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); 466f5fd1fbfSStefano Zampini ierr = ISDestroy(&pcbddc->zerodiag);CHKERRQ(ierr); 467*81d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 468*81d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 469674ae819SStefano Zampini PetscFunctionReturn(0); 470674ae819SStefano Zampini } 471674ae819SStefano Zampini 472674ae819SStefano Zampini #undef __FUNCT__ 473f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 474f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4756bfb1811SStefano Zampini { 4766bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4776bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4786bfb1811SStefano Zampini VecType impVecType; 479e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4806bfb1811SStefano Zampini PetscErrorCode ierr; 4816bfb1811SStefano Zampini 4826bfb1811SStefano Zampini PetscFunctionBegin; 483f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 484f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 485f4ddd8eeSStefano Zampini } 486e7b262bdSStefano Zampini /* get sizes */ 487b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 488b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4896bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 490e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 491e7b262bdSStefano Zampini /* R nodes */ 492e7b262bdSStefano Zampini old_size = -1; 493e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 494e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 495e7b262bdSStefano Zampini } 496e7b262bdSStefano Zampini if (n_R != old_size) { 497e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 498e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4996bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5006bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5016bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5026bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 503e7b262bdSStefano Zampini } 504e7b262bdSStefano Zampini /* local primal dofs */ 505e7b262bdSStefano Zampini old_size = -1; 506e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 507e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 508e7b262bdSStefano Zampini } 509e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 510e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 51183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 512e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5136bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 514e7b262bdSStefano Zampini } 515e7b262bdSStefano Zampini /* local explicit constraints */ 516e7b262bdSStefano Zampini old_size = -1; 517e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 518e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 519e7b262bdSStefano Zampini } 520e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 521e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 52283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 52383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 52483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 52583b7ccabSStefano Zampini } 5266bfb1811SStefano Zampini PetscFunctionReturn(0); 5276bfb1811SStefano Zampini } 5286bfb1811SStefano Zampini 5296bfb1811SStefano Zampini #undef __FUNCT__ 53047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 53147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 53288ebb749SStefano Zampini { 53325084f0cSStefano Zampini PetscErrorCode ierr; 53425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 53588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 53688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 537d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 53825084f0cSStefano Zampini /* submatrices of local problem */ 53980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 54006656605SStefano Zampini /* submatrices of local coarse problem */ 54106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 54225084f0cSStefano Zampini /* working matrices */ 54306656605SStefano Zampini Mat C_CR; 54425084f0cSStefano Zampini /* additional working stuff */ 54506656605SStefano Zampini PC pc_R; 54606656605SStefano Zampini Mat F; 54706656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 54806656605SStefano Zampini 54925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 55006656605SStefano Zampini PetscScalar *work; 55106656605SStefano Zampini PetscInt *idx_V_B; 55206656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 55306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 554b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 55545a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 55688ebb749SStefano Zampini MatType impMatType; 55725084f0cSStefano Zampini /* some shortcuts to scalars */ 55806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 55988ebb749SStefano Zampini 56088ebb749SStefano Zampini PetscFunctionBegin; 561b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 56288ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 56388ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 564b371cd4fSStefano Zampini n_B = pcis->n_B; 565b371cd4fSStefano Zampini n_D = pcis->n - n_B; 56688ebb749SStefano Zampini n_R = pcis->n - n_vertices; 56788ebb749SStefano Zampini 56888ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 56988ebb749SStefano Zampini impMatType = MATSEQDENSE; 57088ebb749SStefano Zampini 57188ebb749SStefano Zampini /* vertices in boundary numbering */ 572785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5730e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 57488ebb749SStefano Zampini if (i != n_vertices) { 57522d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 57688ebb749SStefano Zampini } 57788ebb749SStefano Zampini 57806656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 57906656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 58206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 58306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 58406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 58506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 58606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 58706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 58806656605SStefano Zampini 58906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 59006656605SStefano Zampini /* allocate workspace */ 59106656605SStefano Zampini n = 0; 59206656605SStefano Zampini if (n_constraints) { 59306656605SStefano Zampini n += n_R*n_constraints; 59406656605SStefano Zampini } 59506656605SStefano Zampini if (n_vertices) { 59606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 59780677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 59806656605SStefano Zampini } 5993301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 60006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 60106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 60206656605SStefano Zampini } 60306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 60406656605SStefano Zampini 60506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 60606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 60706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 60806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 60906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 61006656605SStefano Zampini if (isLU || isILU || isCHOL) { 61106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 612d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 613d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 614d62866d3SStefano Zampini MatFactorType type; 615d62866d3SStefano Zampini 6166816873aSStefano Zampini F = reuse_mumps->F; 6176816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 618d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 61906656605SStefano Zampini } else { 62006656605SStefano Zampini F = NULL; 62106656605SStefano Zampini } 62206656605SStefano Zampini 62388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 62488ebb749SStefano Zampini if (n_constraints) { 62506656605SStefano Zampini Mat M1,M2,M3; 62680677318SStefano Zampini Mat auxmat; 62706656605SStefano Zampini IS is_aux; 62880677318SStefano Zampini PetscScalar *array,*array2; 62906656605SStefano Zampini 630f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 63180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 63288ebb749SStefano Zampini 63325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 63425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6358ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 63680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 63788ebb749SStefano Zampini 63880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 63980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 64006656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 64188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 64206656605SStefano Zampini const PetscScalar *row_cmat_values; 64306656605SStefano Zampini const PetscInt *row_cmat_indices; 64406656605SStefano Zampini PetscInt size_of_constraint,j; 64588ebb749SStefano Zampini 64606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 64806656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64906656605SStefano Zampini } 65006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 65106656605SStefano Zampini } 65280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 65306656605SStefano Zampini if (F) { 65406656605SStefano Zampini Mat B; 65506656605SStefano Zampini 65606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 65780677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 65806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 65906656605SStefano Zampini } else { 66080677318SStefano Zampini PetscScalar *marr; 66180677318SStefano Zampini 66280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 66406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 66580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 66606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 66706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 66906656605SStefano Zampini } 67080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 67106656605SStefano Zampini } 67280677318SStefano Zampini if (!pcbddc->switch_static) { 67380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 67580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 67780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 67880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 67980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 68080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 68180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 68280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 68380677318SStefano Zampini } 68480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 68580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 68680677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68780677318SStefano Zampini } else { 68880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 68980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 69025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 69180677318SStefano Zampini } 69280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 69380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 69480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 69506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 69606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 69780677318SStefano Zampini if (isCHOL) { 69880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 69980677318SStefano Zampini } else { 70025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 70180677318SStefano Zampini } 70280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 70306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 70425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 70525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 70625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 70780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 70880677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 70980677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 71006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 71106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 712f4ddd8eeSStefano Zampini } 71388ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 71488ebb749SStefano Zampini if (n_vertices) { 71506656605SStefano Zampini IS is_aux; 7163a50541eSStefano Zampini 7176816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7186816873aSStefano Zampini IS tis; 7196816873aSStefano Zampini 7206816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7216816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7226816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7236816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7246816873aSStefano Zampini } else { 7253a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7266816873aSStefano Zampini } 7279577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7289577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 73025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 73188ebb749SStefano Zampini } 73288ebb749SStefano Zampini 73388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 734f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 73506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 73606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 73706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73806656605SStefano Zampini } 739f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 74006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 74106656605SStefano Zampini PetscScalar *marray; 74206656605SStefano Zampini 74306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 74406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 747f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 748f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 749f4ddd8eeSStefano Zampini } 750f4ddd8eeSStefano Zampini } 75106656605SStefano Zampini 752f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 75306656605SStefano Zampini PetscScalar *marray; 75488ebb749SStefano Zampini 75506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7568eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75888ebb749SStefano Zampini } 7593301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 76006656605SStefano Zampini n *= 2; 76188ebb749SStefano Zampini } 76206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 76306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 76406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7658eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 76706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76888ebb749SStefano Zampini } 7693301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 77006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7718eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 77488ebb749SStefano Zampini } 77588ebb749SStefano Zampini } else { 776c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 777c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7781b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 779c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 780c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 781c0553b1fSStefano Zampini } 78288ebb749SStefano Zampini } 78306656605SStefano Zampini } 78406656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 78506656605SStefano Zampini /* vertices */ 78606656605SStefano Zampini if (n_vertices) { 78716f15bc4SStefano Zampini 78804708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 78904708bb6SStefano Zampini 79016f15bc4SStefano Zampini if (n_R) { 79106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 79206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 79316f15bc4SStefano Zampini PetscScalar *x,*y; 79404708bb6SStefano Zampini PetscBool isseqaij; 79506656605SStefano Zampini 79621eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 79706656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 79804708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7996816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 80006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 80106656605SStefano Zampini } else { 80206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 80406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 80606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 80706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 80906656605SStefano Zampini } 81006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 81106656605SStefano Zampini } 81280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81306656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 81406656605SStefano Zampini if (n_constraints) { 81506656605SStefano Zampini Mat B; 81680677318SStefano Zampini 817b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 81880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 81980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 82080677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82580677318SStefano Zampini } 82680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 82880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 83080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 83406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83506656605SStefano Zampini } 83604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 83704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 83804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 83904708bb6SStefano Zampini } 84006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 84180677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 84506656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84606656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84706656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 84806656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 85016f15bc4SStefano Zampini } else { 85116f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85216f15bc4SStefano Zampini } 85321eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 85406656605SStefano Zampini /* coarse basis functions */ 85506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85616f15bc4SStefano Zampini PetscScalar *y; 85716f15bc4SStefano Zampini 85806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 86106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 86406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86606656605SStefano Zampini 86706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87406656605SStefano Zampini } 87506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87606656605SStefano Zampini } 87704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 87804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 87906656605SStefano Zampini } 88006656605SStefano Zampini 88106656605SStefano Zampini if (n_constraints) { 88206656605SStefano Zampini Mat B; 88306656605SStefano Zampini 88406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88806656605SStefano Zampini if (n_vertices) { 88980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 89080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 89180677318SStefano Zampini } else { 89280677318SStefano Zampini Mat S_VCt; 89380677318SStefano Zampini 89480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 89580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89780677318SStefano Zampini } 89806656605SStefano Zampini } 89906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 90006656605SStefano Zampini /* coarse basis functions */ 90106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 90206656605SStefano Zampini PetscScalar *y; 90306656605SStefano Zampini 90406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 91406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91806656605SStefano Zampini } 91906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini } 92280677318SStefano Zampini if (n_constraints) { 92380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 92480677318SStefano Zampini } 92506656605SStefano Zampini 92606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9273301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 92806656605SStefano Zampini 92906656605SStefano Zampini if (n_constraints) { 93016f15bc4SStefano Zampini Mat S_CCT,B_C; 93106656605SStefano Zampini 93280677318SStefano Zampini /* this is a lazy thing */ 93380677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93806656605SStefano Zampini if (n_vertices) { 93916f15bc4SStefano Zampini Mat B_V,S_VCT; 94006656605SStefano Zampini 94106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 94516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94606656605SStefano Zampini } 94706656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94880677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94980677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 95006656605SStefano Zampini } 95116f15bc4SStefano Zampini if (n_vertices && n_R) { 95206656605SStefano Zampini Mat A_VRT; 95380677318SStefano Zampini PetscScalar *marray; 95406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 95506656605SStefano Zampini 95680677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95880677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 96080677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 96180677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 96216f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 96306656605SStefano Zampini } 96406656605SStefano Zampini 96506656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97206656605SStefano Zampini } 97306656605SStefano Zampini } else { 97406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 98006656605SStefano Zampini } 98106656605SStefano Zampini } 98206656605SStefano Zampini /* coarse basis functions */ 98306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98406656605SStefano Zampini PetscScalar *y; 98506656605SStefano Zampini 98606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99106656605SStefano Zampini if (i<n_vertices) { 99206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 99306656605SStefano Zampini } 99406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99606656605SStefano Zampini 99706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100406656605SStefano Zampini } 100506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100606656605SStefano Zampini } 100706656605SStefano Zampini } 1008d62866d3SStefano Zampini /* free memory */ 100988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 101106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1014d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1015d62866d3SStefano Zampini if (n_vertices) { 1016d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1017d62866d3SStefano Zampini } 1018d62866d3SStefano Zampini if (n_constraints) { 1019d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1020d62866d3SStefano Zampini } 102188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 102288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102425084f0cSStefano Zampini if (pcbddc->dbg_flag) { 102588ebb749SStefano Zampini Mat coarse_sub_mat; 102625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10308bec7fa6SStefano Zampini Mat C_B,CPHI; 10318bec7fa6SStefano Zampini IS is_dummy; 10328bec7fa6SStefano Zampini Vec mones; 103388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 103488ebb749SStefano Zampini PetscReal real_value; 103588ebb749SStefano Zampini 103688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1042c0553b1fSStefano Zampini if (unsymmetric_check) { 104388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 104588ebb749SStefano Zampini } 104688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104788ebb749SStefano Zampini 104825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10493301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 105025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1051c0553b1fSStefano Zampini if (unsymmetric_check) { 105288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106488ebb749SStefano Zampini } else { 106588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107388ebb749SStefano Zampini } 107488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10800fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini 10838bec7fa6SStefano Zampini /* check constraints */ 10848bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10858bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10868bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10888bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10908bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1091bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1092c0553b1fSStefano Zampini if (unsymmetric_check) { 1093bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1094bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1095bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1096bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1097bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 109888ebb749SStefano Zampini } 10998bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 11008bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11018bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11028bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11038bec7fa6SStefano Zampini 110425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110888ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110988ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 111088ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 111188ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 111288ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 111388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 111488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1115c0553b1fSStefano Zampini if (unsymmetric_check) { 111688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111888ebb749SStefano Zampini } 111988ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 112088ebb749SStefano Zampini } 11218629588bSStefano Zampini /* get back data */ 11228629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 112388ebb749SStefano Zampini PetscFunctionReturn(0); 112488ebb749SStefano Zampini } 112588ebb749SStefano Zampini 112688ebb749SStefano Zampini #undef __FUNCT__ 1127d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1128d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1129aa0d41d4SStefano Zampini { 1130d65f70fdSStefano Zampini Mat *work_mat; 1131d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1132d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1133d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1134aa0d41d4SStefano Zampini PetscErrorCode ierr; 1135aa0d41d4SStefano Zampini 1136aa0d41d4SStefano Zampini PetscFunctionBegin; 1137d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1138d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1140d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1141aa0d41d4SStefano Zampini 1142d65f70fdSStefano Zampini if (!rsorted) { 1143906d46d4SStefano Zampini const PetscInt *idxs; 1144906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1145aa0d41d4SStefano Zampini 1146d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1149d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1150aa0d41d4SStefano Zampini } 1151d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1154d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1155aa0d41d4SStefano Zampini } 1156d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini } else { 1159d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini isrow_s = isrow; 1161aa0d41d4SStefano Zampini } 1162906d46d4SStefano Zampini 1163d65f70fdSStefano Zampini if (!csorted) { 1164d65f70fdSStefano Zampini if (isrow == iscol) { 1165d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini iscol_s = isrow_s; 1167d65f70fdSStefano Zampini } else { 1168d65f70fdSStefano Zampini const PetscInt *idxs; 1169d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1170906d46d4SStefano Zampini 1171d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1172d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1173d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1174d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1175d65f70fdSStefano Zampini } 1176d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1177d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1179d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1182d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini } 1184d65f70fdSStefano Zampini } else { 1185d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini iscol_s = iscol; 1187d65f70fdSStefano Zampini } 1188d65f70fdSStefano Zampini 1189d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1190d65f70fdSStefano Zampini 1191d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1192906d46d4SStefano Zampini Mat new_mat; 1193d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1194906d46d4SStefano Zampini 1195d65f70fdSStefano Zampini if (!rsorted) { 1196d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1197d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1199d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1200906d46d4SStefano Zampini } 1201d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1202d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini } else { 1204d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1205906d46d4SStefano Zampini } 1206d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini 1208d65f70fdSStefano Zampini if (!csorted) { 1209d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1210d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1212d65f70fdSStefano Zampini } else { 1213d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1214d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1215d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1216d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1217d65f70fdSStefano Zampini } 1218d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini } 1221d65f70fdSStefano Zampini } else { 1222d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1223d65f70fdSStefano Zampini } 1224d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1225d65f70fdSStefano Zampini 1226d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1228d65f70fdSStefano Zampini work_mat[0] = new_mat; 1229d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini } 1232d65f70fdSStefano Zampini 1233d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1234d65f70fdSStefano Zampini *B = work_mat[0]; 1235d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1236d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1237d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1238d65f70fdSStefano Zampini PetscFunctionReturn(0); 1239d65f70fdSStefano Zampini } 1240d65f70fdSStefano Zampini 1241d65f70fdSStefano Zampini #undef __FUNCT__ 12425e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12435e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1244aa0d41d4SStefano Zampini { 1245aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12463bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 12475e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1248d65f70fdSStefano Zampini Mat new_mat; 12495e8657edSStefano Zampini IS is_local,is_global; 1250d65f70fdSStefano Zampini PetscInt local_size; 1251d65f70fdSStefano Zampini PetscBool isseqaij; 1252aa0d41d4SStefano Zampini PetscErrorCode ierr; 1253aa0d41d4SStefano Zampini 1254aa0d41d4SStefano Zampini PetscFunctionBegin; 1255aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12565e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12575e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 12583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1259aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1260d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1261aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1262906d46d4SStefano Zampini 1263906d46d4SStefano Zampini /* check */ 1264906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1265906d46d4SStefano Zampini Vec x,x_change; 1266906d46d4SStefano Zampini PetscReal error; 1267906d46d4SStefano Zampini 12685e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1269906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12705e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1271e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1272e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1273d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1274e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1275e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1276906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1277906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1278906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1279906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1280906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1281906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1282906d46d4SStefano Zampini } 1283906d46d4SStefano Zampini 128422d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12859b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 128622d5777bSStefano Zampini if (isseqaij) { 1287d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1288aa0d41d4SStefano Zampini } else { 1289aa0d41d4SStefano Zampini Mat work_mat; 1290aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1291d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1292aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1293aa0d41d4SStefano Zampini } 12943301b35fSStefano Zampini if (matis->A->symmetric_set) { 12953301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1296e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12973301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1298e496cd5dSStefano Zampini #endif 12993301b35fSStefano Zampini } 130045a1bb75SStefano Zampini /* 130145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1302d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 130345a1bb75SStefano Zampini */ 1304d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1305aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1306aa0d41d4SStefano Zampini } 1307aa0d41d4SStefano Zampini 1308aa0d41d4SStefano Zampini #undef __FUNCT__ 1309a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13108ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1311a64d13efSStefano Zampini { 1312a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1313a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1314d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 131553892102SStefano Zampini PetscInt *idx_R_local=NULL; 13163a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13173a50541eSStefano Zampini PetscInt vbs,bs; 13186816873aSStefano Zampini PetscBT bitmask=NULL; 1319a64d13efSStefano Zampini PetscErrorCode ierr; 1320a64d13efSStefano Zampini 1321a64d13efSStefano Zampini PetscFunctionBegin; 1322b23d619eSStefano Zampini /* 1323b23d619eSStefano Zampini No need to setup local scatters if 1324b23d619eSStefano Zampini - primal space is unchanged 1325b23d619eSStefano Zampini AND 1326b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1327b23d619eSStefano Zampini AND 1328b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1329b23d619eSStefano Zampini */ 1330b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1331f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1332f4ddd8eeSStefano Zampini } 1333f4ddd8eeSStefano Zampini /* destroy old objects */ 1334f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1335f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1336f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1337a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1338b371cd4fSStefano Zampini n_B = pcis->n_B; 1339b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1340b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13413a50541eSStefano Zampini 1342a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13436816873aSStefano Zampini 134453892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13456816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1346854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1347a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1348a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13490e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1350a64d13efSStefano Zampini } 1351a64d13efSStefano Zampini 1352a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13534641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13546816873aSStefano Zampini idx_R_local[n_R++] = i; 1355a64d13efSStefano Zampini } 1356a64d13efSStefano Zampini } 135753892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13586816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13596816873aSStefano Zampini 136053892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 136153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13626816873aSStefano Zampini } 13633a50541eSStefano Zampini 13643a50541eSStefano Zampini /* Block code */ 13653a50541eSStefano Zampini vbs = 1; 13663a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13673a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13683a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13693a50541eSStefano Zampini PetscInt *vary; 1370d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1371785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13723a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1373d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1374d3df7717SStefano 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 */ 13750e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1376d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13773a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13783a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13793a50541eSStefano Zampini break; 13803a50541eSStefano Zampini } 13813a50541eSStefano Zampini } 1382d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1383d3df7717SStefano Zampini } else { 1384d3df7717SStefano Zampini /* Verify directly the R set */ 1385d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1386d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1387d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1388d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1389d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1390d3df7717SStefano Zampini break; 1391d3df7717SStefano Zampini } 1392d3df7717SStefano Zampini } 1393d3df7717SStefano Zampini } 1394d3df7717SStefano Zampini } 13953a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13963a50541eSStefano Zampini vbs = bs; 13973a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13983a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13993a50541eSStefano Zampini } 14003a50541eSStefano Zampini } 14013a50541eSStefano Zampini } 14023a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 140353892102SStefano Zampini if (sub_schurs->reuse_mumps) { 140453892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 140553892102SStefano Zampini 140653892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140753892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140853892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140953892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 141053892102SStefano Zampini } else { 14113a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 141253892102SStefano Zampini } 1413a64d13efSStefano Zampini 1414a64d13efSStefano Zampini /* print some info if requested */ 1415a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1416a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1417a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14180fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1419a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1420a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14213a50541eSStefano 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); 1422a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1423a64d13efSStefano Zampini } 1424a64d13efSStefano Zampini 1425a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14266816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14276816873aSStefano Zampini IS is_aux1,is_aux2; 14286816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14296816873aSStefano Zampini 14303a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1431854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1432854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1433a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14344641a718SStefano Zampini for (i=0; i<n_D; i++) { 14354641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14364641a718SStefano Zampini } 1437a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1438a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14394641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14404641a718SStefano Zampini aux_array1[j++] = i; 1441a64d13efSStefano Zampini } 1442a64d13efSStefano Zampini } 1443a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1444a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1445a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14464641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14474641a718SStefano Zampini aux_array2[j++] = i; 1448a64d13efSStefano Zampini } 1449a64d13efSStefano Zampini } 1450a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1451a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1452a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1453a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1454a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1455a64d13efSStefano Zampini 14568eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1457785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1458a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14594641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14604641a718SStefano Zampini aux_array1[j++] = i; 1461a64d13efSStefano Zampini } 1462a64d13efSStefano Zampini } 1463a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1464a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1465a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1466a64d13efSStefano Zampini } 14674641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14683a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1469d62866d3SStefano Zampini } else { 147053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14716816873aSStefano Zampini IS tis; 14726816873aSStefano Zampini PetscInt schur_size; 14736816873aSStefano Zampini 147453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14756816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 147653892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14776816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14786816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14796816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14806816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14816816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1482d62866d3SStefano Zampini } 1483d62866d3SStefano Zampini } 1484a64d13efSStefano Zampini PetscFunctionReturn(0); 1485a64d13efSStefano Zampini } 1486a64d13efSStefano Zampini 1487304d26faSStefano Zampini 1488304d26faSStefano Zampini #undef __FUNCT__ 1489304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1490684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1491304d26faSStefano Zampini { 1492304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1493304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1494304d26faSStefano Zampini PC pc_temp; 1495304d26faSStefano Zampini Mat A_RR; 1496f4ddd8eeSStefano Zampini MatReuse reuse; 1497304d26faSStefano Zampini PetscScalar m_one = -1.0; 1498304d26faSStefano Zampini PetscReal value; 149904708bb6SStefano Zampini PetscInt n_D,n_R; 15009577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1501304d26faSStefano Zampini PetscErrorCode ierr; 1502e604994aSStefano Zampini /* prefixes stuff */ 1503312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1504e604994aSStefano Zampini size_t len; 1505304d26faSStefano Zampini 1506304d26faSStefano Zampini PetscFunctionBegin; 1507304d26faSStefano Zampini 1508e604994aSStefano Zampini /* compute prefixes */ 1509e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1510e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1511e604994aSStefano Zampini if (!pcbddc->current_level) { 1512e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1513e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1514e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1515e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1516e604994aSStefano Zampini } else { 1517e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1518312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1519e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1520e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1521312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1522312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 152334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 152434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1525e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1526e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1527e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1528e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1529e604994aSStefano Zampini } 1530e604994aSStefano Zampini 1531304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1532684f6988SStefano Zampini if (dirichlet) { 1533d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15343301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15353301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1536964fefecSStefano Zampini } 1537ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1538964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1539304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1540304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1541304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1542304d26faSStefano Zampini /* default */ 1543304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1544e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15459577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1546304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15479577ea80SStefano Zampini if (issbaij) { 15489577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15499577ea80SStefano Zampini } else { 1550304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15519577ea80SStefano Zampini } 1552304d26faSStefano Zampini /* Allow user's customization */ 1553304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1554304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1555304d26faSStefano Zampini } 1556d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1557d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1558d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1559d62866d3SStefano Zampini 1560d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1561d5574798SStefano Zampini } 1562304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1563304d26faSStefano Zampini if (!n_D) { 1564304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1565304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1566304d26faSStefano Zampini } 1567304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1568304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1569304d26faSStefano Zampini /* set ksp_D into pcis data */ 1570304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1571304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1572304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1573684f6988SStefano Zampini } 1574304d26faSStefano Zampini 1575304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1576684f6988SStefano Zampini A_RR = 0; 1577684f6988SStefano Zampini if (neumann) { 1578d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 157904708bb6SStefano Zampini PetscInt ibs,mbs; 158004708bb6SStefano Zampini PetscBool issbaij; 158104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1582f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15838ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1585f4ddd8eeSStefano Zampini PetscInt nn_R; 158681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1587f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1588f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1589f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1590f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1591f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1592f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1593f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1594727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1595f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1596f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1597f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1598f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1599f4ddd8eeSStefano Zampini } 1600f4ddd8eeSStefano Zampini } 1601f4ddd8eeSStefano Zampini /* last check */ 1602d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1603f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1604f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1605f4ddd8eeSStefano Zampini } 1606f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1607f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1608f4ddd8eeSStefano Zampini } 1609f4ddd8eeSStefano Zampini /* extract A_RR */ 1610af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1611af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 161204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 161304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 161404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 161504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 161604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1617af732b37SStefano Zampini } else { 161804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16196816873aSStefano Zampini } 162004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 162104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 162204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 162304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162404708bb6SStefano Zampini } else { 162504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 162604708bb6SStefano Zampini } 162704708bb6SStefano Zampini } 162804708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1629f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16303301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16313301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16326816873aSStefano Zampini } 16336816873aSStefano Zampini } else { 16346816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16356816873aSStefano Zampini 16366816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16376816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16386816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1639af732b37SStefano Zampini } 1640f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1641304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1642304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1643304d26faSStefano Zampini /* default */ 1644304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1645e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1646304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16479577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16489577ea80SStefano Zampini if (issbaij) { 16499577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16509577ea80SStefano Zampini } else { 1651304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16529577ea80SStefano Zampini } 1653304d26faSStefano Zampini /* Allow user's customization */ 1654304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1655304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1656304d26faSStefano Zampini } 1657d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1658304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1659304d26faSStefano Zampini if (!n_R) { 1660304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1661304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1662304d26faSStefano Zampini } 1663d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1664d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1665d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1666d62866d3SStefano Zampini 1667d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1668d62866d3SStefano Zampini } 1669304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1670304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1671684f6988SStefano Zampini } 16726816873aSStefano Zampini /* free Neumann problem's matrix */ 16736816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1674304d26faSStefano Zampini 1675304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16760fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1677684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1678684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1679684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1680684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1681684f6988SStefano Zampini } 1682684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16830fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16870fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1688304d26faSStefano Zampini /* need to be adapted? */ 1689b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1690b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1691b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1692304d26faSStefano Zampini /* print info */ 1693304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1694e604994aSStefano 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); 1695304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1696304d26faSStefano Zampini } 1697b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1698298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1699304d26faSStefano Zampini } 1700684f6988SStefano Zampini } 1701684f6988SStefano Zampini if (neumann) { /* Neumann */ 17026816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 17030fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 17040fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 17050fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 17060fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17070fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1708304d26faSStefano Zampini /* need to be adapted? */ 1709b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1710b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1711304d26faSStefano Zampini /* print info */ 1712304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1713e604994aSStefano 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); 1714304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1715304d26faSStefano Zampini } 1716b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1717298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1718304d26faSStefano Zampini } 17190fccc4e9SStefano Zampini } 1720684f6988SStefano Zampini } 1721304d26faSStefano Zampini PetscFunctionReturn(0); 1722304d26faSStefano Zampini } 1723304d26faSStefano Zampini 1724304d26faSStefano Zampini #undef __FUNCT__ 1725ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 172680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1727674ae819SStefano Zampini { 1728674ae819SStefano Zampini PetscErrorCode ierr; 1729674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1730be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1731674ae819SStefano Zampini 1732674ae819SStefano Zampini PetscFunctionBegin; 1733be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 173520c7b377SStefano Zampini } 173680677318SStefano Zampini if (!pcbddc->switch_static) { 173780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 173980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 174020c7b377SStefano Zampini } 1741be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174420c7b377SStefano Zampini } else { 1745be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1746be83ff47SStefano Zampini 174753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174920c7b377SStefano Zampini } 1750be83ff47SStefano Zampini } else { 175180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 175680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 175780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1760674ae819SStefano Zampini } 1761674ae819SStefano Zampini } 1762be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176380677318SStefano Zampini if (applytranspose) { 176480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176580677318SStefano Zampini } else { 176680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176780677318SStefano Zampini } 176853892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1769be83ff47SStefano Zampini } else { 1770be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1771be83ff47SStefano Zampini 1772be83ff47SStefano Zampini if (applytranspose) { 177353892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1774be83ff47SStefano Zampini } else { 177553892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1776be83ff47SStefano Zampini } 177753892102SStefano Zampini #endif 1778be83ff47SStefano Zampini } 177980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 178080677318SStefano Zampini if (!pcbddc->switch_static) { 1781be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 178280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1784be83ff47SStefano Zampini } else { 1785be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1786be83ff47SStefano Zampini 178753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1789be83ff47SStefano Zampini } 179080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 179280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 179380677318SStefano Zampini } 179480677318SStefano Zampini } else { 179580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 180080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 180180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 180280677318SStefano Zampini } 180380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1807674ae819SStefano Zampini } 1808674ae819SStefano Zampini PetscFunctionReturn(0); 1809674ae819SStefano Zampini } 1810674ae819SStefano Zampini 1811dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1812674ae819SStefano Zampini #undef __FUNCT__ 1813674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1814dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1815674ae819SStefano Zampini { 1816674ae819SStefano Zampini PetscErrorCode ierr; 1817674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1818674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1819674ae819SStefano Zampini const PetscScalar zero = 0.0; 1820674ae819SStefano Zampini 1821674ae819SStefano Zampini PetscFunctionBegin; 1822dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1823dc359a40SStefano Zampini if (applytranspose) { 1824674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18258eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1826dc359a40SStefano Zampini } else { 1827674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1828674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 182915aaf578SStefano Zampini } 183012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 183112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 183212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183412edc857SStefano Zampini 18359f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 183612edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 183712edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1838964fefecSStefano Zampini Vec rhs,sol; 1839964fefecSStefano Zampini 1840964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1841964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 184212edc857SStefano Zampini if (applytranspose) { 1843964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184412edc857SStefano Zampini } else { 1845964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184612edc857SStefano Zampini } 184712edc857SStefano Zampini } 1848674ae819SStefano Zampini 1849674ae819SStefano Zampini /* Local solution on R nodes */ 185080677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 185180677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18529f00e9b4SStefano Zampini } 1853674ae819SStefano Zampini 18549f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18559f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 185612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1857674ae819SStefano Zampini 1858674ae819SStefano Zampini /* Sum contributions from two levels */ 1859dc359a40SStefano Zampini if (applytranspose) { 1860dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1861dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1862dc359a40SStefano Zampini } else { 1863674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18648eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1865dc359a40SStefano Zampini } 1866674ae819SStefano Zampini PetscFunctionReturn(0); 1867674ae819SStefano Zampini } 1868674ae819SStefano Zampini 1869674ae819SStefano Zampini #undef __FUNCT__ 1870674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 187112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1872674ae819SStefano Zampini { 1873674ae819SStefano Zampini PetscErrorCode ierr; 1874674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187558da7f69SStefano Zampini PetscScalar *array; 187612edc857SStefano Zampini Vec from,to; 1877674ae819SStefano Zampini 1878674ae819SStefano Zampini PetscFunctionBegin; 187912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 188012edc857SStefano Zampini from = pcbddc->coarse_vec; 188112edc857SStefano Zampini to = pcbddc->vec1_P; 188212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 188312edc857SStefano Zampini Vec tvec; 188458da7f69SStefano Zampini 188558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 188712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 188958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 189058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 189112edc857SStefano Zampini } 189212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189312edc857SStefano Zampini from = pcbddc->vec1_P; 189412edc857SStefano Zampini to = pcbddc->coarse_vec; 189512edc857SStefano Zampini } 189612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1897674ae819SStefano Zampini PetscFunctionReturn(0); 1898674ae819SStefano Zampini } 1899674ae819SStefano Zampini 1900674ae819SStefano Zampini #undef __FUNCT__ 1901674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 190212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1903674ae819SStefano Zampini { 1904674ae819SStefano Zampini PetscErrorCode ierr; 1905674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 190658da7f69SStefano Zampini PetscScalar *array; 190712edc857SStefano Zampini Vec from,to; 1908674ae819SStefano Zampini 1909674ae819SStefano Zampini PetscFunctionBegin; 191012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 191112edc857SStefano Zampini from = pcbddc->coarse_vec; 191212edc857SStefano Zampini to = pcbddc->vec1_P; 191312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 191412edc857SStefano Zampini from = pcbddc->vec1_P; 191512edc857SStefano Zampini to = pcbddc->coarse_vec; 191612edc857SStefano Zampini } 191712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 191812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 191912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 192012edc857SStefano Zampini Vec tvec; 192158da7f69SStefano Zampini 192212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 192358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 192458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 192558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 192658da7f69SStefano Zampini } 192758da7f69SStefano Zampini } else { 192858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 192958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 193012edc857SStefano Zampini } 193112edc857SStefano Zampini } 1932674ae819SStefano Zampini PetscFunctionReturn(0); 1933674ae819SStefano Zampini } 1934674ae819SStefano Zampini 1935984c4197SStefano Zampini /* uncomment for testing purposes */ 1936984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1937674ae819SStefano Zampini #undef __FUNCT__ 1938674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1939674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1940674ae819SStefano Zampini { 1941674ae819SStefano Zampini PetscErrorCode ierr; 1942674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1943674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1944674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1945984c4197SStefano Zampini /* one and zero */ 1946984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1947984c4197SStefano Zampini /* space to store constraints and their local indices */ 19489162d606SStefano Zampini PetscScalar *constraints_data; 19499162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19509162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19519162d606SStefano Zampini PetscInt *constraints_n; 1952984c4197SStefano Zampini /* iterators */ 1953b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1954984c4197SStefano Zampini /* BLAS integers */ 1955e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1956e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1957c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1958727cdba6SStefano Zampini /* reuse */ 19590e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19600e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1961984c4197SStefano Zampini /* change of basis */ 1962b3d85658SStefano Zampini PetscBool qr_needed; 19639162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1964984c4197SStefano Zampini /* auxiliary stuff */ 196564efe560SStefano Zampini PetscInt *nnz,*is_indices; 19668a0068c3SStefano Zampini PetscInt ncc; 1967984c4197SStefano Zampini /* some quantities */ 196845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1969a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1970984c4197SStefano Zampini 1971674ae819SStefano Zampini PetscFunctionBegin; 19728e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19738e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19748e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1975088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1976088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19770e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19780e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19790e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19800e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19810e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1982088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1983cf5a6209SStefano Zampini 1984cf5a6209SStefano Zampini /* print some info */ 1985cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1986cf5a6209SStefano Zampini IS vertices; 1987cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1988cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1989cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1990cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1991cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1992cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1993cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1994fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1995fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1996cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1997cf5a6209SStefano Zampini } 1998cf5a6209SStefano Zampini 1999cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 20009162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2001cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2002cf5a6209SStefano Zampini const Vec *nearnullvecs; 2003cf5a6209SStefano Zampini Vec *localnearnullsp; 2004cf5a6209SStefano Zampini PetscScalar *array; 2005cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2006cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2007674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2008b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2009674ae819SStefano Zampini PetscScalar *work; 2010674ae819SStefano Zampini PetscReal *singular_vals; 2011674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2012674ae819SStefano Zampini PetscReal *rwork; 2013674ae819SStefano Zampini #endif 2014674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2015674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2016674ae819SStefano Zampini #else 2017964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2018964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2019674ae819SStefano Zampini #endif 2020674ae819SStefano Zampini 2021674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2022d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2023d06fc5fdSStefano Zampini /* free unneeded index sets */ 2024d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2025d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2026674ae819SStefano Zampini } 2027d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2028d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2029d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2030d06fc5fdSStefano Zampini } 2031d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2032d06fc5fdSStefano Zampini n_ISForEdges = 0; 2033d06fc5fdSStefano Zampini } 2034d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2035d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2036d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2037d06fc5fdSStefano Zampini } 2038d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2039d06fc5fdSStefano Zampini n_ISForFaces = 0; 2040d06fc5fdSStefano Zampini } 204170022509SStefano Zampini 204270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 204370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 204470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 204570022509SStefano Zampini if (pcbddc->NullSpace) { 204670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 204770022509SStefano Zampini 204870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2049b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2050d06fc5fdSStefano Zampini if (!ISForEdges) { 2051d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2052d06fc5fdSStefano Zampini } 2053b8ffe317SStefano Zampini } 2054d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2055d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2056d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2057d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2058d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 205998a51de6SStefano Zampini } 206070022509SStefano Zampini #endif 206108122e43SStefano Zampini 2062674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2063674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2064674ae819SStefano Zampini if (nearnullsp) { 2065674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2066f4ddd8eeSStefano Zampini /* remove any stored info */ 2067f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2068f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2069f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2070f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2071f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2072473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2073f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2074f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2075f4ddd8eeSStefano Zampini } 2076984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2077984c4197SStefano Zampini nnsp_size = 0; 2078674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2079674ae819SStefano Zampini } 2080984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2081984c4197SStefano Zampini max_constraints = nnsp_size; 2082984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2083984c4197SStefano Zampini 2084674ae819SStefano Zampini /* 2085674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20869162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20879162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20889162d606SStefano Zampini There can be multiple constraints per connected component 2089674ae819SStefano Zampini */ 2090674ae819SStefano Zampini n_vertices = 0; 2091674ae819SStefano Zampini if (ISForVertices) { 2092674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2093674ae819SStefano Zampini } 20949162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20959162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20969162d606SStefano Zampini 20979162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20989162d606SStefano Zampini total_counts *= max_constraints; 2099674ae819SStefano Zampini total_counts += n_vertices; 21004641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 21019162d606SStefano Zampini 2102674ae819SStefano Zampini total_counts = 0; 2103674ae819SStefano Zampini max_size_of_constraint = 0; 2104674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 21059162d606SStefano Zampini IS used_is; 2106674ae819SStefano Zampini if (i<n_ISForEdges) { 21079162d606SStefano Zampini used_is = ISForEdges[i]; 2108674ae819SStefano Zampini } else { 21099162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2110674ae819SStefano Zampini } 21119162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2112674ae819SStefano Zampini total_counts += j; 2113674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2114674ae819SStefano Zampini } 21159162d606SStefano 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); 21169162d606SStefano Zampini 2117984c4197SStefano Zampini /* get local part of global near null space vectors */ 2118785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2119984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2120984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2121e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2122e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2123984c4197SStefano Zampini } 2124674ae819SStefano Zampini 2125242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2126242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2127a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2128242a89d7SStefano Zampini 2129984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2130a773dcb8SStefano Zampini if (!skip_lapack) { 2131674ae819SStefano Zampini PetscScalar temp_work; 2132911cabfeSStefano Zampini 2133674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2134984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2135785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2136785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2137785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2138674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2139785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2140674ae819SStefano Zampini #endif 2141674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2142c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2143c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2144674ae819SStefano Zampini lwork = -1; 2145674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2146674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2147c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2148674ae819SStefano Zampini #else 2149c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2150674ae819SStefano Zampini #endif 2151674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2152984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2153674ae819SStefano Zampini #else /* on missing GESVD */ 2154674ae819SStefano Zampini /* SVD */ 2155674ae819SStefano Zampini PetscInt max_n,min_n; 2156674ae819SStefano Zampini max_n = max_size_of_constraint; 2157984c4197SStefano Zampini min_n = max_constraints; 2158984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2159674ae819SStefano Zampini min_n = max_size_of_constraint; 2160984c4197SStefano Zampini max_n = max_constraints; 2161674ae819SStefano Zampini } 2162785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2163674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2164785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2165674ae819SStefano Zampini #endif 2166674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2167674ae819SStefano Zampini lwork = -1; 2168e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2169e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2170b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2171674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2172674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21739162d606SStefano 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)); 2174674ae819SStefano Zampini #else 21759162d606SStefano 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)); 2176674ae819SStefano Zampini #endif 2177674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2178984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2179984c4197SStefano Zampini #endif /* on missing GESVD */ 2180674ae819SStefano Zampini /* Allocate optimal workspace */ 2181674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2182854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2183674ae819SStefano Zampini } 2184674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2185674ae819SStefano Zampini total_counts = 0; 21869162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21879162d606SStefano Zampini constraints_data_ptr[0] = 0; 2188674ae819SStefano Zampini /* vertices */ 21899162d606SStefano Zampini if (n_vertices) { 2190674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21919162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21929162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2193674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21949162d606SStefano Zampini constraints_n[total_counts] = 1; 21959162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21969162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21979162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2198674ae819SStefano Zampini total_counts++; 2199674ae819SStefano Zampini } 2200674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2201984c4197SStefano Zampini PetscBool used_vertex; 2202674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2203674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2204674ae819SStefano Zampini k = 0; 2205674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2206984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2207984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22089162d606SStefano Zampini constraints_n[total_counts] = 1; 22099162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22109162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22119162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22129162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2213674ae819SStefano Zampini total_counts++; 2214674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2215674ae819SStefano Zampini } 2216984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2217674ae819SStefano Zampini k++; 2218674ae819SStefano Zampini } 2219674ae819SStefano Zampini } 2220674ae819SStefano Zampini } 2221674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2222674ae819SStefano Zampini n_vertices = total_counts; 2223674ae819SStefano Zampini } 2224984c4197SStefano Zampini 2225674ae819SStefano Zampini /* edges and faces */ 22269162d606SStefano Zampini total_counts_cc = total_counts; 2227911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22289162d606SStefano Zampini IS used_is; 22299162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22309162d606SStefano Zampini 2231911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22329162d606SStefano Zampini used_is = ISForEdges[ncc]; 2233984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2234674ae819SStefano Zampini } else { 22359162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2236984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2237674ae819SStefano Zampini } 2238674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22399162d606SStefano Zampini 22409162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22419162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2242984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2243984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2244674ae819SStefano Zampini if (nnsp_has_cnst) { 22455b08dc53SStefano Zampini PetscScalar quad_value; 22469162d606SStefano Zampini 22479162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22489162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22499162d606SStefano Zampini 2250a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2251674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2252a773dcb8SStefano Zampini } else { 2253a773dcb8SStefano Zampini quad_value = 1.0; 2254a773dcb8SStefano Zampini } 2255674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22569162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2257674ae819SStefano Zampini } 22589162d606SStefano Zampini temp_constraints++; 2259674ae819SStefano Zampini total_counts++; 2260674ae819SStefano Zampini } 2261674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2262984c4197SStefano Zampini PetscReal real_value; 22639162d606SStefano Zampini PetscScalar *ptr_to_data; 22649162d606SStefano Zampini 2265984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22669162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2267674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22689162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2269674ae819SStefano Zampini } 2270984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2271984c4197SStefano Zampini /* check if array is null on the connected component */ 2272e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22739162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22745b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2275674ae819SStefano Zampini temp_constraints++; 2276674ae819SStefano Zampini total_counts++; 22779162d606SStefano Zampini if (!idxs_copied) { 22789162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22799162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2280674ae819SStefano Zampini } 2281674ae819SStefano Zampini } 22829162d606SStefano Zampini } 22839162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 228445a1bb75SStefano Zampini valid_constraints = temp_constraints; 2285eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2286a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22879162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22889162d606SStefano Zampini 22899162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2290a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22919162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2292a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22939162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2294a773dcb8SStefano Zampini } else { /* perform SVD */ 2295984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22969162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2297674ae819SStefano Zampini 2298674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2299984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2300984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2301984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2302984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2303984c4197SStefano Zampini from that computed using LAPACKgesvd 2304984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2305984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2306984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2307674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2308e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2309984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2310674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2311674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23129162d606SStefano 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)); 2313674ae819SStefano Zampini } 2314674ae819SStefano Zampini } 2315e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2316e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2317e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2318674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2319c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2320674ae819SStefano Zampini #else 2321c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2322674ae819SStefano Zampini #endif 2323674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2324984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2325984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2326674ae819SStefano Zampini j = 0; 2327984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2328674ae819SStefano Zampini total_counts = total_counts-j; 232945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2330e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2331c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2332c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2333c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2334c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2335c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2336c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2337674ae819SStefano Zampini if (j<temp_constraints) { 2338984c4197SStefano Zampini PetscInt ii; 2339984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2340674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23419162d606SStefano 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)); 2342674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2343984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2344674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23459162d606SStefano 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]; 2346674ae819SStefano Zampini } 2347674ae819SStefano Zampini } 2348674ae819SStefano Zampini } 2349674ae819SStefano Zampini #else /* on missing GESVD */ 2350e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2351e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2352b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2353674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2354674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23559162d606SStefano 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)); 2356674ae819SStefano Zampini #else 23579162d606SStefano 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)); 2358674ae819SStefano Zampini #endif 2359984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2360674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2361984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2362e310c8b4SStefano Zampini k = temp_constraints; 2363e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2364674ae819SStefano Zampini j = 0; 2365e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 236645a1bb75SStefano Zampini valid_constraints = k-j; 2367911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2368984c4197SStefano Zampini #endif /* on missing GESVD */ 2369674ae819SStefano Zampini } 2370a773dcb8SStefano Zampini } 23719162d606SStefano Zampini /* update pointers information */ 23729162d606SStefano Zampini if (valid_constraints) { 23739162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23749162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23759162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23769162d606SStefano Zampini /* set change_of_basis flag */ 237745a1bb75SStefano Zampini if (boolforchange) { 2378b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23799162d606SStefano Zampini } 2380b3d85658SStefano Zampini total_counts_cc++; 238145a1bb75SStefano Zampini } 238245a1bb75SStefano Zampini } 2383984c4197SStefano Zampini /* free workspace */ 23848f1c130eSStefano Zampini if (!skip_lapack) { 2385984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2386984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2387984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2388984c4197SStefano Zampini #endif 2389984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2390984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2391984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2392984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2393984c4197SStefano Zampini #endif 2394984c4197SStefano Zampini } 2395984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2396984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2397984c4197SStefano Zampini } 2398984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2399cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2400cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2401cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2402cf5a6209SStefano Zampini } 2403cf5a6209SStefano Zampini if (n_ISForFaces) { 2404cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2405cf5a6209SStefano Zampini } 2406cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2407cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2408cf5a6209SStefano Zampini } 2409cf5a6209SStefano Zampini if (n_ISForEdges) { 2410cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2411cf5a6209SStefano Zampini } 2412cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 241308122e43SStefano Zampini } else { 241408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2415984c4197SStefano Zampini 241608122e43SStefano Zampini total_counts = 0; 241708122e43SStefano Zampini n_vertices = 0; 2418d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2419d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 242008122e43SStefano Zampini } 242108122e43SStefano Zampini max_constraints = 0; 24229162d606SStefano Zampini total_counts_cc = 0; 242308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 242408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24259162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 242608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 242708122e43SStefano Zampini } 24289162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24299162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24309162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24319162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 243274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24339162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24349162d606SStefano Zampini total_counts_cc = 0; 24359162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24369162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24379162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 243808122e43SStefano Zampini } 243908122e43SStefano Zampini } 24409162d606SStefano Zampini #if 0 24419162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24429162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24439162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24449162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24459162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24469162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24479162d606SStefano Zampini } 24489162d606SStefano Zampini printf("\n"); 24499162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24509162d606SStefano Zampini } 24511b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24528bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24531b968477SStefano Zampini } 24541b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24558bec7fa6SStefano 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]); 24561b968477SStefano Zampini } 245708122e43SStefano Zampini #endif 245808122e43SStefano Zampini 24598bec7fa6SStefano Zampini max_size_of_constraint = 0; 24609162d606SStefano 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]); 24619162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 246208122e43SStefano Zampini /* Change of basis */ 2463b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 246408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 246508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 246608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2467b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 246808122e43SStefano Zampini } 246908122e43SStefano Zampini } 247008122e43SStefano Zampini } 247108122e43SStefano Zampini } 2472984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24730e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 247408122e43SStefano Zampini 24759162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24769162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24779162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24789162d606SStefano 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); 247908122e43SStefano Zampini } 2480674ae819SStefano Zampini 2481674ae819SStefano Zampini /* Create constraint matrix */ 2482674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 248316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2484984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2485984c4197SStefano Zampini 2486984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2487a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2488a717540cSStefano Zampini qr_needed = PETSC_FALSE; 248974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2490984c4197SStefano Zampini total_primal_vertices=0; 2491b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24929162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24939162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2494984c4197SStefano Zampini if (size_of_constraint == 1) { 24959162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2496b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 249764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24989162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24999162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2500a717540cSStefano Zampini } 2501b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 250274d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2503a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2504a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2505a717540cSStefano Zampini } 2506fa434743SStefano Zampini } else { 2507b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2508fa434743SStefano Zampini } 2509a717540cSStefano Zampini } 2510b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2511b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2512674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 251370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2514b3d85658SStefano Zampini 2515b3d85658SStefano 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); 25160e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25170e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2518984c4197SStefano Zampini 2519984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 252074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2521785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2522984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 252374d5cdf7SStefano Zampini 2524984c4197SStefano Zampini j = total_primal_vertices; 252574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2526b3d85658SStefano Zampini cum = total_primal_vertices; 25279162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25284641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2529b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2530b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2531b3d85658SStefano Zampini cum++; 25329162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 253374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 253474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 253574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 253674d5cdf7SStefano Zampini } 25379162d606SStefano Zampini j += constraints_n[i]; 2538674ae819SStefano Zampini } 2539674ae819SStefano Zampini } 2540674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2541674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2542088faed8SStefano Zampini 2543674ae819SStefano Zampini /* set values in constraint matrix */ 2544984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25450e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2546674ae819SStefano Zampini } 2547984c4197SStefano Zampini total_counts = total_primal_vertices; 25489162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25494641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25509162d606SStefano Zampini PetscInt *cols; 25519162d606SStefano Zampini 25529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25539162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25549162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25559162d606SStefano Zampini PetscInt row = total_counts+k; 25569162d606SStefano Zampini PetscScalar *vals; 25579162d606SStefano Zampini 25589162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25599162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25609162d606SStefano Zampini } 25619162d606SStefano Zampini total_counts += constraints_n[i]; 2562674ae819SStefano Zampini } 2563674ae819SStefano Zampini } 2564674ae819SStefano Zampini /* assembling */ 2565674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2566674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2567088faed8SStefano Zampini 2568984c4197SStefano Zampini /* 256945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2570984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2571984c4197SStefano Zampini */ 2572674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2573674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2574026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2575984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2576984c4197SStefano Zampini /* working stuff for GEQRF */ 257781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2578984c4197SStefano Zampini PetscBLASInt lqr_work; 2579984c4197SStefano Zampini /* working stuff for UNGQR */ 2580984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2581984c4197SStefano Zampini PetscBLASInt lgqr_work; 2582984c4197SStefano Zampini /* working stuff for TRTRS */ 2583984c4197SStefano Zampini PetscScalar *trs_rhs; 25843f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2585984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2586984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2587984c4197SStefano Zampini PetscScalar *start_vals; 2588984c4197SStefano Zampini /* working stuff for values insertion */ 25894641a718SStefano Zampini PetscBT is_primal; 259064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2591906d46d4SStefano Zampini /* matrix sizes */ 2592906d46d4SStefano Zampini PetscInt global_size,local_size; 2593906d46d4SStefano Zampini /* temporary change of basis */ 2594906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2595cf5a6209SStefano Zampini /* extra space for debugging */ 2596cf5a6209SStefano Zampini PetscScalar *dbg_work; 2597984c4197SStefano Zampini 2598906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2599906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 260016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2601bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2602906d46d4SStefano Zampini /* nonzeros for local mat */ 2603bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2604bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 26059162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2606a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26079162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2608a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26099162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2610a717540cSStefano Zampini } else { 26119162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26129162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2613a717540cSStefano Zampini } 2614a717540cSStefano Zampini } 2615a717540cSStefano Zampini } 2616906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2617bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2618a717540cSStefano Zampini /* Set initial identity in the matrix */ 2619bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2620906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2621a717540cSStefano Zampini } 2622a717540cSStefano Zampini 2623a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2624a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2625a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2626a717540cSStefano Zampini } 2627a717540cSStefano Zampini 2628a717540cSStefano Zampini 2629a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2630a717540cSStefano Zampini /* 2631a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2632a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2633a717540cSStefano Zampini 2634a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2635a717540cSStefano Zampini 2636a6b551f4SStefano 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) 2637a6b551f4SStefano Zampini 2638a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2639a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2640a717540cSStefano Zampini | ... | 2641a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2642a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2643a717540cSStefano Zampini 2644a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2645a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2646a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2647a6b551f4SStefano Zampini 2648a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2649a717540cSStefano Zampini */ 2650a717540cSStefano Zampini if (qr_needed) { 2651984c4197SStefano Zampini /* space to store Q */ 2652854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2653984c4197SStefano Zampini /* first we issue queries for optimal work */ 26543f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26553f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26563f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2657984c4197SStefano Zampini lqr_work = -1; 26583f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2659984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2660984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2661785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2662984c4197SStefano Zampini lgqr_work = -1; 26633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26643f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26653f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26663f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26673f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26683f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2669984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2670984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2671785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2672984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2673785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2674984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2675785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2676a717540cSStefano Zampini /* allocating workspace for check */ 2677a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2678cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2679a717540cSStefano Zampini } 2680a717540cSStefano Zampini } 2681984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26824641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2683473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26840e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 268539e2fb2aSStefano Zampini if (i != total_primal_vertices) { 268639e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26874641a718SStefano Zampini } 268839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 268939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 269039e2fb2aSStefano Zampini } 269139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2692984c4197SStefano Zampini 2693a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26949162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26959162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26964641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2697984c4197SStefano Zampini /* get constraint info */ 26989162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2699984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2700984c4197SStefano Zampini 2701984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27029162d606SStefano 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); 2703674ae819SStefano Zampini } 2704984c4197SStefano Zampini 2705fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2706a717540cSStefano Zampini 2707a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2708a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27099162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2710a717540cSStefano Zampini } 2711984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27129162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2713984c4197SStefano Zampini 2714984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2718674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27193f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2720984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2721674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2722984c4197SStefano Zampini 2723984c4197SStefano Zampini /* explictly compute R^-T */ 2724984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2725984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2730984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27313f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2732984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2733984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2734984c4197SStefano Zampini 2735a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27363f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27383f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27393f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2740984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27413f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2742984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2743984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2744984c4197SStefano Zampini 2745984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2746984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2747984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27483f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27513f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27523f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2754984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27559162d606SStefano 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)); 2756984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27579162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2758984c4197SStefano Zampini 2759984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27609162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2761984c4197SStefano Zampini /* insert cols for primal dofs */ 2762984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2763984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27649162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2765906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2766984c4197SStefano Zampini } 2767984c4197SStefano Zampini /* insert cols for dual dofs */ 2768984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27699162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2770984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27719162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2772906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2773984c4197SStefano Zampini j++; 2774674ae819SStefano Zampini } 2775674ae819SStefano Zampini } 2776984c4197SStefano Zampini 2777984c4197SStefano Zampini /* check change of basis */ 2778984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2779984c4197SStefano Zampini PetscInt ii,jj; 2780984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2781c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2782c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2783c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2784c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2785c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2786c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2787984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2788cf5a6209SStefano 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)); 2789984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2790984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2791984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2792cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2793cf5a6209SStefano 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; 2794674ae819SStefano Zampini } 2795674ae819SStefano Zampini } 2796984c4197SStefano Zampini if (!valid_qr) { 279722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2798984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2799984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2800cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2801cf5a6209SStefano 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])); 2802674ae819SStefano Zampini } 2803cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2804cf5a6209SStefano 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])); 2805984c4197SStefano Zampini } 2806984c4197SStefano Zampini } 2807984c4197SStefano Zampini } 2808674ae819SStefano Zampini } else { 280922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2810674ae819SStefano Zampini } 2811674ae819SStefano Zampini } 2812a717540cSStefano Zampini } else { /* simple transformation block */ 2813a717540cSStefano Zampini PetscInt row,col; 2814a6b551f4SStefano Zampini PetscScalar val,norm; 2815a6b551f4SStefano Zampini 2816a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28179162d606SStefano 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)); 2818a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28199162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28209162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2821bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28229162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2823906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28249162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2825a717540cSStefano Zampini } else { 2826a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28279162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2828a717540cSStefano Zampini if (row != col) { 28299162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2830a717540cSStefano Zampini } else { 28319162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2832a717540cSStefano Zampini } 2833906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2834a717540cSStefano Zampini } 2835a717540cSStefano Zampini } 2836a717540cSStefano Zampini } 283798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 283822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2839a717540cSStefano Zampini } 2840674ae819SStefano Zampini } 2841984c4197SStefano Zampini } else { 2842984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28439162d606SStefano 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); 2844674ae819SStefano Zampini } 2845674ae819SStefano Zampini } 2846674ae819SStefano Zampini } 2847a717540cSStefano Zampini 2848a717540cSStefano Zampini /* free workspace */ 2849a717540cSStefano Zampini if (qr_needed) { 2850984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2851cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2852984c4197SStefano Zampini } 2853984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2854984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2855984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2856984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2857984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2858674ae819SStefano Zampini } 2859a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2860906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2861906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2862906d46d4SStefano Zampini 2863906d46d4SStefano Zampini /* assembling of global change of variable */ 2864bbb9e6c6SStefano Zampini { 2865bbb9e6c6SStefano Zampini Mat tmat; 286616f15bc4SStefano Zampini PetscInt bs; 286716f15bc4SStefano Zampini 2868906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2869906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2870bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2872bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2873bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 287416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 287516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2876906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2879bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2880bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2881bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2882e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2883e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2884bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2885bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2886906d46d4SStefano Zampini } 2887906d46d4SStefano Zampini /* check */ 2888906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2889906d46d4SStefano Zampini PetscReal error; 2890906d46d4SStefano Zampini Vec x,x_change; 2891906d46d4SStefano Zampini 2892906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2893906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2894906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2895906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2896e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2897e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2898bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2899e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2900e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2901906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2902906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2903906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2904906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2905bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2906906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2907906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2908906d46d4SStefano Zampini } 2909b96c3477SStefano Zampini 2910b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2911b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2912b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2913b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2914ac632422SStefano Zampini Mat S_new,tmat; 2915bbb9e6c6SStefano Zampini IS is_all_N; 2916bbb9e6c6SStefano Zampini 2917bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29186816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2919bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2920ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2921b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2922ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2923ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2924ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2925ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2926ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2927b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2928ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2929ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2930ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2931ac632422SStefano Zampini } 2932b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2933b96c3477SStefano Zampini } 2934b96c3477SStefano Zampini } 2935906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2936906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2937b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2938b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2939b9b85e73SStefano Zampini } 2940906d46d4SStefano Zampini 2941906d46d4SStefano Zampini /* set up change of basis context */ 2942906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2943906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2944906d46d4SStefano Zampini 2945906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2946906d46d4SStefano Zampini PetscInt global_size,local_size; 2947906d46d4SStefano Zampini 2948906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2951906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2953906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2954906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2955906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2956906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2957906d46d4SStefano Zampini } else { 2958906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2959906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2960906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2961906d46d4SStefano Zampini } 2962906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2963906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2964906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2965906d46d4SStefano Zampini } else { 2966906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2967906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2968906d46d4SStefano Zampini } 2969906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2970906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2971906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2972906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2973b9b85e73SStefano Zampini } 2974a717540cSStefano Zampini 2975727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2976727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2977727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2978aff50787SStefano 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); 2979c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29800e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2981aff50787SStefano 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); 2982727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2983727cdba6SStefano Zampini } 29840e6343abSStefano Zampini } 29850e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2986727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2987727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2988727cdba6SStefano Zampini 2989a717540cSStefano Zampini /* flush dbg viewer */ 2990b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2991b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2992b8ffe317SStefano Zampini } 2993a717540cSStefano Zampini 2994e310c8b4SStefano Zampini /* free workspace */ 2995a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29964641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 299708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29989162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29999162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 300008122e43SStefano Zampini } else { 30019162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 30029162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 30039162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 300408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 300508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 30069162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30079162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 300808122e43SStefano Zampini } 3009674ae819SStefano Zampini PetscFunctionReturn(0); 3010674ae819SStefano Zampini } 3011674ae819SStefano Zampini 3012674ae819SStefano Zampini #undef __FUNCT__ 3013674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3014674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3015674ae819SStefano Zampini { 3016674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3017674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3018674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30197fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3020674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3021674ae819SStefano Zampini 3022674ae819SStefano Zampini PetscFunctionBegin; 30238e61c736SStefano Zampini /* Reset previously computed graph */ 30248e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3025674ae819SStefano Zampini /* Init local Graph struct */ 30267fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30273bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3028674ae819SStefano Zampini 3029575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3030575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3031575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3032575ad6abSStefano Zampini } 30339577ea80SStefano Zampini 3034674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30354d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30364d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30374d379d7bSStefano Zampini PetscInt nvtxs; 3038e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3039674ae819SStefano Zampini 30404d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30412fffb893SStefano Zampini 30422fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30432fffb893SStefano Zampini if (flg_row) { 30444d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3045b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30462fffb893SStefano Zampini } 30472fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30489b28b941SStefano 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 */ 30494d379d7bSStefano Zampini IS is_dummy; 30504d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30514d379d7bSStefano Zampini PetscInt j,sum; 30524d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30534d379d7bSStefano Zampini const PetscInt *idxs; 30544d379d7bSStefano Zampini PCBDDCGraph graph; 30554d379d7bSStefano Zampini PetscBT is_on_boundary; 30564d379d7bSStefano Zampini 30574d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30584d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30594d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30604d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30617fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30624d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3063e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3064e496cd5dSStefano Zampini if (flg_row) { 30654d379d7bSStefano Zampini graph->xadj = xadj; 30664d379d7bSStefano Zampini graph->adjncy = adjncy; 3067e496cd5dSStefano Zampini } 30684d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30694d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3070e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30714d379d7bSStefano Zampini 30724d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30739b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30744d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30754d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30764d379d7bSStefano Zampini } 30774d379d7bSStefano Zampini } 30784d379d7bSStefano Zampini 3079e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30804d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30814d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30824d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30834d379d7bSStefano Zampini } 30844d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30854d379d7bSStefano Zampini 3086e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30874d379d7bSStefano Zampini sum = 0; 30884d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30894d379d7bSStefano Zampini PetscInt sizecc = 0; 30904d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30914d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30924d379d7bSStefano Zampini sizecc++; 30934d379d7bSStefano Zampini } 30944d379d7bSStefano Zampini } 30954d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30964d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30974d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini } 31004d379d7bSStefano Zampini sum += sizecc*sizecc; 31014d379d7bSStefano Zampini } 31024d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 31034d379d7bSStefano Zampini sum = 0; 3104e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 31054d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 31064d379d7bSStefano Zampini cxadj[i] = sum; 31074d379d7bSStefano Zampini sum += temp; 31084d379d7bSStefano Zampini } 3109e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31104d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31114d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31124d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31134d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31144d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31154d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31164d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31174d379d7bSStefano Zampini sizecc++; 31184d379d7bSStefano Zampini } 31194d379d7bSStefano Zampini } 31204d379d7bSStefano Zampini } 31214d379d7bSStefano Zampini } 31224d379d7bSStefano Zampini } 31239b28b941SStefano Zampini if (sum) { 3124e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31254d379d7bSStefano Zampini } else { 31264d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31274d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31284d379d7bSStefano Zampini } 31294d379d7bSStefano Zampini graph->xadj = 0; 31304d379d7bSStefano Zampini graph->adjncy = 0; 31314d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31324d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31334d379d7bSStefano Zampini } 3134674ae819SStefano Zampini } 31359b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31369b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3137674ae819SStefano Zampini } 3138674ae819SStefano Zampini 313963602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3140674ae819SStefano Zampini vertex_size = 1; 314163602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 314263602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 314395ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 314463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3145e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314663602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3147674ae819SStefano Zampini } 314863602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 314963602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 315063602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3151674ae819SStefano Zampini } 315263602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3153674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 315463602bcaSStefano Zampini } else { 315563602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 315663602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3157854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 315863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 315963602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 316063602bcaSStefano Zampini } 316163602bcaSStefano Zampini } 3162674ae819SStefano Zampini } 3163674ae819SStefano Zampini 3164674ae819SStefano Zampini /* Setup of Graph */ 3165785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3166e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3167785d1243SStefano Zampini } 3168785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3169e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3170785d1243SStefano Zampini } 3171302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3172674ae819SStefano Zampini 3173674ae819SStefano Zampini /* Graph's connected components analysis */ 3174674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3175674ae819SStefano Zampini 3176674ae819SStefano Zampini /* print some info to stdout */ 3177674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3178302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3179674ae819SStefano Zampini } 3180fb180af4SStefano Zampini 3181fb180af4SStefano Zampini /* mark topography has done */ 3182fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3183674ae819SStefano Zampini PetscFunctionReturn(0); 3184674ae819SStefano Zampini } 3185674ae819SStefano Zampini 3186dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3187674ae819SStefano Zampini #undef __FUNCT__ 3188674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3189dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3190674ae819SStefano Zampini { 3191dc456d91SStefano Zampini PetscSF sf; 3192dc456d91SStefano Zampini PetscLayout map; 3193dc456d91SStefano Zampini const PetscInt *idxs; 3194dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3195dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3196dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3197dc456d91SStefano Zampini PetscMPIInt commsize; 3198674ae819SStefano Zampini PetscBool first_found; 3199674ae819SStefano Zampini PetscErrorCode ierr; 3200674ae819SStefano Zampini 3201674ae819SStefano Zampini PetscFunctionBegin; 3202dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3203dc456d91SStefano Zampini if (subset_mult) { 3204dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3205dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3206dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3207674ae819SStefano Zampini } 3208dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3209dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3210dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3211dc456d91SStefano Zampini for (i=0;i<n;i++) { 3212dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3213dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3214674ae819SStefano Zampini } 3215dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3216dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3217dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3218dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3219dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3220dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3221dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3222dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3223dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3224dc456d91SStefano Zampini 3225dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3226dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3227dc456d91SStefano Zampini if (subset_mult) { 3228dc456d91SStefano Zampini const PetscInt* idxs_mult; 3229dc456d91SStefano Zampini 3230dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3231dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3232dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3233674ae819SStefano Zampini } else { 3234dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3235674ae819SStefano Zampini } 3236dc456d91SStefano Zampini /* local size of new subset */ 3237dc456d91SStefano Zampini n_n = 0; 3238dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3239dc456d91SStefano Zampini 3240dc456d91SStefano Zampini /* global indexes in layout */ 3241dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3242dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3243dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3244dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3245dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3246dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3247dc456d91SStefano Zampini 3248dc456d91SStefano Zampini /* reduce from leaves to roots */ 3249dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 325064a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 325164a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3252dc456d91SStefano Zampini 3253dc456d91SStefano Zampini /* count indexes in local part of layout */ 3254674ae819SStefano Zampini nlocals = 0; 3255674ae819SStefano Zampini first_index = -1; 3256674ae819SStefano Zampini first_found = PETSC_FALSE; 3257dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3258dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3259674ae819SStefano Zampini first_found = PETSC_TRUE; 3260674ae819SStefano Zampini first_index = i; 3261674ae819SStefano Zampini } 3262dc456d91SStefano Zampini nlocals += root_data[i]; 3263674ae819SStefano Zampini } 3264dc456d91SStefano Zampini 3265dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32665fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3267dc456d91SStefano Zampini start = 0; 326864a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32695fa240b1SStefano Zampini #else 327064a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32715fa240b1SStefano Zampini start = start-nlocals; 32725fa240b1SStefano Zampini #endif 32735fa240b1SStefano Zampini 3274dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3275dc456d91SStefano Zampini *N_n = start + nlocals; 3276dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3277dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3278674ae819SStefano Zampini } 32795fa240b1SStefano Zampini 32805fa240b1SStefano Zampini /* adapt root data with cumulative */ 3281674ae819SStefano Zampini if (first_found) { 3282dc456d91SStefano Zampini PetscInt old_index; 3283dc456d91SStefano Zampini 3284dc456d91SStefano Zampini root_data[first_index] += start; 3285674ae819SStefano Zampini old_index = first_index; 3286dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3287dc456d91SStefano Zampini if (root_data[i]) { 3288dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3289674ae819SStefano Zampini old_index = i; 3290674ae819SStefano Zampini } 3291674ae819SStefano Zampini } 3292674ae819SStefano Zampini } 3293dc456d91SStefano Zampini 3294dc456d91SStefano Zampini /* from roots to leaves */ 3295dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3296dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3297dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3298dc456d91SStefano Zampini 3299dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3300dc456d91SStefano Zampini if (subset_mult) { 3301dc456d91SStefano Zampini const PetscInt* idxs_mult; 3302dc456d91SStefano Zampini PetscInt cum; 3303dc456d91SStefano Zampini 3304dc456d91SStefano Zampini cum = 0; 3305dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3306dc456d91SStefano Zampini for (i=0;i<n;i++) { 3307dc456d91SStefano Zampini PetscInt j; 3308dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3309674ae819SStefano Zampini } 3310dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3311674ae819SStefano Zampini } else { 3312dc456d91SStefano Zampini for (i=0;i<n;i++) { 3313dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3314674ae819SStefano Zampini } 3315674ae819SStefano Zampini } 3316dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3317dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3318674ae819SStefano Zampini PetscFunctionReturn(0); 3319674ae819SStefano Zampini } 33209a7d3425SStefano Zampini 33219a7d3425SStefano Zampini #undef __FUNCT__ 33229a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33239a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33249a7d3425SStefano Zampini { 33259a7d3425SStefano Zampini PetscInt i,j; 33269a7d3425SStefano Zampini PetscScalar *alphas; 33279a7d3425SStefano Zampini PetscErrorCode ierr; 33289a7d3425SStefano Zampini 33299a7d3425SStefano Zampini PetscFunctionBegin; 33309a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3331785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33329a7d3425SStefano Zampini for (i=0;i<n;i++) { 33339a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33349a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33359a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33369a7d3425SStefano Zampini } 33379a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33389a7d3425SStefano Zampini PetscFunctionReturn(0); 33399a7d3425SStefano Zampini } 33409a7d3425SStefano Zampini 3341e7931f94SStefano Zampini #undef __FUNCT__ 334270cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3343b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3344e7931f94SStefano Zampini { 334552e5ac9dSStefano Zampini IS ranks_send_to; 3346e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3347e7931f94SStefano Zampini PetscMPIInt size,rank,color; 334852e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 334952e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33503837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33512b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3352e7931f94SStefano Zampini PetscSubcomm subcomm; 335352e5ac9dSStefano Zampini PetscErrorCode ierr; 3354a57a6d2fSStefano Zampini 3355e7931f94SStefano Zampini PetscFunctionBegin; 33562b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33572b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33582b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3359e7931f94SStefano Zampini 3360e7931f94SStefano Zampini /* Get info on mapping */ 33613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33623bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3363e7931f94SStefano Zampini 3364e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3365785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3366e7931f94SStefano Zampini xadj[0] = 0; 3367e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3368785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3369785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3370e7931f94SStefano Zampini 33712b510759SStefano Zampini if (threshold) { 3372d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33732b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3374d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3375d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3376d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3377d023bfaeSStefano Zampini xadj_count++; 3378e7931f94SStefano Zampini } 3379e7931f94SStefano Zampini } 3380d023bfaeSStefano Zampini xadj[1] = xadj_count; 3381c8587f34SStefano Zampini } else { 3382e7931f94SStefano Zampini if (xadj[1]) { 3383e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3384e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3385c8587f34SStefano Zampini } 3386e7931f94SStefano Zampini } 33873bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3388e7931f94SStefano Zampini if (use_square) { 3389e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3390e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3391e7931f94SStefano Zampini } 3392e7931f94SStefano Zampini } 3393e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3394e7931f94SStefano Zampini 33953837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3396e7931f94SStefano Zampini 3397e7931f94SStefano Zampini /* 3398e7931f94SStefano Zampini Restrict work on active processes only. 3399e7931f94SStefano Zampini */ 3400e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3401e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3402e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 34032b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3404d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3405e7931f94SStefano Zampini if (color) { 3406e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3407e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3408e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3409c8587f34SStefano Zampini } else { 341052e5ac9dSStefano Zampini Mat subdomain_adj; 341152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 341252e5ac9dSStefano Zampini MatPartitioning partitioner; 341352e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 341452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3415b0c7d250SStefano Zampini PetscBool aggregate; 3416b0c7d250SStefano Zampini 3417306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3418785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3419e7931f94SStefano Zampini prank = rank; 3420306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34218002ef2cSStefano Zampini /* 3422e7931f94SStefano Zampini for (i=0;i<size;i++) { 3423e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3424c8587f34SStefano Zampini } 34258002ef2cSStefano Zampini */ 3426e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3427e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3428c8587f34SStefano Zampini } 3429e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3430b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3431b0c7d250SStefano Zampini if (aggregate) { 3432b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3433b0c7d250SStefano Zampini PetscMPIInt nrank; 3434b0c7d250SStefano Zampini PetscScalar *vals; 3435b0c7d250SStefano Zampini 3436b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3437b0c7d250SStefano Zampini lrows = 0; 3438b0c7d250SStefano Zampini if (nrank<redprocs) { 3439b0c7d250SStefano Zampini lrows = size/redprocs; 3440b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3441b0c7d250SStefano Zampini } 34425fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3443b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3444b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3445b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3446b0c7d250SStefano Zampini row = nrank; 3447b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3448b0c7d250SStefano Zampini cols = adjncy; 3449b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3450b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3451b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3452b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3453b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 345452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 345552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 345652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3457b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3458b0c7d250SStefano Zampini } else { 3459306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3460b0c7d250SStefano Zampini } 346122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3462e7931f94SStefano Zampini 3463e7931f94SStefano Zampini /* Partition */ 3464306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3465e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3466e7931f94SStefano Zampini if (use_vwgt) { 34673837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3468e7931f94SStefano Zampini v_wgt[0] = local_size; 3469e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3470c8587f34SStefano Zampini } 347128143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 347228143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3473e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3474e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 347522b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3476e7931f94SStefano Zampini 347752e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 347852e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 347952e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 348052e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3481b0c7d250SStefano Zampini if (!redprocs) { 3482b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 348328143c3dSStefano Zampini } else { 3484b0c7d250SStefano Zampini PetscInt idxs[1]; 3485b0c7d250SStefano Zampini PetscMPIInt tag; 3486b0c7d250SStefano Zampini MPI_Request *reqs; 3487b0c7d250SStefano Zampini 3488b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3489b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3490b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3491b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 349228143c3dSStefano Zampini } 3493b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3494b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3495b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3496b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3497e7931f94SStefano Zampini } 349852e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3499e7931f94SStefano Zampini /* clean up */ 3500e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 350152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3502e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3503e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3504e7931f94SStefano Zampini } 3505e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3506e7931f94SStefano Zampini 3507e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3508e7931f94SStefano Zampini i = 1; 3509e7931f94SStefano Zampini if (color) i=0; 3510e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3511e7931f94SStefano Zampini /* get back IS */ 3512e7931f94SStefano Zampini *is_sends = ranks_send_to; 3513e7931f94SStefano Zampini PetscFunctionReturn(0); 3514e7931f94SStefano Zampini } 3515e7931f94SStefano Zampini 3516e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3517e7931f94SStefano Zampini 3518e7931f94SStefano Zampini #undef __FUNCT__ 3519e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 352053a05cb3SStefano 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[]) 3521e7931f94SStefano Zampini { 352270cf5478SStefano Zampini Mat local_mat; 3523e7931f94SStefano Zampini IS is_sends_internal; 35249d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 352528143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35269d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3527e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3528e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3529e7931f94SStefano Zampini const PetscInt* is_indices; 3530e7931f94SStefano Zampini MatType new_local_type; 3531e7931f94SStefano Zampini /* buffers */ 3532e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 353328143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35349d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3535e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3536e7931f94SStefano Zampini /* MPI */ 353728143c3dSStefano Zampini MPI_Comm comm,comm_n; 353828143c3dSStefano Zampini PetscSubcomm subcomm; 3539e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 354028143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 354128143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 354228143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 354328143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 354428143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3545e7931f94SStefano Zampini PetscErrorCode ierr; 3546e7931f94SStefano Zampini 3547e7931f94SStefano Zampini PetscFunctionBegin; 354828143c3dSStefano Zampini /* TODO: add missing checks */ 354928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 355028143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 355128143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 355228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3553e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 355428143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3555e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3556e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3557e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3558e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3559e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 356028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 356170cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 356270cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 356328143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 356470cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 356570cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 356670cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 356770cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 356870cf5478SStefano Zampini } 3569e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3570e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3571e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3572e7931f94SStefano Zampini if (!is_sends) { 357328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3574b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3575c8587f34SStefano Zampini } else { 3576e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3577e7931f94SStefano Zampini is_sends_internal = is_sends; 3578c8587f34SStefano Zampini } 3579e7931f94SStefano Zampini 3580e7931f94SStefano Zampini /* get comm */ 3581a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3582e7931f94SStefano Zampini 3583e7931f94SStefano Zampini /* compute number of sends */ 3584e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3585e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3586e7931f94SStefano Zampini 3587e7931f94SStefano Zampini /* compute number of receives */ 3588e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3589785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3590e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3591e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3592e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3593e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3594e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3595e7931f94SStefano Zampini 359628143c3dSStefano Zampini /* restrict comm if requested */ 359728143c3dSStefano Zampini subcomm = 0; 359828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 359928143c3dSStefano Zampini if (restrict_comm) { 3600779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3601779c1cceSStefano Zampini 360228143c3dSStefano Zampini color = 0; 360353a05cb3SStefano Zampini if (restrict_full) { 360453a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 360553a05cb3SStefano Zampini } else { 360653a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 360753a05cb3SStefano Zampini } 360828143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 360928143c3dSStefano Zampini subcommsize = commsize - subcommsize; 361028143c3dSStefano Zampini /* check if reuse has been requested */ 361128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 361228143c3dSStefano Zampini if (*mat_n) { 361328143c3dSStefano Zampini PetscMPIInt subcommsize2; 361428143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 361528143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 361628143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 361728143c3dSStefano Zampini } else { 361828143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 361928143c3dSStefano Zampini } 362028143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3621779c1cceSStefano Zampini PetscMPIInt rank; 3622779c1cceSStefano Zampini 3623779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 362428143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 362528143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 362628143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3627306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 362828143c3dSStefano Zampini } 362928143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 363028143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 363128143c3dSStefano Zampini } else { 363228143c3dSStefano Zampini comm_n = comm; 363328143c3dSStefano Zampini } 363428143c3dSStefano Zampini 3635e7931f94SStefano Zampini /* prepare send/receive buffers */ 3636785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3637e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3638785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3639e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 364028143c3dSStefano Zampini if (nis) { 3641854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 364228143c3dSStefano Zampini } 3643e7931f94SStefano Zampini 364428143c3dSStefano Zampini /* Get data from local matrices */ 3645e7931f94SStefano Zampini if (!isdense) { 3646a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3647e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3648e7931f94SStefano Zampini /* 3649e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3650e7931f94SStefano Zampini send_buffer_idxs should contain: 3651e7931f94SStefano Zampini - MatType_PRIVATE type 3652e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3653e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3654e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3655e7931f94SStefano Zampini */ 3656e7931f94SStefano Zampini } else { 3657e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3659854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3660e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3661e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36623bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3663e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36643bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3665e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3666e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3667e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3668e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3669c8587f34SStefano Zampini } 3670c8587f34SStefano Zampini } 3671e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 367228143c3dSStefano Zampini /* additional is (if any) */ 367328143c3dSStefano Zampini if (nis) { 367428143c3dSStefano Zampini PetscMPIInt psum; 367528143c3dSStefano Zampini PetscInt j; 367628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 367728143c3dSStefano Zampini PetscInt plen; 367828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 367928143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 368028143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 368128143c3dSStefano Zampini } 3682854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 368328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 368428143c3dSStefano Zampini PetscInt plen; 368528143c3dSStefano Zampini const PetscInt *is_array_idxs; 368628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 368728143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 368828143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 368928143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 369028143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 369128143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 369228143c3dSStefano Zampini } 369328143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 369428143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 369528143c3dSStefano Zampini } 369628143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 369728143c3dSStefano Zampini } 369828143c3dSStefano Zampini 3699e7931f94SStefano Zampini buf_size_idxs = 0; 3700e7931f94SStefano Zampini buf_size_vals = 0; 370128143c3dSStefano Zampini buf_size_idxs_is = 0; 3702e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3703e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3704e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 370528143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3706e7931f94SStefano Zampini } 3707785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3708785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 370995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3710e7931f94SStefano Zampini 3711e7931f94SStefano Zampini /* get new tags for clean communications */ 3712e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3713e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 371428143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3715e7931f94SStefano Zampini 3716e7931f94SStefano Zampini /* allocate for requests */ 3717785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3718785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 371995ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3720785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3721785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 372295ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3723e7931f94SStefano Zampini 3724e7931f94SStefano Zampini /* communications */ 3725e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3726e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 372728143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3728e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3729e7931f94SStefano Zampini source_dest = onodes[i]; 3730e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3731e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3732e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3733e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 373428143c3dSStefano Zampini if (nis) { 373528143c3dSStefano 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); 373628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 373728143c3dSStefano Zampini } 3738e7931f94SStefano Zampini } 3739e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3740e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3741e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3742e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 374328143c3dSStefano Zampini if (nis) { 374428143c3dSStefano 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); 374528143c3dSStefano Zampini } 3746e7931f94SStefano Zampini } 3747e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3748e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3749e7931f94SStefano Zampini 3750e7931f94SStefano Zampini /* assemble new l2g map */ 3751e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3752e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37539d30be91SStefano Zampini new_local_rows = 0; 3754e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37559d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3756e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3757e7931f94SStefano Zampini } 37589d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3759e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37609d30be91SStefano Zampini new_local_rows = 0; 3761e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37629d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37639d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3764e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3765e7931f94SStefano Zampini } 37669d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37679d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3768e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3769e7931f94SStefano Zampini 3770e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3771e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3772e7931f94SStefano 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) */ 3773e7931f94SStefano Zampini if (n_recvs) { 377428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3775e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3776e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3777e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3778e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3779e7931f94SStefano Zampini break; 3780e7931f94SStefano Zampini } 3781e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3782e7931f94SStefano Zampini } 3783e7931f94SStefano Zampini switch (new_local_type_private) { 378428143c3dSStefano Zampini case MATDENSE_PRIVATE: 378528143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3786e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3787e7931f94SStefano Zampini bs = 1; 378828143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 378928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 379028143c3dSStefano Zampini bs = 1; 379128143c3dSStefano Zampini } 3792e7931f94SStefano Zampini break; 3793e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3794e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3795e7931f94SStefano Zampini bs = 1; 3796e7931f94SStefano Zampini break; 3797e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3798e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3799e7931f94SStefano Zampini break; 3800e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3801e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3802e7931f94SStefano Zampini break; 3803e7931f94SStefano Zampini default: 38049d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3805e7931f94SStefano Zampini break; 3806e7931f94SStefano Zampini } 380728143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 380828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 380928143c3dSStefano Zampini bs = 1; 3810e7931f94SStefano Zampini } 3811e7931f94SStefano Zampini 381270cf5478SStefano Zampini /* create MATIS object if needed */ 381370cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3814e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3815e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 381670cf5478SStefano Zampini } else { 381770cf5478SStefano Zampini /* it also destroys the local matrices */ 381870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 381970cf5478SStefano Zampini } 382070cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3821e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 38229d30be91SStefano Zampini 38239d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38249d30be91SStefano Zampini 38259d30be91SStefano Zampini /* Global to local map of received indices */ 38269d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38279d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38289d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38299d30be91SStefano Zampini 38309d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38319d30be91SStefano Zampini buf_size_idxs = 0; 38329d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38339d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38349d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38359d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38369d30be91SStefano Zampini } 38379d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38389d30be91SStefano Zampini 38399d30be91SStefano Zampini /* set preallocation */ 38409d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38419d30be91SStefano Zampini if (!newisdense) { 38429d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38439d30be91SStefano Zampini 38449d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38459d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38469d30be91SStefano Zampini if (n_recvs) { 38479d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38489d30be91SStefano Zampini } 38499d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38509d30be91SStefano Zampini PetscInt j; 38519d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38529d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38539d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38549d30be91SStefano Zampini } 38559d30be91SStefano Zampini } else { 38569d30be91SStefano Zampini /* TODO */ 38579d30be91SStefano Zampini } 38589d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38599d30be91SStefano Zampini } 38609d30be91SStefano Zampini if (new_local_nnz) { 38619d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38629d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38639d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38649d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38659d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38669d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38679d30be91SStefano Zampini } else { 38689d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38699d30be91SStefano Zampini } 38709d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38719d30be91SStefano Zampini } else { 38729d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38739d30be91SStefano Zampini } 3874e7931f94SStefano Zampini 3875e7931f94SStefano Zampini /* set values */ 3876e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38779d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3878e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3879e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3880e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38819d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3882e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3883e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3884e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 388528143c3dSStefano Zampini } else { 388628143c3dSStefano Zampini /* TODO */ 3887e7931f94SStefano Zampini } 3888e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3889e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3890e7931f94SStefano Zampini } 3891e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3892e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389370cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389470cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38959d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38969d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3897e7931f94SStefano Zampini 3898dfd14d43SStefano Zampini #if 0 389928143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3900e7931f94SStefano Zampini Vec lvec,rvec; 3901e7931f94SStefano Zampini PetscReal infty_error; 3902e7931f94SStefano Zampini 39032a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3904e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3906e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 390770cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3910e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3911e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3912e7931f94SStefano Zampini } 391328143c3dSStefano Zampini #endif 3914e7931f94SStefano Zampini 391528143c3dSStefano Zampini /* assemble new additional is (if any) */ 391628143c3dSStefano Zampini if (nis) { 391728143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 391828143c3dSStefano Zampini 391928143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3920854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 392128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 392228143c3dSStefano Zampini psum = 0; 392328143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 392428143c3dSStefano Zampini for (j=0;j<nis;j++) { 392528143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 392628143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 392728143c3dSStefano Zampini psum += plen; 392828143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 392928143c3dSStefano Zampini } 393028143c3dSStefano Zampini } 3931854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3932854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 393328143c3dSStefano Zampini for (i=1;i<nis;i++) { 393428143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 393528143c3dSStefano Zampini } 393628143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 393728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 393828143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 393928143c3dSStefano Zampini for (j=0;j<nis;j++) { 394028143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 394128143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 394228143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 394328143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 394428143c3dSStefano Zampini } 394528143c3dSStefano Zampini } 394628143c3dSStefano Zampini for (i=0;i<nis;i++) { 394728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 394828143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 394928143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 395028143c3dSStefano Zampini } 395128143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 395228143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 395328143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 395428143c3dSStefano Zampini } 3955e7931f94SStefano Zampini /* free workspace */ 395628143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3958e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3959e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3960e7931f94SStefano Zampini if (isdense) { 3961e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3962e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3963e7931f94SStefano Zampini } else { 3964e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3965e7931f94SStefano Zampini } 396628143c3dSStefano Zampini if (nis) { 396728143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 396828143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 396928143c3dSStefano Zampini } 3970e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3971e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 397228143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3973e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3974e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 397528143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3976e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3977e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3978e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3979e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3980e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 398128143c3dSStefano Zampini if (nis) { 398228143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 398328143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 398428143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 398528143c3dSStefano Zampini } 398628143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 398728143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 398828143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 398928143c3dSStefano Zampini for (i=0;i<nis;i++) { 399028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 399128143c3dSStefano Zampini } 399253a05cb3SStefano Zampini *mat_n = NULL; 399328143c3dSStefano Zampini } 3994e7931f94SStefano Zampini PetscFunctionReturn(0); 3995e7931f94SStefano Zampini } 3996a57a6d2fSStefano Zampini 399712edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3998af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 399912edc857SStefano Zampini 4000c8587f34SStefano Zampini #undef __FUNCT__ 4001c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4002c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4003c8587f34SStefano Zampini { 4004c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4005c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 400620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 40079881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 400820a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 40096e683305SStefano Zampini IS coarse_is,*isarray; 40106e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 40116e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4012f9eb5b7dSStefano Zampini PC pc_temp; 4013c8587f34SStefano Zampini PCType coarse_pc_type; 4014c8587f34SStefano Zampini KSPType coarse_ksp_type; 4015f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 40164f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 40176e683305SStefano Zampini Mat t_coarse_mat_is; 40186e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 40196e683305SStefano Zampini PetscMPIInt all_procs; 402074e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 402168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 402222bc73bbSStefano Zampini PetscScalar *array; 40239881197aSStefano Zampini PetscErrorCode ierr; 4024fdc09c96SStefano Zampini 4025c8587f34SStefano Zampini PetscFunctionBegin; 4026c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 402768457ee5SStefano 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 */ 4028fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40295a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4030fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4031f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4032f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4033f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4034fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 403551bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 403651bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4037dc4bcba2SStefano Zampini PC pc; 4038dc4bcba2SStefano Zampini PetscBool isbddc; 4039dc4bcba2SStefano Zampini 4040dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4041dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4042dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4043dc4bcba2SStefano Zampini if (isbddc) { 4044dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4045dc4bcba2SStefano Zampini } 4046727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4047fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4048fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4049fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4050f4ddd8eeSStefano Zampini } 4051fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4052fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4053f4ddd8eeSStefano Zampini } 405470cf5478SStefano Zampini /* reset any subassembling information */ 405570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40566e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40576e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4058fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4059f4ddd8eeSStefano Zampini } 4060c8587f34SStefano Zampini 40616e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40622b510759SStefano Zampini im_active = !!(pcis->n); 40632b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40646e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40656e683305SStefano Zampini void_procs = all_procs-active_procs; 40666e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 406774e2c79eSStefano Zampini redist = PETSC_FALSE; 406822bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40696e683305SStefano Zampini csin_ml = PETSC_TRUE; 40706e683305SStefano Zampini ncoarse_ml = void_procs; 4071779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4072779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40736e683305SStefano Zampini csin_ds = PETSC_TRUE; 407418a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 407518a45a71SStefano Zampini redist = PETSC_TRUE; 407618a45a71SStefano Zampini } else { 40776e683305SStefano Zampini csin_ds = PETSC_TRUE; 4078779c1cceSStefano Zampini ncoarse_ds = active_procs; 4079779c1cceSStefano Zampini redist = PETSC_TRUE; 408018a45a71SStefano Zampini } 40816e683305SStefano Zampini } else { 40826e683305SStefano Zampini csin_ml = PETSC_FALSE; 40836e683305SStefano Zampini ncoarse_ml = all_procs; 40846e683305SStefano Zampini if (void_procs) { 40856e683305SStefano Zampini csin_ds = PETSC_TRUE; 40866e683305SStefano Zampini ncoarse_ds = void_procs; 40876e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40886e683305SStefano Zampini } else { 4089779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 409074e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 409174e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 409274e2c79eSStefano Zampini redist = PETSC_TRUE; 409374e2c79eSStefano Zampini } else { 40946e683305SStefano Zampini csin_ds = PETSC_FALSE; 40956e683305SStefano Zampini ncoarse_ds = all_procs; 40966e683305SStefano Zampini } 40976e683305SStefano Zampini } 409874e2c79eSStefano Zampini } 40996e683305SStefano Zampini 41006e683305SStefano Zampini /* 41016e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 41026e683305SStefano Zampini - we have not exceeded the number of levels requested 41036e683305SStefano Zampini - we can actually subassemble the active processes 41046e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 41056e683305SStefano Zampini */ 41066e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 41076e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 41086e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 41096e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 41106e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4111f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 41122b510759SStefano Zampini } else { 4113f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4114c8587f34SStefano Zampini } 4115c8587f34SStefano Zampini } 41166e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 41176e683305SStefano Zampini if (multilevel_allowed) { 41186e683305SStefano Zampini ncoarse = ncoarse_ml; 41196e683305SStefano Zampini csin = csin_ml; 412058da7f69SStefano Zampini redist = PETSC_FALSE; 41216e683305SStefano Zampini } else { 41226e683305SStefano Zampini ncoarse = ncoarse_ds; 41236e683305SStefano Zampini csin = csin_ds; 41246e683305SStefano Zampini } 4125e7931f94SStefano Zampini 4126abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4127abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4128abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4129abbbba34SStefano Zampini 4130abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 413122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 413222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 413322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 413422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4135b9b85e73SStefano Zampini #if 0 4136b9b85e73SStefano Zampini { 4137b9b85e73SStefano Zampini PetscViewer viewer; 4138b9b85e73SStefano Zampini char filename[256]; 4139b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4140b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4141b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4142b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4143b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4144b9b85e73SStefano Zampini } 4145b9b85e73SStefano Zampini #endif 4146e176bc59SStefano 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); 41476e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41486e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41496e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4150abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4151abbbba34SStefano Zampini 41526e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41536e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41546e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41556e683305SStefano Zampini const PetscInt *idxs; 41566e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41576e683305SStefano Zampini 41586e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41590be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41606e683305SStefano Zampini /* allocate space for temporary storage */ 4161854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4162854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41636e683305SStefano Zampini /* allocate for IS array */ 41646e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41656e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41666e683305SStefano Zampini nis = nisdofs + nisneu; 4167854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41686e683305SStefano Zampini /* dofs splitting */ 41696e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41706e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41716e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41726e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41736e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41746e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41756e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41766e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41776e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41786e683305SStefano Zampini } 41796e683305SStefano Zampini /* neumann boundaries */ 41806e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41816e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41826e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41836e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41846e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41856e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41866e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41876e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41886e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41896e683305SStefano Zampini } 41906e683305SStefano Zampini /* free memory */ 41916e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41926e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41936e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41946e683305SStefano Zampini } else { 41956e683305SStefano Zampini nis = 0; 41966e683305SStefano Zampini nisdofs = 0; 41976e683305SStefano Zampini nisneu = 0; 41986e683305SStefano Zampini isarray = NULL; 41996e683305SStefano Zampini } 42006e683305SStefano Zampini /* destroy no longer needed map */ 42016e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 42026e683305SStefano Zampini 42036e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 42046e683305SStefano Zampini coarse_mat_is = NULL; 42056e683305SStefano Zampini if (csin) { 42066e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 420774e2c79eSStefano Zampini if (redist) { 420874e2c79eSStefano Zampini PetscMPIInt rank; 4209779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 421074e2c79eSStefano Zampini 421174e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 421258da7f69SStefano Zampini spc = active_procs/ncoarse; 421358da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4214779c1cceSStefano Zampini if (im_active) { 4215779c1cceSStefano Zampini destsize = 1; 421674e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 421774e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 421874e2c79eSStefano Zampini } else { 421974e2c79eSStefano Zampini dest[0] = rank/(spc+1); 422074e2c79eSStefano Zampini } 422174e2c79eSStefano Zampini } else { 4222779c1cceSStefano Zampini destsize = 0; 42236e683305SStefano Zampini } 4224779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4225779c1cceSStefano Zampini } else if (csin_type_simple) { 42266e683305SStefano Zampini PetscMPIInt rank; 42276e683305SStefano Zampini PetscInt issize,isidx; 4228779c1cceSStefano Zampini 42296e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42306e683305SStefano Zampini if (im_active) { 42316e683305SStefano Zampini issize = 1; 42326e683305SStefano Zampini isidx = (PetscInt)rank; 42336e683305SStefano Zampini } else { 42346e683305SStefano Zampini issize = 0; 42356e683305SStefano Zampini isidx = -1; 42366e683305SStefano Zampini } 42376e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4238779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4239b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42406e683305SStefano Zampini } 4241779c1cceSStefano Zampini 4242779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4243779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4244779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4245779c1cceSStefano Zampini PetscInt *coarse_candidates; 4246779c1cceSStefano Zampini const PetscInt* tisindices; 4247779c1cceSStefano Zampini 4248779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4249779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4250779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4251779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4252779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4253779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4254779c1cceSStefano Zampini } 4255779c1cceSStefano Zampini } 4256779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4257779c1cceSStefano Zampini 4258779c1cceSStefano Zampini 42596e683305SStefano Zampini if (pcbddc->dbg_flag) { 42606e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42626e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4264779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42666e683305SStefano Zampini } 42676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42686e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42696e683305SStefano Zampini } 42706e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42716e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42726e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4273854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42746e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42756e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42766e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42776e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42786e683305SStefano Zampini } 42796e683305SStefano Zampini if (pcbddc->dbg_flag) { 42806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42826e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42836e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42846e683305SStefano Zampini } 4285779c1cceSStefano Zampini } 42866e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 428753a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 428853a05cb3SStefano 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); 428953a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 429053a05cb3SStefano 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); 429153a05cb3SStefano Zampini } 42926e683305SStefano Zampini } else { 42936e683305SStefano Zampini if (pcbddc->dbg_flag) { 42946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42966e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42976e683305SStefano Zampini } 42986e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42996e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 43006e683305SStefano Zampini } 43016e683305SStefano Zampini 43026e683305SStefano Zampini /* create local to global scatters for coarse problem */ 430368457ee5SStefano Zampini if (compute_vecs) { 43046e683305SStefano Zampini PetscInt lrows; 43056e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43066e683305SStefano Zampini if (coarse_mat_is) { 43076e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 43086e683305SStefano Zampini } else { 43096e683305SStefano Zampini lrows = 0; 43106e683305SStefano Zampini } 43116e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 43126e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 43136e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 43146e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43156e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43166e683305SStefano Zampini } 43176e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 43186e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4319c8587f34SStefano Zampini 4320f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4321f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4322f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4323f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4324f9eb5b7dSStefano Zampini } else { 4325f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4326f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4327c8587f34SStefano Zampini } 4328c8587f34SStefano Zampini 43296e683305SStefano Zampini /* print some info if requested */ 43306e683305SStefano Zampini if (pcbddc->dbg_flag) { 43316e683305SStefano Zampini if (!multilevel_allowed) { 43326e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43336e683305SStefano Zampini if (multilevel_requested) { 43346e683305SStefano 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); 43356e683305SStefano Zampini } else if (pcbddc->max_levels) { 43366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43376e683305SStefano Zampini } 43386e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43396e683305SStefano Zampini } 43406e683305SStefano Zampini } 43416e683305SStefano Zampini 4342f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43436e683305SStefano Zampini if (coarse_mat_is) { 43446e683305SStefano Zampini MatReuse coarse_mat_reuse; 43456a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43466e683305SStefano Zampini if (pcbddc->dbg_flag) { 43476e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43486e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43496e683305SStefano Zampini } 4350f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4351312be037SStefano Zampini char prefix[256],str_level[16]; 4352e604994aSStefano Zampini size_t len; 43536e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4354422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4355c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4356f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43575f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4358c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43596e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4360c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4361c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4362e604994aSStefano Zampini /* prefix */ 4363e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4364e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4365e604994aSStefano Zampini if (!pcbddc->current_level) { 4366e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4367e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4368c8587f34SStefano Zampini } else { 4369e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4370312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4371312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 437234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4373312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4374e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4375e604994aSStefano Zampini } 4376e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43773e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43783e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43793e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43803e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4381f9eb5b7dSStefano Zampini /* allow user customization */ 4382f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43833e3c6dadSStefano Zampini } 43843e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 438551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43863e3c6dadSStefano Zampini if (nisdofs) { 43873e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43883e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43893e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43903e3c6dadSStefano Zampini } 43913e3c6dadSStefano Zampini } 43923e3c6dadSStefano Zampini if (nisneu) { 43933e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43943e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4395312be037SStefano Zampini } 4396f9eb5b7dSStefano Zampini 4397f9eb5b7dSStefano Zampini /* get some info after set from options */ 4398f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4399f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 44004f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 44016e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4402f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4403f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4404f9eb5b7dSStefano Zampini } 440539f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 44064f3a063dSStefano Zampini if (isredundant) { 44074f3a063dSStefano Zampini KSP inner_ksp; 44084f3a063dSStefano Zampini PC inner_pc; 44094f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 44104f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 44114f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 44124f3a063dSStefano Zampini } 4413f9eb5b7dSStefano Zampini 4414f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4415fa7f1dd8SStefano Zampini if (coarse_reuse) { 441681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4417fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 44186e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4419fa7f1dd8SStefano Zampini } else { 44206e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4421fa7f1dd8SStefano Zampini } 4422c8587f34SStefano Zampini if (isbddc || isnn) { 442322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 442470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4425b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 442622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44296e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44306e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 443122b6e8a2SStefano Zampini } 443270cf5478SStefano Zampini } 443353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 443470cf5478SStefano Zampini } else { 443522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 443622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 443722bc73bbSStefano Zampini } 443822bc73bbSStefano Zampini } else { 44392e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4440c8587f34SStefano Zampini } 4441c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4442c8587f34SStefano Zampini 44433301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44445a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44453301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44463301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44473301b35fSStefano Zampini } 44483301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44493301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44503301b35fSStefano Zampini } 44513301b35fSStefano Zampini if (pc->pmat->spd_set) { 44523301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44533301b35fSStefano Zampini } 44546e683305SStefano Zampini /* set operators */ 44555f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44566e683305SStefano Zampini if (pcbddc->dbg_flag) { 44576e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44586e683305SStefano Zampini } 44596e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44606e683305SStefano Zampini coarse_mat = 0; 44616e683305SStefano Zampini } 44626e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4463b9b85e73SStefano Zampini #if 0 4464b9b85e73SStefano Zampini { 4465b9b85e73SStefano Zampini PetscViewer viewer; 4466b9b85e73SStefano Zampini char filename[256]; 4467b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4468b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4469b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4470b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4471b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4472b9b85e73SStefano Zampini } 4473b9b85e73SStefano Zampini #endif 4474c8587f34SStefano Zampini 4475c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4476298c0119SStefano Zampini #if 0 4477c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4478c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 447998a51de6SStefano Zampini } 4480298c0119SStefano Zampini #endif 448198a51de6SStefano Zampini 448298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 448398a51de6SStefano Zampini Vec crhs,csol; 448498a51de6SStefano Zampini PetscBool ispreonly; 448504708bb6SStefano Zampini 448698a51de6SStefano Zampini if (CoarseNullSpace) { 4487c8587f34SStefano Zampini if (isbddc) { 4488c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4489c8587f34SStefano Zampini } else { 44905fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4491c8587f34SStefano Zampini } 4492c8587f34SStefano Zampini } 4493f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4494f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4495f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4496f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44976e683305SStefano Zampini /* hack */ 4498f347579bSStefano Zampini if (!csol) { 44992a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4500f9eb5b7dSStefano Zampini } 4501f347579bSStefano Zampini if (!crhs) { 45022a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4503f347579bSStefano Zampini } 4504cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4505cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 45066e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4507c8587f34SStefano Zampini KSP check_ksp; 45082b510759SStefano Zampini KSPType check_ksp_type; 4509c8587f34SStefano Zampini PC check_pc; 45106e683305SStefano Zampini Vec check_vec,coarse_vec; 45116a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 45122b510759SStefano Zampini PetscInt its; 45136e683305SStefano Zampini PetscBool compute_eigs; 45146e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 45156e683305SStefano Zampini PetscInt neigs; 45168e185a42SStefano Zampini const char *prefix; 4517c8587f34SStefano Zampini 45182b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 45196e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4520422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 452123ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4522f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 45232b510759SStefano Zampini if (ispreonly) { 45242b510759SStefano Zampini check_ksp_type = KSPPREONLY; 45256e683305SStefano Zampini compute_eigs = PETSC_FALSE; 45262b510759SStefano Zampini } else { 4527cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45286e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4529c8587f34SStefano Zampini } 4530c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45316e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45326e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45336e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4534a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4535a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4536a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4537a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4538c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4539c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4540c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4541c8587f34SStefano Zampini /* create random vec */ 45426e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45436e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4544c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4545c8587f34SStefano Zampini if (CoarseNullSpace) { 4546c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4547c8587f34SStefano Zampini } 45486e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4549c8587f34SStefano Zampini /* solve coarse problem */ 45506e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4551c8587f34SStefano Zampini if (CoarseNullSpace) { 45526e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4553c8587f34SStefano Zampini } 4554cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45556e683305SStefano Zampini if (compute_eigs) { 4556854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4557854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45586e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45596e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45606e683305SStefano Zampini lambda_min = eigs_r[0]; 45616e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45626e683305SStefano Zampini if (lambda_max>lambda_min) { 4563cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4564cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4565cbcc2c2aSStefano Zampini } 4566c8587f34SStefano Zampini } 4567c8587f34SStefano Zampini } 4568cbcc2c2aSStefano Zampini 4569c8587f34SStefano Zampini /* check coarse problem residual error */ 45706e683305SStefano Zampini if (pcbddc->dbg_flag) { 45716e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45726e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45736e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4574c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45756e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45766e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4577c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4578779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45796e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45806e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45836e683305SStefano Zampini if (compute_eigs) { 45846e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4585deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4586c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45876e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45886e683305SStefano 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); 45896e683305SStefano Zampini for (i=0;i<neigs;i++) { 45906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4591c8587f34SStefano Zampini } 45926e683305SStefano Zampini } 45936e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45946e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45956e683305SStefano Zampini } 4596c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45976e683305SStefano Zampini if (compute_eigs) { 45986e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45996e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4600c8587f34SStefano Zampini } 46016e683305SStefano Zampini } 46026e683305SStefano Zampini } 4603cbcc2c2aSStefano Zampini /* print additional info */ 4604cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 46056e683305SStefano Zampini /* waits until all processes reaches this point */ 46066e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4607cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4608cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4609cbcc2c2aSStefano Zampini } 4610cbcc2c2aSStefano Zampini 46112b510759SStefano Zampini /* free memory */ 4612c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4613fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4614c8587f34SStefano Zampini PetscFunctionReturn(0); 4615c8587f34SStefano Zampini } 4616674ae819SStefano Zampini 4617f34684f1SStefano Zampini #undef __FUNCT__ 4618f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4619f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4620f34684f1SStefano Zampini { 4621f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4622f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4623f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4624dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4625dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 462673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4627dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4628f34684f1SStefano Zampini PetscErrorCode ierr; 4629f34684f1SStefano Zampini 4630f34684f1SStefano Zampini PetscFunctionBegin; 4631f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46320e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 46330e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4634727cdba6SStefano Zampini } 4635dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46363bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4637dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4638dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4639dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4640dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4641dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4642dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46430e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 46440e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 46450e6343abSStefano Zampini } 4646dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4647dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4648dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4649dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4650dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4651f34684f1SStefano Zampini 4652f34684f1SStefano Zampini /* check numbering */ 4653f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4654f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4655dc456d91SStefano Zampini PetscInt i; 4656b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4657f34684f1SStefano Zampini 4658f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4659f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4660f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46610fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4662f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4663f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4664727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4665f34684f1SStefano Zampini } 4666f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4667f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4668f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4669e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4670e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4671e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4672e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4673f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4674f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4675f34684f1SStefano Zampini if (array[i] == 1.0) { 4676b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4677f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4678f34684f1SStefano Zampini } 4679f34684f1SStefano Zampini } 4680b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4681f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4682f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4683f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4684f34684f1SStefano Zampini } 4685f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4686f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4687e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4688e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4689f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4690f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4691b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4692ca8b9ea9SStefano Zampini PetscInt *gidxs; 4693ca8b9ea9SStefano Zampini 4694ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46953bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4696f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4697f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4698f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4699f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 47004bc2dc4bSStefano 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); 4701f34684f1SStefano Zampini } 4702f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4703ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4704f34684f1SStefano Zampini } 4705f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4706302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4707f34684f1SStefano Zampini } 47088bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4709f34684f1SStefano Zampini /* get back data */ 4710f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4711f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4712674ae819SStefano Zampini PetscFunctionReturn(0); 4713674ae819SStefano Zampini } 4714674ae819SStefano Zampini 4715e456f2a8SStefano Zampini #undef __FUNCT__ 4716e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4717a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4718e456f2a8SStefano Zampini { 4719e456f2a8SStefano Zampini IS localis_t; 4720a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4721e456f2a8SStefano Zampini PetscScalar *vals; 4722e456f2a8SStefano Zampini PetscErrorCode ierr; 4723e456f2a8SStefano Zampini 4724e456f2a8SStefano Zampini PetscFunctionBegin; 4725a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4726e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4727854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4728e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4729e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4730a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4731a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47321035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4733a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47341035eff8SStefano Zampini } 4735a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4736e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4737e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4738a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4739a7dc3881SStefano Zampini /* now compute set in local ordering */ 4740a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4741a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4742a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4743a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4744a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4745ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4746e456f2a8SStefano Zampini lsize++; 4747e456f2a8SStefano Zampini } 4748e456f2a8SStefano Zampini } 4749854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4750a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4751ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4752e456f2a8SStefano Zampini idxs[lsize++] = i; 4753e456f2a8SStefano Zampini } 4754e456f2a8SStefano Zampini } 4755a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4756a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4757e456f2a8SStefano Zampini *localis = localis_t; 4758e456f2a8SStefano Zampini PetscFunctionReturn(0); 4759e456f2a8SStefano Zampini } 4760906d46d4SStefano Zampini 4761906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4762906d46d4SStefano Zampini #undef __FUNCT__ 4763906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4764906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4765906d46d4SStefano Zampini { 4766906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4767906d46d4SStefano Zampini PetscErrorCode ierr; 4768906d46d4SStefano Zampini 4769906d46d4SStefano Zampini PetscFunctionBegin; 4770906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4771906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4772906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4773906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4774906d46d4SStefano Zampini PetscFunctionReturn(0); 4775906d46d4SStefano Zampini } 4776906d46d4SStefano Zampini 4777906d46d4SStefano Zampini #undef __FUNCT__ 4778906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4779906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4780906d46d4SStefano Zampini { 4781906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4782906d46d4SStefano Zampini PetscErrorCode ierr; 4783906d46d4SStefano Zampini 4784906d46d4SStefano Zampini PetscFunctionBegin; 4785906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4786906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4787906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4788906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4789906d46d4SStefano Zampini PetscFunctionReturn(0); 4790906d46d4SStefano Zampini } 4791b96c3477SStefano Zampini 4792b96c3477SStefano Zampini #undef __FUNCT__ 4793b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 479408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4795b96c3477SStefano Zampini { 4796a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4797b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4798b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4799a64f4aa4SStefano Zampini Mat S_j; 4800b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4801b96c3477SStefano Zampini PetscBool free_used_adj; 4802b96c3477SStefano Zampini PetscErrorCode ierr; 4803b96c3477SStefano Zampini 4804b96c3477SStefano Zampini PetscFunctionBegin; 4805b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4806b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 480708122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4808b96c3477SStefano Zampini used_xadj = NULL; 4809b96c3477SStefano Zampini used_adjncy = NULL; 4810b96c3477SStefano Zampini } else { 481108122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 481208122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 481308122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 481408122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4815b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4816b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4817b96c3477SStefano Zampini } else { 48182fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4819b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4820b96c3477SStefano Zampini PetscInt nvtxs; 4821b96c3477SStefano Zampini 48222fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 48232fffb893SStefano Zampini if (flg_row) { 4824b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4825b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4826b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4827b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48282fffb893SStefano Zampini } else { 48292fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48302fffb893SStefano Zampini used_xadj = NULL; 48312fffb893SStefano Zampini used_adjncy = NULL; 48322fffb893SStefano Zampini } 48332fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4834b96c3477SStefano Zampini } 4835b96c3477SStefano Zampini } 4836d5574798SStefano Zampini 4837d5574798SStefano Zampini /* setup sub_schurs data */ 4838a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4839a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4840a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4841a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4842862806e4SStefano 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); 4843a64f4aa4SStefano Zampini } else { 48446816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 484504708bb6SStefano Zampini PetscBool isseqaij; 48465feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48475feab87aSStefano Zampini PetscInt n_vertices; 48485feab87aSStefano Zampini 48495feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48502034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48515feab87aSStefano Zampini } 485204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 485304708bb6SStefano Zampini if (!isseqaij) { 485404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 485504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 485604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 485704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 485804708bb6SStefano Zampini } else { 485904708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 486004708bb6SStefano Zampini } 486104708bb6SStefano Zampini } 4862862806e4SStefano 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); 4863a64f4aa4SStefano Zampini } 4864a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4865b96c3477SStefano Zampini 4866b96c3477SStefano Zampini /* free adjacency */ 4867b96c3477SStefano Zampini if (free_used_adj) { 4868b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4869b96c3477SStefano Zampini } 4870b96c3477SStefano Zampini PetscFunctionReturn(0); 4871b96c3477SStefano Zampini } 4872b96c3477SStefano Zampini 4873b96c3477SStefano Zampini #undef __FUNCT__ 4874b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 487508122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4876b96c3477SStefano Zampini { 4877b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4878b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4879b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4880b96c3477SStefano Zampini PCBDDCGraph graph; 4881b96c3477SStefano Zampini PetscErrorCode ierr; 4882b96c3477SStefano Zampini 4883b96c3477SStefano Zampini PetscFunctionBegin; 4884b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 488508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48863301b35fSStefano Zampini IS verticesIS,verticescomm; 48873301b35fSStefano Zampini PetscInt vsize,*idxs; 4888b96c3477SStefano Zampini 4889b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48903301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48913301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48923301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48933301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48943301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4895b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48967fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48973301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48983301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4899b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4900b96c3477SStefano Zampini /* 4901b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4902b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4903b96c3477SStefano Zampini } 4904b96c3477SStefano Zampini */ 4905b96c3477SStefano Zampini } else { 4906b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4907b96c3477SStefano Zampini } 4908b96c3477SStefano Zampini 4909b96c3477SStefano Zampini /* sub_schurs init */ 4910a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4911a64f4aa4SStefano Zampini 4912b96c3477SStefano Zampini /* free graph struct */ 491308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4914b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4915b96c3477SStefano Zampini } 4916b96c3477SStefano Zampini PetscFunctionReturn(0); 4917b96c3477SStefano Zampini } 4918