */ 3511e7931f94SStefano Zampini *is_sends = ranks_send_to; 3512e7931f94SStefano Zampini PetscFunctionReturn(0); 3513e7931f94SStefano Zampini } 3514e7931f94SStefano Zampini 3515e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3516e7931f94SStefano Zampini 3517e7931f94SStefano Zampini #undef __FUNCT__ 3518e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 351953a05cb3SStefano 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[]) 3520e7931f94SStefano Zampini { 352170cf5478SStefano Zampini Mat local_mat; 3522e7931f94SStefano Zampini IS is_sends_internal; 35239d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 352428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35259d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3526e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3527e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3528e7931f94SStefano Zampini const PetscInt* is_indices; 3529e7931f94SStefano Zampini MatType new_local_type; 3530e7931f94SStefano Zampini /* buffers */ 3531e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 353228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35339d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3534e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3535e7931f94SStefano Zampini /* MPI */ 353628143c3dSStefano Zampini MPI_Comm comm,comm_n; 353728143c3dSStefano Zampini PetscSubcomm subcomm; 3538e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 353928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 354028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 354128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 354228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 354328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3544e7931f94SStefano Zampini PetscErrorCode ierr; 3545e7931f94SStefano Zampini 3546e7931f94SStefano Zampini PetscFunctionBegin; 354728143c3dSStefano Zampini /* TODO: add missing checks */ 354828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 354928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 355028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 355128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3552e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 355328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3554e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3555e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3556e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3557e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3558e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 355928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 356070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 356170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 356228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 356370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 356470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 356570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 356670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 356770cf5478SStefano Zampini } 3568e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3569e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3570e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3571e7931f94SStefano Zampini if (!is_sends) { 357228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3573b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3574c8587f34SStefano Zampini } else { 3575e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3576e7931f94SStefano Zampini is_sends_internal = is_sends; 3577c8587f34SStefano Zampini } 3578e7931f94SStefano Zampini 3579e7931f94SStefano Zampini /* get comm */ 3580a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3581e7931f94SStefano Zampini 3582e7931f94SStefano Zampini /* compute number of sends */ 3583e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3584e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3585e7931f94SStefano Zampini 3586e7931f94SStefano Zampini /* compute number of receives */ 3587e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3588785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3589e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3590e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3591e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3592e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3593e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3594e7931f94SStefano Zampini 359528143c3dSStefano Zampini /* restrict comm if requested */ 359628143c3dSStefano Zampini subcomm = 0; 359728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 359828143c3dSStefano Zampini if (restrict_comm) { 3599779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3600779c1cceSStefano Zampini 360128143c3dSStefano Zampini color = 0; 360253a05cb3SStefano Zampini if (restrict_full) { 360353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 360453a05cb3SStefano Zampini } else { 360553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 360653a05cb3SStefano Zampini } 360728143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 360828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 360928143c3dSStefano Zampini /* check if reuse has been requested */ 361028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 361128143c3dSStefano Zampini if (*mat_n) { 361228143c3dSStefano Zampini PetscMPIInt subcommsize2; 361328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 361428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 361528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 361628143c3dSStefano Zampini } else { 361728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 361828143c3dSStefano Zampini } 361928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3620779c1cceSStefano Zampini PetscMPIInt rank; 3621779c1cceSStefano Zampini 3622779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 362328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 362428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 362528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3626306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 362728143c3dSStefano Zampini } 362828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 362928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 363028143c3dSStefano Zampini } else { 363128143c3dSStefano Zampini comm_n = comm; 363228143c3dSStefano Zampini } 363328143c3dSStefano Zampini 3634e7931f94SStefano Zampini /* prepare send/receive buffers */ 3635785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3636e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3637785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3638e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 363928143c3dSStefano Zampini if (nis) { 3640854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 364128143c3dSStefano Zampini } 3642e7931f94SStefano Zampini 364328143c3dSStefano Zampini /* Get data from local matrices */ 3644e7931f94SStefano Zampini if (!isdense) { 3645a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3646e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3647e7931f94SStefano Zampini /* 3648e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3649e7931f94SStefano Zampini send_buffer_idxs should contain: 3650e7931f94SStefano Zampini - MatType_PRIVATE type 3651e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3652e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3653e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3654e7931f94SStefano Zampini */ 3655e7931f94SStefano Zampini } else { 3656e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36573bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3658854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3659e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3660e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3662e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36633bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3664e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3665e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3666e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3667e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3668c8587f34SStefano Zampini } 3669c8587f34SStefano Zampini } 3670e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 367128143c3dSStefano Zampini /* additional is (if any) */ 367228143c3dSStefano Zampini if (nis) { 367328143c3dSStefano Zampini PetscMPIInt psum; 367428143c3dSStefano Zampini PetscInt j; 367528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 367628143c3dSStefano Zampini PetscInt plen; 367728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 367828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 367928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 368028143c3dSStefano Zampini } 3681854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 368228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 368328143c3dSStefano Zampini PetscInt plen; 368428143c3dSStefano Zampini const PetscInt *is_array_idxs; 368528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 368628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 368728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 368828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 368928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 369028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 369128143c3dSStefano Zampini } 369228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 369328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 369428143c3dSStefano Zampini } 369528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 369628143c3dSStefano Zampini } 369728143c3dSStefano Zampini 3698e7931f94SStefano Zampini buf_size_idxs = 0; 3699e7931f94SStefano Zampini buf_size_vals = 0; 370028143c3dSStefano Zampini buf_size_idxs_is = 0; 3701e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3702e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3703e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 370428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3705e7931f94SStefano Zampini } 3706785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3707785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 370895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3709e7931f94SStefano Zampini 3710e7931f94SStefano Zampini /* get new tags for clean communications */ 3711e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3712e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 371328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3714e7931f94SStefano Zampini 3715e7931f94SStefano Zampini /* allocate for requests */ 3716785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3717785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 371895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3719785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3720785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 372195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3722e7931f94SStefano Zampini 3723e7931f94SStefano Zampini /* communications */ 3724e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3725e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 372628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3727e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3728e7931f94SStefano Zampini source_dest = onodes[i]; 3729e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3730e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3731e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3732e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 373328143c3dSStefano Zampini if (nis) { 373428143c3dSStefano 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); 373528143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 373628143c3dSStefano Zampini } 3737e7931f94SStefano Zampini } 3738e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3739e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3740e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3741e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 374228143c3dSStefano Zampini if (nis) { 374328143c3dSStefano 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); 374428143c3dSStefano Zampini } 3745e7931f94SStefano Zampini } 3746e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3747e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3748e7931f94SStefano Zampini 3749e7931f94SStefano Zampini /* assemble new l2g map */ 3750e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3751e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37529d30be91SStefano Zampini new_local_rows = 0; 3753e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37549d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3755e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3756e7931f94SStefano Zampini } 37579d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3758e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37599d30be91SStefano Zampini new_local_rows = 0; 3760e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37619d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37629d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3763e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3764e7931f94SStefano Zampini } 37659d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37669d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3767e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3768e7931f94SStefano Zampini 3769e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3770e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3771e7931f94SStefano 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) */ 3772e7931f94SStefano Zampini if (n_recvs) { 377328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3774e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3775e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3776e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3777e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3778e7931f94SStefano Zampini break; 3779e7931f94SStefano Zampini } 3780e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3781e7931f94SStefano Zampini } 3782e7931f94SStefano Zampini switch (new_local_type_private) { 378328143c3dSStefano Zampini case MATDENSE_PRIVATE: 378428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3785e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3786e7931f94SStefano Zampini bs = 1; 378728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 378828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 378928143c3dSStefano Zampini bs = 1; 379028143c3dSStefano Zampini } 3791e7931f94SStefano Zampini break; 3792e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3793e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3794e7931f94SStefano Zampini bs = 1; 3795e7931f94SStefano Zampini break; 3796e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3797e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3798e7931f94SStefano Zampini break; 3799e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3800e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3801e7931f94SStefano Zampini break; 3802e7931f94SStefano Zampini default: 38039d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3804e7931f94SStefano Zampini break; 3805e7931f94SStefano Zampini } 380628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 380728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 380828143c3dSStefano Zampini bs = 1; 3809e7931f94SStefano Zampini } 3810e7931f94SStefano Zampini 381170cf5478SStefano Zampini /* create MATIS object if needed */ 381270cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3813e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3814e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 381570cf5478SStefano Zampini } else { 381670cf5478SStefano Zampini /* it also destroys the local matrices */ 381770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 381870cf5478SStefano Zampini } 381970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3820e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 38219d30be91SStefano Zampini 38229d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38239d30be91SStefano Zampini 38249d30be91SStefano Zampini /* Global to local map of received indices */ 38259d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38269d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38279d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38289d30be91SStefano Zampini 38299d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38309d30be91SStefano Zampini buf_size_idxs = 0; 38319d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38329d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38339d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38349d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38359d30be91SStefano Zampini } 38369d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38379d30be91SStefano Zampini 38389d30be91SStefano Zampini /* set preallocation */ 38399d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38409d30be91SStefano Zampini if (!newisdense) { 38419d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38429d30be91SStefano Zampini 38439d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38449d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38459d30be91SStefano Zampini if (n_recvs) { 38469d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38479d30be91SStefano Zampini } 38489d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38499d30be91SStefano Zampini PetscInt j; 38509d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38519d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38529d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38539d30be91SStefano Zampini } 38549d30be91SStefano Zampini } else { 38559d30be91SStefano Zampini /* TODO */ 38569d30be91SStefano Zampini } 38579d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38589d30be91SStefano Zampini } 38599d30be91SStefano Zampini if (new_local_nnz) { 38609d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38619d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38629d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38639d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38649d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38659d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38669d30be91SStefano Zampini } else { 38679d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38689d30be91SStefano Zampini } 38699d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38709d30be91SStefano Zampini } else { 38719d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38729d30be91SStefano Zampini } 3873e7931f94SStefano Zampini 3874e7931f94SStefano Zampini /* set values */ 3875e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38769d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3877e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3878e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3879e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38809d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3881e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3882e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3883e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 388428143c3dSStefano Zampini } else { 388528143c3dSStefano Zampini /* TODO */ 3886e7931f94SStefano Zampini } 3887e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3888e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3889e7931f94SStefano Zampini } 3890e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3891e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38949d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38959d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3896e7931f94SStefano Zampini 3897dfd14d43SStefano Zampini #if 0 389828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3899e7931f94SStefano Zampini Vec lvec,rvec; 3900e7931f94SStefano Zampini PetscReal infty_error; 3901e7931f94SStefano Zampini 39022a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3903e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3904e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 390670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3907e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3909e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3910e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3911e7931f94SStefano Zampini } 391228143c3dSStefano Zampini #endif 3913e7931f94SStefano Zampini 391428143c3dSStefano Zampini /* assemble new additional is (if any) */ 391528143c3dSStefano Zampini if (nis) { 391628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 391728143c3dSStefano Zampini 391828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3919854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 392028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 392128143c3dSStefano Zampini psum = 0; 392228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 392328143c3dSStefano Zampini for (j=0;j<nis;j++) { 392428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 392528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 392628143c3dSStefano Zampini psum += plen; 392728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 392828143c3dSStefano Zampini } 392928143c3dSStefano Zampini } 3930854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3931854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 393228143c3dSStefano Zampini for (i=1;i<nis;i++) { 393328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 393428143c3dSStefano Zampini } 393528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 393628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 393728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 393828143c3dSStefano Zampini for (j=0;j<nis;j++) { 393928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 394028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 394128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 394228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 394328143c3dSStefano Zampini } 394428143c3dSStefano Zampini } 394528143c3dSStefano Zampini for (i=0;i<nis;i++) { 394628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 394728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 394828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 394928143c3dSStefano Zampini } 395028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 395128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 395228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 395328143c3dSStefano Zampini } 3954e7931f94SStefano Zampini /* free workspace */ 395528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3956e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3958e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3959e7931f94SStefano Zampini if (isdense) { 3960e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3961e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3962e7931f94SStefano Zampini } else { 3963e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3964e7931f94SStefano Zampini } 396528143c3dSStefano Zampini if (nis) { 396628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 396728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 396828143c3dSStefano Zampini } 3969e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3970e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 397128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3972e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3973e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 397428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3975e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3976e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3977e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3978e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3979e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 398028143c3dSStefano Zampini if (nis) { 398128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 398228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 398328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 398428143c3dSStefano Zampini } 398528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 398628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 398728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 398828143c3dSStefano Zampini for (i=0;i<nis;i++) { 398928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 399028143c3dSStefano Zampini } 399153a05cb3SStefano Zampini *mat_n = NULL; 399228143c3dSStefano Zampini } 3993e7931f94SStefano Zampini PetscFunctionReturn(0); 3994e7931f94SStefano Zampini } 3995a57a6d2fSStefano Zampini 399612edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3997af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 399812edc857SStefano Zampini 3999c8587f34SStefano Zampini #undef __FUNCT__ 4000c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4001c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4002c8587f34SStefano Zampini { 4003c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4004c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 400520a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 40069881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 400720a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 40086e683305SStefano Zampini IS coarse_is,*isarray; 40096e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 40106e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4011f9eb5b7dSStefano Zampini PC pc_temp; 4012c8587f34SStefano Zampini PCType coarse_pc_type; 4013c8587f34SStefano Zampini KSPType coarse_ksp_type; 4014f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 40154f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 40166e683305SStefano Zampini Mat t_coarse_mat_is; 40176e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 40186e683305SStefano Zampini PetscMPIInt all_procs; 401974e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 402068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 402122bc73bbSStefano Zampini PetscScalar *array; 40229881197aSStefano Zampini PetscErrorCode ierr; 4023fdc09c96SStefano Zampini 4024c8587f34SStefano Zampini PetscFunctionBegin; 4025c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 402668457ee5SStefano 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 */ 4027fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40285a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4029fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4030f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4031f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4032f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4033fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 403451bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 403551bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4036dc4bcba2SStefano Zampini PC pc; 4037dc4bcba2SStefano Zampini PetscBool isbddc; 4038dc4bcba2SStefano Zampini 4039dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4040dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4041dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4042dc4bcba2SStefano Zampini if (isbddc) { 4043dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4044dc4bcba2SStefano Zampini } 4045727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4046fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4047fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4048fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4049f4ddd8eeSStefano Zampini } 4050fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4051fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4052f4ddd8eeSStefano Zampini } 405370cf5478SStefano Zampini /* reset any subassembling information */ 405470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40556e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40566e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4057fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4058f4ddd8eeSStefano Zampini } 4059c8587f34SStefano Zampini 40606e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40612b510759SStefano Zampini im_active = !!(pcis->n); 40622b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40636e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40646e683305SStefano Zampini void_procs = all_procs-active_procs; 40656e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 406674e2c79eSStefano Zampini redist = PETSC_FALSE; 406722bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40686e683305SStefano Zampini csin_ml = PETSC_TRUE; 40696e683305SStefano Zampini ncoarse_ml = void_procs; 4070779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4071779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40726e683305SStefano Zampini csin_ds = PETSC_TRUE; 407318a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 407418a45a71SStefano Zampini redist = PETSC_TRUE; 407518a45a71SStefano Zampini } else { 40766e683305SStefano Zampini csin_ds = PETSC_TRUE; 4077779c1cceSStefano Zampini ncoarse_ds = active_procs; 4078779c1cceSStefano Zampini redist = PETSC_TRUE; 407918a45a71SStefano Zampini } 40806e683305SStefano Zampini } else { 40816e683305SStefano Zampini csin_ml = PETSC_FALSE; 40826e683305SStefano Zampini ncoarse_ml = all_procs; 40836e683305SStefano Zampini if (void_procs) { 40846e683305SStefano Zampini csin_ds = PETSC_TRUE; 40856e683305SStefano Zampini ncoarse_ds = void_procs; 40866e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40876e683305SStefano Zampini } else { 4088779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 408974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 409074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 409174e2c79eSStefano Zampini redist = PETSC_TRUE; 409274e2c79eSStefano Zampini } else { 40936e683305SStefano Zampini csin_ds = PETSC_FALSE; 40946e683305SStefano Zampini ncoarse_ds = all_procs; 40956e683305SStefano Zampini } 40966e683305SStefano Zampini } 409774e2c79eSStefano Zampini } 40986e683305SStefano Zampini 40996e683305SStefano Zampini /* 41006e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 41016e683305SStefano Zampini - we have not exceeded the number of levels requested 41026e683305SStefano Zampini - we can actually subassemble the active processes 41036e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 41046e683305SStefano Zampini */ 41056e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 41066e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 41076e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 41086e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 41096e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4110f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 41112b510759SStefano Zampini } else { 4112f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4113c8587f34SStefano Zampini } 4114c8587f34SStefano Zampini } 41156e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 41166e683305SStefano Zampini if (multilevel_allowed) { 41176e683305SStefano Zampini ncoarse = ncoarse_ml; 41186e683305SStefano Zampini csin = csin_ml; 411958da7f69SStefano Zampini redist = PETSC_FALSE; 41206e683305SStefano Zampini } else { 41216e683305SStefano Zampini ncoarse = ncoarse_ds; 41226e683305SStefano Zampini csin = csin_ds; 41236e683305SStefano Zampini } 4124e7931f94SStefano Zampini 4125abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4126abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4127abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4128abbbba34SStefano Zampini 4129abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 413022bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 413122bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 413222bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 413322bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4134b9b85e73SStefano Zampini #if 0 4135b9b85e73SStefano Zampini { 4136b9b85e73SStefano Zampini PetscViewer viewer; 4137b9b85e73SStefano Zampini char filename[256]; 4138b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4139b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4140b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4141b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4142b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4143b9b85e73SStefano Zampini } 4144b9b85e73SStefano Zampini #endif 4145e176bc59SStefano 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); 41466e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41476e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41486e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4149abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4150abbbba34SStefano Zampini 41516e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41526e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41536e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41546e683305SStefano Zampini const PetscInt *idxs; 41556e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41566e683305SStefano Zampini 41576e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41580be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41596e683305SStefano Zampini /* allocate space for temporary storage */ 4160854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4161854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41626e683305SStefano Zampini /* allocate for IS array */ 41636e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41646e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41656e683305SStefano Zampini nis = nisdofs + nisneu; 4166854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41676e683305SStefano Zampini /* dofs splitting */ 41686e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41696e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41706e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41716e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41726e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41736e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41746e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41756e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41766e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41776e683305SStefano Zampini } 41786e683305SStefano Zampini /* neumann boundaries */ 41796e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41806e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41816e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41836e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41846e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41856e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41866e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41876e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41886e683305SStefano Zampini } 41896e683305SStefano Zampini /* free memory */ 41906e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41916e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41926e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41936e683305SStefano Zampini } else { 41946e683305SStefano Zampini nis = 0; 41956e683305SStefano Zampini nisdofs = 0; 41966e683305SStefano Zampini nisneu = 0; 41976e683305SStefano Zampini isarray = NULL; 41986e683305SStefano Zampini } 41996e683305SStefano Zampini /* destroy no longer needed map */ 42006e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 42016e683305SStefano Zampini 42026e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 42036e683305SStefano Zampini coarse_mat_is = NULL; 42046e683305SStefano Zampini if (csin) { 42056e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 420674e2c79eSStefano Zampini if (redist) { 420774e2c79eSStefano Zampini PetscMPIInt rank; 4208779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 420974e2c79eSStefano Zampini 421074e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 421158da7f69SStefano Zampini spc = active_procs/ncoarse; 421258da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4213779c1cceSStefano Zampini if (im_active) { 4214779c1cceSStefano Zampini destsize = 1; 421574e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 421674e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 421774e2c79eSStefano Zampini } else { 421874e2c79eSStefano Zampini dest[0] = rank/(spc+1); 421974e2c79eSStefano Zampini } 422074e2c79eSStefano Zampini } else { 4221779c1cceSStefano Zampini destsize = 0; 42226e683305SStefano Zampini } 4223779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4224779c1cceSStefano Zampini } else if (csin_type_simple) { 42256e683305SStefano Zampini PetscMPIInt rank; 42266e683305SStefano Zampini PetscInt issize,isidx; 4227779c1cceSStefano Zampini 42286e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42296e683305SStefano Zampini if (im_active) { 42306e683305SStefano Zampini issize = 1; 42316e683305SStefano Zampini isidx = (PetscInt)rank; 42326e683305SStefano Zampini } else { 42336e683305SStefano Zampini issize = 0; 42346e683305SStefano Zampini isidx = -1; 42356e683305SStefano Zampini } 42366e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4237779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4238b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42396e683305SStefano Zampini } 4240779c1cceSStefano Zampini 4241779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4242779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4243779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4244779c1cceSStefano Zampini PetscInt *coarse_candidates; 4245779c1cceSStefano Zampini const PetscInt* tisindices; 4246779c1cceSStefano Zampini 4247779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4248779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4249779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4250779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4251779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4252779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4253779c1cceSStefano Zampini } 4254779c1cceSStefano Zampini } 4255779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4256779c1cceSStefano Zampini 4257779c1cceSStefano Zampini 42586e683305SStefano Zampini if (pcbddc->dbg_flag) { 42596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42606e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42616e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4263779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42656e683305SStefano Zampini } 42666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42676e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42686e683305SStefano Zampini } 42696e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42706e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42716e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4272854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42736e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42746e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42756e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42766e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42776e683305SStefano Zampini } 42786e683305SStefano Zampini if (pcbddc->dbg_flag) { 42796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42816e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42826e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42836e683305SStefano Zampini } 4284779c1cceSStefano Zampini } 42856e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 428653a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 428753a05cb3SStefano 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); 428853a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 428953a05cb3SStefano 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); 429053a05cb3SStefano Zampini } 42916e683305SStefano Zampini } else { 42926e683305SStefano Zampini if (pcbddc->dbg_flag) { 42936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42956e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42966e683305SStefano Zampini } 42976e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42986e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42996e683305SStefano Zampini } 43006e683305SStefano Zampini 43016e683305SStefano Zampini /* create local to global scatters for coarse problem */ 430268457ee5SStefano Zampini if (compute_vecs) { 43036e683305SStefano Zampini PetscInt lrows; 43046e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43056e683305SStefano Zampini if (coarse_mat_is) { 43066e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 43076e683305SStefano Zampini } else { 43086e683305SStefano Zampini lrows = 0; 43096e683305SStefano Zampini } 43106e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 43116e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 43126e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 43136e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43146e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43156e683305SStefano Zampini } 43166e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 43176e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4318c8587f34SStefano Zampini 4319f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4320f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4321f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4322f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4323f9eb5b7dSStefano Zampini } else { 4324f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4325f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4326c8587f34SStefano Zampini } 4327c8587f34SStefano Zampini 43286e683305SStefano Zampini /* print some info if requested */ 43296e683305SStefano Zampini if (pcbddc->dbg_flag) { 43306e683305SStefano Zampini if (!multilevel_allowed) { 43316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43326e683305SStefano Zampini if (multilevel_requested) { 43336e683305SStefano 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); 43346e683305SStefano Zampini } else if (pcbddc->max_levels) { 43356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43366e683305SStefano Zampini } 43376e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43386e683305SStefano Zampini } 43396e683305SStefano Zampini } 43406e683305SStefano Zampini 4341f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43426e683305SStefano Zampini if (coarse_mat_is) { 43436e683305SStefano Zampini MatReuse coarse_mat_reuse; 43446a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43456e683305SStefano Zampini if (pcbddc->dbg_flag) { 43466e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43476e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43486e683305SStefano Zampini } 4349f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4350312be037SStefano Zampini char prefix[256],str_level[16]; 4351e604994aSStefano Zampini size_t len; 43526e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4353422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4354c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4355f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43565f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4357c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43586e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4359c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4360c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4361e604994aSStefano Zampini /* prefix */ 4362e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4363e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4364e604994aSStefano Zampini if (!pcbddc->current_level) { 4365e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4366e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4367c8587f34SStefano Zampini } else { 4368e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4369312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4370312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 437134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4372312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4373e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4374e604994aSStefano Zampini } 4375e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43763e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43773e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43783e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43793e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4380f9eb5b7dSStefano Zampini /* allow user customization */ 4381f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43823e3c6dadSStefano Zampini } 43833e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 438451bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43853e3c6dadSStefano Zampini if (nisdofs) { 43863e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43873e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43883e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43893e3c6dadSStefano Zampini } 43903e3c6dadSStefano Zampini } 43913e3c6dadSStefano Zampini if (nisneu) { 43923e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43933e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4394312be037SStefano Zampini } 4395f9eb5b7dSStefano Zampini 4396f9eb5b7dSStefano Zampini /* get some info after set from options */ 4397f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4398f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43994f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 44006e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4401f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4402f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4403f9eb5b7dSStefano Zampini } 440439f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 44054f3a063dSStefano Zampini if (isredundant) { 44064f3a063dSStefano Zampini KSP inner_ksp; 44074f3a063dSStefano Zampini PC inner_pc; 44084f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 44094f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 44104f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 44114f3a063dSStefano Zampini } 4412f9eb5b7dSStefano Zampini 4413f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4414fa7f1dd8SStefano Zampini if (coarse_reuse) { 441581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4416fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 44176e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4418fa7f1dd8SStefano Zampini } else { 44196e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4420fa7f1dd8SStefano Zampini } 4421c8587f34SStefano Zampini if (isbddc || isnn) { 442222bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 442370cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4424b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 442522b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44296e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 443022b6e8a2SStefano Zampini } 443170cf5478SStefano Zampini } 443253a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 443370cf5478SStefano Zampini } else { 443422bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 443522bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 443622bc73bbSStefano Zampini } 443722bc73bbSStefano Zampini } else { 44382e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4439c8587f34SStefano Zampini } 4440c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4441c8587f34SStefano Zampini 44423301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44435a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44443301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44453301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44463301b35fSStefano Zampini } 44473301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44483301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44493301b35fSStefano Zampini } 44503301b35fSStefano Zampini if (pc->pmat->spd_set) { 44513301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44523301b35fSStefano Zampini } 44536e683305SStefano Zampini /* set operators */ 44545f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44556e683305SStefano Zampini if (pcbddc->dbg_flag) { 44566e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44576e683305SStefano Zampini } 44586e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44596e683305SStefano Zampini coarse_mat = 0; 44606e683305SStefano Zampini } 44616e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4462b9b85e73SStefano Zampini #if 0 4463b9b85e73SStefano Zampini { 4464b9b85e73SStefano Zampini PetscViewer viewer; 4465b9b85e73SStefano Zampini char filename[256]; 4466b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4467b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4468b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4469b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4470b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4471b9b85e73SStefano Zampini } 4472b9b85e73SStefano Zampini #endif 4473c8587f34SStefano Zampini 4474c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4475298c0119SStefano Zampini #if 0 4476c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4477c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 447898a51de6SStefano Zampini } 4479298c0119SStefano Zampini #endif 448098a51de6SStefano Zampini 448198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 448298a51de6SStefano Zampini Vec crhs,csol; 448398a51de6SStefano Zampini PetscBool ispreonly; 448404708bb6SStefano Zampini 448598a51de6SStefano Zampini if (CoarseNullSpace) { 4486c8587f34SStefano Zampini if (isbddc) { 4487c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4488c8587f34SStefano Zampini } else { 44895fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4490c8587f34SStefano Zampini } 4491c8587f34SStefano Zampini } 4492f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4493f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4494f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4495f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44966e683305SStefano Zampini /* hack */ 4497f347579bSStefano Zampini if (!csol) { 44982a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4499f9eb5b7dSStefano Zampini } 4500f347579bSStefano Zampini if (!crhs) { 45012a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4502f347579bSStefano Zampini } 4503cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4504cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 45056e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4506c8587f34SStefano Zampini KSP check_ksp; 45072b510759SStefano Zampini KSPType check_ksp_type; 4508c8587f34SStefano Zampini PC check_pc; 45096e683305SStefano Zampini Vec check_vec,coarse_vec; 45106a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 45112b510759SStefano Zampini PetscInt its; 45126e683305SStefano Zampini PetscBool compute_eigs; 45136e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 45146e683305SStefano Zampini PetscInt neigs; 45158e185a42SStefano Zampini const char *prefix; 4516c8587f34SStefano Zampini 45172b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 45186e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4519422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 452023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4521f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 45222b510759SStefano Zampini if (ispreonly) { 45232b510759SStefano Zampini check_ksp_type = KSPPREONLY; 45246e683305SStefano Zampini compute_eigs = PETSC_FALSE; 45252b510759SStefano Zampini } else { 4526cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45276e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4528c8587f34SStefano Zampini } 4529c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45306e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45316e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45326e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4533a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4534a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4535a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4536a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4537c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4538c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4539c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4540c8587f34SStefano Zampini /* create random vec */ 45416e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45426e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4543c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4544c8587f34SStefano Zampini if (CoarseNullSpace) { 4545c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4546c8587f34SStefano Zampini } 45476e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4548c8587f34SStefano Zampini /* solve coarse problem */ 45496e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4550c8587f34SStefano Zampini if (CoarseNullSpace) { 45516e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4552c8587f34SStefano Zampini } 4553cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45546e683305SStefano Zampini if (compute_eigs) { 4555854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4556854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45576e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45586e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45596e683305SStefano Zampini lambda_min = eigs_r[0]; 45606e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45616e683305SStefano Zampini if (lambda_max>lambda_min) { 4562cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4563cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4564cbcc2c2aSStefano Zampini } 4565c8587f34SStefano Zampini } 4566c8587f34SStefano Zampini } 4567cbcc2c2aSStefano Zampini 4568c8587f34SStefano Zampini /* check coarse problem residual error */ 45696e683305SStefano Zampini if (pcbddc->dbg_flag) { 45706e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45716e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45726e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4573c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45746e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45756e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4576c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4577779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45786e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45796e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45826e683305SStefano Zampini if (compute_eigs) { 45836e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4584deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4585c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45866e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45876e683305SStefano 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); 45886e683305SStefano Zampini for (i=0;i<neigs;i++) { 45896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4590c8587f34SStefano Zampini } 45916e683305SStefano Zampini } 45926e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45936e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45946e683305SStefano Zampini } 4595c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45966e683305SStefano Zampini if (compute_eigs) { 45976e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45986e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4599c8587f34SStefano Zampini } 46006e683305SStefano Zampini } 46016e683305SStefano Zampini } 4602cbcc2c2aSStefano Zampini /* print additional info */ 4603cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 46046e683305SStefano Zampini /* waits until all processes reaches this point */ 46056e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4606cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4607cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4608cbcc2c2aSStefano Zampini } 4609cbcc2c2aSStefano Zampini 46102b510759SStefano Zampini /* free memory */ 4611c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4612fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4613c8587f34SStefano Zampini PetscFunctionReturn(0); 4614c8587f34SStefano Zampini } 4615674ae819SStefano Zampini 4616f34684f1SStefano Zampini #undef __FUNCT__ 4617f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4618f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4619f34684f1SStefano Zampini { 4620f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4621f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4622f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4623dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4624dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 462573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4626dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4627f34684f1SStefano Zampini PetscErrorCode ierr; 4628f34684f1SStefano Zampini 4629f34684f1SStefano Zampini PetscFunctionBegin; 4630f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46310e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 46320e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4633727cdba6SStefano Zampini } 4634dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4636dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4637dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4638dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4639dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4640dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4641dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46420e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 46430e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 46440e6343abSStefano Zampini } 4645dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4646dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4647dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4648dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4649dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4650f34684f1SStefano Zampini 4651f34684f1SStefano Zampini /* check numbering */ 4652f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4653f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4654dc456d91SStefano Zampini PetscInt i; 4655b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4656f34684f1SStefano Zampini 4657f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4658f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4659f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46600fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4661f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4662f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4663727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4664f34684f1SStefano Zampini } 4665f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4666f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4667f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4668e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4669e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4670e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4671e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4672f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4673f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4674f34684f1SStefano Zampini if (array[i] == 1.0) { 4675b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4676f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4677f34684f1SStefano Zampini } 4678f34684f1SStefano Zampini } 4679b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4680f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4681f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4682f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4683f34684f1SStefano Zampini } 4684f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4685f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4686e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4687e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4688f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4689f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4690b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4691ca8b9ea9SStefano Zampini PetscInt *gidxs; 4692ca8b9ea9SStefano Zampini 4693ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46943bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4695f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4696f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4697f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4698f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46994bc2dc4bSStefano 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); 4700f34684f1SStefano Zampini } 4701f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4702ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4703f34684f1SStefano Zampini } 4704f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4705302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4706f34684f1SStefano Zampini } 47078bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4708f34684f1SStefano Zampini /* get back data */ 4709f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4710f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4711674ae819SStefano Zampini PetscFunctionReturn(0); 4712674ae819SStefano Zampini } 4713674ae819SStefano Zampini 4714e456f2a8SStefano Zampini #undef __FUNCT__ 4715e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4716a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4717e456f2a8SStefano Zampini { 4718e456f2a8SStefano Zampini IS localis_t; 4719a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4720e456f2a8SStefano Zampini PetscScalar *vals; 4721e456f2a8SStefano Zampini PetscErrorCode ierr; 4722e456f2a8SStefano Zampini 4723e456f2a8SStefano Zampini PetscFunctionBegin; 4724a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4725e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4726854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4727e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4728e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4729a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4730a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47311035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4732a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47331035eff8SStefano Zampini } 4734a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4735e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4736e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4737a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4738a7dc3881SStefano Zampini /* now compute set in local ordering */ 4739a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4740a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4741a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4742a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4743a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4744ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4745e456f2a8SStefano Zampini lsize++; 4746e456f2a8SStefano Zampini } 4747e456f2a8SStefano Zampini } 4748854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4749a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4750ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4751e456f2a8SStefano Zampini idxs[lsize++] = i; 4752e456f2a8SStefano Zampini } 4753e456f2a8SStefano Zampini } 4754a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4755a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4756e456f2a8SStefano Zampini *localis = localis_t; 4757e456f2a8SStefano Zampini PetscFunctionReturn(0); 4758e456f2a8SStefano Zampini } 4759906d46d4SStefano Zampini 4760906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4761906d46d4SStefano Zampini #undef __FUNCT__ 4762906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4763906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4764906d46d4SStefano Zampini { 4765906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4766906d46d4SStefano Zampini PetscErrorCode ierr; 4767906d46d4SStefano Zampini 4768906d46d4SStefano Zampini PetscFunctionBegin; 4769906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4770906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4771906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4772906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4773906d46d4SStefano Zampini PetscFunctionReturn(0); 4774906d46d4SStefano Zampini } 4775906d46d4SStefano Zampini 4776906d46d4SStefano Zampini #undef __FUNCT__ 4777906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4778906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4779906d46d4SStefano Zampini { 4780906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4781906d46d4SStefano Zampini PetscErrorCode ierr; 4782906d46d4SStefano Zampini 4783906d46d4SStefano Zampini PetscFunctionBegin; 4784906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4785906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4786906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4787906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4788906d46d4SStefano Zampini PetscFunctionReturn(0); 4789906d46d4SStefano Zampini } 4790b96c3477SStefano Zampini 4791b96c3477SStefano Zampini #undef __FUNCT__ 4792b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 479308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4794b96c3477SStefano Zampini { 4795a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4796b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4797b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4798a64f4aa4SStefano Zampini Mat S_j; 4799b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4800b96c3477SStefano Zampini PetscBool free_used_adj; 4801b96c3477SStefano Zampini PetscErrorCode ierr; 4802b96c3477SStefano Zampini 4803b96c3477SStefano Zampini PetscFunctionBegin; 4804b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4805b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 480608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4807b96c3477SStefano Zampini used_xadj = NULL; 4808b96c3477SStefano Zampini used_adjncy = NULL; 4809b96c3477SStefano Zampini } else { 481008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 481108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 481208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 481308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4814b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4815b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4816b96c3477SStefano Zampini } else { 48172fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4818b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4819b96c3477SStefano Zampini PetscInt nvtxs; 4820b96c3477SStefano Zampini 48212fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 48222fffb893SStefano Zampini if (flg_row) { 4823b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4824b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4825b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4826b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48272fffb893SStefano Zampini } else { 48282fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48292fffb893SStefano Zampini used_xadj = NULL; 48302fffb893SStefano Zampini used_adjncy = NULL; 48312fffb893SStefano Zampini } 48322fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4833b96c3477SStefano Zampini } 4834b96c3477SStefano Zampini } 4835d5574798SStefano Zampini 4836d5574798SStefano Zampini /* setup sub_schurs data */ 4837a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4838a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4839a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4840a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4841862806e4SStefano 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); 4842a64f4aa4SStefano Zampini } else { 48436816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 484404708bb6SStefano Zampini PetscBool isseqaij; 48455feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48465feab87aSStefano Zampini PetscInt n_vertices; 48475feab87aSStefano Zampini 48485feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48492034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48505feab87aSStefano Zampini } 485104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 485204708bb6SStefano Zampini if (!isseqaij) { 485304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 485404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 485504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 485604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 485704708bb6SStefano Zampini } else { 485804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 485904708bb6SStefano Zampini } 486004708bb6SStefano Zampini } 4861862806e4SStefano 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); 4862a64f4aa4SStefano Zampini } 4863a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4864b96c3477SStefano Zampini 4865b96c3477SStefano Zampini /* free adjacency */ 4866b96c3477SStefano Zampini if (free_used_adj) { 4867b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4868b96c3477SStefano Zampini } 4869b96c3477SStefano Zampini PetscFunctionReturn(0); 4870b96c3477SStefano Zampini } 4871b96c3477SStefano Zampini 4872b96c3477SStefano Zampini #undef __FUNCT__ 4873b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 487408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4875b96c3477SStefano Zampini { 4876b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4877b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4878b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4879b96c3477SStefano Zampini PCBDDCGraph graph; 4880b96c3477SStefano Zampini PetscErrorCode ierr; 4881b96c3477SStefano Zampini 4882b96c3477SStefano Zampini PetscFunctionBegin; 4883b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 488408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48853301b35fSStefano Zampini IS verticesIS,verticescomm; 48863301b35fSStefano Zampini PetscInt vsize,*idxs; 4887b96c3477SStefano Zampini 4888b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48893301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48903301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48913301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48923301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48933301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4894b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48957fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48963301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48973301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4898b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4899b96c3477SStefano Zampini /* 4900b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4901b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4902b96c3477SStefano Zampini } 4903b96c3477SStefano Zampini */ 4904b96c3477SStefano Zampini } else { 4905b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4906b96c3477SStefano Zampini } 4907b96c3477SStefano Zampini 4908b96c3477SStefano Zampini /* sub_schurs init */ 4909a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4910a64f4aa4SStefano Zampini 4911b96c3477SStefano Zampini /* free graph struct */ 491208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4913b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4914b96c3477SStefano Zampini } 4915b96c3477SStefano Zampini PetscFunctionReturn(0); 4916b96c3477SStefano Zampini } 4917