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; 286c4ed002SBarry Smith if (!sub_schurs->use_mumps) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 2908122e43SStefano Zampini 30fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 31fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 33fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 341575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 35fd14bc51SStefano Zampini } 36fd14bc51SStefano Zampini 37e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 38e496cd5dSStefano 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); 39e496cd5dSStefano Zampini } 40e496cd5dSStefano Zampini 416c4ed002SBarry Smith if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) 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); 42e496cd5dSStefano Zampini 4308122e43SStefano Zampini /* max size of subsets */ 4408122e43SStefano Zampini mss = 0; 4508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4608122e43SStefano Zampini PetscInt subset_size; 47862806e4SStefano Zampini 4808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5008122e43SStefano Zampini } 5108122e43SStefano Zampini 5208122e43SStefano Zampini /* min/max and threshold */ 5308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 54f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 5508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 56f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 57f6f667cfSStefano Zampini if (nmin) { 58f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 59f6f667cfSStefano Zampini } 6008122e43SStefano Zampini 6108122e43SStefano Zampini /* allocate lapack workspace */ 6208122e43SStefano Zampini cum = cum2 = 0; 6308122e43SStefano Zampini maxneigs = 0; 6408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6508122e43SStefano Zampini PetscInt n,subset_size; 66f6f667cfSStefano Zampini 6708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 699162d606SStefano Zampini cum += subset_size; 709162d606SStefano Zampini cum2 += subset_size*n; 7108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7208122e43SStefano Zampini } 7308122e43SStefano Zampini if (mss) { 749ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7508122e43SStefano Zampini PetscBLASInt B_itype = 1; 7608122e43SStefano Zampini PetscBLASInt B_N = mss; 774c6709b3SStefano Zampini PetscReal zero = 0.0; 784c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 7908122e43SStefano Zampini 8008122e43SStefano Zampini B_lwork = -1; 8108122e43SStefano Zampini S = NULL; 8208122e43SStefano Zampini St = NULL; 83a58a30b4SStefano Zampini eigs = NULL; 84a58a30b4SStefano Zampini eigv = NULL; 85a58a30b4SStefano Zampini B_iwork = NULL; 86a58a30b4SStefano Zampini B_ifail = NULL; 87d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 88d1710679SStefano Zampini rwork = NULL; 89d1710679SStefano Zampini #endif 908bec7fa6SStefano Zampini thresh = 1.0; 9108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9308122e43SStefano 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)); 9408122e43SStefano Zampini #else 9508122e43SStefano 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)); 9608122e43SStefano Zampini #endif 9708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 9808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 9908122e43SStefano Zampini } else { 10008122e43SStefano Zampini /* TODO */ 10108122e43SStefano Zampini } 10208122e43SStefano Zampini } else { 10308122e43SStefano Zampini lwork = 0; 10408122e43SStefano Zampini } 10508122e43SStefano Zampini 10608122e43SStefano Zampini nv = 0; 107d62866d3SStefano 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) */ 108d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 10908122e43SStefano Zampini } 1104c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 111f6f667cfSStefano Zampini if (allocated_S_St) { 112f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 113f6f667cfSStefano Zampini } 114f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 11708122e43SStefano Zampini #endif 1189162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1199162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1209162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1229162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12408122e43SStefano Zampini 12508122e43SStefano Zampini maxneigs = 0; 126c43ebad9SStefano Zampini cum = cumarray = 0; 1279162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1289162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 129d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13008122e43SStefano Zampini const PetscInt *idxs; 13108122e43SStefano Zampini 132d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 13608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1379162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1389162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 13908122e43SStefano Zampini } 140d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14108122e43SStefano Zampini } 14208122e43SStefano Zampini 14308122e43SStefano Zampini if (mss) { /* multilevel */ 14408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 14508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 14608122e43SStefano Zampini } 14708122e43SStefano Zampini 14808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 14908122e43SStefano Zampini 15008122e43SStefano Zampini const PetscInt *idxs; 151f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 152862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15308122e43SStefano Zampini PetscBLASInt B_N; 154aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 15508122e43SStefano Zampini 156862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 157f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 158f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1599ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 160aff50787SStefano Zampini PetscInt j,k; 161aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 162aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 163aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 16408122e43SStefano Zampini } 16508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 166aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 167aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 168aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 169aff50787SStefano Zampini } 17008122e43SStefano Zampini } 17108122e43SStefano Zampini } else { 17208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17408122e43SStefano Zampini } 1758bec7fa6SStefano Zampini } else { 176f6f667cfSStefano Zampini S = Sarray + cumarray; 177f6f667cfSStefano Zampini St = Starray + cumarray; 1788bec7fa6SStefano Zampini } 17908122e43SStefano Zampini 180f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 181aff50787SStefano Zampini /* see if we can save some work */ 182aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 183aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 184aff50787SStefano Zampini } 185aff50787SStefano Zampini 186aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 187aff50787SStefano Zampini B_neigs = 0; 188aff50787SStefano Zampini } else { 189aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 190f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 191f6f667cfSStefano Zampini 1929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 19308122e43SStefano Zampini PetscBLASInt B_itype = 1; 194f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1954c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1969552c7c7SStefano Zampini PetscInt nmin_s; 19708122e43SStefano Zampini 198f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 199fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2008bec7fa6SStefano 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]]); 201fd14bc51SStefano Zampini } 20208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 20308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 204f6f667cfSStefano 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)); 20508122e43SStefano Zampini #else 206f6f667cfSStefano 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)); 20708122e43SStefano Zampini #endif 20808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 20908122e43SStefano Zampini if (B_ierr) { 2106c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 2116c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 2126c4ed002SBarry Smith else 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); 21308122e43SStefano Zampini } 21408122e43SStefano Zampini 21508122e43SStefano Zampini if (B_neigs > nmax) { 216fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 217fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 218fd14bc51SStefano Zampini } 219f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 22008122e43SStefano Zampini B_neigs = nmax; 22108122e43SStefano Zampini } 22208122e43SStefano Zampini 2239552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2249552c7c7SStefano Zampini if (B_neigs < nmin_s) { 22508122e43SStefano Zampini PetscBLASInt B_neigs2; 22608122e43SStefano Zampini 227f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 228f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 229fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 230fd14bc51SStefano 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); 231fd14bc51SStefano Zampini } 2329ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23308122e43SStefano Zampini PetscInt j; 23408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23508122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23608122e43SStefano Zampini } 23708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23808122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23908122e43SStefano Zampini } 24008122e43SStefano Zampini } else { 24108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24308122e43SStefano Zampini } 24408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 246f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 24708122e43SStefano Zampini #else 248f6f667cfSStefano 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)); 24908122e43SStefano Zampini #endif 25008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 25108122e43SStefano Zampini B_neigs += B_neigs2; 25208122e43SStefano Zampini } 25308122e43SStefano Zampini if (B_ierr) { 2546c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 2556c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 2566c4ed002SBarry Smith else 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); 25708122e43SStefano Zampini } 258fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 26008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 26108122e43SStefano Zampini if (eigs[j] == 0.0) { 262ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 26308122e43SStefano Zampini } else { 264ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 265fd14bc51SStefano Zampini } 26608122e43SStefano Zampini } 26708122e43SStefano Zampini } 26808122e43SStefano Zampini } else { 26908122e43SStefano Zampini /* TODO */ 27008122e43SStefano Zampini } 271aff50787SStefano Zampini } 2728bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2738bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2749162d606SStefano Zampini if (B_neigs) { 2759162d606SStefano 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); 276fd14bc51SStefano Zampini 277fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2789552c7c7SStefano Zampini PetscInt ii; 2799552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 280ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 2819552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 282ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 283ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 284ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 285ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 286ac47001eSStefano Zampini #else 287ac47001eSStefano 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); 288ac47001eSStefano Zampini #endif 2899552c7c7SStefano Zampini } 2909552c7c7SStefano Zampini } 291fd14bc51SStefano Zampini } 29208122e43SStefano Zampini #if 0 2939162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29408122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29508122e43SStefano Zampini PetscScalar norm; 29608122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 2979162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 2989162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 2993d3eaba7SBarry Smith if (pcbddc->adaptive_constraints_data[cum] > 0.0) { 30008122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30108122e43SStefano Zampini } else { 30208122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 30308122e43SStefano Zampini } 3049162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 305b1b3d7a2SStefano Zampini } 306b1b3d7a2SStefano Zampini #endif 3079162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3089162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3099162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3109162d606SStefano Zampini cum++; 31108122e43SStefano Zampini } 31208122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 31308122e43SStefano Zampini /* shift for next computation */ 31408122e43SStefano Zampini cumarray += subset_size*subset_size; 31508122e43SStefano Zampini } 316fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 317fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 318fd14bc51SStefano Zampini } 31908122e43SStefano Zampini 32008122e43SStefano Zampini if (mss) { 32108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 32208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 323f6f667cfSStefano Zampini /* destroy matrices (junk) */ 324f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 325f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 32608122e43SStefano Zampini } 327f6f667cfSStefano Zampini if (allocated_S_St) { 328f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 329f6f667cfSStefano Zampini } 330f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 33108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 33208122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33308122e43SStefano Zampini #endif 33408122e43SStefano Zampini if (pcbddc->dbg_flag) { 3351b968477SStefano Zampini PetscInt maxneigs_r; 336b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3379b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33808122e43SStefano Zampini } 33908122e43SStefano Zampini PetscFunctionReturn(0); 34008122e43SStefano Zampini } 341b1b3d7a2SStefano Zampini 342674ae819SStefano Zampini #undef __FUNCT__ 343c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 344c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 345c8587f34SStefano Zampini { 3468629588bSStefano Zampini PetscScalar *coarse_submat_vals; 347c8587f34SStefano Zampini PetscErrorCode ierr; 348c8587f34SStefano Zampini 349c8587f34SStefano Zampini PetscFunctionBegin; 350f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3515e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 352c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 353c8587f34SStefano Zampini 354684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3550fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 356684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 357c8587f34SStefano Zampini 3588629588bSStefano Zampini /* 3598629588bSStefano Zampini Setup local correction and local part of coarse basis. 3608629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3618629588bSStefano Zampini */ 36247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3638629588bSStefano Zampini 3648629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3658629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3668629588bSStefano Zampini 3678629588bSStefano Zampini /* free */ 3688629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 369c8587f34SStefano Zampini PetscFunctionReturn(0); 370c8587f34SStefano Zampini } 371c8587f34SStefano Zampini 372c8587f34SStefano Zampini #undef __FUNCT__ 373674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 374674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 375674ae819SStefano Zampini { 376674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 377674ae819SStefano Zampini PetscErrorCode ierr; 378674ae819SStefano Zampini 379674ae819SStefano Zampini PetscFunctionBegin; 380674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 381674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 382674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 383785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 384674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 385f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 386f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 387785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 38863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 38963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 390674ae819SStefano Zampini PetscFunctionReturn(0); 391674ae819SStefano Zampini } 392674ae819SStefano Zampini 393674ae819SStefano Zampini #undef __FUNCT__ 394674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 395674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 396674ae819SStefano Zampini { 397674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 398674ae819SStefano Zampini PetscErrorCode ierr; 399674ae819SStefano Zampini 400674ae819SStefano Zampini PetscFunctionBegin; 401b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 402674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 403674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 404674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 405b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 406674ae819SStefano Zampini PetscFunctionReturn(0); 407674ae819SStefano Zampini } 408674ae819SStefano Zampini 409674ae819SStefano Zampini #undef __FUNCT__ 410674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 411674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 412674ae819SStefano Zampini { 413674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 41406656605SStefano Zampini PetscScalar *array; 415674ae819SStefano Zampini PetscErrorCode ierr; 416674ae819SStefano Zampini 417674ae819SStefano Zampini PetscFunctionBegin; 418674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42258da7f69SStefano Zampini } 423674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 424674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 42515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 42615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 427674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 428674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 429674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 431674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 432674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4338ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 434674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 437f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 438f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 439f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 440f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 441727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4420e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 443f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 44470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4456e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 446674ae819SStefano Zampini PetscFunctionReturn(0); 447674ae819SStefano Zampini } 448674ae819SStefano Zampini 449674ae819SStefano Zampini #undef __FUNCT__ 450f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 451f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4526bfb1811SStefano Zampini { 4536bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4546bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4556bfb1811SStefano Zampini VecType impVecType; 456e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4576bfb1811SStefano Zampini PetscErrorCode ierr; 4586bfb1811SStefano Zampini 4596bfb1811SStefano Zampini PetscFunctionBegin; 4606c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 461e7b262bdSStefano Zampini /* get sizes */ 462b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 463b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4646bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 465e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 466e7b262bdSStefano Zampini /* R nodes */ 467e7b262bdSStefano Zampini old_size = -1; 468e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 469e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 470e7b262bdSStefano Zampini } 471e7b262bdSStefano Zampini if (n_R != old_size) { 472e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 473e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4746bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4756bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4766bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4776bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 478e7b262bdSStefano Zampini } 479e7b262bdSStefano Zampini /* local primal dofs */ 480e7b262bdSStefano Zampini old_size = -1; 481e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 482e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 483e7b262bdSStefano Zampini } 484e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 485e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 48683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 487e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4886bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e7b262bdSStefano Zampini /* local explicit constraints */ 491e7b262bdSStefano Zampini old_size = -1; 492e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 493e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 494e7b262bdSStefano Zampini } 495e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 496e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 49783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 49883b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 49983b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50083b7ccabSStefano Zampini } 5016bfb1811SStefano Zampini PetscFunctionReturn(0); 5026bfb1811SStefano Zampini } 5036bfb1811SStefano Zampini 5046bfb1811SStefano Zampini #undef __FUNCT__ 50547f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 50647f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 50788ebb749SStefano Zampini { 50825084f0cSStefano Zampini PetscErrorCode ierr; 50925084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51088ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51188ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 512d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 51325084f0cSStefano Zampini /* submatrices of local problem */ 51480677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 51506656605SStefano Zampini /* submatrices of local coarse problem */ 51606656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 51725084f0cSStefano Zampini /* working matrices */ 51806656605SStefano Zampini Mat C_CR; 51925084f0cSStefano Zampini /* additional working stuff */ 52006656605SStefano Zampini PC pc_R; 52106656605SStefano Zampini Mat F; 52206656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 52306656605SStefano Zampini 52425084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 52506656605SStefano Zampini PetscScalar *work; 52606656605SStefano Zampini PetscInt *idx_V_B; 52706656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 52806656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 529b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53045a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53188ebb749SStefano Zampini MatType impMatType; 53225084f0cSStefano Zampini /* some shortcuts to scalars */ 53306656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 53488ebb749SStefano Zampini 53588ebb749SStefano Zampini PetscFunctionBegin; 536b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 53788ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 53888ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 539b371cd4fSStefano Zampini n_B = pcis->n_B; 540b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54188ebb749SStefano Zampini n_R = pcis->n - n_vertices; 54288ebb749SStefano Zampini 54388ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 54488ebb749SStefano Zampini impMatType = MATSEQDENSE; 54588ebb749SStefano Zampini 54688ebb749SStefano Zampini /* vertices in boundary numbering */ 547785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5480e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 5496c4ed002SBarry Smith if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",n_vertices,i); 55088ebb749SStefano Zampini 55106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 55206656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 55306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 55406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 55506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 55606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 55706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 55806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 55906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56106656605SStefano Zampini 56206656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 56306656605SStefano Zampini /* allocate workspace */ 56406656605SStefano Zampini n = 0; 56506656605SStefano Zampini if (n_constraints) { 56606656605SStefano Zampini n += n_R*n_constraints; 56706656605SStefano Zampini } 56806656605SStefano Zampini if (n_vertices) { 56906656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57080677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57106656605SStefano Zampini } 5723301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 57306656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 57406656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 57506656605SStefano Zampini } 57606656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 57706656605SStefano Zampini 57806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 57906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 58206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 58306656605SStefano Zampini if (isLU || isILU || isCHOL) { 58406656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 585d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 586d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 587d62866d3SStefano Zampini MatFactorType type; 588d62866d3SStefano Zampini 5896816873aSStefano Zampini F = reuse_mumps->F; 5906816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 591d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 59206656605SStefano Zampini } else { 59306656605SStefano Zampini F = NULL; 59406656605SStefano Zampini } 59506656605SStefano Zampini 59688ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 59788ebb749SStefano Zampini if (n_constraints) { 59806656605SStefano Zampini Mat M1,M2,M3; 59980677318SStefano Zampini Mat auxmat; 60006656605SStefano Zampini IS is_aux; 60180677318SStefano Zampini PetscScalar *array,*array2; 60206656605SStefano Zampini 603f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 60480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 60588ebb749SStefano Zampini 60625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 60725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6088ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 60980677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61088ebb749SStefano Zampini 61180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 61280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 61306656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 61488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 61506656605SStefano Zampini const PetscScalar *row_cmat_values; 61606656605SStefano Zampini const PetscInt *row_cmat_indices; 61706656605SStefano Zampini PetscInt size_of_constraint,j; 61888ebb749SStefano Zampini 61906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62106656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 62206656605SStefano Zampini } 62306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62406656605SStefano Zampini } 62580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 62606656605SStefano Zampini if (F) { 62706656605SStefano Zampini Mat B; 62806656605SStefano Zampini 62906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 63080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 63206656605SStefano Zampini } else { 63380677318SStefano Zampini PetscScalar *marr; 63480677318SStefano Zampini 63580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 63606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 63706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 63880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 64206656605SStefano Zampini } 64380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini } 64580677318SStefano Zampini if (!pcbddc->switch_static) { 64680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 64780677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 64880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 64980677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65080677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 65280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65680677318SStefano Zampini } 65780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65880677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66080677318SStefano Zampini } else { 66180677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 66280677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 66325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66480677318SStefano Zampini } 66580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 66680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 66780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 66806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 66906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 67080677318SStefano Zampini if (isCHOL) { 67180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 67280677318SStefano Zampini } else { 67325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 67480677318SStefano Zampini } 67580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 67725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 67825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 67925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 68080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68180677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 68280677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 68306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 68406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 685f4ddd8eeSStefano Zampini } 68688ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 68788ebb749SStefano Zampini if (n_vertices) { 68806656605SStefano Zampini IS is_aux; 6893a50541eSStefano Zampini 6906816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 6916816873aSStefano Zampini IS tis; 6926816873aSStefano Zampini 6936816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 6946816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 6956816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 6966816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 6976816873aSStefano Zampini } else { 6983a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 6996816873aSStefano Zampini } 7009577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7019577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 70204708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 70325084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 70488ebb749SStefano Zampini } 70588ebb749SStefano Zampini 70688ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 707f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 70806656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 70906656605SStefano Zampini if (pcbddc->coarse_phi_D) { 71006656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 71106656605SStefano Zampini } 712f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 71306656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 71406656605SStefano Zampini PetscScalar *marray; 71506656605SStefano Zampini 71606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 71706656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 718f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 719f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 720f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 721f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 722f4ddd8eeSStefano Zampini } 723f4ddd8eeSStefano Zampini } 72406656605SStefano Zampini 725f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 72606656605SStefano Zampini PetscScalar *marray; 72788ebb749SStefano Zampini 72806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7298eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 73188ebb749SStefano Zampini } 7323301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 73306656605SStefano Zampini n *= 2; 73488ebb749SStefano Zampini } 73506656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 73606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 73706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7388eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 74006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74188ebb749SStefano Zampini } 7423301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 74306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 74606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 74788ebb749SStefano Zampini } 74888ebb749SStefano Zampini } else { 749c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 750c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7511b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 752c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 753c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 754c0553b1fSStefano Zampini } 75588ebb749SStefano Zampini } 75606656605SStefano Zampini } 75706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 75806656605SStefano Zampini /* vertices */ 75906656605SStefano Zampini if (n_vertices) { 76016f15bc4SStefano Zampini 761511c6705SHong Zhang ierr = MatConvert(A_VV,impMatType,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 76204708bb6SStefano Zampini 76316f15bc4SStefano Zampini if (n_R) { 76406656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 76506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 76616f15bc4SStefano Zampini PetscScalar *x,*y; 76704708bb6SStefano Zampini PetscBool isseqaij; 76806656605SStefano Zampini 76921eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 770511c6705SHong Zhang ierr = MatConvert(A_RV,impMatType,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 77104708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7726816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 77306656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 77406656605SStefano Zampini } else { 77506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 77606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 77706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 77806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 77906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 78006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 78106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 78206656605SStefano Zampini } 78306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 78406656605SStefano Zampini } 78580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 78606656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 78706656605SStefano Zampini if (n_constraints) { 78806656605SStefano Zampini Mat B; 78980677318SStefano Zampini 790b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 79180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 79280677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 79380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 79480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 79580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 79680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 79780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 79880677318SStefano Zampini } 79980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 80080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 80180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 80206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 80380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 80606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 80706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 80806656605SStefano Zampini } 80904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 81004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 811511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 81204708bb6SStefano Zampini } 81306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 81480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 81506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 81606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 81706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 81806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 81906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 82106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 82206656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 82316f15bc4SStefano Zampini } else { 82416f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 82516f15bc4SStefano Zampini } 82621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 82706656605SStefano Zampini /* coarse basis functions */ 82806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 82916f15bc4SStefano Zampini PetscScalar *y; 83016f15bc4SStefano Zampini 83106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 83306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 83406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 83706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 83806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 83906656605SStefano Zampini 84006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 84106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 84706656605SStefano Zampini } 84806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 84906656605SStefano Zampini } 85004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 85104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 85206656605SStefano Zampini } 85306656605SStefano Zampini 85406656605SStefano Zampini if (n_constraints) { 85506656605SStefano Zampini Mat B; 85606656605SStefano Zampini 85706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 85980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86106656605SStefano Zampini if (n_vertices) { 86280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 86380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 86480677318SStefano Zampini } else { 86580677318SStefano Zampini Mat S_VCt; 86680677318SStefano Zampini 86780677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 86880677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 86980677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 87080677318SStefano Zampini } 87106656605SStefano Zampini } 87206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 87306656605SStefano Zampini /* coarse basis functions */ 87406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 87506656605SStefano Zampini PetscScalar *y; 87606656605SStefano Zampini 87706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 87806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 87906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 88006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 88406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 88506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 88606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 88706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 89006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89106656605SStefano Zampini } 89206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 89306656605SStefano Zampini } 89406656605SStefano Zampini } 89580677318SStefano Zampini if (n_constraints) { 89680677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 89780677318SStefano Zampini } 89806656605SStefano Zampini 89906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9003301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 90106656605SStefano Zampini 90206656605SStefano Zampini if (n_constraints) { 90316f15bc4SStefano Zampini Mat S_CCT,B_C; 90406656605SStefano Zampini 90580677318SStefano Zampini /* this is a lazy thing */ 906511c6705SHong Zhang ierr = MatConvert(C_CR,impMatType,MAT_INPLACE_MATRIX,&C_CR);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 91016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 91106656605SStefano Zampini if (n_vertices) { 91216f15bc4SStefano Zampini Mat B_V,S_VCT; 91306656605SStefano Zampini 91406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 91816f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 91906656605SStefano Zampini } 92006656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 92180677318SStefano Zampini } else { /* if there are no constraints, reset work */ 92280677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 92306656605SStefano Zampini } 92416f15bc4SStefano Zampini if (n_vertices && n_R) { 92506656605SStefano Zampini Mat A_VRT; 92680677318SStefano Zampini PetscScalar *marray; 92706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 92806656605SStefano Zampini 92980677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 930511c6705SHong Zhang ierr = MatConvert(A_VRT,impMatType,MAT_INPLACE_MATRIX,&A_VRT);CHKERRQ(ierr); 93180677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 93206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 93380677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 93480677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 93516f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 93606656605SStefano Zampini } 93706656605SStefano Zampini 93806656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 93906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 94006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 94506656605SStefano Zampini } 94606656605SStefano Zampini } else { 94706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 94806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95306656605SStefano Zampini } 95406656605SStefano Zampini } 95506656605SStefano Zampini /* coarse basis functions */ 95606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95706656605SStefano Zampini PetscScalar *y; 95806656605SStefano Zampini 95906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 96006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96406656605SStefano Zampini if (i<n_vertices) { 96506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 96606656605SStefano Zampini } 96706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 96906656605SStefano Zampini 97006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 97106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 97306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 97706656605SStefano Zampini } 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97906656605SStefano Zampini } 98006656605SStefano Zampini } 981d62866d3SStefano Zampini /* free memory */ 98288ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 98506656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 987d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 988d62866d3SStefano Zampini if (n_vertices) { 989d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 990d62866d3SStefano Zampini } 991d62866d3SStefano Zampini if (n_constraints) { 992d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 993d62866d3SStefano Zampini } 99488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 99588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99725084f0cSStefano Zampini if (pcbddc->dbg_flag) { 99888ebb749SStefano Zampini Mat coarse_sub_mat; 99925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 100088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 100188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 100288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10038bec7fa6SStefano Zampini Mat C_B,CPHI; 10048bec7fa6SStefano Zampini IS is_dummy; 10058bec7fa6SStefano Zampini Vec mones; 100688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 100788ebb749SStefano Zampini PetscReal real_value; 100888ebb749SStefano Zampini 100988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 101088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 101188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 101288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 101388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 101488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1015c0553b1fSStefano Zampini if (unsymmetric_check) { 101688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 101788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 101888ebb749SStefano Zampini } 101988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 102088ebb749SStefano Zampini 102125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10223301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 102325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1024c0553b1fSStefano Zampini if (unsymmetric_check) { 102588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 102688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 102788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 103088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103788ebb749SStefano Zampini } else { 103888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104688ebb749SStefano Zampini } 104788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 1050511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 105181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10528bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10531575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 105406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10558bec7fa6SStefano Zampini 10568bec7fa6SStefano Zampini /* check constraints */ 10578bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10583d3eaba7SBarry Smith ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 10598bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10608bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10618bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10628bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10638bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1064bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1065c0553b1fSStefano Zampini if (unsymmetric_check) { 1066bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1067bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1068bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1069bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1070bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 107188ebb749SStefano Zampini } 10728bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10738bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10748bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10758bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10768bec7fa6SStefano Zampini 107725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 107888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 107988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 108088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 108188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 108288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 108388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 108488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 108588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 108688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 108788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1088c0553b1fSStefano Zampini if (unsymmetric_check) { 108988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 109188ebb749SStefano Zampini } 109288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 109388ebb749SStefano Zampini } 10948629588bSStefano Zampini /* get back data */ 10958629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 109688ebb749SStefano Zampini PetscFunctionReturn(0); 109788ebb749SStefano Zampini } 109888ebb749SStefano Zampini 109988ebb749SStefano Zampini #undef __FUNCT__ 1100d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1101d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1102aa0d41d4SStefano Zampini { 1103d65f70fdSStefano Zampini Mat *work_mat; 1104d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1105d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1106c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 1107aa0d41d4SStefano Zampini PetscErrorCode ierr; 1108aa0d41d4SStefano Zampini 1109aa0d41d4SStefano Zampini PetscFunctionBegin; 1110d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1111d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1112d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1113d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1114aa0d41d4SStefano Zampini 1115d65f70fdSStefano Zampini if (!rsorted) { 1116906d46d4SStefano Zampini const PetscInt *idxs; 1117906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1118aa0d41d4SStefano Zampini 1119d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1120d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1121d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1122d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1123aa0d41d4SStefano Zampini } 1124d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1125d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1126d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1127d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1128aa0d41d4SStefano Zampini } 1129d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1130d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini } else { 1132d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini isrow_s = isrow; 1134aa0d41d4SStefano Zampini } 1135906d46d4SStefano Zampini 1136d65f70fdSStefano Zampini if (!csorted) { 1137d65f70fdSStefano Zampini if (isrow == iscol) { 1138d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini iscol_s = isrow_s; 1140d65f70fdSStefano Zampini } else { 1141d65f70fdSStefano Zampini const PetscInt *idxs; 1142d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1143906d46d4SStefano Zampini 1144d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1147d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1148d65f70fdSStefano Zampini } 1149d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1150d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1152d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1153d65f70fdSStefano Zampini } 1154d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1155d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1156d65f70fdSStefano Zampini } 1157d65f70fdSStefano Zampini } else { 1158d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini iscol_s = iscol; 1160d65f70fdSStefano Zampini } 1161d65f70fdSStefano Zampini 1162d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1163d65f70fdSStefano Zampini 1164d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1165906d46d4SStefano Zampini Mat new_mat; 1166d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1167906d46d4SStefano Zampini 1168d65f70fdSStefano Zampini if (!rsorted) { 1169d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1170d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1172d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1173906d46d4SStefano Zampini } 1174d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1175d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini } else { 1177d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1178906d46d4SStefano Zampini } 1179d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini 1181d65f70fdSStefano Zampini if (!csorted) { 1182d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1183d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1185d65f70fdSStefano Zampini } else { 1186d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1187d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1188d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1189d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1190d65f70fdSStefano Zampini } 1191d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1192d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini } else { 1195d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini } 1197d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini 1199d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1201d65f70fdSStefano Zampini work_mat[0] = new_mat; 1202d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini } 1205d65f70fdSStefano Zampini 1206d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini *B = work_mat[0]; 1208d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1210d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini PetscFunctionReturn(0); 1212d65f70fdSStefano Zampini } 1213d65f70fdSStefano Zampini 1214d65f70fdSStefano Zampini #undef __FUNCT__ 12155e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12165e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1217aa0d41d4SStefano Zampini { 1218aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12195e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1220d65f70fdSStefano Zampini Mat new_mat; 12215e8657edSStefano Zampini IS is_local,is_global; 1222d65f70fdSStefano Zampini PetscInt local_size; 1223d65f70fdSStefano Zampini PetscBool isseqaij; 1224aa0d41d4SStefano Zampini PetscErrorCode ierr; 1225aa0d41d4SStefano Zampini 1226aa0d41d4SStefano Zampini PetscFunctionBegin; 1227aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12285e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12295e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1230df1311c0SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1231aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1232d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1233aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1234906d46d4SStefano Zampini 1235906d46d4SStefano Zampini /* check */ 1236906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1237906d46d4SStefano Zampini Vec x,x_change; 1238906d46d4SStefano Zampini PetscReal error; 1239906d46d4SStefano Zampini 12405e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1241906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12425e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1243e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1244e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1245d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1246e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1247e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1248906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1249906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1250906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1252906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1253906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1254906d46d4SStefano Zampini } 1255906d46d4SStefano Zampini 125622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12579b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 125822d5777bSStefano Zampini if (isseqaij) { 1259d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1260aa0d41d4SStefano Zampini } else { 1261aa0d41d4SStefano Zampini Mat work_mat; 1262aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1263d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1264aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1265aa0d41d4SStefano Zampini } 12663301b35fSStefano Zampini if (matis->A->symmetric_set) { 12673301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1268e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12693301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1270e496cd5dSStefano Zampini #endif 12713301b35fSStefano Zampini } 1272d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1273aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1274aa0d41d4SStefano Zampini } 1275aa0d41d4SStefano Zampini 1276aa0d41d4SStefano Zampini #undef __FUNCT__ 1277a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12788ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1279a64d13efSStefano Zampini { 1280a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1281a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1282d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 128353892102SStefano Zampini PetscInt *idx_R_local=NULL; 12843a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12853a50541eSStefano Zampini PetscInt vbs,bs; 12866816873aSStefano Zampini PetscBT bitmask=NULL; 1287a64d13efSStefano Zampini PetscErrorCode ierr; 1288a64d13efSStefano Zampini 1289a64d13efSStefano Zampini PetscFunctionBegin; 1290b23d619eSStefano Zampini /* 1291b23d619eSStefano Zampini No need to setup local scatters if 1292b23d619eSStefano Zampini - primal space is unchanged 1293b23d619eSStefano Zampini AND 1294b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1295b23d619eSStefano Zampini AND 1296b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1297b23d619eSStefano Zampini */ 1298b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1299f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1300f4ddd8eeSStefano Zampini } 1301f4ddd8eeSStefano Zampini /* destroy old objects */ 1302f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1303f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1304f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1305a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1306b371cd4fSStefano Zampini n_B = pcis->n_B; 1307b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1308b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13093a50541eSStefano Zampini 1310a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13116816873aSStefano Zampini 131253892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13136816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1314854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1315a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1316a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13170e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1318a64d13efSStefano Zampini } 1319a64d13efSStefano Zampini 1320a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13214641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13226816873aSStefano Zampini idx_R_local[n_R++] = i; 1323a64d13efSStefano Zampini } 1324a64d13efSStefano Zampini } 132553892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13266816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13276816873aSStefano Zampini 132853892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 132953892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13306816873aSStefano Zampini } 13313a50541eSStefano Zampini 13323a50541eSStefano Zampini /* Block code */ 13333a50541eSStefano Zampini vbs = 1; 13343a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13353a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13363a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13373a50541eSStefano Zampini PetscInt *vary; 1338d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1339785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13403a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1341d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1342d3df7717SStefano 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 */ 13430e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1344d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13453a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13463a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13473a50541eSStefano Zampini break; 13483a50541eSStefano Zampini } 13493a50541eSStefano Zampini } 1350d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1351d3df7717SStefano Zampini } else { 1352d3df7717SStefano Zampini /* Verify directly the R set */ 1353d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1354d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1355d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1356d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1357d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1358d3df7717SStefano Zampini break; 1359d3df7717SStefano Zampini } 1360d3df7717SStefano Zampini } 1361d3df7717SStefano Zampini } 1362d3df7717SStefano Zampini } 13633a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13643a50541eSStefano Zampini vbs = bs; 13653a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13663a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13673a50541eSStefano Zampini } 13683a50541eSStefano Zampini } 13693a50541eSStefano Zampini } 13703a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 137153892102SStefano Zampini if (sub_schurs->reuse_mumps) { 137253892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 137353892102SStefano Zampini 137453892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 137553892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 137653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 137753892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 137853892102SStefano Zampini } else { 13793a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 138053892102SStefano Zampini } 1381a64d13efSStefano Zampini 1382a64d13efSStefano Zampini /* print some info if requested */ 1383a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1384a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1385a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13861575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1387a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1388a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13893a50541eSStefano 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); 1390a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1391a64d13efSStefano Zampini } 1392a64d13efSStefano Zampini 1393a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13946816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 13956816873aSStefano Zampini IS is_aux1,is_aux2; 13966816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 13976816873aSStefano Zampini 13983a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1399854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1400854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1401a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14024641a718SStefano Zampini for (i=0; i<n_D; i++) { 14034641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14044641a718SStefano Zampini } 1405a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1406a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14074641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14084641a718SStefano Zampini aux_array1[j++] = i; 1409a64d13efSStefano Zampini } 1410a64d13efSStefano Zampini } 1411a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1412a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1413a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14144641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14154641a718SStefano Zampini aux_array2[j++] = i; 1416a64d13efSStefano Zampini } 1417a64d13efSStefano Zampini } 1418a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1419a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1420a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1421a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1422a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1423a64d13efSStefano Zampini 14248eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1425785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1426a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14274641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14284641a718SStefano Zampini aux_array1[j++] = i; 1429a64d13efSStefano Zampini } 1430a64d13efSStefano Zampini } 1431a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1432a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1433a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1434a64d13efSStefano Zampini } 14354641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14363a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1437d62866d3SStefano Zampini } else { 143853892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14396816873aSStefano Zampini IS tis; 14406816873aSStefano Zampini PetscInt schur_size; 14416816873aSStefano Zampini 144253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14436816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 144453892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14456816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14466816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14476816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14486816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14496816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1450d62866d3SStefano Zampini } 1451d62866d3SStefano Zampini } 1452a64d13efSStefano Zampini PetscFunctionReturn(0); 1453a64d13efSStefano Zampini } 1454a64d13efSStefano Zampini 1455304d26faSStefano Zampini 1456304d26faSStefano Zampini #undef __FUNCT__ 1457304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1458684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1459304d26faSStefano Zampini { 1460304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1461304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1462304d26faSStefano Zampini PC pc_temp; 1463304d26faSStefano Zampini Mat A_RR; 1464f4ddd8eeSStefano Zampini MatReuse reuse; 1465304d26faSStefano Zampini PetscScalar m_one = -1.0; 1466304d26faSStefano Zampini PetscReal value; 146704708bb6SStefano Zampini PetscInt n_D,n_R; 1468*c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 1469304d26faSStefano Zampini PetscErrorCode ierr; 1470e604994aSStefano Zampini /* prefixes stuff */ 1471312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1472e604994aSStefano Zampini size_t len; 1473304d26faSStefano Zampini 1474304d26faSStefano Zampini PetscFunctionBegin; 1475304d26faSStefano Zampini 1476e604994aSStefano Zampini /* compute prefixes */ 1477e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1478e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1479e604994aSStefano Zampini if (!pcbddc->current_level) { 1480e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1481e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1482e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1483e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1484e604994aSStefano Zampini } else { 1485e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1486312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1487e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1488e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1489312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1490312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 149134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 149234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1493e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1494e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1495e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1496e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1497e604994aSStefano Zampini } 1498e604994aSStefano Zampini 1499304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1500684f6988SStefano Zampini if (dirichlet) { 1501d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15023301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15033301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1504964fefecSStefano Zampini } 1505ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1506964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1507304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1508304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1509304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1510304d26faSStefano Zampini /* default */ 1511304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1512e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15139577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1514304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15159577ea80SStefano Zampini if (issbaij) { 15169577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15179577ea80SStefano Zampini } else { 1518304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15199577ea80SStefano Zampini } 1520304d26faSStefano Zampini /* Allow user's customization */ 1521304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1522304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1523304d26faSStefano Zampini } 1524d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1525d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1526d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1527d62866d3SStefano Zampini 1528d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1529d5574798SStefano Zampini } 1530304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1531304d26faSStefano Zampini if (!n_D) { 1532304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1533304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1534304d26faSStefano Zampini } 1535304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1536304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1537304d26faSStefano Zampini /* set ksp_D into pcis data */ 1538304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1539304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1540304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1541684f6988SStefano Zampini } 1542304d26faSStefano Zampini 1543304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1544684f6988SStefano Zampini A_RR = 0; 1545684f6988SStefano Zampini if (neumann) { 1546d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 154704708bb6SStefano Zampini PetscInt ibs,mbs; 154804708bb6SStefano Zampini PetscBool issbaij; 154904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1550f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15518ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1552f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1553f4ddd8eeSStefano Zampini PetscInt nn_R; 155481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1555f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1556f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1557f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1558f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1559f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1560f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1561f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1562727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1563f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1564f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1565f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1566f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1567f4ddd8eeSStefano Zampini } 1568f4ddd8eeSStefano Zampini } 1569f4ddd8eeSStefano Zampini /* last check */ 1570d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1571f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1572f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1573f4ddd8eeSStefano Zampini } 1574f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1575f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1576f4ddd8eeSStefano Zampini } 1577f4ddd8eeSStefano Zampini /* extract A_RR */ 1578af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1579af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 158004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 158104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 158204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 158304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 158404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1585af732b37SStefano Zampini } else { 1586511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 15876816873aSStefano Zampini } 158804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 158904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 159004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 159104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 159204708bb6SStefano Zampini } else { 1593511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 159404708bb6SStefano Zampini } 159504708bb6SStefano Zampini } 159604708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1597f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 15983301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15993301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16006816873aSStefano Zampini } 16016816873aSStefano Zampini } else { 16026816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16036816873aSStefano Zampini 16046816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16056816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16066816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1607af732b37SStefano Zampini } 1608f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1609304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1610304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1611304d26faSStefano Zampini /* default */ 1612304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1613e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1614304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16159577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16169577ea80SStefano Zampini if (issbaij) { 16179577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16189577ea80SStefano Zampini } else { 1619304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16209577ea80SStefano Zampini } 1621304d26faSStefano Zampini /* Allow user's customization */ 1622304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1623304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1624304d26faSStefano Zampini } 1625d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1626304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1627304d26faSStefano Zampini if (!n_R) { 1628304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1629304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1630304d26faSStefano Zampini } 1631d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1632d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1633d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1634d62866d3SStefano Zampini 1635d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1636d62866d3SStefano Zampini } 1637304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1638304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1639684f6988SStefano Zampini } 16406816873aSStefano Zampini /* free Neumann problem's matrix */ 16416816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1642304d26faSStefano Zampini 1643684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1644684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16451575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1646684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1647684f6988SStefano Zampini } 1648*c7017625SStefano Zampini 1649*c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 1650*c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 1651*c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 1652*c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 1653*c7017625SStefano Zampini } 1654*c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 1655*c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 1656*c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 1657*c7017625SStefano Zampini } 1658*c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 1659*c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 1660*c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 1661*c7017625SStefano Zampini } 1662*c7017625SStefano Zampini 1663*c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 1664*c7017625SStefano Zampini if (pcbddc->dbg_flag) { 1665684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16660fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16670fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16680fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16690fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16700fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1671e604994aSStefano 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); 1672*c7017625SStefano Zampini if (check_corr[0]) { 1673*c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 1674*c7017625SStefano Zampini } 1675304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1676304d26faSStefano Zampini } 1677684f6988SStefano Zampini if (neumann) { /* Neumann */ 16786816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16790fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16800fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16810fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16820fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1684e604994aSStefano 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); 1685*c7017625SStefano Zampini if (check_corr[1]) { 1686*c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 1687*c7017625SStefano Zampini } 1688304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1689304d26faSStefano Zampini } 1690684f6988SStefano Zampini } 1691304d26faSStefano Zampini PetscFunctionReturn(0); 1692304d26faSStefano Zampini } 1693304d26faSStefano Zampini 1694304d26faSStefano Zampini #undef __FUNCT__ 1695ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 169680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1697674ae819SStefano Zampini { 1698674ae819SStefano Zampini PetscErrorCode ierr; 1699674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1700be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1701674ae819SStefano Zampini 1702674ae819SStefano Zampini PetscFunctionBegin; 1703be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 170480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 170520c7b377SStefano Zampini } 170680677318SStefano Zampini if (!pcbddc->switch_static) { 170780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 170880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 170980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 171020c7b377SStefano Zampini } 1711be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171420c7b377SStefano Zampini } else { 1715be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1716be83ff47SStefano Zampini 171753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 171920c7b377SStefano Zampini } 1720be83ff47SStefano Zampini } else { 172180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 172680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 172780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1730674ae819SStefano Zampini } 1731674ae819SStefano Zampini } 1732be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173380677318SStefano Zampini if (applytranspose) { 173480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 173580677318SStefano Zampini } else { 173680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 173780677318SStefano Zampini } 1738be83ff47SStefano Zampini } else { 1739be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1740be83ff47SStefano Zampini 1741be83ff47SStefano Zampini if (applytranspose) { 17425a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1743be83ff47SStefano Zampini } else { 17445a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1745be83ff47SStefano Zampini } 1746be83ff47SStefano Zampini } 174780677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 174880677318SStefano Zampini if (!pcbddc->switch_static) { 1749be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 175180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1752be83ff47SStefano Zampini } else { 1753be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1754be83ff47SStefano Zampini 175553892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175653892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1757be83ff47SStefano Zampini } 175880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 175980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 176080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 176180677318SStefano Zampini } 176280677318SStefano Zampini } else { 176380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 176880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 176980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 177080677318SStefano Zampini } 177180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1775674ae819SStefano Zampini } 1776674ae819SStefano Zampini PetscFunctionReturn(0); 1777674ae819SStefano Zampini } 1778674ae819SStefano Zampini 1779dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1780674ae819SStefano Zampini #undef __FUNCT__ 1781674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1782dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1783674ae819SStefano Zampini { 1784674ae819SStefano Zampini PetscErrorCode ierr; 1785674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1786674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1787674ae819SStefano Zampini const PetscScalar zero = 0.0; 1788674ae819SStefano Zampini 1789674ae819SStefano Zampini PetscFunctionBegin; 1790dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1791dc359a40SStefano Zampini if (applytranspose) { 1792674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 17938eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1794dc359a40SStefano Zampini } else { 1795674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1796674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 179715aaf578SStefano Zampini } 179812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 179912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 180012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180212edc857SStefano Zampini 18039f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 180412edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 180512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1806964fefecSStefano Zampini Vec rhs,sol; 1807964fefecSStefano Zampini 1808964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1809964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 181012edc857SStefano Zampini if (applytranspose) { 1811964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 181212edc857SStefano Zampini } else { 1813964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 181412edc857SStefano Zampini } 181512edc857SStefano Zampini } 1816674ae819SStefano Zampini 1817674ae819SStefano Zampini /* Local solution on R nodes */ 181880677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 181980677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18209f00e9b4SStefano Zampini } 1821674ae819SStefano Zampini 18229f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18239f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 182412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1825674ae819SStefano Zampini 1826674ae819SStefano Zampini /* Sum contributions from two levels */ 1827dc359a40SStefano Zampini if (applytranspose) { 1828dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1829dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1830dc359a40SStefano Zampini } else { 1831674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18328eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1833dc359a40SStefano Zampini } 1834674ae819SStefano Zampini PetscFunctionReturn(0); 1835674ae819SStefano Zampini } 1836674ae819SStefano Zampini 1837674ae819SStefano Zampini #undef __FUNCT__ 1838674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 183912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1840674ae819SStefano Zampini { 1841674ae819SStefano Zampini PetscErrorCode ierr; 1842674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 184358da7f69SStefano Zampini PetscScalar *array; 184412edc857SStefano Zampini Vec from,to; 1845674ae819SStefano Zampini 1846674ae819SStefano Zampini PetscFunctionBegin; 184712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 184812edc857SStefano Zampini from = pcbddc->coarse_vec; 184912edc857SStefano Zampini to = pcbddc->vec1_P; 185012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 185112edc857SStefano Zampini Vec tvec; 185258da7f69SStefano Zampini 185358da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 185458da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 185512edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 185658da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 185758da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 185858da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 185912edc857SStefano Zampini } 186012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 186112edc857SStefano Zampini from = pcbddc->vec1_P; 186212edc857SStefano Zampini to = pcbddc->coarse_vec; 186312edc857SStefano Zampini } 186412edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1865674ae819SStefano Zampini PetscFunctionReturn(0); 1866674ae819SStefano Zampini } 1867674ae819SStefano Zampini 1868674ae819SStefano Zampini #undef __FUNCT__ 1869674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 187012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1871674ae819SStefano Zampini { 1872674ae819SStefano Zampini PetscErrorCode ierr; 1873674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187458da7f69SStefano Zampini PetscScalar *array; 187512edc857SStefano Zampini Vec from,to; 1876674ae819SStefano Zampini 1877674ae819SStefano Zampini PetscFunctionBegin; 187812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187912edc857SStefano Zampini from = pcbddc->coarse_vec; 188012edc857SStefano Zampini to = pcbddc->vec1_P; 188112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188212edc857SStefano Zampini from = pcbddc->vec1_P; 188312edc857SStefano Zampini to = pcbddc->coarse_vec; 188412edc857SStefano Zampini } 188512edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 188612edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 188712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 188812edc857SStefano Zampini Vec tvec; 188958da7f69SStefano Zampini 189012edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 189158da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 189258da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 189358da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 189458da7f69SStefano Zampini } 189558da7f69SStefano Zampini } else { 189658da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 189758da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 189812edc857SStefano Zampini } 189912edc857SStefano Zampini } 1900674ae819SStefano Zampini PetscFunctionReturn(0); 1901674ae819SStefano Zampini } 1902674ae819SStefano Zampini 1903984c4197SStefano Zampini /* uncomment for testing purposes */ 1904984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1905674ae819SStefano Zampini #undef __FUNCT__ 1906674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1907674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1908674ae819SStefano Zampini { 1909674ae819SStefano Zampini PetscErrorCode ierr; 1910674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1911674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1912674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1913984c4197SStefano Zampini /* one and zero */ 1914984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1915984c4197SStefano Zampini /* space to store constraints and their local indices */ 19169162d606SStefano Zampini PetscScalar *constraints_data; 19179162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19189162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19199162d606SStefano Zampini PetscInt *constraints_n; 1920984c4197SStefano Zampini /* iterators */ 1921b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1922984c4197SStefano Zampini /* BLAS integers */ 1923e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1924e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1925c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1926727cdba6SStefano Zampini /* reuse */ 19270e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19280e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1929984c4197SStefano Zampini /* change of basis */ 1930b3d85658SStefano Zampini PetscBool qr_needed; 19319162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1932984c4197SStefano Zampini /* auxiliary stuff */ 193364efe560SStefano Zampini PetscInt *nnz,*is_indices; 19348a0068c3SStefano Zampini PetscInt ncc; 1935984c4197SStefano Zampini /* some quantities */ 193645a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1937a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1938984c4197SStefano Zampini 1939674ae819SStefano Zampini PetscFunctionBegin; 19408e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19418e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19428e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1943088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1944088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19450e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19460e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19470e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19480e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19490e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1950088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1951cf5a6209SStefano Zampini 1952cf5a6209SStefano Zampini /* print some info */ 1953cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1954cf5a6209SStefano Zampini IS vertices; 1955cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1956cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1957cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1958cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 19591575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1960cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1961cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1962fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1963fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1964cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 19651575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1966cf5a6209SStefano Zampini } 1967cf5a6209SStefano Zampini 1968cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19699162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1970cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1971cf5a6209SStefano Zampini const Vec *nearnullvecs; 1972cf5a6209SStefano Zampini Vec *localnearnullsp; 1973cf5a6209SStefano Zampini PetscScalar *array; 1974cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1975cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1976674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1977b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1978674ae819SStefano Zampini PetscScalar *work; 1979674ae819SStefano Zampini PetscReal *singular_vals; 1980674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1981674ae819SStefano Zampini PetscReal *rwork; 1982674ae819SStefano Zampini #endif 1983674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1984674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1985674ae819SStefano Zampini #else 1986964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1987964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1988674ae819SStefano Zampini #endif 1989674ae819SStefano Zampini 1990674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1991d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1992d06fc5fdSStefano Zampini /* free unneeded index sets */ 1993d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1994d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1995674ae819SStefano Zampini } 1996d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1997d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1998d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1999d06fc5fdSStefano Zampini } 2000d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2001d06fc5fdSStefano Zampini n_ISForEdges = 0; 2002d06fc5fdSStefano Zampini } 2003d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2004d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2005d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2006d06fc5fdSStefano Zampini } 2007d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2008d06fc5fdSStefano Zampini n_ISForFaces = 0; 2009d06fc5fdSStefano Zampini } 201070022509SStefano Zampini 2011674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2012674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2013674ae819SStefano Zampini if (nearnullsp) { 2014674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2015f4ddd8eeSStefano Zampini /* remove any stored info */ 2016f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2017f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2018f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2019f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2020f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2021473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2022f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2023f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2024f4ddd8eeSStefano Zampini } 2025984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2026984c4197SStefano Zampini nnsp_size = 0; 2027674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2028674ae819SStefano Zampini } 2029984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2030984c4197SStefano Zampini max_constraints = nnsp_size; 2031984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2032984c4197SStefano Zampini 2033674ae819SStefano Zampini /* 2034674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20359162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20369162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20379162d606SStefano Zampini There can be multiple constraints per connected component 2038674ae819SStefano Zampini */ 2039674ae819SStefano Zampini n_vertices = 0; 2040674ae819SStefano Zampini if (ISForVertices) { 2041674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2042674ae819SStefano Zampini } 20439162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20449162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20459162d606SStefano Zampini 20469162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20479162d606SStefano Zampini total_counts *= max_constraints; 2048674ae819SStefano Zampini total_counts += n_vertices; 20494641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20509162d606SStefano Zampini 2051674ae819SStefano Zampini total_counts = 0; 2052674ae819SStefano Zampini max_size_of_constraint = 0; 2053674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20549162d606SStefano Zampini IS used_is; 2055674ae819SStefano Zampini if (i<n_ISForEdges) { 20569162d606SStefano Zampini used_is = ISForEdges[i]; 2057674ae819SStefano Zampini } else { 20589162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2059674ae819SStefano Zampini } 20609162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2061674ae819SStefano Zampini total_counts += j; 2062674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2063674ae819SStefano Zampini } 20649162d606SStefano 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); 20659162d606SStefano Zampini 2066984c4197SStefano Zampini /* get local part of global near null space vectors */ 2067785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2068984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2069984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2070e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2071e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2072984c4197SStefano Zampini } 2073674ae819SStefano Zampini 2074242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2075242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2076a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2077242a89d7SStefano Zampini 2078984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2079a773dcb8SStefano Zampini if (!skip_lapack) { 2080674ae819SStefano Zampini PetscScalar temp_work; 2081911cabfeSStefano Zampini 2082674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2083984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2084785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2085785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2086785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2087674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2088785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2089674ae819SStefano Zampini #endif 2090674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2091c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2092c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2093674ae819SStefano Zampini lwork = -1; 2094674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2095674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2096c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2097674ae819SStefano Zampini #else 2098c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2099674ae819SStefano Zampini #endif 2100674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2101984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2102674ae819SStefano Zampini #else /* on missing GESVD */ 2103674ae819SStefano Zampini /* SVD */ 2104674ae819SStefano Zampini PetscInt max_n,min_n; 2105674ae819SStefano Zampini max_n = max_size_of_constraint; 2106984c4197SStefano Zampini min_n = max_constraints; 2107984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2108674ae819SStefano Zampini min_n = max_size_of_constraint; 2109984c4197SStefano Zampini max_n = max_constraints; 2110674ae819SStefano Zampini } 2111785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2112674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2113785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2114674ae819SStefano Zampini #endif 2115674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2116674ae819SStefano Zampini lwork = -1; 2117e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2118e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2119b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2120674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2121674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21229162d606SStefano 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)); 2123674ae819SStefano Zampini #else 21249162d606SStefano 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)); 2125674ae819SStefano Zampini #endif 2126674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2127984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2128984c4197SStefano Zampini #endif /* on missing GESVD */ 2129674ae819SStefano Zampini /* Allocate optimal workspace */ 2130674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2131854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2132674ae819SStefano Zampini } 2133674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2134674ae819SStefano Zampini total_counts = 0; 21359162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21369162d606SStefano Zampini constraints_data_ptr[0] = 0; 2137674ae819SStefano Zampini /* vertices */ 21389162d606SStefano Zampini if (n_vertices) { 2139674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21409162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2142674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21439162d606SStefano Zampini constraints_n[total_counts] = 1; 21449162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21459162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21469162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2147674ae819SStefano Zampini total_counts++; 2148674ae819SStefano Zampini } 2149674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2150984c4197SStefano Zampini PetscBool used_vertex; 2151674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2152674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2153674ae819SStefano Zampini k = 0; 2154674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2155984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2156984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21579162d606SStefano Zampini constraints_n[total_counts] = 1; 21589162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21599162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21609162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21619162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2162674ae819SStefano Zampini total_counts++; 2163674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2164674ae819SStefano Zampini } 2165984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2166674ae819SStefano Zampini k++; 2167674ae819SStefano Zampini } 2168674ae819SStefano Zampini } 2169674ae819SStefano Zampini } 2170674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2171674ae819SStefano Zampini n_vertices = total_counts; 2172674ae819SStefano Zampini } 2173984c4197SStefano Zampini 2174674ae819SStefano Zampini /* edges and faces */ 21759162d606SStefano Zampini total_counts_cc = total_counts; 2176911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 21779162d606SStefano Zampini IS used_is; 21789162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 21799162d606SStefano Zampini 2180911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 21819162d606SStefano Zampini used_is = ISForEdges[ncc]; 2182984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2183674ae819SStefano Zampini } else { 21849162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2185984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2186674ae819SStefano Zampini } 2187674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 21889162d606SStefano Zampini 21899162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 21909162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2191984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2192984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2193674ae819SStefano Zampini if (nnsp_has_cnst) { 21945b08dc53SStefano Zampini PetscScalar quad_value; 21959162d606SStefano Zampini 21969162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21979162d606SStefano Zampini idxs_copied = PETSC_TRUE; 21989162d606SStefano Zampini 2199a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2200674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2201a773dcb8SStefano Zampini } else { 2202a773dcb8SStefano Zampini quad_value = 1.0; 2203a773dcb8SStefano Zampini } 2204674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22059162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2206674ae819SStefano Zampini } 22079162d606SStefano Zampini temp_constraints++; 2208674ae819SStefano Zampini total_counts++; 2209674ae819SStefano Zampini } 2210674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2211984c4197SStefano Zampini PetscReal real_value; 22129162d606SStefano Zampini PetscScalar *ptr_to_data; 22139162d606SStefano Zampini 2214984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22159162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2216674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22179162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2218674ae819SStefano Zampini } 2219984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2220984c4197SStefano Zampini /* check if array is null on the connected component */ 2221e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22229162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22235b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2224674ae819SStefano Zampini temp_constraints++; 2225674ae819SStefano Zampini total_counts++; 22269162d606SStefano Zampini if (!idxs_copied) { 22279162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22289162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2229674ae819SStefano Zampini } 2230674ae819SStefano Zampini } 22319162d606SStefano Zampini } 22329162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 223345a1bb75SStefano Zampini valid_constraints = temp_constraints; 2234eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2235a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22369162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22379162d606SStefano Zampini 22389162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2239a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22409162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2241a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22429162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2243a773dcb8SStefano Zampini } else { /* perform SVD */ 2244984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22459162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2246674ae819SStefano Zampini 2247674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2248984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2249984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2250984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2251984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2252984c4197SStefano Zampini from that computed using LAPACKgesvd 2253984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2254984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2255984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2256674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2257e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2258984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2259674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2260674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22619162d606SStefano 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)); 2262674ae819SStefano Zampini } 2263674ae819SStefano Zampini } 2264e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2265e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2266e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2267674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2268c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2269674ae819SStefano Zampini #else 2270c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2271674ae819SStefano Zampini #endif 2272674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2273984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2274984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2275674ae819SStefano Zampini j = 0; 2276984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2277674ae819SStefano Zampini total_counts = total_counts-j; 227845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2279e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2280c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2281c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2282c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2283c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2284c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2285c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2286674ae819SStefano Zampini if (j<temp_constraints) { 2287984c4197SStefano Zampini PetscInt ii; 2288984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2289674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 22909162d606SStefano 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)); 2291674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2292984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2293674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 22949162d606SStefano 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]; 2295674ae819SStefano Zampini } 2296674ae819SStefano Zampini } 2297674ae819SStefano Zampini } 2298674ae819SStefano Zampini #else /* on missing GESVD */ 2299e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2300e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2301b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2302674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2303674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23049162d606SStefano 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)); 2305674ae819SStefano Zampini #else 23069162d606SStefano 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)); 2307674ae819SStefano Zampini #endif 2308984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2309674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2310984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2311e310c8b4SStefano Zampini k = temp_constraints; 2312e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2313674ae819SStefano Zampini j = 0; 2314e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 231545a1bb75SStefano Zampini valid_constraints = k-j; 2316911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2317984c4197SStefano Zampini #endif /* on missing GESVD */ 2318674ae819SStefano Zampini } 2319a773dcb8SStefano Zampini } 23209162d606SStefano Zampini /* update pointers information */ 23219162d606SStefano Zampini if (valid_constraints) { 23229162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23239162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23249162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23259162d606SStefano Zampini /* set change_of_basis flag */ 232645a1bb75SStefano Zampini if (boolforchange) { 2327b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23289162d606SStefano Zampini } 2329b3d85658SStefano Zampini total_counts_cc++; 233045a1bb75SStefano Zampini } 233145a1bb75SStefano Zampini } 2332984c4197SStefano Zampini /* free workspace */ 23338f1c130eSStefano Zampini if (!skip_lapack) { 2334984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2335984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2336984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2337984c4197SStefano Zampini #endif 2338984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2339984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2340984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2341984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2342984c4197SStefano Zampini #endif 2343984c4197SStefano Zampini } 2344984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2345984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2346984c4197SStefano Zampini } 2347984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2348cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2349cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2350cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2351cf5a6209SStefano Zampini } 2352cf5a6209SStefano Zampini if (n_ISForFaces) { 2353cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2354cf5a6209SStefano Zampini } 2355cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2356cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2357cf5a6209SStefano Zampini } 2358cf5a6209SStefano Zampini if (n_ISForEdges) { 2359cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2360cf5a6209SStefano Zampini } 2361cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 236208122e43SStefano Zampini } else { 236308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2364984c4197SStefano Zampini 236508122e43SStefano Zampini total_counts = 0; 236608122e43SStefano Zampini n_vertices = 0; 2367d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2368d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 236908122e43SStefano Zampini } 237008122e43SStefano Zampini max_constraints = 0; 23719162d606SStefano Zampini total_counts_cc = 0; 237208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 237308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 23749162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 237508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 237608122e43SStefano Zampini } 23779162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 23789162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 23799162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 23809162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 238174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 23829162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 23839162d606SStefano Zampini total_counts_cc = 0; 23849162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 23859162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 23869162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 238708122e43SStefano Zampini } 238808122e43SStefano Zampini } 23899162d606SStefano Zampini #if 0 23909162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 23919162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23929162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 23939162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 23949162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 23959162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 23969162d606SStefano Zampini } 23979162d606SStefano Zampini printf("\n"); 23989162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 23999162d606SStefano Zampini } 24001b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24018bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24021b968477SStefano Zampini } 24031b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24048bec7fa6SStefano 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]); 24051b968477SStefano Zampini } 240608122e43SStefano Zampini #endif 240708122e43SStefano Zampini 24088bec7fa6SStefano Zampini max_size_of_constraint = 0; 24099162d606SStefano 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]); 24109162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 241108122e43SStefano Zampini /* Change of basis */ 2412b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 241308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 241408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 241508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2416b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 241708122e43SStefano Zampini } 241808122e43SStefano Zampini } 241908122e43SStefano Zampini } 242008122e43SStefano Zampini } 2421984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24220e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 242308122e43SStefano Zampini 24249162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24259162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24266c4ed002SBarry Smith if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %D != %D\n",constraints_idxs_ptr[total_counts_cc],i); 2427674ae819SStefano Zampini 2428674ae819SStefano Zampini /* Create constraint matrix */ 2429674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 243016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2431984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2432984c4197SStefano Zampini 2433984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2434a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2435a717540cSStefano Zampini qr_needed = PETSC_FALSE; 243674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2437984c4197SStefano Zampini total_primal_vertices=0; 2438b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24399162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2441984c4197SStefano Zampini if (size_of_constraint == 1) { 24429162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2443b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 244464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24459162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24469162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2447a717540cSStefano Zampini } 2448b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 244974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2450a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2451a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2452a717540cSStefano Zampini } 2453fa434743SStefano Zampini } else { 2454b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2455fa434743SStefano Zampini } 2456a717540cSStefano Zampini } 2457b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2458b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2459674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 246070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2461b3d85658SStefano Zampini 2462b3d85658SStefano 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); 24630e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24640e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2465984c4197SStefano Zampini 2466984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 246774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2468785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2469984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 247074d5cdf7SStefano Zampini 2471984c4197SStefano Zampini j = total_primal_vertices; 247274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2473b3d85658SStefano Zampini cum = total_primal_vertices; 24749162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24754641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2476b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2477b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2478b3d85658SStefano Zampini cum++; 24799162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 248074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 248174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 248274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 248374d5cdf7SStefano Zampini } 24849162d606SStefano Zampini j += constraints_n[i]; 2485674ae819SStefano Zampini } 2486674ae819SStefano Zampini } 2487674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2488674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2489088faed8SStefano Zampini 2490674ae819SStefano Zampini /* set values in constraint matrix */ 2491984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 24920e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2493674ae819SStefano Zampini } 2494984c4197SStefano Zampini total_counts = total_primal_vertices; 24959162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24964641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 24979162d606SStefano Zampini PetscInt *cols; 24989162d606SStefano Zampini 24999162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25009162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25019162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25029162d606SStefano Zampini PetscInt row = total_counts+k; 25039162d606SStefano Zampini PetscScalar *vals; 25049162d606SStefano Zampini 25059162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25069162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25079162d606SStefano Zampini } 25089162d606SStefano Zampini total_counts += constraints_n[i]; 2509674ae819SStefano Zampini } 2510674ae819SStefano Zampini } 2511674ae819SStefano Zampini /* assembling */ 2512674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2513674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2514088faed8SStefano Zampini 2515984c4197SStefano Zampini /* 25166a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2517984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2518f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 2519984c4197SStefano Zampini */ 2520674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2521674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2522026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2523984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2524984c4197SStefano Zampini /* working stuff for GEQRF */ 252581d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2526984c4197SStefano Zampini PetscBLASInt lqr_work; 2527984c4197SStefano Zampini /* working stuff for UNGQR */ 2528984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2529984c4197SStefano Zampini PetscBLASInt lgqr_work; 2530984c4197SStefano Zampini /* working stuff for TRTRS */ 2531984c4197SStefano Zampini PetscScalar *trs_rhs; 25323f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2533984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2534984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2535984c4197SStefano Zampini PetscScalar *start_vals; 2536984c4197SStefano Zampini /* working stuff for values insertion */ 25374641a718SStefano Zampini PetscBT is_primal; 253864efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2539906d46d4SStefano Zampini /* matrix sizes */ 2540906d46d4SStefano Zampini PetscInt global_size,local_size; 2541906d46d4SStefano Zampini /* temporary change of basis */ 2542906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2543cf5a6209SStefano Zampini /* extra space for debugging */ 2544cf5a6209SStefano Zampini PetscScalar *dbg_work; 2545984c4197SStefano Zampini 2546906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2547906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 254816f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2549bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2550906d46d4SStefano Zampini /* nonzeros for local mat */ 2551bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2552bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25539162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2554a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2556a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25579162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2558a717540cSStefano Zampini } else { 25599162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25609162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2561a717540cSStefano Zampini } 2562a717540cSStefano Zampini } 2563a717540cSStefano Zampini } 2564906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2565bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2566a717540cSStefano Zampini /* Set initial identity in the matrix */ 2567bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2568906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2569a717540cSStefano Zampini } 2570a717540cSStefano Zampini 2571a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2572a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2573a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2574a717540cSStefano Zampini } 2575a717540cSStefano Zampini 2576a717540cSStefano Zampini 2577a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2578a717540cSStefano Zampini /* 2579a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2580a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2581a717540cSStefano Zampini 2582a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2583a717540cSStefano Zampini 2584a6b551f4SStefano 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) 2585a6b551f4SStefano Zampini 2586a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2587a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2588a717540cSStefano Zampini | ... | 2589a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2590a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2591a717540cSStefano Zampini 2592a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2593a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2594a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2595a6b551f4SStefano Zampini 2596a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2597a717540cSStefano Zampini */ 2598a717540cSStefano Zampini if (qr_needed) { 2599984c4197SStefano Zampini /* space to store Q */ 2600854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2601984c4197SStefano Zampini /* first we issue queries for optimal work */ 26023f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2605984c4197SStefano Zampini lqr_work = -1; 26063f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2607984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2608984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2609785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2610984c4197SStefano Zampini lgqr_work = -1; 26113f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26123f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26153f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26163f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2617984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2618984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2619785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2620984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2621785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2622984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2623785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2624a717540cSStefano Zampini /* allocating workspace for check */ 2625a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2626cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2627a717540cSStefano Zampini } 2628a717540cSStefano Zampini } 2629984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26304641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2631473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 26336c4ed002SBarry Smith if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %D != %D\n",total_primal_vertices,i); 263439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 263539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 263639e2fb2aSStefano Zampini } 263739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2638984c4197SStefano Zampini 2639a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26409162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26419162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26424641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2643984c4197SStefano Zampini /* get constraint info */ 26449162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2645984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2646984c4197SStefano Zampini 2647984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26489162d606SStefano 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); 2649674ae819SStefano Zampini } 2650984c4197SStefano Zampini 2651fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2652a717540cSStefano Zampini 2653a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2654a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26559162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2656a717540cSStefano Zampini } 2657984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26589162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2659984c4197SStefano Zampini 2660984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26613f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26623f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26633f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2664674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2666984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2667674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2668984c4197SStefano Zampini 2669984c4197SStefano Zampini /* explictly compute R^-T */ 2670984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2671984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 26723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 26743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2676984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26773f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2678984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2679984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2680984c4197SStefano Zampini 2681a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 26823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26833f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26843f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26853f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2686984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26873f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2688984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2689984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2690984c4197SStefano Zampini 2691984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2692984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2693984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 26943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26963f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 26993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2700984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27019162d606SStefano 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)); 2702984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27039162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2704984c4197SStefano Zampini 2705984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27069162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2707984c4197SStefano Zampini /* insert cols for primal dofs */ 2708984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2709984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27109162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2711906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2712984c4197SStefano Zampini } 2713984c4197SStefano Zampini /* insert cols for dual dofs */ 2714984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27159162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2716984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27179162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2718906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2719984c4197SStefano Zampini j++; 2720674ae819SStefano Zampini } 2721674ae819SStefano Zampini } 2722984c4197SStefano Zampini 2723984c4197SStefano Zampini /* check change of basis */ 2724984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2725984c4197SStefano Zampini PetscInt ii,jj; 2726984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2727c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2728c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2729c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2730c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2731c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2732c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2733984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2734cf5a6209SStefano 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)); 2735984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2736984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2737984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2738cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2739cf5a6209SStefano 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; 2740674ae819SStefano Zampini } 2741674ae819SStefano Zampini } 2742984c4197SStefano Zampini if (!valid_qr) { 274322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2744984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2745984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2746cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2747cf5a6209SStefano 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])); 2748674ae819SStefano Zampini } 2749cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2750cf5a6209SStefano 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])); 2751984c4197SStefano Zampini } 2752984c4197SStefano Zampini } 2753984c4197SStefano Zampini } 2754674ae819SStefano Zampini } else { 275522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2756674ae819SStefano Zampini } 2757674ae819SStefano Zampini } 2758a717540cSStefano Zampini } else { /* simple transformation block */ 2759a717540cSStefano Zampini PetscInt row,col; 2760a6b551f4SStefano Zampini PetscScalar val,norm; 2761a6b551f4SStefano Zampini 2762a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27639162d606SStefano 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)); 2764a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27659162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27669162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2767bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 27689162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2769906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 27709162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2771a717540cSStefano Zampini } else { 2772a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 27739162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2774a717540cSStefano Zampini if (row != col) { 27759162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2776a717540cSStefano Zampini } else { 27779162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2778a717540cSStefano Zampini } 2779906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2780a717540cSStefano Zampini } 2781a717540cSStefano Zampini } 2782a717540cSStefano Zampini } 278398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 278422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2785a717540cSStefano Zampini } 2786674ae819SStefano Zampini } 2787984c4197SStefano Zampini } else { 2788984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27899162d606SStefano 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); 2790674ae819SStefano Zampini } 2791674ae819SStefano Zampini } 2792674ae819SStefano Zampini } 2793a717540cSStefano Zampini 2794a717540cSStefano Zampini /* free workspace */ 2795a717540cSStefano Zampini if (qr_needed) { 2796984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2797cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2798984c4197SStefano Zampini } 2799984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2800984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2801984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2802984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2803984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2804674ae819SStefano Zampini } 2805a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2806906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2807906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2808906d46d4SStefano Zampini 2809906d46d4SStefano Zampini /* assembling of global change of variable */ 2810bbb9e6c6SStefano Zampini { 2811bbb9e6c6SStefano Zampini Mat tmat; 281216f15bc4SStefano Zampini PetscInt bs; 281316f15bc4SStefano Zampini 2814906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2815906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2816bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2817bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2818bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2819bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 282016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 282116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2822906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2823bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2824bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2825bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2826bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2827bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2828e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2829e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2830bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2831bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2832906d46d4SStefano Zampini } 2833906d46d4SStefano Zampini /* check */ 2834906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2835906d46d4SStefano Zampini PetscReal error; 2836906d46d4SStefano Zampini Vec x,x_change; 2837906d46d4SStefano Zampini 2838906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2839906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2840906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2841906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2842e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2843e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2844bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2845e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2846e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2847906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2848906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2849906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2850906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2851bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2852906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2853906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2854906d46d4SStefano Zampini } 2855b96c3477SStefano Zampini 2856b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2857b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2858b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2859b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2860ac632422SStefano Zampini Mat S_new,tmat; 2861bbb9e6c6SStefano Zampini IS is_all_N; 2862bbb9e6c6SStefano Zampini 2863bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28646816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2866ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2867b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2868ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2869ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2870ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2871ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2872ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2873b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2874ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2875ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2876ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2877ac632422SStefano Zampini } 2878b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2879b96c3477SStefano Zampini } 2880b96c3477SStefano Zampini } 2881906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2882906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2883b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2884b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2885b9b85e73SStefano Zampini } 2886906d46d4SStefano Zampini 2887906d46d4SStefano Zampini /* set up change of basis context */ 2888906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2889906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2890906d46d4SStefano Zampini 2891906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2892906d46d4SStefano Zampini PetscInt global_size,local_size; 2893906d46d4SStefano Zampini 2894906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2895906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2896906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2897906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2898906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2899906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2900906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2901906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2902906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2903906d46d4SStefano Zampini } else { 2904906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2905906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2906906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2907906d46d4SStefano Zampini } 2908906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2909906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2910906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2911906d46d4SStefano Zampini } else { 2912906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2913906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2914906d46d4SStefano Zampini } 2915906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2916906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2917906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2918906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2919b9b85e73SStefano Zampini } 2920a717540cSStefano Zampini 2921727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2922727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2923727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2924aff50787SStefano 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); 2925c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29260e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2927aff50787SStefano 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); 2928727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2929727cdba6SStefano Zampini } 29300e6343abSStefano Zampini } 29310e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2932727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2933b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2934727cdba6SStefano Zampini 2935a717540cSStefano Zampini /* flush dbg viewer */ 2936b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2937b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2938b8ffe317SStefano Zampini } 2939a717540cSStefano Zampini 2940e310c8b4SStefano Zampini /* free workspace */ 2941a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29424641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 294308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29449162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29459162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 294608122e43SStefano Zampini } else { 29479162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29489162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29499162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 295008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 295108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29529162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29539162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 295408122e43SStefano Zampini } 2955674ae819SStefano Zampini PetscFunctionReturn(0); 2956674ae819SStefano Zampini } 2957674ae819SStefano Zampini 2958674ae819SStefano Zampini #undef __FUNCT__ 2959674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2960674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2961674ae819SStefano Zampini { 2962674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2963674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2964674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29657fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2966674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2967674ae819SStefano Zampini 2968674ae819SStefano Zampini PetscFunctionBegin; 29698e61c736SStefano Zampini /* Reset previously computed graph */ 29708e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2971674ae819SStefano Zampini /* Init local Graph struct */ 29727fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 29733bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 2974674ae819SStefano Zampini 2975575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2976575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2977575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2978575ad6abSStefano Zampini } 29799577ea80SStefano Zampini 2980674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 29814d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 29824d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 29834d379d7bSStefano Zampini PetscInt nvtxs; 2984e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 2985674ae819SStefano Zampini 29864d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 29872fffb893SStefano Zampini 29882fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29892fffb893SStefano Zampini if (flg_row) { 29904d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2991b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 29922fffb893SStefano Zampini } 29932fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29949b28b941SStefano 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 */ 29954d379d7bSStefano Zampini IS is_dummy; 29964d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 29974d379d7bSStefano Zampini PetscInt j,sum; 29984d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 29994d379d7bSStefano Zampini const PetscInt *idxs; 30004d379d7bSStefano Zampini PCBDDCGraph graph; 30014d379d7bSStefano Zampini PetscBT is_on_boundary; 30024d379d7bSStefano Zampini 30034d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30044d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30054d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30064d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30077fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30084d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3009e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3010e496cd5dSStefano Zampini if (flg_row) { 30114d379d7bSStefano Zampini graph->xadj = xadj; 30124d379d7bSStefano Zampini graph->adjncy = adjncy; 3013e496cd5dSStefano Zampini } 30144d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30154d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3016e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30174d379d7bSStefano Zampini 30184d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30199b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30204d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30214d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30224d379d7bSStefano Zampini } 30234d379d7bSStefano Zampini } 30244d379d7bSStefano Zampini 3025e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30264d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30274d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30284d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30294d379d7bSStefano Zampini } 30304d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30314d379d7bSStefano Zampini 3032e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30334d379d7bSStefano Zampini sum = 0; 30344d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30354d379d7bSStefano Zampini PetscInt sizecc = 0; 30364d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30374d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30384d379d7bSStefano Zampini sizecc++; 30394d379d7bSStefano Zampini } 30404d379d7bSStefano Zampini } 30414d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30424d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30434d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30444d379d7bSStefano Zampini } 30454d379d7bSStefano Zampini } 30464d379d7bSStefano Zampini sum += sizecc*sizecc; 30474d379d7bSStefano Zampini } 30484d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30494d379d7bSStefano Zampini sum = 0; 3050e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30514d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30524d379d7bSStefano Zampini cxadj[i] = sum; 30534d379d7bSStefano Zampini sum += temp; 30544d379d7bSStefano Zampini } 3055e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30564d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30574d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30584d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30594d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30604d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30614d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30624d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30634d379d7bSStefano Zampini sizecc++; 30644d379d7bSStefano Zampini } 30654d379d7bSStefano Zampini } 30664d379d7bSStefano Zampini } 30674d379d7bSStefano Zampini } 30684d379d7bSStefano Zampini } 30699b28b941SStefano Zampini if (sum) { 3070e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 30714d379d7bSStefano Zampini } else { 30724d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 30734d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 30744d379d7bSStefano Zampini } 30754d379d7bSStefano Zampini graph->xadj = 0; 30764d379d7bSStefano Zampini graph->adjncy = 0; 30774d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 30784d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 30794d379d7bSStefano Zampini } 3080674ae819SStefano Zampini } 30819b28b941SStefano Zampini if (pcbddc->dbg_flag) { 30829b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3083674ae819SStefano Zampini } 3084674ae819SStefano Zampini 308563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3086674ae819SStefano Zampini vertex_size = 1; 308763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 308863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 308995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 309063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3091e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 309263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3093674ae819SStefano Zampini } 309463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 309563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 309663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3097674ae819SStefano Zampini } 309863602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3099674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 310063602bcaSStefano Zampini } else { 310163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 310263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3103854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 310463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 310563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 310663602bcaSStefano Zampini } 310763602bcaSStefano Zampini } 3108674ae819SStefano Zampini } 3109674ae819SStefano Zampini 3110674ae819SStefano Zampini /* Setup of Graph */ 3111785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3112e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3113785d1243SStefano Zampini } 3114785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3115e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3116785d1243SStefano Zampini } 3117302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3118674ae819SStefano Zampini 3119674ae819SStefano Zampini /* Graph's connected components analysis */ 3120674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3121674ae819SStefano Zampini 3122674ae819SStefano Zampini /* print some info to stdout */ 3123674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3124302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3125674ae819SStefano Zampini } 3126fb180af4SStefano Zampini 3127fb180af4SStefano Zampini /* mark topography has done */ 3128fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3129674ae819SStefano Zampini PetscFunctionReturn(0); 3130674ae819SStefano Zampini } 3131674ae819SStefano Zampini 3132dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3133674ae819SStefano Zampini #undef __FUNCT__ 3134674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3135dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3136674ae819SStefano Zampini { 3137dc456d91SStefano Zampini PetscSF sf; 3138dc456d91SStefano Zampini PetscLayout map; 3139dc456d91SStefano Zampini const PetscInt *idxs; 3140dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3141dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3142dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3143dc456d91SStefano Zampini PetscMPIInt commsize; 3144674ae819SStefano Zampini PetscBool first_found; 3145674ae819SStefano Zampini PetscErrorCode ierr; 3146674ae819SStefano Zampini 3147674ae819SStefano Zampini PetscFunctionBegin; 3148dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3149dc456d91SStefano Zampini if (subset_mult) { 3150dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3151dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3152dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3153674ae819SStefano Zampini } 3154dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3155dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3156dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3157dc456d91SStefano Zampini for (i=0;i<n;i++) { 3158dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3159dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3160674ae819SStefano Zampini } 3161dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3162b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3163dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3164dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3165dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3166dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3167dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3168dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3169dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3170dc456d91SStefano Zampini 3171dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3172dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3173dc456d91SStefano Zampini if (subset_mult) { 3174dc456d91SStefano Zampini const PetscInt* idxs_mult; 3175dc456d91SStefano Zampini 3176dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3177dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3178dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3179674ae819SStefano Zampini } else { 3180dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3181674ae819SStefano Zampini } 3182dc456d91SStefano Zampini /* local size of new subset */ 3183dc456d91SStefano Zampini n_n = 0; 3184dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3185dc456d91SStefano Zampini 3186dc456d91SStefano Zampini /* global indexes in layout */ 3187dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3188dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3189dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3190dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3191dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3192dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3193dc456d91SStefano Zampini 3194dc456d91SStefano Zampini /* reduce from leaves to roots */ 3195dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 319664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 319764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3198dc456d91SStefano Zampini 3199dc456d91SStefano Zampini /* count indexes in local part of layout */ 3200674ae819SStefano Zampini nlocals = 0; 3201674ae819SStefano Zampini first_index = -1; 3202674ae819SStefano Zampini first_found = PETSC_FALSE; 3203dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3204dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3205674ae819SStefano Zampini first_found = PETSC_TRUE; 3206674ae819SStefano Zampini first_index = i; 3207674ae819SStefano Zampini } 3208dc456d91SStefano Zampini nlocals += root_data[i]; 3209674ae819SStefano Zampini } 3210dc456d91SStefano Zampini 3211dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32125fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3213dc456d91SStefano Zampini start = 0; 321464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32155fa240b1SStefano Zampini #else 321664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32175fa240b1SStefano Zampini start = start-nlocals; 32185fa240b1SStefano Zampini #endif 32195fa240b1SStefano Zampini 3220dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3221dc456d91SStefano Zampini *N_n = start + nlocals; 3222dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3223dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3224674ae819SStefano Zampini } 32255fa240b1SStefano Zampini 32265fa240b1SStefano Zampini /* adapt root data with cumulative */ 3227674ae819SStefano Zampini if (first_found) { 3228dc456d91SStefano Zampini PetscInt old_index; 3229dc456d91SStefano Zampini 3230dc456d91SStefano Zampini root_data[first_index] += start; 3231674ae819SStefano Zampini old_index = first_index; 3232dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3233dc456d91SStefano Zampini if (root_data[i]) { 3234dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3235674ae819SStefano Zampini old_index = i; 3236674ae819SStefano Zampini } 3237674ae819SStefano Zampini } 3238674ae819SStefano Zampini } 3239dc456d91SStefano Zampini 3240dc456d91SStefano Zampini /* from roots to leaves */ 3241dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3242dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3243dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3244dc456d91SStefano Zampini 3245dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3246dc456d91SStefano Zampini if (subset_mult) { 3247dc456d91SStefano Zampini const PetscInt* idxs_mult; 3248dc456d91SStefano Zampini PetscInt cum; 3249dc456d91SStefano Zampini 3250dc456d91SStefano Zampini cum = 0; 3251dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3252dc456d91SStefano Zampini for (i=0;i<n;i++) { 3253dc456d91SStefano Zampini PetscInt j; 3254dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3255674ae819SStefano Zampini } 3256dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3257674ae819SStefano Zampini } else { 3258dc456d91SStefano Zampini for (i=0;i<n;i++) { 3259dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3260674ae819SStefano Zampini } 3261674ae819SStefano Zampini } 3262dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3263dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3264674ae819SStefano Zampini PetscFunctionReturn(0); 3265674ae819SStefano Zampini } 32669a7d3425SStefano Zampini 32679a7d3425SStefano Zampini #undef __FUNCT__ 32689a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32699a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32709a7d3425SStefano Zampini { 32719a7d3425SStefano Zampini PetscInt i,j; 32729a7d3425SStefano Zampini PetscScalar *alphas; 32739a7d3425SStefano Zampini PetscErrorCode ierr; 32749a7d3425SStefano Zampini 32759a7d3425SStefano Zampini PetscFunctionBegin; 32769a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3277785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 32789a7d3425SStefano Zampini for (i=0;i<n;i++) { 32799a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 32809a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 32819a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 32829a7d3425SStefano Zampini } 32839a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 32849a7d3425SStefano Zampini PetscFunctionReturn(0); 32859a7d3425SStefano Zampini } 32869a7d3425SStefano Zampini 3287e7931f94SStefano Zampini #undef __FUNCT__ 328870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3289b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3290e7931f94SStefano Zampini { 329152e5ac9dSStefano Zampini IS ranks_send_to; 3292e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3293e7931f94SStefano Zampini PetscMPIInt size,rank,color; 329452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 329552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 32963837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 32972b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3298e7931f94SStefano Zampini PetscSubcomm subcomm; 329952e5ac9dSStefano Zampini PetscErrorCode ierr; 3300a57a6d2fSStefano Zampini 3301e7931f94SStefano Zampini PetscFunctionBegin; 3302c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 3303c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 3304c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3305e7931f94SStefano Zampini 3306e7931f94SStefano Zampini /* Get info on mapping */ 33073bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33083bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3309e7931f94SStefano Zampini 3310e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3311785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3312e7931f94SStefano Zampini xadj[0] = 0; 3313e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3314785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3315785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3316e7931f94SStefano Zampini 33172b510759SStefano Zampini if (threshold) { 3318d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33192b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3320d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3321d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3322d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3323d023bfaeSStefano Zampini xadj_count++; 3324e7931f94SStefano Zampini } 3325e7931f94SStefano Zampini } 3326d023bfaeSStefano Zampini xadj[1] = xadj_count; 3327c8587f34SStefano Zampini } else { 3328e7931f94SStefano Zampini if (xadj[1]) { 3329e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3330e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3331c8587f34SStefano Zampini } 3332e7931f94SStefano Zampini } 33333bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3334e7931f94SStefano Zampini if (use_square) { 3335e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3336e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3337e7931f94SStefano Zampini } 3338e7931f94SStefano Zampini } 3339e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3340e7931f94SStefano Zampini 33413837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3342e7931f94SStefano Zampini 3343e7931f94SStefano Zampini /* 3344e7931f94SStefano Zampini Restrict work on active processes only. 3345e7931f94SStefano Zampini */ 3346e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3347e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3348e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33492b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3350d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3351e7931f94SStefano Zampini if (color) { 3352e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3353e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3354e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3355c8587f34SStefano Zampini } else { 335652e5ac9dSStefano Zampini Mat subdomain_adj; 335752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 335852e5ac9dSStefano Zampini MatPartitioning partitioner; 335952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 336052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3361b0c7d250SStefano Zampini PetscBool aggregate; 3362b0c7d250SStefano Zampini 3363306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3364785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3365e7931f94SStefano Zampini prank = rank; 3366306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33678002ef2cSStefano Zampini /* 3368e7931f94SStefano Zampini for (i=0;i<size;i++) { 3369e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3370c8587f34SStefano Zampini } 33718002ef2cSStefano Zampini */ 3372e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3373e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3374c8587f34SStefano Zampini } 3375e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3376b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3377b0c7d250SStefano Zampini if (aggregate) { 3378b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3379b0c7d250SStefano Zampini PetscMPIInt nrank; 3380b0c7d250SStefano Zampini PetscScalar *vals; 3381b0c7d250SStefano Zampini 3382b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3383b0c7d250SStefano Zampini lrows = 0; 3384b0c7d250SStefano Zampini if (nrank<redprocs) { 3385b0c7d250SStefano Zampini lrows = size/redprocs; 3386b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3387b0c7d250SStefano Zampini } 33885fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3389b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3390b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3391b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3392b0c7d250SStefano Zampini row = nrank; 3393b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3394b0c7d250SStefano Zampini cols = adjncy; 3395b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3396b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3397b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3398b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3399b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 340052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 340152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 340252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3403b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3404b0c7d250SStefano Zampini } else { 3405306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3406b0c7d250SStefano Zampini } 340722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3408e7931f94SStefano Zampini 3409e7931f94SStefano Zampini /* Partition */ 3410306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3411e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3412e7931f94SStefano Zampini if (use_vwgt) { 34133837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3414e7931f94SStefano Zampini v_wgt[0] = local_size; 3415e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3416c8587f34SStefano Zampini } 341728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 341828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3419e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3420e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 342122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3422e7931f94SStefano Zampini 342352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 342452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 342552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 342652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3427b0c7d250SStefano Zampini if (!redprocs) { 3428b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 342928143c3dSStefano Zampini } else { 3430b0c7d250SStefano Zampini PetscInt idxs[1]; 3431b0c7d250SStefano Zampini PetscMPIInt tag; 3432b0c7d250SStefano Zampini MPI_Request *reqs; 3433b0c7d250SStefano Zampini 3434b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3435b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3436b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3437b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 343828143c3dSStefano Zampini } 3439b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3440b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3441b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3442b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3443e7931f94SStefano Zampini } 344452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3445e7931f94SStefano Zampini /* clean up */ 3446e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 344752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3448e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3449e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3450e7931f94SStefano Zampini } 3451e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3452e7931f94SStefano Zampini 3453e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3454e7931f94SStefano Zampini i = 1; 3455e7931f94SStefano Zampini if (color) i=0; 3456e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3457e7931f94SStefano Zampini /* get back IS */ 3458e7931f94SStefano Zampini *is_sends = ranks_send_to; 3459e7931f94SStefano Zampini PetscFunctionReturn(0); 3460e7931f94SStefano Zampini } 3461e7931f94SStefano Zampini 3462e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3463e7931f94SStefano Zampini 3464e7931f94SStefano Zampini #undef __FUNCT__ 3465e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 346653a05cb3SStefano 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[]) 3467e7931f94SStefano Zampini { 346870cf5478SStefano Zampini Mat local_mat; 3469e7931f94SStefano Zampini IS is_sends_internal; 34709d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 347128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34729d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3473e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3474e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3475e7931f94SStefano Zampini const PetscInt* is_indices; 3476e7931f94SStefano Zampini MatType new_local_type; 3477e7931f94SStefano Zampini /* buffers */ 3478e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 347928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34809d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3481e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3482e7931f94SStefano Zampini /* MPI */ 348328143c3dSStefano Zampini MPI_Comm comm,comm_n; 348428143c3dSStefano Zampini PetscSubcomm subcomm; 3485e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 348628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 348728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 348828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 348928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 349028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3491e7931f94SStefano Zampini PetscErrorCode ierr; 3492e7931f94SStefano Zampini 3493e7931f94SStefano Zampini PetscFunctionBegin; 349428143c3dSStefano Zampini /* TODO: add missing checks */ 349528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 349628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 349728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 349828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3499e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 350028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3501e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3502e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3503e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3504e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3505e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 350628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 350770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 350870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 350928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 351070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 351170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 351270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 351370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 351470cf5478SStefano Zampini } 3515e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3516e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3517e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3518e7931f94SStefano Zampini if (!is_sends) { 351928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3520b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3521c8587f34SStefano Zampini } else { 3522e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3523e7931f94SStefano Zampini is_sends_internal = is_sends; 3524c8587f34SStefano Zampini } 3525e7931f94SStefano Zampini 3526e7931f94SStefano Zampini /* get comm */ 3527a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3528e7931f94SStefano Zampini 3529e7931f94SStefano Zampini /* compute number of sends */ 3530e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3531e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3532e7931f94SStefano Zampini 3533e7931f94SStefano Zampini /* compute number of receives */ 3534e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3535785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3536e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3537e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3538e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3539e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3540e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3541e7931f94SStefano Zampini 354228143c3dSStefano Zampini /* restrict comm if requested */ 354328143c3dSStefano Zampini subcomm = 0; 354428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 354528143c3dSStefano Zampini if (restrict_comm) { 3546779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3547779c1cceSStefano Zampini 354828143c3dSStefano Zampini color = 0; 354953a05cb3SStefano Zampini if (restrict_full) { 355053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 355153a05cb3SStefano Zampini } else { 355253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 355353a05cb3SStefano Zampini } 3554b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 355528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 355628143c3dSStefano Zampini /* check if reuse has been requested */ 355728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 355828143c3dSStefano Zampini if (*mat_n) { 355928143c3dSStefano Zampini PetscMPIInt subcommsize2; 356028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 356128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 356228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 356328143c3dSStefano Zampini } else { 356428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 356528143c3dSStefano Zampini } 356628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3567779c1cceSStefano Zampini PetscMPIInt rank; 3568779c1cceSStefano Zampini 3569779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 357028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 357128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 357228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3573306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 357428143c3dSStefano Zampini } 357528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 357628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 357728143c3dSStefano Zampini } else { 357828143c3dSStefano Zampini comm_n = comm; 357928143c3dSStefano Zampini } 358028143c3dSStefano Zampini 3581e7931f94SStefano Zampini /* prepare send/receive buffers */ 3582785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3583e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3584785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3585e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 358628143c3dSStefano Zampini if (nis) { 3587854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 358828143c3dSStefano Zampini } 3589e7931f94SStefano Zampini 359028143c3dSStefano Zampini /* Get data from local matrices */ 35916c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3592e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3593e7931f94SStefano Zampini /* 3594e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3595e7931f94SStefano Zampini send_buffer_idxs should contain: 3596e7931f94SStefano Zampini - MatType_PRIVATE type 3597e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3598e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3599e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3600e7931f94SStefano Zampini */ 36016c4ed002SBarry Smith else { 3602e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36033bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3604854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3605e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3606e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36073bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3608e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36093bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3610e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3611e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3612e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3613e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3614c8587f34SStefano Zampini } 3615c8587f34SStefano Zampini } 3616e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 361728143c3dSStefano Zampini /* additional is (if any) */ 361828143c3dSStefano Zampini if (nis) { 361928143c3dSStefano Zampini PetscMPIInt psum; 362028143c3dSStefano Zampini PetscInt j; 362128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 362228143c3dSStefano Zampini PetscInt plen; 362328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 362428143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 362528143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 362628143c3dSStefano Zampini } 3627854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 362828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 362928143c3dSStefano Zampini PetscInt plen; 363028143c3dSStefano Zampini const PetscInt *is_array_idxs; 363128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 363228143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 363328143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363428143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 363528143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363628143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 363728143c3dSStefano Zampini } 363828143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 363928143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 364028143c3dSStefano Zampini } 364128143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 364228143c3dSStefano Zampini } 364328143c3dSStefano Zampini 3644e7931f94SStefano Zampini buf_size_idxs = 0; 3645e7931f94SStefano Zampini buf_size_vals = 0; 364628143c3dSStefano Zampini buf_size_idxs_is = 0; 3647e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3648e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3649e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 365028143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3651e7931f94SStefano Zampini } 3652785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3653785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 365495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3655e7931f94SStefano Zampini 3656e7931f94SStefano Zampini /* get new tags for clean communications */ 3657e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3658e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 365928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3660e7931f94SStefano Zampini 3661e7931f94SStefano Zampini /* allocate for requests */ 3662785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3663785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 366495ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3665785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3666785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 366795ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3668e7931f94SStefano Zampini 3669e7931f94SStefano Zampini /* communications */ 3670e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3671e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 367228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3673e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3674e7931f94SStefano Zampini source_dest = onodes[i]; 3675e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3676e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3677e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3678e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 367928143c3dSStefano Zampini if (nis) { 368028143c3dSStefano 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); 368128143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 368228143c3dSStefano Zampini } 3683e7931f94SStefano Zampini } 3684e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3685e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3686e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3687e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 368828143c3dSStefano Zampini if (nis) { 368928143c3dSStefano 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); 369028143c3dSStefano Zampini } 3691e7931f94SStefano Zampini } 3692e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3693e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3694e7931f94SStefano Zampini 3695e7931f94SStefano Zampini /* assemble new l2g map */ 3696e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3697e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36989d30be91SStefano Zampini new_local_rows = 0; 3699e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37009d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3701e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3702e7931f94SStefano Zampini } 37039d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3704e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37059d30be91SStefano Zampini new_local_rows = 0; 3706e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37079d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37089d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3709e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3710e7931f94SStefano Zampini } 37119d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37129d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3713e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3714e7931f94SStefano Zampini 3715e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3716e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3717e7931f94SStefano 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) */ 3718e7931f94SStefano Zampini if (n_recvs) { 371928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3720e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3721e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3722e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3723e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3724e7931f94SStefano Zampini break; 3725e7931f94SStefano Zampini } 3726e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3727e7931f94SStefano Zampini } 3728e7931f94SStefano Zampini switch (new_local_type_private) { 372928143c3dSStefano Zampini case MATDENSE_PRIVATE: 373028143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3731e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3732e7931f94SStefano Zampini bs = 1; 373328143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 373428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 373528143c3dSStefano Zampini bs = 1; 373628143c3dSStefano Zampini } 3737e7931f94SStefano Zampini break; 3738e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3739e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3740e7931f94SStefano Zampini bs = 1; 3741e7931f94SStefano Zampini break; 3742e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3743e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3744e7931f94SStefano Zampini break; 3745e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3746e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3747e7931f94SStefano Zampini break; 3748e7931f94SStefano Zampini default: 37499d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3750e7931f94SStefano Zampini break; 3751e7931f94SStefano Zampini } 375228143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 375328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 375428143c3dSStefano Zampini bs = 1; 3755e7931f94SStefano Zampini } 3756e7931f94SStefano Zampini 375770cf5478SStefano Zampini /* create MATIS object if needed */ 375870cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3759e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3760e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 376170cf5478SStefano Zampini } else { 376270cf5478SStefano Zampini /* it also destroys the local matrices */ 376370cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 376470cf5478SStefano Zampini } 376570cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3766e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37679d30be91SStefano Zampini 37689d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37699d30be91SStefano Zampini 37709d30be91SStefano Zampini /* Global to local map of received indices */ 37719d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37729d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37739d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37749d30be91SStefano Zampini 37759d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37769d30be91SStefano Zampini buf_size_idxs = 0; 37779d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37789d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37799d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37809d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37819d30be91SStefano Zampini } 37829d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37839d30be91SStefano Zampini 37849d30be91SStefano Zampini /* set preallocation */ 37859d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37869d30be91SStefano Zampini if (!newisdense) { 37879d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37889d30be91SStefano Zampini 37899d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37909d30be91SStefano Zampini if (n_recvs) { 37919d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37929d30be91SStefano Zampini } 37939d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37949d30be91SStefano Zampini PetscInt j; 37959d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37969d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37979d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37989d30be91SStefano Zampini } 37999d30be91SStefano Zampini } else { 38009d30be91SStefano Zampini /* TODO */ 38019d30be91SStefano Zampini } 38029d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38039d30be91SStefano Zampini } 38049d30be91SStefano Zampini if (new_local_nnz) { 38059d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38069d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38079d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38089d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38099d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38109d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38119d30be91SStefano Zampini } else { 38129d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38139d30be91SStefano Zampini } 38149d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38159d30be91SStefano Zampini } else { 38169d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38179d30be91SStefano Zampini } 3818e7931f94SStefano Zampini 3819e7931f94SStefano Zampini /* set values */ 3820e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38219d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3822e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3823e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3824e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38259d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3826e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3827e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3828e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 382928143c3dSStefano Zampini } else { 383028143c3dSStefano Zampini /* TODO */ 3831e7931f94SStefano Zampini } 3832e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3833e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3834e7931f94SStefano Zampini } 3835e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3836e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38399d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38409d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3841e7931f94SStefano Zampini 3842dfd14d43SStefano Zampini #if 0 384328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3844e7931f94SStefano Zampini Vec lvec,rvec; 3845e7931f94SStefano Zampini PetscReal infty_error; 3846e7931f94SStefano Zampini 38472a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3848e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3849e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3850e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 385170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3852e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3853e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3854e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3855e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3856e7931f94SStefano Zampini } 385728143c3dSStefano Zampini #endif 3858e7931f94SStefano Zampini 385928143c3dSStefano Zampini /* assemble new additional is (if any) */ 386028143c3dSStefano Zampini if (nis) { 386128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 386228143c3dSStefano Zampini 386328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3864854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 386528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 386628143c3dSStefano Zampini psum = 0; 386728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 386828143c3dSStefano Zampini for (j=0;j<nis;j++) { 386928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 387128143c3dSStefano Zampini psum += plen; 387228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387328143c3dSStefano Zampini } 387428143c3dSStefano Zampini } 3875854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3876854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 387728143c3dSStefano Zampini for (i=1;i<nis;i++) { 387828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 387928143c3dSStefano Zampini } 388028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 388128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 388228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 388328143c3dSStefano Zampini for (j=0;j<nis;j++) { 388428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 388528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 388628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 388728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 388828143c3dSStefano Zampini } 388928143c3dSStefano Zampini } 389028143c3dSStefano Zampini for (i=0;i<nis;i++) { 389128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 389228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 389328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 389428143c3dSStefano Zampini } 389528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 389628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 389728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 389828143c3dSStefano Zampini } 3899e7931f94SStefano Zampini /* free workspace */ 390028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3901e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3902e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3903e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3904e7931f94SStefano Zampini if (isdense) { 3905e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3906e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3907e7931f94SStefano Zampini } else { 3908e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3909e7931f94SStefano Zampini } 391028143c3dSStefano Zampini if (nis) { 391128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 391228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 391328143c3dSStefano Zampini } 3914e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3915e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 391628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3917e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3918e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 391928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3920e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3921e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3922e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3923e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3924e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 392528143c3dSStefano Zampini if (nis) { 392628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 392728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 392828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 392928143c3dSStefano Zampini } 393028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 393128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 393228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 393328143c3dSStefano Zampini for (i=0;i<nis;i++) { 393428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 393528143c3dSStefano Zampini } 393653a05cb3SStefano Zampini *mat_n = NULL; 393728143c3dSStefano Zampini } 3938e7931f94SStefano Zampini PetscFunctionReturn(0); 3939e7931f94SStefano Zampini } 3940a57a6d2fSStefano Zampini 394112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3942af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 394312edc857SStefano Zampini 3944c8587f34SStefano Zampini #undef __FUNCT__ 3945c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3946c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3947c8587f34SStefano Zampini { 3948c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3949c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 395020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 395120a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39526e683305SStefano Zampini IS coarse_is,*isarray; 39536e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39546e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3955f9eb5b7dSStefano Zampini PC pc_temp; 3956c8587f34SStefano Zampini PCType coarse_pc_type; 3957c8587f34SStefano Zampini KSPType coarse_ksp_type; 3958f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39594f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39606e683305SStefano Zampini Mat t_coarse_mat_is; 39616e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39626e683305SStefano Zampini PetscMPIInt all_procs; 396374e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 396468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 396522bc73bbSStefano Zampini PetscScalar *array; 39669881197aSStefano Zampini PetscErrorCode ierr; 3967fdc09c96SStefano Zampini 3968c8587f34SStefano Zampini PetscFunctionBegin; 3969c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 397068457ee5SStefano 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 */ 3971fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39725a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3973fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3974f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3975f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3976f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3977fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 397851bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 397951bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 3980dc4bcba2SStefano Zampini PC pc; 3981dc4bcba2SStefano Zampini PetscBool isbddc; 3982dc4bcba2SStefano Zampini 3983dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 3984dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 3985dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 3986dc4bcba2SStefano Zampini if (isbddc) { 3987dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 3988dc4bcba2SStefano Zampini } 3989727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3990fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3991fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3992fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3993f4ddd8eeSStefano Zampini } 3994fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3995fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3996f4ddd8eeSStefano Zampini } 399770cf5478SStefano Zampini /* reset any subassembling information */ 399870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39996e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40006e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4001fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4002f4ddd8eeSStefano Zampini } 4003c8587f34SStefano Zampini 40046e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40052b510759SStefano Zampini im_active = !!(pcis->n); 4006b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40076e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40086e683305SStefano Zampini void_procs = all_procs-active_procs; 40096e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 401074e2c79eSStefano Zampini redist = PETSC_FALSE; 401122bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40126e683305SStefano Zampini csin_ml = PETSC_TRUE; 40136e683305SStefano Zampini ncoarse_ml = void_procs; 4014779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4015779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40166e683305SStefano Zampini csin_ds = PETSC_TRUE; 401718a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401818a45a71SStefano Zampini redist = PETSC_TRUE; 401918a45a71SStefano Zampini } else { 40206e683305SStefano Zampini csin_ds = PETSC_TRUE; 4021779c1cceSStefano Zampini ncoarse_ds = active_procs; 4022779c1cceSStefano Zampini redist = PETSC_TRUE; 402318a45a71SStefano Zampini } 40246e683305SStefano Zampini } else { 40256e683305SStefano Zampini csin_ml = PETSC_FALSE; 40266e683305SStefano Zampini ncoarse_ml = all_procs; 40276e683305SStefano Zampini if (void_procs) { 40286e683305SStefano Zampini csin_ds = PETSC_TRUE; 40296e683305SStefano Zampini ncoarse_ds = void_procs; 40306e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40316e683305SStefano Zampini } else { 4032779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 403374e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 403474e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 403574e2c79eSStefano Zampini redist = PETSC_TRUE; 403674e2c79eSStefano Zampini } else { 40376e683305SStefano Zampini csin_ds = PETSC_FALSE; 40386e683305SStefano Zampini ncoarse_ds = all_procs; 40396e683305SStefano Zampini } 40406e683305SStefano Zampini } 404174e2c79eSStefano Zampini } 40426e683305SStefano Zampini 40436e683305SStefano Zampini /* 40446e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40456e683305SStefano Zampini - we have not exceeded the number of levels requested 40466e683305SStefano Zampini - we can actually subassemble the active processes 40476e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40486e683305SStefano Zampini */ 40496e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40506e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40516e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40526e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40536e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4054f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40552b510759SStefano Zampini } else { 4056f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4057c8587f34SStefano Zampini } 4058c8587f34SStefano Zampini } 40596e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40606e683305SStefano Zampini if (multilevel_allowed) { 40616e683305SStefano Zampini ncoarse = ncoarse_ml; 40626e683305SStefano Zampini csin = csin_ml; 406358da7f69SStefano Zampini redist = PETSC_FALSE; 40646e683305SStefano Zampini } else { 40656e683305SStefano Zampini ncoarse = ncoarse_ds; 40666e683305SStefano Zampini csin = csin_ds; 40676e683305SStefano Zampini } 4068e7931f94SStefano Zampini 4069abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4070abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4071abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4072abbbba34SStefano Zampini 4073abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 407422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 407522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 407622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 407722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4078b9b85e73SStefano Zampini #if 0 4079b9b85e73SStefano Zampini { 4080b9b85e73SStefano Zampini PetscViewer viewer; 4081b9b85e73SStefano Zampini char filename[256]; 4082b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4083b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 40846a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4085b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4086f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 4087b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4088b9b85e73SStefano Zampini } 4089b9b85e73SStefano Zampini #endif 4090e176bc59SStefano 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); 40916e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40926e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40936e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4094abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4095abbbba34SStefano Zampini 40966e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40976e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40986e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40996e683305SStefano Zampini const PetscInt *idxs; 41006e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41016e683305SStefano Zampini 41026e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41030be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41046e683305SStefano Zampini /* allocate space for temporary storage */ 4105854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4106854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41076e683305SStefano Zampini /* allocate for IS array */ 41086e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41096e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41106e683305SStefano Zampini nis = nisdofs + nisneu; 4111854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41126e683305SStefano Zampini /* dofs splitting */ 41136e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41146e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41156e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41166e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41176e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41186e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41196e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41206e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41216e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41226e683305SStefano Zampini } 41236e683305SStefano Zampini /* neumann boundaries */ 41246e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41256e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41266e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41276e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41286e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41296e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41306e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41316e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41326e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41336e683305SStefano Zampini } 41346e683305SStefano Zampini /* free memory */ 41356e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41366e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41376e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41386e683305SStefano Zampini } else { 41396e683305SStefano Zampini nis = 0; 41406e683305SStefano Zampini nisdofs = 0; 41416e683305SStefano Zampini nisneu = 0; 41426e683305SStefano Zampini isarray = NULL; 41436e683305SStefano Zampini } 41446e683305SStefano Zampini /* destroy no longer needed map */ 41456e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41466e683305SStefano Zampini 41476e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41486e683305SStefano Zampini coarse_mat_is = NULL; 41496e683305SStefano Zampini if (csin) { 41506e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 415174e2c79eSStefano Zampini if (redist) { 415274e2c79eSStefano Zampini PetscMPIInt rank; 4153779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 415474e2c79eSStefano Zampini 415574e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 415658da7f69SStefano Zampini spc = active_procs/ncoarse; 415758da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4158779c1cceSStefano Zampini if (im_active) { 4159779c1cceSStefano Zampini destsize = 1; 416074e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 416174e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 416274e2c79eSStefano Zampini } else { 416374e2c79eSStefano Zampini dest[0] = rank/(spc+1); 416474e2c79eSStefano Zampini } 416574e2c79eSStefano Zampini } else { 4166779c1cceSStefano Zampini destsize = 0; 41676e683305SStefano Zampini } 4168779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4169779c1cceSStefano Zampini } else if (csin_type_simple) { 41706e683305SStefano Zampini PetscMPIInt rank; 41716e683305SStefano Zampini PetscInt issize,isidx; 4172779c1cceSStefano Zampini 41736e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41746e683305SStefano Zampini if (im_active) { 41756e683305SStefano Zampini issize = 1; 41766e683305SStefano Zampini isidx = (PetscInt)rank; 41776e683305SStefano Zampini } else { 41786e683305SStefano Zampini issize = 0; 41796e683305SStefano Zampini isidx = -1; 41806e683305SStefano Zampini } 41816e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4182779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4183b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41846e683305SStefano Zampini } 4185779c1cceSStefano Zampini 4186779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4187779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4188779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4189779c1cceSStefano Zampini PetscInt *coarse_candidates; 4190779c1cceSStefano Zampini const PetscInt* tisindices; 4191779c1cceSStefano Zampini 4192779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4193779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4194779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4195779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4196779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4197779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4198779c1cceSStefano Zampini } 4199779c1cceSStefano Zampini } 4200779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4201779c1cceSStefano Zampini 4202779c1cceSStefano Zampini 42036e683305SStefano Zampini if (pcbddc->dbg_flag) { 42046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42066e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4208779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42106e683305SStefano Zampini } 42116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42126e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42136e683305SStefano Zampini } 42146e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42156e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42166e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4217854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42186e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42196e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42206e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42216e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42226e683305SStefano Zampini } 42236e683305SStefano Zampini if (pcbddc->dbg_flag) { 42246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42266e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42276e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42286e683305SStefano Zampini } 4229779c1cceSStefano Zampini } 42306e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 423153a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 423253a05cb3SStefano 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); 423353a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 423453a05cb3SStefano 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); 423553a05cb3SStefano Zampini } 42366e683305SStefano Zampini } else { 42376e683305SStefano Zampini if (pcbddc->dbg_flag) { 42386e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42406e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42416e683305SStefano Zampini } 42426e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42436e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42446e683305SStefano Zampini } 42456e683305SStefano Zampini 42466e683305SStefano Zampini /* create local to global scatters for coarse problem */ 424768457ee5SStefano Zampini if (compute_vecs) { 42486e683305SStefano Zampini PetscInt lrows; 42496e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42506e683305SStefano Zampini if (coarse_mat_is) { 42516e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42526e683305SStefano Zampini } else { 42536e683305SStefano Zampini lrows = 0; 42546e683305SStefano Zampini } 42556e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42566e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42576e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42586e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42596e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42606e683305SStefano Zampini } 42616e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42626e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4263c8587f34SStefano Zampini 4264f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4265f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4266f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4267f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4268f9eb5b7dSStefano Zampini } else { 4269f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4270f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4271c8587f34SStefano Zampini } 4272c8587f34SStefano Zampini 42736e683305SStefano Zampini /* print some info if requested */ 42746e683305SStefano Zampini if (pcbddc->dbg_flag) { 42756e683305SStefano Zampini if (!multilevel_allowed) { 42766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42776e683305SStefano Zampini if (multilevel_requested) { 42786e683305SStefano 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); 42796e683305SStefano Zampini } else if (pcbddc->max_levels) { 42806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42816e683305SStefano Zampini } 42826e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42836e683305SStefano Zampini } 42846e683305SStefano Zampini } 42856e683305SStefano Zampini 4286f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42876e683305SStefano Zampini if (coarse_mat_is) { 42886e683305SStefano Zampini MatReuse coarse_mat_reuse; 42896a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42906e683305SStefano Zampini if (pcbddc->dbg_flag) { 42916e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42926e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42936e683305SStefano Zampini } 4294f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4295312be037SStefano Zampini char prefix[256],str_level[16]; 4296e604994aSStefano Zampini size_t len; 42976e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4298422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4299c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4300f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43015f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4302c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43036e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4304c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4305c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4306e604994aSStefano Zampini /* prefix */ 4307e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4308e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4309e604994aSStefano Zampini if (!pcbddc->current_level) { 4310e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4311e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4312c8587f34SStefano Zampini } else { 4313e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4314312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4315312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 431634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4317312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4318e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4319e604994aSStefano Zampini } 4320e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43213e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43223e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43233e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43243e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4325f9eb5b7dSStefano Zampini /* allow user customization */ 4326f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43273e3c6dadSStefano Zampini } 43283e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 432951bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43303e3c6dadSStefano Zampini if (nisdofs) { 43313e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43323e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43333e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43343e3c6dadSStefano Zampini } 43353e3c6dadSStefano Zampini } 43363e3c6dadSStefano Zampini if (nisneu) { 43373e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43383e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4339312be037SStefano Zampini } 4340f9eb5b7dSStefano Zampini 4341f9eb5b7dSStefano Zampini /* get some info after set from options */ 4342f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4343f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43444f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43456e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4346f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4347f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4348f9eb5b7dSStefano Zampini } 434939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 43504f3a063dSStefano Zampini if (isredundant) { 43514f3a063dSStefano Zampini KSP inner_ksp; 43524f3a063dSStefano Zampini PC inner_pc; 43534f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43544f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43554f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43564f3a063dSStefano Zampini } 4357f9eb5b7dSStefano Zampini 4358f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4359fa7f1dd8SStefano Zampini if (coarse_reuse) { 436081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4361fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43626e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4363fa7f1dd8SStefano Zampini } else { 43646e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4365fa7f1dd8SStefano Zampini } 4366c8587f34SStefano Zampini if (isbddc || isnn) { 436722bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 436870cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4369b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 437022b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43736e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43746e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 437522b6e8a2SStefano Zampini } 437670cf5478SStefano Zampini } 437753a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 437870cf5478SStefano Zampini } else { 437922bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 438022bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 438122bc73bbSStefano Zampini } 438222bc73bbSStefano Zampini } else { 43832e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4384c8587f34SStefano Zampini } 4385c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4386c8587f34SStefano Zampini 43873301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43885a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 43893301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 43903301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 43913301b35fSStefano Zampini } 43923301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 43933301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 43943301b35fSStefano Zampini } 43953301b35fSStefano Zampini if (pc->pmat->spd_set) { 43963301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 43973301b35fSStefano Zampini } 43986e683305SStefano Zampini /* set operators */ 43995f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44006e683305SStefano Zampini if (pcbddc->dbg_flag) { 44016e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44026e683305SStefano Zampini } 44036e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44046e683305SStefano Zampini coarse_mat = 0; 44056e683305SStefano Zampini } 44066e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4407b9b85e73SStefano Zampini #if 0 4408b9b85e73SStefano Zampini { 4409b9b85e73SStefano Zampini PetscViewer viewer; 4410b9b85e73SStefano Zampini char filename[256]; 4411b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4412b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 44136a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4414b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4415f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 4416b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4417b9b85e73SStefano Zampini } 4418b9b85e73SStefano Zampini #endif 4419c8587f34SStefano Zampini 442098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 442198a51de6SStefano Zampini Vec crhs,csol; 442298a51de6SStefano Zampini PetscBool ispreonly; 442304708bb6SStefano Zampini 4424f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4425f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4426f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4427f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44286e683305SStefano Zampini /* hack */ 4429f347579bSStefano Zampini if (!csol) { 44302a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4431f9eb5b7dSStefano Zampini } 4432f347579bSStefano Zampini if (!crhs) { 44332a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4434f347579bSStefano Zampini } 4435cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4436cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44376e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4438c8587f34SStefano Zampini KSP check_ksp; 44392b510759SStefano Zampini KSPType check_ksp_type; 4440c8587f34SStefano Zampini PC check_pc; 44416e683305SStefano Zampini Vec check_vec,coarse_vec; 44426a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44432b510759SStefano Zampini PetscInt its; 44446e683305SStefano Zampini PetscBool compute_eigs; 44456e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44466e683305SStefano Zampini PetscInt neigs; 44478e185a42SStefano Zampini const char *prefix; 4448c8587f34SStefano Zampini 44492b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44506e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4451422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 445223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4453f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44542b510759SStefano Zampini if (ispreonly) { 44552b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44566e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44572b510759SStefano Zampini } else { 4458cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44596e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4460c8587f34SStefano Zampini } 4461c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44626e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44636e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44646e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4465a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4466a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4467a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4468a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4469c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4470c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4471c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4472c8587f34SStefano Zampini /* create random vec */ 44736e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44746e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4475c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4477c8587f34SStefano Zampini /* solve coarse problem */ 44786e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4479cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44806e683305SStefano Zampini if (compute_eigs) { 4481854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4482854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44836e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44846e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44856e683305SStefano Zampini lambda_min = eigs_r[0]; 44866e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44876e683305SStefano Zampini if (lambda_max>lambda_min) { 4488cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4489cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4490cbcc2c2aSStefano Zampini } 4491c8587f34SStefano Zampini } 4492c8587f34SStefano Zampini } 4493cbcc2c2aSStefano Zampini 4494c8587f34SStefano Zampini /* check coarse problem residual error */ 44956e683305SStefano Zampini if (pcbddc->dbg_flag) { 44966e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44976e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44986e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4499c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45006e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45016e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4502c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4503779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45046e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45056e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45086e683305SStefano Zampini if (compute_eigs) { 45096e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4510deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4511c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45126e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45136e683305SStefano 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); 45146e683305SStefano Zampini for (i=0;i<neigs;i++) { 45156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4516c8587f34SStefano Zampini } 45176e683305SStefano Zampini } 45186e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45196e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45206e683305SStefano Zampini } 4521c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45226e683305SStefano Zampini if (compute_eigs) { 45236e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45246e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4525c8587f34SStefano Zampini } 45266e683305SStefano Zampini } 45276e683305SStefano Zampini } 4528cbcc2c2aSStefano Zampini /* print additional info */ 4529cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45306e683305SStefano Zampini /* waits until all processes reaches this point */ 45316e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4532cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4533cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4534cbcc2c2aSStefano Zampini } 4535cbcc2c2aSStefano Zampini 45362b510759SStefano Zampini /* free memory */ 4537fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4538c8587f34SStefano Zampini PetscFunctionReturn(0); 4539c8587f34SStefano Zampini } 4540674ae819SStefano Zampini 4541f34684f1SStefano Zampini #undef __FUNCT__ 4542f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4543f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4544f34684f1SStefano Zampini { 4545f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4546f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4547f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4548dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4549dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 455073be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4551dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4552f34684f1SStefano Zampini PetscErrorCode ierr; 4553f34684f1SStefano Zampini 4554f34684f1SStefano Zampini PetscFunctionBegin; 4555f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45566c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4557dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 45583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4559dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4560dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4561dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4562dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4563dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4564dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 45656c4ed002SBarry Smith if (local_size != pcbddc->local_primal_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %D != %D",local_size,pcbddc->local_primal_size); 4566dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4567dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4568dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4569dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4570dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4571f34684f1SStefano Zampini 4572f34684f1SStefano Zampini /* check numbering */ 4573f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4574f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4575dc456d91SStefano Zampini PetscInt i; 4576b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4577f34684f1SStefano Zampini 4578f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4579f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4580f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45811575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4582f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4583f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4584727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4585f34684f1SStefano Zampini } 4586f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4587f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4588f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4589e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4590e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4591e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4592e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4593f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4594f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4595f34684f1SStefano Zampini if (array[i] == 1.0) { 4596b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4597f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4598f34684f1SStefano Zampini } 4599f34684f1SStefano Zampini } 4600b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4601f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4602f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4603f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4604f34684f1SStefano Zampini } 4605f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4606f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4607e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4608e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4609f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4610f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4611b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4612ca8b9ea9SStefano Zampini PetscInt *gidxs; 4613ca8b9ea9SStefano Zampini 4614ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4616f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4617f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4618f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4619f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46204bc2dc4bSStefano 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); 4621f34684f1SStefano Zampini } 4622f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4623ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4624f34684f1SStefano Zampini } 4625f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46261575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4627302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4628f34684f1SStefano Zampini } 46298bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4630f34684f1SStefano Zampini /* get back data */ 4631f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4632f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4633674ae819SStefano Zampini PetscFunctionReturn(0); 4634674ae819SStefano Zampini } 4635674ae819SStefano Zampini 4636e456f2a8SStefano Zampini #undef __FUNCT__ 4637e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4638a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4639e456f2a8SStefano Zampini { 4640e456f2a8SStefano Zampini IS localis_t; 4641a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4642e456f2a8SStefano Zampini PetscScalar *vals; 4643e456f2a8SStefano Zampini PetscErrorCode ierr; 4644e456f2a8SStefano Zampini 4645e456f2a8SStefano Zampini PetscFunctionBegin; 4646a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4647e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4648854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4649e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4650e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4651a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4652a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46531035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4654a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46551035eff8SStefano Zampini } 4656a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4657e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4658e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4659a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4660a7dc3881SStefano Zampini /* now compute set in local ordering */ 4661a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4662a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4663a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4664a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4665a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4666ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4667e456f2a8SStefano Zampini lsize++; 4668e456f2a8SStefano Zampini } 4669e456f2a8SStefano Zampini } 4670854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4671a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4672ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4673e456f2a8SStefano Zampini idxs[lsize++] = i; 4674e456f2a8SStefano Zampini } 4675e456f2a8SStefano Zampini } 4676a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4677a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4678e456f2a8SStefano Zampini *localis = localis_t; 4679e456f2a8SStefano Zampini PetscFunctionReturn(0); 4680e456f2a8SStefano Zampini } 4681906d46d4SStefano Zampini 4682906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4683906d46d4SStefano Zampini #undef __FUNCT__ 4684906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4685906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4686906d46d4SStefano Zampini { 4687906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4688906d46d4SStefano Zampini PetscErrorCode ierr; 4689906d46d4SStefano Zampini 4690906d46d4SStefano Zampini PetscFunctionBegin; 4691906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4692906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4693906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4694906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4695906d46d4SStefano Zampini PetscFunctionReturn(0); 4696906d46d4SStefano Zampini } 4697906d46d4SStefano Zampini 4698906d46d4SStefano Zampini #undef __FUNCT__ 4699906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4700906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4701906d46d4SStefano Zampini { 4702906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4703906d46d4SStefano Zampini PetscErrorCode ierr; 4704906d46d4SStefano Zampini 4705906d46d4SStefano Zampini PetscFunctionBegin; 4706906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4707906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4708906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4709906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4710906d46d4SStefano Zampini PetscFunctionReturn(0); 4711906d46d4SStefano Zampini } 4712b96c3477SStefano Zampini 4713b96c3477SStefano Zampini #undef __FUNCT__ 4714b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 471508122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4716b96c3477SStefano Zampini { 4717a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4718b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4719b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4720a64f4aa4SStefano Zampini Mat S_j; 4721b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4722b96c3477SStefano Zampini PetscBool free_used_adj; 4723b96c3477SStefano Zampini PetscErrorCode ierr; 4724b96c3477SStefano Zampini 4725b96c3477SStefano Zampini PetscFunctionBegin; 4726b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4727b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 472808122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4729b96c3477SStefano Zampini used_xadj = NULL; 4730b96c3477SStefano Zampini used_adjncy = NULL; 4731b96c3477SStefano Zampini } else { 473208122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 473308122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 473408122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 473508122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4736b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4737b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4738b96c3477SStefano Zampini } else { 47392fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4740b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4741b96c3477SStefano Zampini PetscInt nvtxs; 4742b96c3477SStefano Zampini 47432fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47442fffb893SStefano Zampini if (flg_row) { 4745b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4746b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4747b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4748b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47492fffb893SStefano Zampini } else { 47502fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47512fffb893SStefano Zampini used_xadj = NULL; 47522fffb893SStefano Zampini used_adjncy = NULL; 47532fffb893SStefano Zampini } 47542fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4755b96c3477SStefano Zampini } 4756b96c3477SStefano Zampini } 4757d5574798SStefano Zampini 4758d5574798SStefano Zampini /* setup sub_schurs data */ 4759a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4760a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4761a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4762a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4763862806e4SStefano 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); 4764a64f4aa4SStefano Zampini } else { 47656816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 476604708bb6SStefano Zampini PetscBool isseqaij; 47675feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 47685feab87aSStefano Zampini PetscInt n_vertices; 47695feab87aSStefano Zampini 47705feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 47712034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 47725feab87aSStefano Zampini } 477304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 477404708bb6SStefano Zampini if (!isseqaij) { 477504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 477604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 477704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 477804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 477904708bb6SStefano Zampini } else { 4780511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 478104708bb6SStefano Zampini } 478204708bb6SStefano Zampini } 4783862806e4SStefano 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); 4784a64f4aa4SStefano Zampini } 4785a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4786b96c3477SStefano Zampini 4787b96c3477SStefano Zampini /* free adjacency */ 4788b96c3477SStefano Zampini if (free_used_adj) { 4789b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4790b96c3477SStefano Zampini } 4791b96c3477SStefano Zampini PetscFunctionReturn(0); 4792b96c3477SStefano Zampini } 4793b96c3477SStefano Zampini 4794b96c3477SStefano Zampini #undef __FUNCT__ 4795b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 479608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4797b96c3477SStefano Zampini { 4798b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4799b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4800b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4801b96c3477SStefano Zampini PCBDDCGraph graph; 4802b96c3477SStefano Zampini PetscErrorCode ierr; 4803b96c3477SStefano Zampini 4804b96c3477SStefano Zampini PetscFunctionBegin; 4805b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 480608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48073301b35fSStefano Zampini IS verticesIS,verticescomm; 48083301b35fSStefano Zampini PetscInt vsize,*idxs; 4809b96c3477SStefano Zampini 4810b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48113301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48123301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48133301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48143301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48153301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4816b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48177fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48183301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48193301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4820b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4821b96c3477SStefano Zampini /* 4822b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4823b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4824b96c3477SStefano Zampini } 4825b96c3477SStefano Zampini */ 4826b96c3477SStefano Zampini } else { 4827b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4828b96c3477SStefano Zampini } 4829b96c3477SStefano Zampini 4830b96c3477SStefano Zampini /* sub_schurs init */ 4831a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4832a64f4aa4SStefano Zampini 4833b96c3477SStefano Zampini /* free graph struct */ 483408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4835b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4836b96c3477SStefano Zampini } 4837b96c3477SStefano Zampini PetscFunctionReturn(0); 4838b96c3477SStefano Zampini } 4839