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; 126*c43ebad9SStefano 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)); 29908122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 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 { 346c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3478629588bSStefano Zampini PetscScalar *coarse_submat_vals; 348c8587f34SStefano Zampini PetscErrorCode ierr; 349c8587f34SStefano Zampini 350c8587f34SStefano Zampini PetscFunctionBegin; 351f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3525e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 353c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 354c8587f34SStefano Zampini 355684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3560fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 357684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 358c8587f34SStefano Zampini 359c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 360b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 361c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 362c8587f34SStefano Zampini } 363c8587f34SStefano Zampini 3648629588bSStefano Zampini /* 3658629588bSStefano Zampini Setup local correction and local part of coarse basis. 3668629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3678629588bSStefano Zampini */ 36847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3698629588bSStefano Zampini 3708629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3718629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3728629588bSStefano Zampini 3738629588bSStefano Zampini /* free */ 3748629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 375c8587f34SStefano Zampini PetscFunctionReturn(0); 376c8587f34SStefano Zampini } 377c8587f34SStefano Zampini 378c8587f34SStefano Zampini #undef __FUNCT__ 379674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 380674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 381674ae819SStefano Zampini { 382674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 383674ae819SStefano Zampini PetscErrorCode ierr; 384674ae819SStefano Zampini 385674ae819SStefano Zampini PetscFunctionBegin; 386674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 387674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 388674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 389674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 390785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 391674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 392f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 393f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 394785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 39663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 397674ae819SStefano Zampini PetscFunctionReturn(0); 398674ae819SStefano Zampini } 399674ae819SStefano Zampini 400674ae819SStefano Zampini #undef __FUNCT__ 401674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 402674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 403674ae819SStefano Zampini { 404674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 405674ae819SStefano Zampini PetscErrorCode ierr; 406674ae819SStefano Zampini 407674ae819SStefano Zampini PetscFunctionBegin; 408b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 409674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 410674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 411674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 412b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 413674ae819SStefano Zampini PetscFunctionReturn(0); 414674ae819SStefano Zampini } 415674ae819SStefano Zampini 416674ae819SStefano Zampini #undef __FUNCT__ 417674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 418674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 419674ae819SStefano Zampini { 420674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 42106656605SStefano Zampini PetscScalar *array; 422674ae819SStefano Zampini PetscErrorCode ierr; 423674ae819SStefano Zampini 424674ae819SStefano Zampini PetscFunctionBegin; 425674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42658da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42958da7f69SStefano Zampini } 430674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 431674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 43215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 434674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4408ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 441674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 442674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 443674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 444f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 445f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 446f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 448727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4490e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 450f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 45170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4526e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 453674ae819SStefano Zampini PetscFunctionReturn(0); 454674ae819SStefano Zampini } 455674ae819SStefano Zampini 456674ae819SStefano Zampini #undef __FUNCT__ 457f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 458f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4596bfb1811SStefano Zampini { 4606bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4616bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4626bfb1811SStefano Zampini VecType impVecType; 463e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4646bfb1811SStefano Zampini PetscErrorCode ierr; 4656bfb1811SStefano Zampini 4666bfb1811SStefano Zampini PetscFunctionBegin; 4676c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 468e7b262bdSStefano Zampini /* get sizes */ 469b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 470b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4716bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 472e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 473e7b262bdSStefano Zampini /* R nodes */ 474e7b262bdSStefano Zampini old_size = -1; 475e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 476e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 477e7b262bdSStefano Zampini } 478e7b262bdSStefano Zampini if (n_R != old_size) { 479e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 480e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4816bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4826bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4846bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 485e7b262bdSStefano Zampini } 486e7b262bdSStefano Zampini /* local primal dofs */ 487e7b262bdSStefano Zampini old_size = -1; 488e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 489e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 490e7b262bdSStefano Zampini } 491e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 492e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 494e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4956bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 496e7b262bdSStefano Zampini } 497e7b262bdSStefano Zampini /* local explicit constraints */ 498e7b262bdSStefano Zampini old_size = -1; 499e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 500e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 501e7b262bdSStefano Zampini } 502e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 503e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50783b7ccabSStefano Zampini } 5086bfb1811SStefano Zampini PetscFunctionReturn(0); 5096bfb1811SStefano Zampini } 5106bfb1811SStefano Zampini 5116bfb1811SStefano Zampini #undef __FUNCT__ 51247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51488ebb749SStefano Zampini { 51525084f0cSStefano Zampini PetscErrorCode ierr; 51625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 519d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 52025084f0cSStefano Zampini /* submatrices of local problem */ 52180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52206656605SStefano Zampini /* submatrices of local coarse problem */ 52306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52425084f0cSStefano Zampini /* working matrices */ 52506656605SStefano Zampini Mat C_CR; 52625084f0cSStefano Zampini /* additional working stuff */ 52706656605SStefano Zampini PC pc_R; 52806656605SStefano Zampini Mat F; 52906656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 53006656605SStefano Zampini 53125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53206656605SStefano Zampini PetscScalar *work; 53306656605SStefano Zampini PetscInt *idx_V_B; 53406656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 536b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53745a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53888ebb749SStefano Zampini MatType impMatType; 53925084f0cSStefano Zampini /* some shortcuts to scalars */ 54006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54188ebb749SStefano Zampini 54288ebb749SStefano Zampini PetscFunctionBegin; 543b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54488ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54588ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 546b371cd4fSStefano Zampini n_B = pcis->n_B; 547b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 54988ebb749SStefano Zampini 55088ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55188ebb749SStefano Zampini impMatType = MATSEQDENSE; 55288ebb749SStefano Zampini 55388ebb749SStefano Zampini /* vertices in boundary numbering */ 554785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5550e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 5566c4ed002SBarry 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); 55788ebb749SStefano Zampini 55806656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 55906656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56806656605SStefano Zampini 56906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57006656605SStefano Zampini /* allocate workspace */ 57106656605SStefano Zampini n = 0; 57206656605SStefano Zampini if (n_constraints) { 57306656605SStefano Zampini n += n_R*n_constraints; 57406656605SStefano Zampini } 57506656605SStefano Zampini if (n_vertices) { 57606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57780677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57806656605SStefano Zampini } 5793301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 58006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58206656605SStefano Zampini } 58306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58406656605SStefano Zampini 58506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59006656605SStefano Zampini if (isLU || isILU || isCHOL) { 59106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 592d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 593d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 594d62866d3SStefano Zampini MatFactorType type; 595d62866d3SStefano Zampini 5966816873aSStefano Zampini F = reuse_mumps->F; 5976816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 598d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 59906656605SStefano Zampini } else { 60006656605SStefano Zampini F = NULL; 60106656605SStefano Zampini } 60206656605SStefano Zampini 60388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 60488ebb749SStefano Zampini if (n_constraints) { 60506656605SStefano Zampini Mat M1,M2,M3; 60680677318SStefano Zampini Mat auxmat; 60706656605SStefano Zampini IS is_aux; 60880677318SStefano Zampini PetscScalar *array,*array2; 60906656605SStefano Zampini 610f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 61288ebb749SStefano Zampini 61325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 61425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6158ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 61680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61788ebb749SStefano Zampini 61880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 61980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62006656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 62206656605SStefano Zampini const PetscScalar *row_cmat_values; 62306656605SStefano Zampini const PetscInt *row_cmat_indices; 62406656605SStefano Zampini PetscInt size_of_constraint,j; 62588ebb749SStefano Zampini 62606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62806656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 62906656605SStefano Zampini } 63006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63106656605SStefano Zampini } 63280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 63306656605SStefano Zampini if (F) { 63406656605SStefano Zampini Mat B; 63506656605SStefano Zampini 63606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 63780677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 63906656605SStefano Zampini } else { 64080677318SStefano Zampini PetscScalar *marr; 64180677318SStefano Zampini 64280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 64406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 64580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 64606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 64906656605SStefano Zampini } 65080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65106656605SStefano Zampini } 65280677318SStefano Zampini if (!pcbddc->switch_static) { 65380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 65480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66380677318SStefano Zampini } 66480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66780677318SStefano Zampini } else { 66880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 66980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67180677318SStefano Zampini } 67280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 67380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 67480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 67506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 67780677318SStefano Zampini if (isCHOL) { 67880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 67980677318SStefano Zampini } else { 68025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68180677318SStefano Zampini } 68280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 68306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 68425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 68525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 68625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 68780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68880677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 68980677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 692f4ddd8eeSStefano Zampini } 69388ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 69488ebb749SStefano Zampini if (n_vertices) { 69506656605SStefano Zampini IS is_aux; 6963a50541eSStefano Zampini 6976816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 6986816873aSStefano Zampini IS tis; 6996816873aSStefano Zampini 7006816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7016816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7026816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7036816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7046816873aSStefano Zampini } else { 7053a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7066816873aSStefano Zampini } 7079577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7089577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 70904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 71025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 71188ebb749SStefano Zampini } 71288ebb749SStefano Zampini 71388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 714f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 71506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 71606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 71706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 71806656605SStefano Zampini } 719f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 72006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 72106656605SStefano Zampini PetscScalar *marray; 72206656605SStefano Zampini 72306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 72406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 725f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 726f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 727f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 728f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 729f4ddd8eeSStefano Zampini } 730f4ddd8eeSStefano Zampini } 73106656605SStefano Zampini 732f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 73306656605SStefano Zampini PetscScalar *marray; 73488ebb749SStefano Zampini 73506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7368eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 73888ebb749SStefano Zampini } 7393301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 74006656605SStefano Zampini n *= 2; 74188ebb749SStefano Zampini } 74206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 74306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 74406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7458eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 74706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74888ebb749SStefano Zampini } 7493301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7518eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 75488ebb749SStefano Zampini } 75588ebb749SStefano Zampini } else { 756c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 757c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7581b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 759c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 760c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 761c0553b1fSStefano Zampini } 76288ebb749SStefano Zampini } 76306656605SStefano Zampini } 76406656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 76506656605SStefano Zampini /* vertices */ 76606656605SStefano Zampini if (n_vertices) { 76716f15bc4SStefano Zampini 768511c6705SHong Zhang ierr = MatConvert(A_VV,impMatType,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 76904708bb6SStefano Zampini 77016f15bc4SStefano Zampini if (n_R) { 77106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 77206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 77316f15bc4SStefano Zampini PetscScalar *x,*y; 77404708bb6SStefano Zampini PetscBool isseqaij; 77506656605SStefano Zampini 77621eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 777511c6705SHong Zhang ierr = MatConvert(A_RV,impMatType,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 77804708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7796816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 78006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 78106656605SStefano Zampini } else { 78206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 78306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 78406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 78506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 78606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 78706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 78806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 78906656605SStefano Zampini } 79006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 79106656605SStefano Zampini } 79280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 79306656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 79406656605SStefano Zampini if (n_constraints) { 79506656605SStefano Zampini Mat B; 79680677318SStefano Zampini 797b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 79880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 79980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 80080677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 80180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 80280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 80380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 80480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80580677318SStefano Zampini } 80680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 80780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 80880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 80906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 81106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 81206656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 81306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 81406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81506656605SStefano Zampini } 81604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 81704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 818511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 81904708bb6SStefano Zampini } 82006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 82206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 82306656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 82406656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 82506656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 82606656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 83016f15bc4SStefano Zampini } else { 83116f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83216f15bc4SStefano Zampini } 83321eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 83406656605SStefano Zampini /* coarse basis functions */ 83506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 83616f15bc4SStefano Zampini PetscScalar *y; 83716f15bc4SStefano Zampini 83806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 83906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 84406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 84606656605SStefano Zampini 84706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 84806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85406656605SStefano Zampini } 85506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 85606656605SStefano Zampini } 85704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 85804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 85906656605SStefano Zampini } 86006656605SStefano Zampini 86106656605SStefano Zampini if (n_constraints) { 86206656605SStefano Zampini Mat B; 86306656605SStefano Zampini 86406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86806656605SStefano Zampini if (n_vertices) { 86980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 87080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 87180677318SStefano Zampini } else { 87280677318SStefano Zampini Mat S_VCt; 87380677318SStefano Zampini 87480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 87580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 87680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 87780677318SStefano Zampini } 87806656605SStefano Zampini } 87906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88006656605SStefano Zampini /* coarse basis functions */ 88106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 88206656605SStefano Zampini PetscScalar *y; 88306656605SStefano Zampini 88406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 88706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 89106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 89206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 89406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89806656605SStefano Zampini } 89906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90006656605SStefano Zampini } 90106656605SStefano Zampini } 90280677318SStefano Zampini if (n_constraints) { 90380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 90480677318SStefano Zampini } 90506656605SStefano Zampini 90606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9073301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 90806656605SStefano Zampini 90906656605SStefano Zampini if (n_constraints) { 91016f15bc4SStefano Zampini Mat S_CCT,B_C; 91106656605SStefano Zampini 91280677318SStefano Zampini /* this is a lazy thing */ 913511c6705SHong Zhang ierr = MatConvert(C_CR,impMatType,MAT_INPLACE_MATRIX,&C_CR);CHKERRQ(ierr); 91406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 91716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 91806656605SStefano Zampini if (n_vertices) { 91916f15bc4SStefano Zampini Mat B_V,S_VCT; 92006656605SStefano Zampini 92106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 92206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 92516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 92606656605SStefano Zampini } 92706656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 92880677318SStefano Zampini } else { /* if there are no constraints, reset work */ 92980677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 93006656605SStefano Zampini } 93116f15bc4SStefano Zampini if (n_vertices && n_R) { 93206656605SStefano Zampini Mat A_VRT; 93380677318SStefano Zampini PetscScalar *marray; 93406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 93506656605SStefano Zampini 93680677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 937511c6705SHong Zhang ierr = MatConvert(A_VRT,impMatType,MAT_INPLACE_MATRIX,&A_VRT);CHKERRQ(ierr); 93880677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 93906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 94080677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 94180677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 94216f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 94306656605SStefano Zampini } 94406656605SStefano Zampini 94506656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 94606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 94706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95206656605SStefano Zampini } 95306656605SStefano Zampini } else { 95406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95706656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96006656605SStefano Zampini } 96106656605SStefano Zampini } 96206656605SStefano Zampini /* coarse basis functions */ 96306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96406656605SStefano Zampini PetscScalar *y; 96506656605SStefano Zampini 96606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 96706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97106656605SStefano Zampini if (i<n_vertices) { 97206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 97306656605SStefano Zampini } 97406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 97606656605SStefano Zampini 97706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 97806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98406656605SStefano Zampini } 98506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98606656605SStefano Zampini } 98706656605SStefano Zampini } 988d62866d3SStefano Zampini /* free memory */ 98988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 994d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 995d62866d3SStefano Zampini if (n_vertices) { 996d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 997d62866d3SStefano Zampini } 998d62866d3SStefano Zampini if (n_constraints) { 999d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1000d62866d3SStefano Zampini } 100188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 100288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 100388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 100425084f0cSStefano Zampini if (pcbddc->dbg_flag) { 100588ebb749SStefano Zampini Mat coarse_sub_mat; 100625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 100788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 100888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 100988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10108bec7fa6SStefano Zampini Mat C_B,CPHI; 10118bec7fa6SStefano Zampini IS is_dummy; 10128bec7fa6SStefano Zampini Vec mones; 101388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 101488ebb749SStefano Zampini PetscReal real_value; 101588ebb749SStefano Zampini 101688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 101788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 101888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 101988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 102088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 102188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1022c0553b1fSStefano Zampini if (unsymmetric_check) { 102388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 102488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 102588ebb749SStefano Zampini } 102688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 102788ebb749SStefano Zampini 102825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10293301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 103025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1031c0553b1fSStefano Zampini if (unsymmetric_check) { 103288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104488ebb749SStefano Zampini } else { 104588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini } 105488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 1057511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 105881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10598bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10601575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 106106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10628bec7fa6SStefano Zampini 10638bec7fa6SStefano Zampini /* check constraints */ 10648bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 1065*c43ebad9SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 10668bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10678bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10688bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10698bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1071bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1072c0553b1fSStefano Zampini if (unsymmetric_check) { 1073bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1074bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1075bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1076bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1077bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 107888ebb749SStefano Zampini } 10798bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10808bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10818bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10838bec7fa6SStefano Zampini 108425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 108588ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 108688ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 108788ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 108888ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 108988ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 109188ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 109288ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 109388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 109488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1095c0553b1fSStefano Zampini if (unsymmetric_check) { 109688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 109788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 109888ebb749SStefano Zampini } 109988ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 110088ebb749SStefano Zampini } 11018629588bSStefano Zampini /* get back data */ 11028629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 110388ebb749SStefano Zampini PetscFunctionReturn(0); 110488ebb749SStefano Zampini } 110588ebb749SStefano Zampini 110688ebb749SStefano Zampini #undef __FUNCT__ 1107d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1108d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1109aa0d41d4SStefano Zampini { 1110d65f70fdSStefano Zampini Mat *work_mat; 1111d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1112d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1113*c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 1114aa0d41d4SStefano Zampini PetscErrorCode ierr; 1115aa0d41d4SStefano Zampini 1116aa0d41d4SStefano Zampini PetscFunctionBegin; 1117d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1118d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1119d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1120d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1121aa0d41d4SStefano Zampini 1122d65f70fdSStefano Zampini if (!rsorted) { 1123906d46d4SStefano Zampini const PetscInt *idxs; 1124906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1125aa0d41d4SStefano Zampini 1126d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1127d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1128d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1129d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1130aa0d41d4SStefano Zampini } 1131d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1134d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1135aa0d41d4SStefano Zampini } 1136d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1137d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1138d65f70fdSStefano Zampini } else { 1139d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1140d65f70fdSStefano Zampini isrow_s = isrow; 1141aa0d41d4SStefano Zampini } 1142906d46d4SStefano Zampini 1143d65f70fdSStefano Zampini if (!csorted) { 1144d65f70fdSStefano Zampini if (isrow == iscol) { 1145d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini iscol_s = isrow_s; 1147d65f70fdSStefano Zampini } else { 1148d65f70fdSStefano Zampini const PetscInt *idxs; 1149d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1150906d46d4SStefano Zampini 1151d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1154d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1155d65f70fdSStefano Zampini } 1156d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1159d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1160d65f70fdSStefano Zampini } 1161d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1162d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1163d65f70fdSStefano Zampini } 1164d65f70fdSStefano Zampini } else { 1165d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini iscol_s = iscol; 1167d65f70fdSStefano Zampini } 1168d65f70fdSStefano Zampini 1169d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini 1171d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1172906d46d4SStefano Zampini Mat new_mat; 1173d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1174906d46d4SStefano Zampini 1175d65f70fdSStefano Zampini if (!rsorted) { 1176d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1177d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1179d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1180906d46d4SStefano Zampini } 1181d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1182d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini } else { 1184d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1185906d46d4SStefano Zampini } 1186d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1187d65f70fdSStefano Zampini 1188d65f70fdSStefano Zampini if (!csorted) { 1189d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1190d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1192d65f70fdSStefano Zampini } else { 1193d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1194d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1196d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1197d65f70fdSStefano Zampini } 1198d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini } 1201d65f70fdSStefano Zampini } else { 1202d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini } 1204d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1205d65f70fdSStefano Zampini 1206d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini work_mat[0] = new_mat; 1209d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1210d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini } 1212d65f70fdSStefano Zampini 1213d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini *B = work_mat[0]; 1215d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1217d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini PetscFunctionReturn(0); 1219d65f70fdSStefano Zampini } 1220d65f70fdSStefano Zampini 1221d65f70fdSStefano Zampini #undef __FUNCT__ 12225e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12235e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1224aa0d41d4SStefano Zampini { 1225aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12265e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1227d65f70fdSStefano Zampini Mat new_mat; 12285e8657edSStefano Zampini IS is_local,is_global; 1229d65f70fdSStefano Zampini PetscInt local_size; 1230d65f70fdSStefano Zampini PetscBool isseqaij; 1231aa0d41d4SStefano Zampini PetscErrorCode ierr; 1232aa0d41d4SStefano Zampini 1233aa0d41d4SStefano Zampini PetscFunctionBegin; 1234aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12355e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12365e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1237df1311c0SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1238aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1239d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1240aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1241906d46d4SStefano Zampini 1242906d46d4SStefano Zampini /* check */ 1243906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1244906d46d4SStefano Zampini Vec x,x_change; 1245906d46d4SStefano Zampini PetscReal error; 1246906d46d4SStefano Zampini 12475e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1248906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1250e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1251e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1252d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1253e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1254e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1255906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1256906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1257906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1258906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1259906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1260906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1261906d46d4SStefano Zampini } 1262906d46d4SStefano Zampini 126322d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12649b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 126522d5777bSStefano Zampini if (isseqaij) { 1266d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1267aa0d41d4SStefano Zampini } else { 1268aa0d41d4SStefano Zampini Mat work_mat; 1269aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1270d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1271aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1272aa0d41d4SStefano Zampini } 12733301b35fSStefano Zampini if (matis->A->symmetric_set) { 12743301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1275e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12763301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1277e496cd5dSStefano Zampini #endif 12783301b35fSStefano Zampini } 1279d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1280aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1281aa0d41d4SStefano Zampini } 1282aa0d41d4SStefano Zampini 1283aa0d41d4SStefano Zampini #undef __FUNCT__ 1284a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12858ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1286a64d13efSStefano Zampini { 1287a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1288a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1289d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 129053892102SStefano Zampini PetscInt *idx_R_local=NULL; 12913a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12923a50541eSStefano Zampini PetscInt vbs,bs; 12936816873aSStefano Zampini PetscBT bitmask=NULL; 1294a64d13efSStefano Zampini PetscErrorCode ierr; 1295a64d13efSStefano Zampini 1296a64d13efSStefano Zampini PetscFunctionBegin; 1297b23d619eSStefano Zampini /* 1298b23d619eSStefano Zampini No need to setup local scatters if 1299b23d619eSStefano Zampini - primal space is unchanged 1300b23d619eSStefano Zampini AND 1301b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1302b23d619eSStefano Zampini AND 1303b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1304b23d619eSStefano Zampini */ 1305b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1306f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1307f4ddd8eeSStefano Zampini } 1308f4ddd8eeSStefano Zampini /* destroy old objects */ 1309f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1310f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1311f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1312a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1313b371cd4fSStefano Zampini n_B = pcis->n_B; 1314b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1315b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13163a50541eSStefano Zampini 1317a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13186816873aSStefano Zampini 131953892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13206816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1321854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1322a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1323a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13240e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1325a64d13efSStefano Zampini } 1326a64d13efSStefano Zampini 1327a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13284641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13296816873aSStefano Zampini idx_R_local[n_R++] = i; 1330a64d13efSStefano Zampini } 1331a64d13efSStefano Zampini } 133253892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13336816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13346816873aSStefano Zampini 133553892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 133653892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13376816873aSStefano Zampini } 13383a50541eSStefano Zampini 13393a50541eSStefano Zampini /* Block code */ 13403a50541eSStefano Zampini vbs = 1; 13413a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13423a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13433a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13443a50541eSStefano Zampini PetscInt *vary; 1345d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1346785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13473a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1348d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1349d3df7717SStefano 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 */ 13500e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1351d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13523a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13533a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13543a50541eSStefano Zampini break; 13553a50541eSStefano Zampini } 13563a50541eSStefano Zampini } 1357d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1358d3df7717SStefano Zampini } else { 1359d3df7717SStefano Zampini /* Verify directly the R set */ 1360d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1361d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1362d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1363d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1364d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1365d3df7717SStefano Zampini break; 1366d3df7717SStefano Zampini } 1367d3df7717SStefano Zampini } 1368d3df7717SStefano Zampini } 1369d3df7717SStefano Zampini } 13703a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13713a50541eSStefano Zampini vbs = bs; 13723a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13733a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13743a50541eSStefano Zampini } 13753a50541eSStefano Zampini } 13763a50541eSStefano Zampini } 13773a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 137853892102SStefano Zampini if (sub_schurs->reuse_mumps) { 137953892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 138053892102SStefano Zampini 138153892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 138253892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 138353892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 138453892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 138553892102SStefano Zampini } else { 13863a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 138753892102SStefano Zampini } 1388a64d13efSStefano Zampini 1389a64d13efSStefano Zampini /* print some info if requested */ 1390a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1391a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1392a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13931575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1394a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1395a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13963a50541eSStefano 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); 1397a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1398a64d13efSStefano Zampini } 1399a64d13efSStefano Zampini 1400a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14016816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14026816873aSStefano Zampini IS is_aux1,is_aux2; 14036816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14046816873aSStefano Zampini 14053a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1406854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1407854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1408a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14094641a718SStefano Zampini for (i=0; i<n_D; i++) { 14104641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14114641a718SStefano Zampini } 1412a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1413a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14144641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14154641a718SStefano Zampini aux_array1[j++] = i; 1416a64d13efSStefano Zampini } 1417a64d13efSStefano Zampini } 1418a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1419a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1420a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14214641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14224641a718SStefano Zampini aux_array2[j++] = i; 1423a64d13efSStefano Zampini } 1424a64d13efSStefano Zampini } 1425a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1426a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1427a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1428a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1429a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1430a64d13efSStefano Zampini 14318eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1432785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1433a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14344641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14354641a718SStefano Zampini aux_array1[j++] = i; 1436a64d13efSStefano Zampini } 1437a64d13efSStefano Zampini } 1438a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1439a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1440a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1441a64d13efSStefano Zampini } 14424641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14433a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1444d62866d3SStefano Zampini } else { 144553892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14466816873aSStefano Zampini IS tis; 14476816873aSStefano Zampini PetscInt schur_size; 14486816873aSStefano Zampini 144953892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14506816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 145153892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14526816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14536816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14546816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14556816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14566816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1457d62866d3SStefano Zampini } 1458d62866d3SStefano Zampini } 1459a64d13efSStefano Zampini PetscFunctionReturn(0); 1460a64d13efSStefano Zampini } 1461a64d13efSStefano Zampini 1462304d26faSStefano Zampini 1463304d26faSStefano Zampini #undef __FUNCT__ 1464304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1465684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1466304d26faSStefano Zampini { 1467304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1468304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1469304d26faSStefano Zampini PC pc_temp; 1470304d26faSStefano Zampini Mat A_RR; 1471f4ddd8eeSStefano Zampini MatReuse reuse; 1472304d26faSStefano Zampini PetscScalar m_one = -1.0; 1473304d26faSStefano Zampini PetscReal value; 147404708bb6SStefano Zampini PetscInt n_D,n_R; 14759577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1476304d26faSStefano Zampini PetscErrorCode ierr; 1477e604994aSStefano Zampini /* prefixes stuff */ 1478312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1479e604994aSStefano Zampini size_t len; 1480304d26faSStefano Zampini 1481304d26faSStefano Zampini PetscFunctionBegin; 1482304d26faSStefano Zampini 1483e604994aSStefano Zampini /* compute prefixes */ 1484e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1485e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1486e604994aSStefano Zampini if (!pcbddc->current_level) { 1487e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1488e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1489e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1490e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1491e604994aSStefano Zampini } else { 1492e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1493312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1494e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1495e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1496312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1497312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 149834d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 149934d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1500e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1501e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1502e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1503e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1504e604994aSStefano Zampini } 1505e604994aSStefano Zampini 1506304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1507684f6988SStefano Zampini if (dirichlet) { 1508d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15093301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15103301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1511964fefecSStefano Zampini } 1512ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1513964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1514304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1515304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1516304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1517304d26faSStefano Zampini /* default */ 1518304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15209577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1521304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15229577ea80SStefano Zampini if (issbaij) { 15239577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15249577ea80SStefano Zampini } else { 1525304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15269577ea80SStefano Zampini } 1527304d26faSStefano Zampini /* Allow user's customization */ 1528304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1529304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1530304d26faSStefano Zampini } 1531d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1532d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1533d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1534d62866d3SStefano Zampini 1535d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1536d5574798SStefano Zampini } 1537304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1538304d26faSStefano Zampini if (!n_D) { 1539304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1540304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1541304d26faSStefano Zampini } 1542304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1543304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1544304d26faSStefano Zampini /* set ksp_D into pcis data */ 1545304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1546304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1547304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1548684f6988SStefano Zampini } 1549304d26faSStefano Zampini 1550304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1551684f6988SStefano Zampini A_RR = 0; 1552684f6988SStefano Zampini if (neumann) { 1553d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 155404708bb6SStefano Zampini PetscInt ibs,mbs; 155504708bb6SStefano Zampini PetscBool issbaij; 155604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1557f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15588ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1559f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1560f4ddd8eeSStefano Zampini PetscInt nn_R; 156181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1562f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1563f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1564f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1565f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1566f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1567f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1568f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1569727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1570f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1571f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1572f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1573f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1574f4ddd8eeSStefano Zampini } 1575f4ddd8eeSStefano Zampini } 1576f4ddd8eeSStefano Zampini /* last check */ 1577d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1578f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1579f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1580f4ddd8eeSStefano Zampini } 1581f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1582f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1583f4ddd8eeSStefano Zampini } 1584f4ddd8eeSStefano Zampini /* extract A_RR */ 1585af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1586af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 158704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 158804708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 158904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 159004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 159104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1592af732b37SStefano Zampini } else { 1593511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 15946816873aSStefano Zampini } 159504708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 159604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 159704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 159804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 159904708bb6SStefano Zampini } else { 1600511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 160104708bb6SStefano Zampini } 160204708bb6SStefano Zampini } 160304708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1604f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16053301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16063301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16076816873aSStefano Zampini } 16086816873aSStefano Zampini } else { 16096816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16106816873aSStefano Zampini 16116816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16126816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16136816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1614af732b37SStefano Zampini } 1615f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1616304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1617304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1618304d26faSStefano Zampini /* default */ 1619304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1620e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1621304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16229577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16239577ea80SStefano Zampini if (issbaij) { 16249577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16259577ea80SStefano Zampini } else { 1626304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16279577ea80SStefano Zampini } 1628304d26faSStefano Zampini /* Allow user's customization */ 1629304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1630304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1631304d26faSStefano Zampini } 1632d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1633304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1634304d26faSStefano Zampini if (!n_R) { 1635304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1636304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1637304d26faSStefano Zampini } 1638d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1639d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1640d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1641d62866d3SStefano Zampini 1642d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1643d62866d3SStefano Zampini } 1644304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1645304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1646684f6988SStefano Zampini } 16476816873aSStefano Zampini /* free Neumann problem's matrix */ 16486816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1649304d26faSStefano Zampini 1650304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16510fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1652684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1653684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1655684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1656684f6988SStefano Zampini } 1657684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16580fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16590fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16600fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16610fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16620fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1663304d26faSStefano Zampini /* need to be adapted? */ 1664b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1665b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1666b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1667304d26faSStefano Zampini /* print info */ 1668304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1669e604994aSStefano 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); 1670304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1671304d26faSStefano Zampini } 1672b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1673298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1674304d26faSStefano Zampini } 1675684f6988SStefano Zampini } 1676684f6988SStefano Zampini if (neumann) { /* Neumann */ 16776816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16780fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16790fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16800fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16810fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16820fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1683304d26faSStefano Zampini /* need to be adapted? */ 1684b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1685b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1686304d26faSStefano Zampini /* print info */ 1687304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1688e604994aSStefano 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); 1689304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1690304d26faSStefano Zampini } 1691b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1692298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1693304d26faSStefano Zampini } 16940fccc4e9SStefano Zampini } 1695684f6988SStefano Zampini } 1696304d26faSStefano Zampini PetscFunctionReturn(0); 1697304d26faSStefano Zampini } 1698304d26faSStefano Zampini 1699304d26faSStefano Zampini #undef __FUNCT__ 1700ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 170180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1702674ae819SStefano Zampini { 1703674ae819SStefano Zampini PetscErrorCode ierr; 1704674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1705be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1706674ae819SStefano Zampini 1707674ae819SStefano Zampini PetscFunctionBegin; 1708be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 170980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 171020c7b377SStefano Zampini } 171180677318SStefano Zampini if (!pcbddc->switch_static) { 171280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 171380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 171480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 171520c7b377SStefano Zampini } 1716be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171920c7b377SStefano Zampini } else { 1720be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1721be83ff47SStefano Zampini 172253892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172353892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172420c7b377SStefano Zampini } 1725be83ff47SStefano Zampini } else { 172680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 173280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1735674ae819SStefano Zampini } 1736674ae819SStefano Zampini } 1737be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173880677318SStefano Zampini if (applytranspose) { 173980677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174080677318SStefano Zampini } else { 174180677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174280677318SStefano Zampini } 1743be83ff47SStefano Zampini } else { 1744be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1745be83ff47SStefano Zampini 1746be83ff47SStefano Zampini if (applytranspose) { 17475a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1748be83ff47SStefano Zampini } else { 17495a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1750be83ff47SStefano Zampini } 1751be83ff47SStefano Zampini } 175280677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 175380677318SStefano Zampini if (!pcbddc->switch_static) { 1754be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 175680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1757be83ff47SStefano Zampini } else { 1758be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1759be83ff47SStefano Zampini 176053892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 176153892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1762be83ff47SStefano Zampini } 176380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 176480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 176580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 176680677318SStefano Zampini } 176780677318SStefano Zampini } else { 176880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 177580677318SStefano Zampini } 177680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1780674ae819SStefano Zampini } 1781674ae819SStefano Zampini PetscFunctionReturn(0); 1782674ae819SStefano Zampini } 1783674ae819SStefano Zampini 1784dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1785674ae819SStefano Zampini #undef __FUNCT__ 1786674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1787dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1788674ae819SStefano Zampini { 1789674ae819SStefano Zampini PetscErrorCode ierr; 1790674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1791674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1792674ae819SStefano Zampini const PetscScalar zero = 0.0; 1793674ae819SStefano Zampini 1794674ae819SStefano Zampini PetscFunctionBegin; 1795dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1796dc359a40SStefano Zampini if (applytranspose) { 1797674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 17988eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1799dc359a40SStefano Zampini } else { 1800674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1801674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 180215aaf578SStefano Zampini } 180312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 180412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 180512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180712edc857SStefano Zampini 18089f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 180912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 181012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1811964fefecSStefano Zampini Vec rhs,sol; 1812964fefecSStefano Zampini 1813964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1814964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 181512edc857SStefano Zampini if (applytranspose) { 1816964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 181712edc857SStefano Zampini } else { 1818964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 181912edc857SStefano Zampini } 182012edc857SStefano Zampini } 1821674ae819SStefano Zampini 1822674ae819SStefano Zampini /* Local solution on R nodes */ 182380677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 182480677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18259f00e9b4SStefano Zampini } 1826674ae819SStefano Zampini 18279f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18289f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 182912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1830674ae819SStefano Zampini 1831674ae819SStefano Zampini /* Sum contributions from two levels */ 1832dc359a40SStefano Zampini if (applytranspose) { 1833dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1834dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1835dc359a40SStefano Zampini } else { 1836674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18378eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1838dc359a40SStefano Zampini } 1839674ae819SStefano Zampini PetscFunctionReturn(0); 1840674ae819SStefano Zampini } 1841674ae819SStefano Zampini 1842674ae819SStefano Zampini #undef __FUNCT__ 1843674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 184412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1845674ae819SStefano Zampini { 1846674ae819SStefano Zampini PetscErrorCode ierr; 1847674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 184858da7f69SStefano Zampini PetscScalar *array; 184912edc857SStefano Zampini Vec from,to; 1850674ae819SStefano Zampini 1851674ae819SStefano Zampini PetscFunctionBegin; 185212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 185312edc857SStefano Zampini from = pcbddc->coarse_vec; 185412edc857SStefano Zampini to = pcbddc->vec1_P; 185512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 185612edc857SStefano Zampini Vec tvec; 185758da7f69SStefano Zampini 185858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 185958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 186012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186158da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 186258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 186358da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 186412edc857SStefano Zampini } 186512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 186612edc857SStefano Zampini from = pcbddc->vec1_P; 186712edc857SStefano Zampini to = pcbddc->coarse_vec; 186812edc857SStefano Zampini } 186912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1870674ae819SStefano Zampini PetscFunctionReturn(0); 1871674ae819SStefano Zampini } 1872674ae819SStefano Zampini 1873674ae819SStefano Zampini #undef __FUNCT__ 1874674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 187512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1876674ae819SStefano Zampini { 1877674ae819SStefano Zampini PetscErrorCode ierr; 1878674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187958da7f69SStefano Zampini PetscScalar *array; 188012edc857SStefano Zampini Vec from,to; 1881674ae819SStefano Zampini 1882674ae819SStefano Zampini PetscFunctionBegin; 188312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 188412edc857SStefano Zampini from = pcbddc->coarse_vec; 188512edc857SStefano Zampini to = pcbddc->vec1_P; 188612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188712edc857SStefano Zampini from = pcbddc->vec1_P; 188812edc857SStefano Zampini to = pcbddc->coarse_vec; 188912edc857SStefano Zampini } 189012edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 189112edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 189212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 189312edc857SStefano Zampini Vec tvec; 189458da7f69SStefano Zampini 189512edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 189658da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 189758da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 189858da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 189958da7f69SStefano Zampini } 190058da7f69SStefano Zampini } else { 190158da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 190258da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 190312edc857SStefano Zampini } 190412edc857SStefano Zampini } 1905674ae819SStefano Zampini PetscFunctionReturn(0); 1906674ae819SStefano Zampini } 1907674ae819SStefano Zampini 1908984c4197SStefano Zampini /* uncomment for testing purposes */ 1909984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1910674ae819SStefano Zampini #undef __FUNCT__ 1911674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1912674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1913674ae819SStefano Zampini { 1914674ae819SStefano Zampini PetscErrorCode ierr; 1915674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1916674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1917674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1918984c4197SStefano Zampini /* one and zero */ 1919984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1920984c4197SStefano Zampini /* space to store constraints and their local indices */ 19219162d606SStefano Zampini PetscScalar *constraints_data; 19229162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19239162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19249162d606SStefano Zampini PetscInt *constraints_n; 1925984c4197SStefano Zampini /* iterators */ 1926b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1927984c4197SStefano Zampini /* BLAS integers */ 1928e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1929e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1930c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1931727cdba6SStefano Zampini /* reuse */ 19320e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19330e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1934984c4197SStefano Zampini /* change of basis */ 1935b3d85658SStefano Zampini PetscBool qr_needed; 19369162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1937984c4197SStefano Zampini /* auxiliary stuff */ 193864efe560SStefano Zampini PetscInt *nnz,*is_indices; 19398a0068c3SStefano Zampini PetscInt ncc; 1940984c4197SStefano Zampini /* some quantities */ 194145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1942a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1943984c4197SStefano Zampini 1944674ae819SStefano Zampini PetscFunctionBegin; 19458e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19468e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19478e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1948088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1949088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19500e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19510e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19520e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19530e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19540e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1955088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1956cf5a6209SStefano Zampini 1957cf5a6209SStefano Zampini /* print some info */ 1958cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1959cf5a6209SStefano Zampini IS vertices; 1960cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1961cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1962cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1963cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 19641575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1965cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1966cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1967fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1968fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1969cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 19701575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1971cf5a6209SStefano Zampini } 1972cf5a6209SStefano Zampini 1973cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19749162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1975cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1976cf5a6209SStefano Zampini const Vec *nearnullvecs; 1977cf5a6209SStefano Zampini Vec *localnearnullsp; 1978cf5a6209SStefano Zampini PetscScalar *array; 1979cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1980cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1981674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1982b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1983674ae819SStefano Zampini PetscScalar *work; 1984674ae819SStefano Zampini PetscReal *singular_vals; 1985674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1986674ae819SStefano Zampini PetscReal *rwork; 1987674ae819SStefano Zampini #endif 1988674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1989674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1990674ae819SStefano Zampini #else 1991964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1992964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1993674ae819SStefano Zampini #endif 1994674ae819SStefano Zampini 1995674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1996d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1997d06fc5fdSStefano Zampini /* free unneeded index sets */ 1998d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1999d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2000674ae819SStefano Zampini } 2001d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2002d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2003d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2004d06fc5fdSStefano Zampini } 2005d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2006d06fc5fdSStefano Zampini n_ISForEdges = 0; 2007d06fc5fdSStefano Zampini } 2008d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2009d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2010d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2011d06fc5fdSStefano Zampini } 2012d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2013d06fc5fdSStefano Zampini n_ISForFaces = 0; 2014d06fc5fdSStefano Zampini } 201570022509SStefano Zampini 201670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 201770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 201870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 201970022509SStefano Zampini if (pcbddc->NullSpace) { 202070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 202170022509SStefano Zampini 202270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2023b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2024d06fc5fdSStefano Zampini if (!ISForEdges) { 2025d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2026d06fc5fdSStefano Zampini } 2027b8ffe317SStefano Zampini } 2028d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2029d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2030b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2031d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2032d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 203398a51de6SStefano Zampini } 203470022509SStefano Zampini #endif 203508122e43SStefano Zampini 2036674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2037674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2038674ae819SStefano Zampini if (nearnullsp) { 2039674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2040f4ddd8eeSStefano Zampini /* remove any stored info */ 2041f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2042f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2043f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2044f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2045f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2046473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2047f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2048f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2049f4ddd8eeSStefano Zampini } 2050984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2051984c4197SStefano Zampini nnsp_size = 0; 2052674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2053674ae819SStefano Zampini } 2054984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2055984c4197SStefano Zampini max_constraints = nnsp_size; 2056984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2057984c4197SStefano Zampini 2058674ae819SStefano Zampini /* 2059674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20609162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20619162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20629162d606SStefano Zampini There can be multiple constraints per connected component 2063674ae819SStefano Zampini */ 2064674ae819SStefano Zampini n_vertices = 0; 2065674ae819SStefano Zampini if (ISForVertices) { 2066674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2067674ae819SStefano Zampini } 20689162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20699162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20709162d606SStefano Zampini 20719162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20729162d606SStefano Zampini total_counts *= max_constraints; 2073674ae819SStefano Zampini total_counts += n_vertices; 20744641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20759162d606SStefano Zampini 2076674ae819SStefano Zampini total_counts = 0; 2077674ae819SStefano Zampini max_size_of_constraint = 0; 2078674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20799162d606SStefano Zampini IS used_is; 2080674ae819SStefano Zampini if (i<n_ISForEdges) { 20819162d606SStefano Zampini used_is = ISForEdges[i]; 2082674ae819SStefano Zampini } else { 20839162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2084674ae819SStefano Zampini } 20859162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2086674ae819SStefano Zampini total_counts += j; 2087674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2088674ae819SStefano Zampini } 20899162d606SStefano 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); 20909162d606SStefano Zampini 2091984c4197SStefano Zampini /* get local part of global near null space vectors */ 2092785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2093984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2094984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2095e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2096e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2097984c4197SStefano Zampini } 2098674ae819SStefano Zampini 2099242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2100242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2101a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2102242a89d7SStefano Zampini 2103984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2104a773dcb8SStefano Zampini if (!skip_lapack) { 2105674ae819SStefano Zampini PetscScalar temp_work; 2106911cabfeSStefano Zampini 2107674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2108984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2109785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2110785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2111785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2112674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2113785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2114674ae819SStefano Zampini #endif 2115674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2116c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2117c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2118674ae819SStefano Zampini lwork = -1; 2119674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2120674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2121c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2122674ae819SStefano Zampini #else 2123c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2124674ae819SStefano Zampini #endif 2125674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2126984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2127674ae819SStefano Zampini #else /* on missing GESVD */ 2128674ae819SStefano Zampini /* SVD */ 2129674ae819SStefano Zampini PetscInt max_n,min_n; 2130674ae819SStefano Zampini max_n = max_size_of_constraint; 2131984c4197SStefano Zampini min_n = max_constraints; 2132984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2133674ae819SStefano Zampini min_n = max_size_of_constraint; 2134984c4197SStefano Zampini max_n = max_constraints; 2135674ae819SStefano Zampini } 2136785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2137674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2138785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2139674ae819SStefano Zampini #endif 2140674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2141674ae819SStefano Zampini lwork = -1; 2142e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2143e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2144b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2145674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2146674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21479162d606SStefano 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)); 2148674ae819SStefano Zampini #else 21499162d606SStefano 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)); 2150674ae819SStefano Zampini #endif 2151674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2152984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2153984c4197SStefano Zampini #endif /* on missing GESVD */ 2154674ae819SStefano Zampini /* Allocate optimal workspace */ 2155674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2156854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2157674ae819SStefano Zampini } 2158674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2159674ae819SStefano Zampini total_counts = 0; 21609162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21619162d606SStefano Zampini constraints_data_ptr[0] = 0; 2162674ae819SStefano Zampini /* vertices */ 21639162d606SStefano Zampini if (n_vertices) { 2164674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21659162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21669162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2167674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21689162d606SStefano Zampini constraints_n[total_counts] = 1; 21699162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21709162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21719162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2172674ae819SStefano Zampini total_counts++; 2173674ae819SStefano Zampini } 2174674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2175984c4197SStefano Zampini PetscBool used_vertex; 2176674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2177674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2178674ae819SStefano Zampini k = 0; 2179674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2180984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2181984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21829162d606SStefano Zampini constraints_n[total_counts] = 1; 21839162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21849162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21859162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21869162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2187674ae819SStefano Zampini total_counts++; 2188674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2189674ae819SStefano Zampini } 2190984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2191674ae819SStefano Zampini k++; 2192674ae819SStefano Zampini } 2193674ae819SStefano Zampini } 2194674ae819SStefano Zampini } 2195674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2196674ae819SStefano Zampini n_vertices = total_counts; 2197674ae819SStefano Zampini } 2198984c4197SStefano Zampini 2199674ae819SStefano Zampini /* edges and faces */ 22009162d606SStefano Zampini total_counts_cc = total_counts; 2201911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22029162d606SStefano Zampini IS used_is; 22039162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22049162d606SStefano Zampini 2205911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22069162d606SStefano Zampini used_is = ISForEdges[ncc]; 2207984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2208674ae819SStefano Zampini } else { 22099162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2210984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2211674ae819SStefano Zampini } 2212674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22139162d606SStefano Zampini 22149162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22159162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2216984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2217984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2218674ae819SStefano Zampini if (nnsp_has_cnst) { 22195b08dc53SStefano Zampini PetscScalar quad_value; 22209162d606SStefano Zampini 22219162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22229162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22239162d606SStefano Zampini 2224a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2225674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2226a773dcb8SStefano Zampini } else { 2227a773dcb8SStefano Zampini quad_value = 1.0; 2228a773dcb8SStefano Zampini } 2229674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22309162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2231674ae819SStefano Zampini } 22329162d606SStefano Zampini temp_constraints++; 2233674ae819SStefano Zampini total_counts++; 2234674ae819SStefano Zampini } 2235674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2236984c4197SStefano Zampini PetscReal real_value; 22379162d606SStefano Zampini PetscScalar *ptr_to_data; 22389162d606SStefano Zampini 2239984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22409162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2241674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22429162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2243674ae819SStefano Zampini } 2244984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2245984c4197SStefano Zampini /* check if array is null on the connected component */ 2246e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22479162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22485b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2249674ae819SStefano Zampini temp_constraints++; 2250674ae819SStefano Zampini total_counts++; 22519162d606SStefano Zampini if (!idxs_copied) { 22529162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22539162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2254674ae819SStefano Zampini } 2255674ae819SStefano Zampini } 22569162d606SStefano Zampini } 22579162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 225845a1bb75SStefano Zampini valid_constraints = temp_constraints; 2259eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2260a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22619162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22629162d606SStefano Zampini 22639162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2264a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22659162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2266a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22679162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2268a773dcb8SStefano Zampini } else { /* perform SVD */ 2269984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22709162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2271674ae819SStefano Zampini 2272674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2273984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2274984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2275984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2276984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2277984c4197SStefano Zampini from that computed using LAPACKgesvd 2278984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2279984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2280984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2281674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2282e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2283984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2284674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2285674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22869162d606SStefano 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)); 2287674ae819SStefano Zampini } 2288674ae819SStefano Zampini } 2289e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2290e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2291e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2292674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2293c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2294674ae819SStefano Zampini #else 2295c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2296674ae819SStefano Zampini #endif 2297674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2298984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2299984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2300674ae819SStefano Zampini j = 0; 2301984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2302674ae819SStefano Zampini total_counts = total_counts-j; 230345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2304e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2305c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2306c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2307c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2308c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2309c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2310c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2311674ae819SStefano Zampini if (j<temp_constraints) { 2312984c4197SStefano Zampini PetscInt ii; 2313984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2314674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23159162d606SStefano 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)); 2316674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2317984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2318674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23199162d606SStefano 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]; 2320674ae819SStefano Zampini } 2321674ae819SStefano Zampini } 2322674ae819SStefano Zampini } 2323674ae819SStefano Zampini #else /* on missing GESVD */ 2324e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2325e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2326b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2327674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2328674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23299162d606SStefano 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)); 2330674ae819SStefano Zampini #else 23319162d606SStefano 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)); 2332674ae819SStefano Zampini #endif 2333984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2334674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2335984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2336e310c8b4SStefano Zampini k = temp_constraints; 2337e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2338674ae819SStefano Zampini j = 0; 2339e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 234045a1bb75SStefano Zampini valid_constraints = k-j; 2341911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2342984c4197SStefano Zampini #endif /* on missing GESVD */ 2343674ae819SStefano Zampini } 2344a773dcb8SStefano Zampini } 23459162d606SStefano Zampini /* update pointers information */ 23469162d606SStefano Zampini if (valid_constraints) { 23479162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23489162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23499162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23509162d606SStefano Zampini /* set change_of_basis flag */ 235145a1bb75SStefano Zampini if (boolforchange) { 2352b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23539162d606SStefano Zampini } 2354b3d85658SStefano Zampini total_counts_cc++; 235545a1bb75SStefano Zampini } 235645a1bb75SStefano Zampini } 2357984c4197SStefano Zampini /* free workspace */ 23588f1c130eSStefano Zampini if (!skip_lapack) { 2359984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2360984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2361984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2362984c4197SStefano Zampini #endif 2363984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2364984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2365984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2366984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2367984c4197SStefano Zampini #endif 2368984c4197SStefano Zampini } 2369984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2370984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2371984c4197SStefano Zampini } 2372984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2373cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2374cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2375cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2376cf5a6209SStefano Zampini } 2377cf5a6209SStefano Zampini if (n_ISForFaces) { 2378cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2379cf5a6209SStefano Zampini } 2380cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2381cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2382cf5a6209SStefano Zampini } 2383cf5a6209SStefano Zampini if (n_ISForEdges) { 2384cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2385cf5a6209SStefano Zampini } 2386cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 238708122e43SStefano Zampini } else { 238808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2389984c4197SStefano Zampini 239008122e43SStefano Zampini total_counts = 0; 239108122e43SStefano Zampini n_vertices = 0; 2392d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2393d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 239408122e43SStefano Zampini } 239508122e43SStefano Zampini max_constraints = 0; 23969162d606SStefano Zampini total_counts_cc = 0; 239708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 239808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 23999162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 240008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 240108122e43SStefano Zampini } 24029162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24039162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24049162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24059162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 240674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24079162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24089162d606SStefano Zampini total_counts_cc = 0; 24099162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24109162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24119162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 241208122e43SStefano Zampini } 241308122e43SStefano Zampini } 24149162d606SStefano Zampini #if 0 24159162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24169162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24179162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24189162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24199162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24209162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24219162d606SStefano Zampini } 24229162d606SStefano Zampini printf("\n"); 24239162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24249162d606SStefano Zampini } 24251b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24268bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24271b968477SStefano Zampini } 24281b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24298bec7fa6SStefano 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]); 24301b968477SStefano Zampini } 243108122e43SStefano Zampini #endif 243208122e43SStefano Zampini 24338bec7fa6SStefano Zampini max_size_of_constraint = 0; 24349162d606SStefano 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]); 24359162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 243608122e43SStefano Zampini /* Change of basis */ 2437b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 243808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 243908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 244008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2441b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 244208122e43SStefano Zampini } 244308122e43SStefano Zampini } 244408122e43SStefano Zampini } 244508122e43SStefano Zampini } 2446984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24470e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 244808122e43SStefano Zampini 24499162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24509162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24516c4ed002SBarry 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); 2452674ae819SStefano Zampini 2453674ae819SStefano Zampini /* Create constraint matrix */ 2454674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 245516f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2456984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2457984c4197SStefano Zampini 2458984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2459a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2460a717540cSStefano Zampini qr_needed = PETSC_FALSE; 246174d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2462984c4197SStefano Zampini total_primal_vertices=0; 2463b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24649162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24659162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2466984c4197SStefano Zampini if (size_of_constraint == 1) { 24679162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2468b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 246964efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24709162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24719162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2472a717540cSStefano Zampini } 2473b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 247474d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2475a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2476a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2477a717540cSStefano Zampini } 2478fa434743SStefano Zampini } else { 2479b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2480fa434743SStefano Zampini } 2481a717540cSStefano Zampini } 2482b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2483b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2484674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 248570022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2486b3d85658SStefano Zampini 2487b3d85658SStefano 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); 24880e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24890e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2490984c4197SStefano Zampini 2491984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 249274d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2493785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2494984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 249574d5cdf7SStefano Zampini 2496984c4197SStefano Zampini j = total_primal_vertices; 249774d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2498b3d85658SStefano Zampini cum = total_primal_vertices; 24999162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25004641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2501b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2502b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2503b3d85658SStefano Zampini cum++; 25049162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 250574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 250674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 250774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 250874d5cdf7SStefano Zampini } 25099162d606SStefano Zampini j += constraints_n[i]; 2510674ae819SStefano Zampini } 2511674ae819SStefano Zampini } 2512674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2513674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2514088faed8SStefano Zampini 2515674ae819SStefano Zampini /* set values in constraint matrix */ 2516984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25170e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2518674ae819SStefano Zampini } 2519984c4197SStefano Zampini total_counts = total_primal_vertices; 25209162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25214641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25229162d606SStefano Zampini PetscInt *cols; 25239162d606SStefano Zampini 25249162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25259162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25269162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25279162d606SStefano Zampini PetscInt row = total_counts+k; 25289162d606SStefano Zampini PetscScalar *vals; 25299162d606SStefano Zampini 25309162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25319162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25329162d606SStefano Zampini } 25339162d606SStefano Zampini total_counts += constraints_n[i]; 2534674ae819SStefano Zampini } 2535674ae819SStefano Zampini } 2536674ae819SStefano Zampini /* assembling */ 2537674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2538674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2539088faed8SStefano Zampini 2540984c4197SStefano Zampini /* 25416a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2542984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2543f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 2544984c4197SStefano Zampini */ 2545674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2546674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2547026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2548984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2549984c4197SStefano Zampini /* working stuff for GEQRF */ 255081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2551984c4197SStefano Zampini PetscBLASInt lqr_work; 2552984c4197SStefano Zampini /* working stuff for UNGQR */ 2553984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2554984c4197SStefano Zampini PetscBLASInt lgqr_work; 2555984c4197SStefano Zampini /* working stuff for TRTRS */ 2556984c4197SStefano Zampini PetscScalar *trs_rhs; 25573f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2558984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2559984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2560984c4197SStefano Zampini PetscScalar *start_vals; 2561984c4197SStefano Zampini /* working stuff for values insertion */ 25624641a718SStefano Zampini PetscBT is_primal; 256364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2564906d46d4SStefano Zampini /* matrix sizes */ 2565906d46d4SStefano Zampini PetscInt global_size,local_size; 2566906d46d4SStefano Zampini /* temporary change of basis */ 2567906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2568cf5a6209SStefano Zampini /* extra space for debugging */ 2569cf5a6209SStefano Zampini PetscScalar *dbg_work; 2570984c4197SStefano Zampini 2571906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2572906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 257316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2574bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2575906d46d4SStefano Zampini /* nonzeros for local mat */ 2576bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2577bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25789162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2579a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25809162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2581a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25829162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2583a717540cSStefano Zampini } else { 25849162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25859162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2586a717540cSStefano Zampini } 2587a717540cSStefano Zampini } 2588a717540cSStefano Zampini } 2589906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2590bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2591a717540cSStefano Zampini /* Set initial identity in the matrix */ 2592bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2593906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2594a717540cSStefano Zampini } 2595a717540cSStefano Zampini 2596a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2597a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2598a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2599a717540cSStefano Zampini } 2600a717540cSStefano Zampini 2601a717540cSStefano Zampini 2602a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2603a717540cSStefano Zampini /* 2604a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2605a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2606a717540cSStefano Zampini 2607a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2608a717540cSStefano Zampini 2609a6b551f4SStefano 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) 2610a6b551f4SStefano Zampini 2611a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2612a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2613a717540cSStefano Zampini | ... | 2614a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2615a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2616a717540cSStefano Zampini 2617a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2618a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2619a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2620a6b551f4SStefano Zampini 2621a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2622a717540cSStefano Zampini */ 2623a717540cSStefano Zampini if (qr_needed) { 2624984c4197SStefano Zampini /* space to store Q */ 2625854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2626984c4197SStefano Zampini /* first we issue queries for optimal work */ 26273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2630984c4197SStefano Zampini lqr_work = -1; 26313f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2632984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2633984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2634785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2635984c4197SStefano Zampini lgqr_work = -1; 26363f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26373f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26383f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26393f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26403f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26413f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2642984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2643984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2644785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2645984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2646785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2647984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2648785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2649a717540cSStefano Zampini /* allocating workspace for check */ 2650a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2651cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2652a717540cSStefano Zampini } 2653a717540cSStefano Zampini } 2654984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26554641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2656473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26570e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 26586c4ed002SBarry 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); 265939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 266039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 266139e2fb2aSStefano Zampini } 266239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2663984c4197SStefano Zampini 2664a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26659162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26669162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26674641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2668984c4197SStefano Zampini /* get constraint info */ 26699162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2670984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2671984c4197SStefano Zampini 2672984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26739162d606SStefano 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); 2674674ae819SStefano Zampini } 2675984c4197SStefano Zampini 2676fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2677a717540cSStefano Zampini 2678a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2679a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26809162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2681a717540cSStefano Zampini } 2682984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26839162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2684984c4197SStefano Zampini 2685984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26863f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26873f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26883f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2689674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26903f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2691984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2692674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2693984c4197SStefano Zampini 2694984c4197SStefano Zampini /* explictly compute R^-T */ 2695984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2696984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 26973f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 26993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27003f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2701984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27023f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2703984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2704984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2705984c4197SStefano Zampini 2706a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27073f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2711984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27123f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2713984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2714984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2715984c4197SStefano Zampini 2716984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2717984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2718984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27193f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27203f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2725984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27269162d606SStefano 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)); 2727984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27289162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2729984c4197SStefano Zampini 2730984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27319162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2732984c4197SStefano Zampini /* insert cols for primal dofs */ 2733984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2734984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27359162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2736906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2737984c4197SStefano Zampini } 2738984c4197SStefano Zampini /* insert cols for dual dofs */ 2739984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27409162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2741984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27429162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2743906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2744984c4197SStefano Zampini j++; 2745674ae819SStefano Zampini } 2746674ae819SStefano Zampini } 2747984c4197SStefano Zampini 2748984c4197SStefano Zampini /* check change of basis */ 2749984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2750984c4197SStefano Zampini PetscInt ii,jj; 2751984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2752c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2753c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2754c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2755c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2756c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2757c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2758984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2759cf5a6209SStefano 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)); 2760984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2761984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2762984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2763cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2764cf5a6209SStefano 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; 2765674ae819SStefano Zampini } 2766674ae819SStefano Zampini } 2767984c4197SStefano Zampini if (!valid_qr) { 276822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2769984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2770984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2771cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2772cf5a6209SStefano 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])); 2773674ae819SStefano Zampini } 2774cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2775cf5a6209SStefano 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])); 2776984c4197SStefano Zampini } 2777984c4197SStefano Zampini } 2778984c4197SStefano Zampini } 2779674ae819SStefano Zampini } else { 278022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2781674ae819SStefano Zampini } 2782674ae819SStefano Zampini } 2783a717540cSStefano Zampini } else { /* simple transformation block */ 2784a717540cSStefano Zampini PetscInt row,col; 2785a6b551f4SStefano Zampini PetscScalar val,norm; 2786a6b551f4SStefano Zampini 2787a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27889162d606SStefano 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)); 2789a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27909162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27919162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2792bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 27939162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2794906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 27959162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2796a717540cSStefano Zampini } else { 2797a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 27989162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2799a717540cSStefano Zampini if (row != col) { 28009162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2801a717540cSStefano Zampini } else { 28029162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2803a717540cSStefano Zampini } 2804906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2805a717540cSStefano Zampini } 2806a717540cSStefano Zampini } 2807a717540cSStefano Zampini } 280898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 280922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2810a717540cSStefano Zampini } 2811674ae819SStefano Zampini } 2812984c4197SStefano Zampini } else { 2813984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28149162d606SStefano 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); 2815674ae819SStefano Zampini } 2816674ae819SStefano Zampini } 2817674ae819SStefano Zampini } 2818a717540cSStefano Zampini 2819a717540cSStefano Zampini /* free workspace */ 2820a717540cSStefano Zampini if (qr_needed) { 2821984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2822cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2823984c4197SStefano Zampini } 2824984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2825984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2826984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2827984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2828984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2829674ae819SStefano Zampini } 2830a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2831906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2832906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2833906d46d4SStefano Zampini 2834906d46d4SStefano Zampini /* assembling of global change of variable */ 2835bbb9e6c6SStefano Zampini { 2836bbb9e6c6SStefano Zampini Mat tmat; 283716f15bc4SStefano Zampini PetscInt bs; 283816f15bc4SStefano Zampini 2839906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2840906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2841bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2842bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2843bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2844bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 284516f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 284616f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2847906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2848bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2850bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2851bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2852bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2853e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2854e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2855bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2857906d46d4SStefano Zampini } 2858906d46d4SStefano Zampini /* check */ 2859906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2860906d46d4SStefano Zampini PetscReal error; 2861906d46d4SStefano Zampini Vec x,x_change; 2862906d46d4SStefano Zampini 2863906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2864906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2865906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2866906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2867e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2868e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2870e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2871e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2874906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2875906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2877906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2878906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2879906d46d4SStefano Zampini } 2880b96c3477SStefano Zampini 2881b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2882b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2883b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2884b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2885ac632422SStefano Zampini Mat S_new,tmat; 2886bbb9e6c6SStefano Zampini IS is_all_N; 2887bbb9e6c6SStefano Zampini 2888bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28896816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2890bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2891ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2892b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2893ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2894ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2895ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2896ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2897ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2898b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2899ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2900ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2901ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2902ac632422SStefano Zampini } 2903b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2904b96c3477SStefano Zampini } 2905b96c3477SStefano Zampini } 2906906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2907906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2908b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2909b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2910b9b85e73SStefano Zampini } 2911906d46d4SStefano Zampini 2912906d46d4SStefano Zampini /* set up change of basis context */ 2913906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2914906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2915906d46d4SStefano Zampini 2916906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2917906d46d4SStefano Zampini PetscInt global_size,local_size; 2918906d46d4SStefano Zampini 2919906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2920906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2921906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2922906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2923906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2924906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2925906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2926906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2927906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2928906d46d4SStefano Zampini } else { 2929906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2930906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2931906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2932906d46d4SStefano Zampini } 2933906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2934906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2935906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2936906d46d4SStefano Zampini } else { 2937906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2938906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2939906d46d4SStefano Zampini } 2940906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2941906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2942906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2943906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2944b9b85e73SStefano Zampini } 2945a717540cSStefano Zampini 2946727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2947727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2948727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2949aff50787SStefano 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); 2950c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29510e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2952aff50787SStefano 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); 2953727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2954727cdba6SStefano Zampini } 29550e6343abSStefano Zampini } 29560e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2957727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2958b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2959727cdba6SStefano Zampini 2960a717540cSStefano Zampini /* flush dbg viewer */ 2961b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2962b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2963b8ffe317SStefano Zampini } 2964a717540cSStefano Zampini 2965e310c8b4SStefano Zampini /* free workspace */ 2966a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29674641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 296808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29699162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29709162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 297108122e43SStefano Zampini } else { 29729162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 297508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 297608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29779162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29789162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 297908122e43SStefano Zampini } 2980674ae819SStefano Zampini PetscFunctionReturn(0); 2981674ae819SStefano Zampini } 2982674ae819SStefano Zampini 2983674ae819SStefano Zampini #undef __FUNCT__ 2984674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2985674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2986674ae819SStefano Zampini { 2987674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2988674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2989674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29907fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2991674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2992674ae819SStefano Zampini 2993674ae819SStefano Zampini PetscFunctionBegin; 29948e61c736SStefano Zampini /* Reset previously computed graph */ 29958e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2996674ae819SStefano Zampini /* Init local Graph struct */ 29977fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 29983bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 2999674ae819SStefano Zampini 3000575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3001575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3002575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3003575ad6abSStefano Zampini } 30049577ea80SStefano Zampini 3005674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30064d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30074d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30084d379d7bSStefano Zampini PetscInt nvtxs; 3009e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3010674ae819SStefano Zampini 30114d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30122fffb893SStefano Zampini 30132fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30142fffb893SStefano Zampini if (flg_row) { 30154d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3016b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30172fffb893SStefano Zampini } 30182fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30199b28b941SStefano 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 */ 30204d379d7bSStefano Zampini IS is_dummy; 30214d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30224d379d7bSStefano Zampini PetscInt j,sum; 30234d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30244d379d7bSStefano Zampini const PetscInt *idxs; 30254d379d7bSStefano Zampini PCBDDCGraph graph; 30264d379d7bSStefano Zampini PetscBT is_on_boundary; 30274d379d7bSStefano Zampini 30284d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30294d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30304d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30314d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30327fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30334d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3034e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3035e496cd5dSStefano Zampini if (flg_row) { 30364d379d7bSStefano Zampini graph->xadj = xadj; 30374d379d7bSStefano Zampini graph->adjncy = adjncy; 3038e496cd5dSStefano Zampini } 30394d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30404d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3041e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30424d379d7bSStefano Zampini 30434d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30449b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30454d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30464d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30474d379d7bSStefano Zampini } 30484d379d7bSStefano Zampini } 30494d379d7bSStefano Zampini 3050e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30514d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30524d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30534d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30544d379d7bSStefano Zampini } 30554d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30564d379d7bSStefano Zampini 3057e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30584d379d7bSStefano Zampini sum = 0; 30594d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30604d379d7bSStefano Zampini PetscInt sizecc = 0; 30614d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30624d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30634d379d7bSStefano Zampini sizecc++; 30644d379d7bSStefano Zampini } 30654d379d7bSStefano Zampini } 30664d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30674d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30684d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30694d379d7bSStefano Zampini } 30704d379d7bSStefano Zampini } 30714d379d7bSStefano Zampini sum += sizecc*sizecc; 30724d379d7bSStefano Zampini } 30734d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30744d379d7bSStefano Zampini sum = 0; 3075e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30764d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30774d379d7bSStefano Zampini cxadj[i] = sum; 30784d379d7bSStefano Zampini sum += temp; 30794d379d7bSStefano Zampini } 3080e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30814d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30824d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30834d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30844d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30854d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30864d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30874d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30884d379d7bSStefano Zampini sizecc++; 30894d379d7bSStefano Zampini } 30904d379d7bSStefano Zampini } 30914d379d7bSStefano Zampini } 30924d379d7bSStefano Zampini } 30934d379d7bSStefano Zampini } 30949b28b941SStefano Zampini if (sum) { 3095e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 30964d379d7bSStefano Zampini } else { 30974d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 30984d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 30994d379d7bSStefano Zampini } 31004d379d7bSStefano Zampini graph->xadj = 0; 31014d379d7bSStefano Zampini graph->adjncy = 0; 31024d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31034d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31044d379d7bSStefano Zampini } 3105674ae819SStefano Zampini } 31069b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31079b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3108674ae819SStefano Zampini } 3109674ae819SStefano Zampini 311063602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3111674ae819SStefano Zampini vertex_size = 1; 311263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 311363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 311495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 311563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3116e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 311763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3118674ae819SStefano Zampini } 311963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 312063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 312163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3122674ae819SStefano Zampini } 312363602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3124674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 312563602bcaSStefano Zampini } else { 312663602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 312763602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3128854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 313063602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313163602bcaSStefano Zampini } 313263602bcaSStefano Zampini } 3133674ae819SStefano Zampini } 3134674ae819SStefano Zampini 3135674ae819SStefano Zampini /* Setup of Graph */ 3136785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3137e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3138785d1243SStefano Zampini } 3139785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3140e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3141785d1243SStefano Zampini } 3142302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3143674ae819SStefano Zampini 3144674ae819SStefano Zampini /* Graph's connected components analysis */ 3145674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3146674ae819SStefano Zampini 3147674ae819SStefano Zampini /* print some info to stdout */ 3148674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3149302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3150674ae819SStefano Zampini } 3151fb180af4SStefano Zampini 3152fb180af4SStefano Zampini /* mark topography has done */ 3153fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3154674ae819SStefano Zampini PetscFunctionReturn(0); 3155674ae819SStefano Zampini } 3156674ae819SStefano Zampini 3157dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3158674ae819SStefano Zampini #undef __FUNCT__ 3159674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3160dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3161674ae819SStefano Zampini { 3162dc456d91SStefano Zampini PetscSF sf; 3163dc456d91SStefano Zampini PetscLayout map; 3164dc456d91SStefano Zampini const PetscInt *idxs; 3165dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3166dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3167dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3168dc456d91SStefano Zampini PetscMPIInt commsize; 3169674ae819SStefano Zampini PetscBool first_found; 3170674ae819SStefano Zampini PetscErrorCode ierr; 3171674ae819SStefano Zampini 3172674ae819SStefano Zampini PetscFunctionBegin; 3173dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3174dc456d91SStefano Zampini if (subset_mult) { 3175dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3176dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3177dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3178674ae819SStefano Zampini } 3179dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3180dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3181dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3182dc456d91SStefano Zampini for (i=0;i<n;i++) { 3183dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3184dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3185674ae819SStefano Zampini } 3186dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3187b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3188dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3189dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3190dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3191dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3192dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3193dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3194dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3195dc456d91SStefano Zampini 3196dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3197dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3198dc456d91SStefano Zampini if (subset_mult) { 3199dc456d91SStefano Zampini const PetscInt* idxs_mult; 3200dc456d91SStefano Zampini 3201dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3202dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3203dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3204674ae819SStefano Zampini } else { 3205dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3206674ae819SStefano Zampini } 3207dc456d91SStefano Zampini /* local size of new subset */ 3208dc456d91SStefano Zampini n_n = 0; 3209dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3210dc456d91SStefano Zampini 3211dc456d91SStefano Zampini /* global indexes in layout */ 3212dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3213dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3214dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3215dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3216dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3217dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3218dc456d91SStefano Zampini 3219dc456d91SStefano Zampini /* reduce from leaves to roots */ 3220dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 322164a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 322264a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3223dc456d91SStefano Zampini 3224dc456d91SStefano Zampini /* count indexes in local part of layout */ 3225674ae819SStefano Zampini nlocals = 0; 3226674ae819SStefano Zampini first_index = -1; 3227674ae819SStefano Zampini first_found = PETSC_FALSE; 3228dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3229dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3230674ae819SStefano Zampini first_found = PETSC_TRUE; 3231674ae819SStefano Zampini first_index = i; 3232674ae819SStefano Zampini } 3233dc456d91SStefano Zampini nlocals += root_data[i]; 3234674ae819SStefano Zampini } 3235dc456d91SStefano Zampini 3236dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32375fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3238dc456d91SStefano Zampini start = 0; 323964a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32405fa240b1SStefano Zampini #else 324164a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32425fa240b1SStefano Zampini start = start-nlocals; 32435fa240b1SStefano Zampini #endif 32445fa240b1SStefano Zampini 3245dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3246dc456d91SStefano Zampini *N_n = start + nlocals; 3247dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3248dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3249674ae819SStefano Zampini } 32505fa240b1SStefano Zampini 32515fa240b1SStefano Zampini /* adapt root data with cumulative */ 3252674ae819SStefano Zampini if (first_found) { 3253dc456d91SStefano Zampini PetscInt old_index; 3254dc456d91SStefano Zampini 3255dc456d91SStefano Zampini root_data[first_index] += start; 3256674ae819SStefano Zampini old_index = first_index; 3257dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3258dc456d91SStefano Zampini if (root_data[i]) { 3259dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3260674ae819SStefano Zampini old_index = i; 3261674ae819SStefano Zampini } 3262674ae819SStefano Zampini } 3263674ae819SStefano Zampini } 3264dc456d91SStefano Zampini 3265dc456d91SStefano Zampini /* from roots to leaves */ 3266dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3267dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3268dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3269dc456d91SStefano Zampini 3270dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3271dc456d91SStefano Zampini if (subset_mult) { 3272dc456d91SStefano Zampini const PetscInt* idxs_mult; 3273dc456d91SStefano Zampini PetscInt cum; 3274dc456d91SStefano Zampini 3275dc456d91SStefano Zampini cum = 0; 3276dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3277dc456d91SStefano Zampini for (i=0;i<n;i++) { 3278dc456d91SStefano Zampini PetscInt j; 3279dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3280674ae819SStefano Zampini } 3281dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3282674ae819SStefano Zampini } else { 3283dc456d91SStefano Zampini for (i=0;i<n;i++) { 3284dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3285674ae819SStefano Zampini } 3286674ae819SStefano Zampini } 3287dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3288dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3289674ae819SStefano Zampini PetscFunctionReturn(0); 3290674ae819SStefano Zampini } 32919a7d3425SStefano Zampini 32929a7d3425SStefano Zampini #undef __FUNCT__ 32939a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32949a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32959a7d3425SStefano Zampini { 32969a7d3425SStefano Zampini PetscInt i,j; 32979a7d3425SStefano Zampini PetscScalar *alphas; 32989a7d3425SStefano Zampini PetscErrorCode ierr; 32999a7d3425SStefano Zampini 33009a7d3425SStefano Zampini PetscFunctionBegin; 33019a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3302785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33039a7d3425SStefano Zampini for (i=0;i<n;i++) { 33049a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33059a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33069a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33079a7d3425SStefano Zampini } 33089a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33099a7d3425SStefano Zampini PetscFunctionReturn(0); 33109a7d3425SStefano Zampini } 33119a7d3425SStefano Zampini 3312e7931f94SStefano Zampini #undef __FUNCT__ 331370cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3314b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3315e7931f94SStefano Zampini { 331652e5ac9dSStefano Zampini IS ranks_send_to; 3317e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3318e7931f94SStefano Zampini PetscMPIInt size,rank,color; 331952e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 332052e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33213837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33222b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3323e7931f94SStefano Zampini PetscSubcomm subcomm; 332452e5ac9dSStefano Zampini PetscErrorCode ierr; 3325a57a6d2fSStefano Zampini 3326e7931f94SStefano Zampini PetscFunctionBegin; 3327c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 3328c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 3329c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3330e7931f94SStefano Zampini 3331e7931f94SStefano Zampini /* Get info on mapping */ 33323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33333bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3334e7931f94SStefano Zampini 3335e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3336785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3337e7931f94SStefano Zampini xadj[0] = 0; 3338e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3339785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3340785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3341e7931f94SStefano Zampini 33422b510759SStefano Zampini if (threshold) { 3343d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33442b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3345d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3346d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3347d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3348d023bfaeSStefano Zampini xadj_count++; 3349e7931f94SStefano Zampini } 3350e7931f94SStefano Zampini } 3351d023bfaeSStefano Zampini xadj[1] = xadj_count; 3352c8587f34SStefano Zampini } else { 3353e7931f94SStefano Zampini if (xadj[1]) { 3354e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3355e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3356c8587f34SStefano Zampini } 3357e7931f94SStefano Zampini } 33583bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3359e7931f94SStefano Zampini if (use_square) { 3360e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3361e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3362e7931f94SStefano Zampini } 3363e7931f94SStefano Zampini } 3364e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3365e7931f94SStefano Zampini 33663837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3367e7931f94SStefano Zampini 3368e7931f94SStefano Zampini /* 3369e7931f94SStefano Zampini Restrict work on active processes only. 3370e7931f94SStefano Zampini */ 3371e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3372e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3373e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33742b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3375d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3376e7931f94SStefano Zampini if (color) { 3377e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3378e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3379e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3380c8587f34SStefano Zampini } else { 338152e5ac9dSStefano Zampini Mat subdomain_adj; 338252e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 338352e5ac9dSStefano Zampini MatPartitioning partitioner; 338452e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 338552e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3386b0c7d250SStefano Zampini PetscBool aggregate; 3387b0c7d250SStefano Zampini 3388306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3389785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3390e7931f94SStefano Zampini prank = rank; 3391306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33928002ef2cSStefano Zampini /* 3393e7931f94SStefano Zampini for (i=0;i<size;i++) { 3394e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3395c8587f34SStefano Zampini } 33968002ef2cSStefano Zampini */ 3397e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3398e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3399c8587f34SStefano Zampini } 3400e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3401b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3402b0c7d250SStefano Zampini if (aggregate) { 3403b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3404b0c7d250SStefano Zampini PetscMPIInt nrank; 3405b0c7d250SStefano Zampini PetscScalar *vals; 3406b0c7d250SStefano Zampini 3407b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3408b0c7d250SStefano Zampini lrows = 0; 3409b0c7d250SStefano Zampini if (nrank<redprocs) { 3410b0c7d250SStefano Zampini lrows = size/redprocs; 3411b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3412b0c7d250SStefano Zampini } 34135fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3414b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3415b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3416b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3417b0c7d250SStefano Zampini row = nrank; 3418b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3419b0c7d250SStefano Zampini cols = adjncy; 3420b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3421b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3422b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3423b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3424b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 342552e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 342652e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 342752e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3428b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3429b0c7d250SStefano Zampini } else { 3430306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3431b0c7d250SStefano Zampini } 343222b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3433e7931f94SStefano Zampini 3434e7931f94SStefano Zampini /* Partition */ 3435306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3436e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3437e7931f94SStefano Zampini if (use_vwgt) { 34383837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3439e7931f94SStefano Zampini v_wgt[0] = local_size; 3440e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3441c8587f34SStefano Zampini } 344228143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 344328143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3444e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3445e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 344622b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3447e7931f94SStefano Zampini 344852e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 344952e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 345052e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 345152e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3452b0c7d250SStefano Zampini if (!redprocs) { 3453b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 345428143c3dSStefano Zampini } else { 3455b0c7d250SStefano Zampini PetscInt idxs[1]; 3456b0c7d250SStefano Zampini PetscMPIInt tag; 3457b0c7d250SStefano Zampini MPI_Request *reqs; 3458b0c7d250SStefano Zampini 3459b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3460b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3461b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3462b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 346328143c3dSStefano Zampini } 3464b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3465b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3466b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3467b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3468e7931f94SStefano Zampini } 346952e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3470e7931f94SStefano Zampini /* clean up */ 3471e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 347252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3473e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3474e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3475e7931f94SStefano Zampini } 3476e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3477e7931f94SStefano Zampini 3478e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3479e7931f94SStefano Zampini i = 1; 3480e7931f94SStefano Zampini if (color) i=0; 3481e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3482e7931f94SStefano Zampini /* get back IS */ 3483e7931f94SStefano Zampini *is_sends = ranks_send_to; 3484e7931f94SStefano Zampini PetscFunctionReturn(0); 3485e7931f94SStefano Zampini } 3486e7931f94SStefano Zampini 3487e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3488e7931f94SStefano Zampini 3489e7931f94SStefano Zampini #undef __FUNCT__ 3490e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 349153a05cb3SStefano 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[]) 3492e7931f94SStefano Zampini { 349370cf5478SStefano Zampini Mat local_mat; 3494e7931f94SStefano Zampini IS is_sends_internal; 34959d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 349628143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34979d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3498e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3499e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3500e7931f94SStefano Zampini const PetscInt* is_indices; 3501e7931f94SStefano Zampini MatType new_local_type; 3502e7931f94SStefano Zampini /* buffers */ 3503e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 350428143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35059d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3506e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3507e7931f94SStefano Zampini /* MPI */ 350828143c3dSStefano Zampini MPI_Comm comm,comm_n; 350928143c3dSStefano Zampini PetscSubcomm subcomm; 3510e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 351128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 351228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 351328143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 351428143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 351528143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3516e7931f94SStefano Zampini PetscErrorCode ierr; 3517e7931f94SStefano Zampini 3518e7931f94SStefano Zampini PetscFunctionBegin; 351928143c3dSStefano Zampini /* TODO: add missing checks */ 352028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 352128143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 352228143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 352328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3524e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 352528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3526e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3527e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3528e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3529e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3530e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 353128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 353270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 353370cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 353428143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 353570cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 353670cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 353770cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 353870cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 353970cf5478SStefano Zampini } 3540e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3541e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3542e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3543e7931f94SStefano Zampini if (!is_sends) { 354428143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3545b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3546c8587f34SStefano Zampini } else { 3547e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3548e7931f94SStefano Zampini is_sends_internal = is_sends; 3549c8587f34SStefano Zampini } 3550e7931f94SStefano Zampini 3551e7931f94SStefano Zampini /* get comm */ 3552a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3553e7931f94SStefano Zampini 3554e7931f94SStefano Zampini /* compute number of sends */ 3555e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3556e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3557e7931f94SStefano Zampini 3558e7931f94SStefano Zampini /* compute number of receives */ 3559e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3560785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3561e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3562e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3563e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3564e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3565e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3566e7931f94SStefano Zampini 356728143c3dSStefano Zampini /* restrict comm if requested */ 356828143c3dSStefano Zampini subcomm = 0; 356928143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 357028143c3dSStefano Zampini if (restrict_comm) { 3571779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3572779c1cceSStefano Zampini 357328143c3dSStefano Zampini color = 0; 357453a05cb3SStefano Zampini if (restrict_full) { 357553a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 357653a05cb3SStefano Zampini } else { 357753a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 357853a05cb3SStefano Zampini } 3579b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 358028143c3dSStefano Zampini subcommsize = commsize - subcommsize; 358128143c3dSStefano Zampini /* check if reuse has been requested */ 358228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 358328143c3dSStefano Zampini if (*mat_n) { 358428143c3dSStefano Zampini PetscMPIInt subcommsize2; 358528143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 358628143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 358728143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 358828143c3dSStefano Zampini } else { 358928143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 359028143c3dSStefano Zampini } 359128143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3592779c1cceSStefano Zampini PetscMPIInt rank; 3593779c1cceSStefano Zampini 3594779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 359528143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 359628143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 359728143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3598306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 359928143c3dSStefano Zampini } 360028143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 360128143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 360228143c3dSStefano Zampini } else { 360328143c3dSStefano Zampini comm_n = comm; 360428143c3dSStefano Zampini } 360528143c3dSStefano Zampini 3606e7931f94SStefano Zampini /* prepare send/receive buffers */ 3607785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3608e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3609785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3610e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 361128143c3dSStefano Zampini if (nis) { 3612854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 361328143c3dSStefano Zampini } 3614e7931f94SStefano Zampini 361528143c3dSStefano Zampini /* Get data from local matrices */ 36166c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3617e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3618e7931f94SStefano Zampini /* 3619e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3620e7931f94SStefano Zampini send_buffer_idxs should contain: 3621e7931f94SStefano Zampini - MatType_PRIVATE type 3622e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3623e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3624e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3625e7931f94SStefano Zampini */ 36266c4ed002SBarry Smith else { 3627e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3629854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3630e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3631e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3633e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36343bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3635e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3636e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3637e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3638e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3639c8587f34SStefano Zampini } 3640c8587f34SStefano Zampini } 3641e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 364228143c3dSStefano Zampini /* additional is (if any) */ 364328143c3dSStefano Zampini if (nis) { 364428143c3dSStefano Zampini PetscMPIInt psum; 364528143c3dSStefano Zampini PetscInt j; 364628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 364728143c3dSStefano Zampini PetscInt plen; 364828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 364928143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 365028143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 365128143c3dSStefano Zampini } 3652854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 365328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 365428143c3dSStefano Zampini PetscInt plen; 365528143c3dSStefano Zampini const PetscInt *is_array_idxs; 365628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 365728143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 365828143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 365928143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 366028143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 366128143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 366228143c3dSStefano Zampini } 366328143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 366428143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 366528143c3dSStefano Zampini } 366628143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 366728143c3dSStefano Zampini } 366828143c3dSStefano Zampini 3669e7931f94SStefano Zampini buf_size_idxs = 0; 3670e7931f94SStefano Zampini buf_size_vals = 0; 367128143c3dSStefano Zampini buf_size_idxs_is = 0; 3672e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3673e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3674e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 367528143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3676e7931f94SStefano Zampini } 3677785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3678785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 367995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3680e7931f94SStefano Zampini 3681e7931f94SStefano Zampini /* get new tags for clean communications */ 3682e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3683e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 368428143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3685e7931f94SStefano Zampini 3686e7931f94SStefano Zampini /* allocate for requests */ 3687785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3688785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 368995ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3690785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3691785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 369295ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3693e7931f94SStefano Zampini 3694e7931f94SStefano Zampini /* communications */ 3695e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3696e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 369728143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3698e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3699e7931f94SStefano Zampini source_dest = onodes[i]; 3700e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3701e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3702e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3703e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 370428143c3dSStefano Zampini if (nis) { 370528143c3dSStefano 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); 370628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 370728143c3dSStefano Zampini } 3708e7931f94SStefano Zampini } 3709e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3710e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3711e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3712e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 371328143c3dSStefano Zampini if (nis) { 371428143c3dSStefano 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); 371528143c3dSStefano Zampini } 3716e7931f94SStefano Zampini } 3717e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3718e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3719e7931f94SStefano Zampini 3720e7931f94SStefano Zampini /* assemble new l2g map */ 3721e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3722e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37239d30be91SStefano Zampini new_local_rows = 0; 3724e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37259d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3726e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3727e7931f94SStefano Zampini } 37289d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3729e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37309d30be91SStefano Zampini new_local_rows = 0; 3731e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37329d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37339d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3734e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3735e7931f94SStefano Zampini } 37369d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37379d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3738e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3739e7931f94SStefano Zampini 3740e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3741e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3742e7931f94SStefano 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) */ 3743e7931f94SStefano Zampini if (n_recvs) { 374428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3745e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3746e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3747e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3748e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3749e7931f94SStefano Zampini break; 3750e7931f94SStefano Zampini } 3751e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3752e7931f94SStefano Zampini } 3753e7931f94SStefano Zampini switch (new_local_type_private) { 375428143c3dSStefano Zampini case MATDENSE_PRIVATE: 375528143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3756e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3757e7931f94SStefano Zampini bs = 1; 375828143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 375928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 376028143c3dSStefano Zampini bs = 1; 376128143c3dSStefano Zampini } 3762e7931f94SStefano Zampini break; 3763e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3764e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3765e7931f94SStefano Zampini bs = 1; 3766e7931f94SStefano Zampini break; 3767e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3768e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3769e7931f94SStefano Zampini break; 3770e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3771e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3772e7931f94SStefano Zampini break; 3773e7931f94SStefano Zampini default: 37749d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3775e7931f94SStefano Zampini break; 3776e7931f94SStefano Zampini } 377728143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 377828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 377928143c3dSStefano Zampini bs = 1; 3780e7931f94SStefano Zampini } 3781e7931f94SStefano Zampini 378270cf5478SStefano Zampini /* create MATIS object if needed */ 378370cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3784e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3785e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 378670cf5478SStefano Zampini } else { 378770cf5478SStefano Zampini /* it also destroys the local matrices */ 378870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 378970cf5478SStefano Zampini } 379070cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3791e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37929d30be91SStefano Zampini 37939d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37949d30be91SStefano Zampini 37959d30be91SStefano Zampini /* Global to local map of received indices */ 37969d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37979d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37989d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37999d30be91SStefano Zampini 38009d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38019d30be91SStefano Zampini buf_size_idxs = 0; 38029d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38039d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38049d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38059d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38069d30be91SStefano Zampini } 38079d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38089d30be91SStefano Zampini 38099d30be91SStefano Zampini /* set preallocation */ 38109d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38119d30be91SStefano Zampini if (!newisdense) { 38129d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38139d30be91SStefano Zampini 38149d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38159d30be91SStefano Zampini if (n_recvs) { 38169d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38179d30be91SStefano Zampini } 38189d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38199d30be91SStefano Zampini PetscInt j; 38209d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38219d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38229d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38239d30be91SStefano Zampini } 38249d30be91SStefano Zampini } else { 38259d30be91SStefano Zampini /* TODO */ 38269d30be91SStefano Zampini } 38279d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38289d30be91SStefano Zampini } 38299d30be91SStefano Zampini if (new_local_nnz) { 38309d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38319d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38329d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38339d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38349d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38359d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38369d30be91SStefano Zampini } else { 38379d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38389d30be91SStefano Zampini } 38399d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38409d30be91SStefano Zampini } else { 38419d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38429d30be91SStefano Zampini } 3843e7931f94SStefano Zampini 3844e7931f94SStefano Zampini /* set values */ 3845e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38469d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3847e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3848e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3849e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38509d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3851e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3852e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3853e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 385428143c3dSStefano Zampini } else { 385528143c3dSStefano Zampini /* TODO */ 3856e7931f94SStefano Zampini } 3857e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3858e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3859e7931f94SStefano Zampini } 3860e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3861e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 386270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 386370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38649d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38659d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3866e7931f94SStefano Zampini 3867dfd14d43SStefano Zampini #if 0 386828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3869e7931f94SStefano Zampini Vec lvec,rvec; 3870e7931f94SStefano Zampini PetscReal infty_error; 3871e7931f94SStefano Zampini 38722a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3873e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3874e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3875e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 387670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3877e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3878e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3879e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3880e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3881e7931f94SStefano Zampini } 388228143c3dSStefano Zampini #endif 3883e7931f94SStefano Zampini 388428143c3dSStefano Zampini /* assemble new additional is (if any) */ 388528143c3dSStefano Zampini if (nis) { 388628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 388728143c3dSStefano Zampini 388828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3889854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 389028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 389128143c3dSStefano Zampini psum = 0; 389228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 389328143c3dSStefano Zampini for (j=0;j<nis;j++) { 389428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 389528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 389628143c3dSStefano Zampini psum += plen; 389728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 389828143c3dSStefano Zampini } 389928143c3dSStefano Zampini } 3900854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3901854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 390228143c3dSStefano Zampini for (i=1;i<nis;i++) { 390328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 390428143c3dSStefano Zampini } 390528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 390628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 390728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 390828143c3dSStefano Zampini for (j=0;j<nis;j++) { 390928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 391028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 391128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 391228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 391328143c3dSStefano Zampini } 391428143c3dSStefano Zampini } 391528143c3dSStefano Zampini for (i=0;i<nis;i++) { 391628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 391728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 391828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 391928143c3dSStefano Zampini } 392028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 392128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 392228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 392328143c3dSStefano Zampini } 3924e7931f94SStefano Zampini /* free workspace */ 392528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3926e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3927e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3928e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3929e7931f94SStefano Zampini if (isdense) { 3930e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3931e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3932e7931f94SStefano Zampini } else { 3933e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3934e7931f94SStefano Zampini } 393528143c3dSStefano Zampini if (nis) { 393628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 393728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 393828143c3dSStefano Zampini } 3939e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3940e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 394128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3942e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3943e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 394428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3945e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3946e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3947e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3948e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3949e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 395028143c3dSStefano Zampini if (nis) { 395128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 395228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 395328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 395428143c3dSStefano Zampini } 395528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 395628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 395728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 395828143c3dSStefano Zampini for (i=0;i<nis;i++) { 395928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 396028143c3dSStefano Zampini } 396153a05cb3SStefano Zampini *mat_n = NULL; 396228143c3dSStefano Zampini } 3963e7931f94SStefano Zampini PetscFunctionReturn(0); 3964e7931f94SStefano Zampini } 3965a57a6d2fSStefano Zampini 396612edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3967af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 396812edc857SStefano Zampini 3969c8587f34SStefano Zampini #undef __FUNCT__ 3970c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3971c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3972c8587f34SStefano Zampini { 3973c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3974c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 397520a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39769881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 397720a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39786e683305SStefano Zampini IS coarse_is,*isarray; 39796e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39806e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3981f9eb5b7dSStefano Zampini PC pc_temp; 3982c8587f34SStefano Zampini PCType coarse_pc_type; 3983c8587f34SStefano Zampini KSPType coarse_ksp_type; 3984f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39854f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39866e683305SStefano Zampini Mat t_coarse_mat_is; 39876e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39886e683305SStefano Zampini PetscMPIInt all_procs; 398974e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 399068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 399122bc73bbSStefano Zampini PetscScalar *array; 39929881197aSStefano Zampini PetscErrorCode ierr; 3993fdc09c96SStefano Zampini 3994c8587f34SStefano Zampini PetscFunctionBegin; 3995c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 399668457ee5SStefano 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 */ 3997fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39985a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3999fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4000f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4001f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4002f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4003fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 400451bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 400551bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4006dc4bcba2SStefano Zampini PC pc; 4007dc4bcba2SStefano Zampini PetscBool isbddc; 4008dc4bcba2SStefano Zampini 4009dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4010dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4011dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4012dc4bcba2SStefano Zampini if (isbddc) { 4013dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4014dc4bcba2SStefano Zampini } 4015727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4016fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4017fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4018fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4019f4ddd8eeSStefano Zampini } 4020fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4021fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4022f4ddd8eeSStefano Zampini } 402370cf5478SStefano Zampini /* reset any subassembling information */ 402470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40256e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40266e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4027fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4028f4ddd8eeSStefano Zampini } 4029c8587f34SStefano Zampini 40306e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40312b510759SStefano Zampini im_active = !!(pcis->n); 4032b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40336e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40346e683305SStefano Zampini void_procs = all_procs-active_procs; 40356e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 403674e2c79eSStefano Zampini redist = PETSC_FALSE; 403722bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40386e683305SStefano Zampini csin_ml = PETSC_TRUE; 40396e683305SStefano Zampini ncoarse_ml = void_procs; 4040779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4041779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40426e683305SStefano Zampini csin_ds = PETSC_TRUE; 404318a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 404418a45a71SStefano Zampini redist = PETSC_TRUE; 404518a45a71SStefano Zampini } else { 40466e683305SStefano Zampini csin_ds = PETSC_TRUE; 4047779c1cceSStefano Zampini ncoarse_ds = active_procs; 4048779c1cceSStefano Zampini redist = PETSC_TRUE; 404918a45a71SStefano Zampini } 40506e683305SStefano Zampini } else { 40516e683305SStefano Zampini csin_ml = PETSC_FALSE; 40526e683305SStefano Zampini ncoarse_ml = all_procs; 40536e683305SStefano Zampini if (void_procs) { 40546e683305SStefano Zampini csin_ds = PETSC_TRUE; 40556e683305SStefano Zampini ncoarse_ds = void_procs; 40566e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40576e683305SStefano Zampini } else { 4058779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 405974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 406074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 406174e2c79eSStefano Zampini redist = PETSC_TRUE; 406274e2c79eSStefano Zampini } else { 40636e683305SStefano Zampini csin_ds = PETSC_FALSE; 40646e683305SStefano Zampini ncoarse_ds = all_procs; 40656e683305SStefano Zampini } 40666e683305SStefano Zampini } 406774e2c79eSStefano Zampini } 40686e683305SStefano Zampini 40696e683305SStefano Zampini /* 40706e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40716e683305SStefano Zampini - we have not exceeded the number of levels requested 40726e683305SStefano Zampini - we can actually subassemble the active processes 40736e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40746e683305SStefano Zampini */ 40756e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40766e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40776e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40786e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40796e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4080f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40812b510759SStefano Zampini } else { 4082f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4083c8587f34SStefano Zampini } 4084c8587f34SStefano Zampini } 40856e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40866e683305SStefano Zampini if (multilevel_allowed) { 40876e683305SStefano Zampini ncoarse = ncoarse_ml; 40886e683305SStefano Zampini csin = csin_ml; 408958da7f69SStefano Zampini redist = PETSC_FALSE; 40906e683305SStefano Zampini } else { 40916e683305SStefano Zampini ncoarse = ncoarse_ds; 40926e683305SStefano Zampini csin = csin_ds; 40936e683305SStefano Zampini } 4094e7931f94SStefano Zampini 4095abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4096abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4097abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4098abbbba34SStefano Zampini 4099abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 410022bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 410122bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 410222bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 410322bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4104b9b85e73SStefano Zampini #if 0 4105b9b85e73SStefano Zampini { 4106b9b85e73SStefano Zampini PetscViewer viewer; 4107b9b85e73SStefano Zampini char filename[256]; 4108b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4109b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 41106a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4111b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4112f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 4113b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4114b9b85e73SStefano Zampini } 4115b9b85e73SStefano Zampini #endif 4116e176bc59SStefano 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); 41176e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41186e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41196e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4120abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4121abbbba34SStefano Zampini 41226e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41236e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41246e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41256e683305SStefano Zampini const PetscInt *idxs; 41266e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41276e683305SStefano Zampini 41286e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41290be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41306e683305SStefano Zampini /* allocate space for temporary storage */ 4131854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4132854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41336e683305SStefano Zampini /* allocate for IS array */ 41346e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41356e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41366e683305SStefano Zampini nis = nisdofs + nisneu; 4137854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41386e683305SStefano Zampini /* dofs splitting */ 41396e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41406e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41416e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41426e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41436e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41446e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41456e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41466e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41476e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41486e683305SStefano Zampini } 41496e683305SStefano Zampini /* neumann boundaries */ 41506e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41516e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41526e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41536e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41546e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41556e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41566e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41576e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41586e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41596e683305SStefano Zampini } 41606e683305SStefano Zampini /* free memory */ 41616e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41626e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41636e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41646e683305SStefano Zampini } else { 41656e683305SStefano Zampini nis = 0; 41666e683305SStefano Zampini nisdofs = 0; 41676e683305SStefano Zampini nisneu = 0; 41686e683305SStefano Zampini isarray = NULL; 41696e683305SStefano Zampini } 41706e683305SStefano Zampini /* destroy no longer needed map */ 41716e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41726e683305SStefano Zampini 41736e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41746e683305SStefano Zampini coarse_mat_is = NULL; 41756e683305SStefano Zampini if (csin) { 41766e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 417774e2c79eSStefano Zampini if (redist) { 417874e2c79eSStefano Zampini PetscMPIInt rank; 4179779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 418074e2c79eSStefano Zampini 418174e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 418258da7f69SStefano Zampini spc = active_procs/ncoarse; 418358da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4184779c1cceSStefano Zampini if (im_active) { 4185779c1cceSStefano Zampini destsize = 1; 418674e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 418774e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 418874e2c79eSStefano Zampini } else { 418974e2c79eSStefano Zampini dest[0] = rank/(spc+1); 419074e2c79eSStefano Zampini } 419174e2c79eSStefano Zampini } else { 4192779c1cceSStefano Zampini destsize = 0; 41936e683305SStefano Zampini } 4194779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4195779c1cceSStefano Zampini } else if (csin_type_simple) { 41966e683305SStefano Zampini PetscMPIInt rank; 41976e683305SStefano Zampini PetscInt issize,isidx; 4198779c1cceSStefano Zampini 41996e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42006e683305SStefano Zampini if (im_active) { 42016e683305SStefano Zampini issize = 1; 42026e683305SStefano Zampini isidx = (PetscInt)rank; 42036e683305SStefano Zampini } else { 42046e683305SStefano Zampini issize = 0; 42056e683305SStefano Zampini isidx = -1; 42066e683305SStefano Zampini } 42076e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4208779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4209b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42106e683305SStefano Zampini } 4211779c1cceSStefano Zampini 4212779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4213779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4214779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4215779c1cceSStefano Zampini PetscInt *coarse_candidates; 4216779c1cceSStefano Zampini const PetscInt* tisindices; 4217779c1cceSStefano Zampini 4218779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4219779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4220779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4221779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4222779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4223779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4224779c1cceSStefano Zampini } 4225779c1cceSStefano Zampini } 4226779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4227779c1cceSStefano Zampini 4228779c1cceSStefano Zampini 42296e683305SStefano Zampini if (pcbddc->dbg_flag) { 42306e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42326e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4234779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42366e683305SStefano Zampini } 42376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42386e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42396e683305SStefano Zampini } 42406e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42416e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42426e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4243854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42446e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42456e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42466e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42476e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42486e683305SStefano Zampini } 42496e683305SStefano Zampini if (pcbddc->dbg_flag) { 42506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42526e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42536e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42546e683305SStefano Zampini } 4255779c1cceSStefano Zampini } 42566e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 425753a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 425853a05cb3SStefano 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); 425953a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 426053a05cb3SStefano 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); 426153a05cb3SStefano Zampini } 42626e683305SStefano Zampini } else { 42636e683305SStefano Zampini if (pcbddc->dbg_flag) { 42646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42666e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42676e683305SStefano Zampini } 42686e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42696e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42706e683305SStefano Zampini } 42716e683305SStefano Zampini 42726e683305SStefano Zampini /* create local to global scatters for coarse problem */ 427368457ee5SStefano Zampini if (compute_vecs) { 42746e683305SStefano Zampini PetscInt lrows; 42756e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42766e683305SStefano Zampini if (coarse_mat_is) { 42776e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42786e683305SStefano Zampini } else { 42796e683305SStefano Zampini lrows = 0; 42806e683305SStefano Zampini } 42816e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42826e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42836e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42846e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42856e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42866e683305SStefano Zampini } 42876e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42886e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4289c8587f34SStefano Zampini 4290f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4291f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4292f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4293f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4294f9eb5b7dSStefano Zampini } else { 4295f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4296f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4297c8587f34SStefano Zampini } 4298c8587f34SStefano Zampini 42996e683305SStefano Zampini /* print some info if requested */ 43006e683305SStefano Zampini if (pcbddc->dbg_flag) { 43016e683305SStefano Zampini if (!multilevel_allowed) { 43026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43036e683305SStefano Zampini if (multilevel_requested) { 43046e683305SStefano 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); 43056e683305SStefano Zampini } else if (pcbddc->max_levels) { 43066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43076e683305SStefano Zampini } 43086e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43096e683305SStefano Zampini } 43106e683305SStefano Zampini } 43116e683305SStefano Zampini 4312f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43136e683305SStefano Zampini if (coarse_mat_is) { 43146e683305SStefano Zampini MatReuse coarse_mat_reuse; 43156a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43166e683305SStefano Zampini if (pcbddc->dbg_flag) { 43176e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43186e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43196e683305SStefano Zampini } 4320f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4321312be037SStefano Zampini char prefix[256],str_level[16]; 4322e604994aSStefano Zampini size_t len; 43236e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4324422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4325c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4326f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43275f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4328c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43296e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4330c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4331c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4332e604994aSStefano Zampini /* prefix */ 4333e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4334e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4335e604994aSStefano Zampini if (!pcbddc->current_level) { 4336e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4337e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4338c8587f34SStefano Zampini } else { 4339e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4340312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4341312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 434234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4343312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4344e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4345e604994aSStefano Zampini } 4346e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43473e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43483e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43493e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43503e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4351f9eb5b7dSStefano Zampini /* allow user customization */ 4352f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43533e3c6dadSStefano Zampini } 43543e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 435551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43563e3c6dadSStefano Zampini if (nisdofs) { 43573e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43583e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43593e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43603e3c6dadSStefano Zampini } 43613e3c6dadSStefano Zampini } 43623e3c6dadSStefano Zampini if (nisneu) { 43633e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43643e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4365312be037SStefano Zampini } 4366f9eb5b7dSStefano Zampini 4367f9eb5b7dSStefano Zampini /* get some info after set from options */ 4368f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4369f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43704f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43716e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4372f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4373f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4374f9eb5b7dSStefano Zampini } 437539f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 43764f3a063dSStefano Zampini if (isredundant) { 43774f3a063dSStefano Zampini KSP inner_ksp; 43784f3a063dSStefano Zampini PC inner_pc; 43794f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43804f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43814f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43824f3a063dSStefano Zampini } 4383f9eb5b7dSStefano Zampini 4384f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4385fa7f1dd8SStefano Zampini if (coarse_reuse) { 438681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4387fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43886e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4389fa7f1dd8SStefano Zampini } else { 43906e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4391fa7f1dd8SStefano Zampini } 4392c8587f34SStefano Zampini if (isbddc || isnn) { 439322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 439470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4395b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 439622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43976e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43996e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44006e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 440122b6e8a2SStefano Zampini } 440270cf5478SStefano Zampini } 440353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 440470cf5478SStefano Zampini } else { 440522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 440622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 440722bc73bbSStefano Zampini } 440822bc73bbSStefano Zampini } else { 44092e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4410c8587f34SStefano Zampini } 4411c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4412c8587f34SStefano Zampini 44133301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44145a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44153301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44163301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44173301b35fSStefano Zampini } 44183301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44193301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44203301b35fSStefano Zampini } 44213301b35fSStefano Zampini if (pc->pmat->spd_set) { 44223301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44233301b35fSStefano Zampini } 44246e683305SStefano Zampini /* set operators */ 44255f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44266e683305SStefano Zampini if (pcbddc->dbg_flag) { 44276e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44286e683305SStefano Zampini } 44296e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44306e683305SStefano Zampini coarse_mat = 0; 44316e683305SStefano Zampini } 44326e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4433b9b85e73SStefano Zampini #if 0 4434b9b85e73SStefano Zampini { 4435b9b85e73SStefano Zampini PetscViewer viewer; 4436b9b85e73SStefano Zampini char filename[256]; 4437b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4438b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 44396a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4440b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4441f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 4442b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4443b9b85e73SStefano Zampini } 4444b9b85e73SStefano Zampini #endif 4445c8587f34SStefano Zampini 4446c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4447298c0119SStefano Zampini #if 0 4448c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4449c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 445098a51de6SStefano Zampini } 4451298c0119SStefano Zampini #endif 445298a51de6SStefano Zampini 445398a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 445498a51de6SStefano Zampini Vec crhs,csol; 445598a51de6SStefano Zampini PetscBool ispreonly; 445604708bb6SStefano Zampini 445798a51de6SStefano Zampini if (CoarseNullSpace) { 4458c8587f34SStefano Zampini if (isbddc) { 4459c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4460c8587f34SStefano Zampini } else { 44615fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4462c8587f34SStefano Zampini } 4463c8587f34SStefano Zampini } 4464f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4465f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4466f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4467f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44686e683305SStefano Zampini /* hack */ 4469f347579bSStefano Zampini if (!csol) { 44702a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4471f9eb5b7dSStefano Zampini } 4472f347579bSStefano Zampini if (!crhs) { 44732a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4474f347579bSStefano Zampini } 4475cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4476cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44776e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4478c8587f34SStefano Zampini KSP check_ksp; 44792b510759SStefano Zampini KSPType check_ksp_type; 4480c8587f34SStefano Zampini PC check_pc; 44816e683305SStefano Zampini Vec check_vec,coarse_vec; 44826a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44832b510759SStefano Zampini PetscInt its; 44846e683305SStefano Zampini PetscBool compute_eigs; 44856e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44866e683305SStefano Zampini PetscInt neigs; 44878e185a42SStefano Zampini const char *prefix; 4488c8587f34SStefano Zampini 44892b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44906e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4491422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 449223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4493f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44942b510759SStefano Zampini if (ispreonly) { 44952b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44966e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44972b510759SStefano Zampini } else { 4498cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44996e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4500c8587f34SStefano Zampini } 4501c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45026e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45036e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45046e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4505a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4506a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4507a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4508a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4509c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4510c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4511c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4512c8587f34SStefano Zampini /* create random vec */ 45136e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45146e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4515c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4516c8587f34SStefano Zampini if (CoarseNullSpace) { 4517c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4518c8587f34SStefano Zampini } 45196e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4520c8587f34SStefano Zampini /* solve coarse problem */ 45216e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4522c8587f34SStefano Zampini if (CoarseNullSpace) { 45236e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4524c8587f34SStefano Zampini } 4525cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45266e683305SStefano Zampini if (compute_eigs) { 4527854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4528854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45296e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45306e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45316e683305SStefano Zampini lambda_min = eigs_r[0]; 45326e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45336e683305SStefano Zampini if (lambda_max>lambda_min) { 4534cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4535cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4536cbcc2c2aSStefano Zampini } 4537c8587f34SStefano Zampini } 4538c8587f34SStefano Zampini } 4539cbcc2c2aSStefano Zampini 4540c8587f34SStefano Zampini /* check coarse problem residual error */ 45416e683305SStefano Zampini if (pcbddc->dbg_flag) { 45426e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45436e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45446e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4545c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45466e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45476e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4548c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4549779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45506e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45516e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45546e683305SStefano Zampini if (compute_eigs) { 45556e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4556deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4557c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45586e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45596e683305SStefano 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); 45606e683305SStefano Zampini for (i=0;i<neigs;i++) { 45616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4562c8587f34SStefano Zampini } 45636e683305SStefano Zampini } 45646e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45656e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45666e683305SStefano Zampini } 4567c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45686e683305SStefano Zampini if (compute_eigs) { 45696e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45706e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4571c8587f34SStefano Zampini } 45726e683305SStefano Zampini } 45736e683305SStefano Zampini } 4574cbcc2c2aSStefano Zampini /* print additional info */ 4575cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45766e683305SStefano Zampini /* waits until all processes reaches this point */ 45776e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4578cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4579cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4580cbcc2c2aSStefano Zampini } 4581cbcc2c2aSStefano Zampini 45822b510759SStefano Zampini /* free memory */ 4583c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4584fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4585c8587f34SStefano Zampini PetscFunctionReturn(0); 4586c8587f34SStefano Zampini } 4587674ae819SStefano Zampini 4588f34684f1SStefano Zampini #undef __FUNCT__ 4589f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4590f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4591f34684f1SStefano Zampini { 4592f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4593f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4594f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4595dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4596dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 459773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4598dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4599f34684f1SStefano Zampini PetscErrorCode ierr; 4600f34684f1SStefano Zampini 4601f34684f1SStefano Zampini PetscFunctionBegin; 4602f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46036c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4604dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46053bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4606dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4607dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4608dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4609dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4610dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4611dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46126c4ed002SBarry 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); 4613dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4614dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4615dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4616dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4617dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4618f34684f1SStefano Zampini 4619f34684f1SStefano Zampini /* check numbering */ 4620f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4621f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4622dc456d91SStefano Zampini PetscInt i; 4623b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4624f34684f1SStefano Zampini 4625f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4626f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4627f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46281575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4629f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4630f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4631727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4632f34684f1SStefano Zampini } 4633f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4634f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4635f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4636e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4637e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4638e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4639e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4640f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4641f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4642f34684f1SStefano Zampini if (array[i] == 1.0) { 4643b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4644f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4645f34684f1SStefano Zampini } 4646f34684f1SStefano Zampini } 4647b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4648f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4649f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4650f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4651f34684f1SStefano Zampini } 4652f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4653f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4654e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4655e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4656f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4657f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4658b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4659ca8b9ea9SStefano Zampini PetscInt *gidxs; 4660ca8b9ea9SStefano Zampini 4661ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46623bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4663f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4664f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4665f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4666f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46674bc2dc4bSStefano 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); 4668f34684f1SStefano Zampini } 4669f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4670ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4671f34684f1SStefano Zampini } 4672f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46731575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4674302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4675f34684f1SStefano Zampini } 46768bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4677f34684f1SStefano Zampini /* get back data */ 4678f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4679f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4680674ae819SStefano Zampini PetscFunctionReturn(0); 4681674ae819SStefano Zampini } 4682674ae819SStefano Zampini 4683e456f2a8SStefano Zampini #undef __FUNCT__ 4684e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4685a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4686e456f2a8SStefano Zampini { 4687e456f2a8SStefano Zampini IS localis_t; 4688a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4689e456f2a8SStefano Zampini PetscScalar *vals; 4690e456f2a8SStefano Zampini PetscErrorCode ierr; 4691e456f2a8SStefano Zampini 4692e456f2a8SStefano Zampini PetscFunctionBegin; 4693a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4694e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4695854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4696e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4697e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4698a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4699a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47001035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4701a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47021035eff8SStefano Zampini } 4703a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4704e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4705e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4706a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4707a7dc3881SStefano Zampini /* now compute set in local ordering */ 4708a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4709a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4710a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4711a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4712a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4713ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4714e456f2a8SStefano Zampini lsize++; 4715e456f2a8SStefano Zampini } 4716e456f2a8SStefano Zampini } 4717854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4718a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4719ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4720e456f2a8SStefano Zampini idxs[lsize++] = i; 4721e456f2a8SStefano Zampini } 4722e456f2a8SStefano Zampini } 4723a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4724a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4725e456f2a8SStefano Zampini *localis = localis_t; 4726e456f2a8SStefano Zampini PetscFunctionReturn(0); 4727e456f2a8SStefano Zampini } 4728906d46d4SStefano Zampini 4729906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4730906d46d4SStefano Zampini #undef __FUNCT__ 4731906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4732906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4733906d46d4SStefano Zampini { 4734906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4735906d46d4SStefano Zampini PetscErrorCode ierr; 4736906d46d4SStefano Zampini 4737906d46d4SStefano Zampini PetscFunctionBegin; 4738906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4739906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4740906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4741906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4742906d46d4SStefano Zampini PetscFunctionReturn(0); 4743906d46d4SStefano Zampini } 4744906d46d4SStefano Zampini 4745906d46d4SStefano Zampini #undef __FUNCT__ 4746906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4747906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4748906d46d4SStefano Zampini { 4749906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4750906d46d4SStefano Zampini PetscErrorCode ierr; 4751906d46d4SStefano Zampini 4752906d46d4SStefano Zampini PetscFunctionBegin; 4753906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4754906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4755906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4756906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4757906d46d4SStefano Zampini PetscFunctionReturn(0); 4758906d46d4SStefano Zampini } 4759b96c3477SStefano Zampini 4760b96c3477SStefano Zampini #undef __FUNCT__ 4761b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 476208122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4763b96c3477SStefano Zampini { 4764a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4765b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4766b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4767a64f4aa4SStefano Zampini Mat S_j; 4768b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4769b96c3477SStefano Zampini PetscBool free_used_adj; 4770b96c3477SStefano Zampini PetscErrorCode ierr; 4771b96c3477SStefano Zampini 4772b96c3477SStefano Zampini PetscFunctionBegin; 4773b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4774b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 477508122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4776b96c3477SStefano Zampini used_xadj = NULL; 4777b96c3477SStefano Zampini used_adjncy = NULL; 4778b96c3477SStefano Zampini } else { 477908122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 478008122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 478108122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 478208122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4783b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4784b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4785b96c3477SStefano Zampini } else { 47862fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4787b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4788b96c3477SStefano Zampini PetscInt nvtxs; 4789b96c3477SStefano Zampini 47902fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47912fffb893SStefano Zampini if (flg_row) { 4792b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4793b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4794b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4795b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47962fffb893SStefano Zampini } else { 47972fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47982fffb893SStefano Zampini used_xadj = NULL; 47992fffb893SStefano Zampini used_adjncy = NULL; 48002fffb893SStefano Zampini } 48012fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4802b96c3477SStefano Zampini } 4803b96c3477SStefano Zampini } 4804d5574798SStefano Zampini 4805d5574798SStefano Zampini /* setup sub_schurs data */ 4806a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4807a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4808a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4809a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4810862806e4SStefano 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); 4811a64f4aa4SStefano Zampini } else { 48126816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 481304708bb6SStefano Zampini PetscBool isseqaij; 48145feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48155feab87aSStefano Zampini PetscInt n_vertices; 48165feab87aSStefano Zampini 48175feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48182034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48195feab87aSStefano Zampini } 482004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 482104708bb6SStefano Zampini if (!isseqaij) { 482204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 482304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 482404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 482504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 482604708bb6SStefano Zampini } else { 4827511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 482804708bb6SStefano Zampini } 482904708bb6SStefano Zampini } 4830862806e4SStefano 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); 4831a64f4aa4SStefano Zampini } 4832a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4833b96c3477SStefano Zampini 4834b96c3477SStefano Zampini /* free adjacency */ 4835b96c3477SStefano Zampini if (free_used_adj) { 4836b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4837b96c3477SStefano Zampini } 4838b96c3477SStefano Zampini PetscFunctionReturn(0); 4839b96c3477SStefano Zampini } 4840b96c3477SStefano Zampini 4841b96c3477SStefano Zampini #undef __FUNCT__ 4842b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 484308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4844b96c3477SStefano Zampini { 4845b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4846b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4847b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4848b96c3477SStefano Zampini PCBDDCGraph graph; 4849b96c3477SStefano Zampini PetscErrorCode ierr; 4850b96c3477SStefano Zampini 4851b96c3477SStefano Zampini PetscFunctionBegin; 4852b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 485308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48543301b35fSStefano Zampini IS verticesIS,verticescomm; 48553301b35fSStefano Zampini PetscInt vsize,*idxs; 4856b96c3477SStefano Zampini 4857b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48583301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48593301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48603301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48613301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48623301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4863b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48647fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48653301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48663301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4867b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4868b96c3477SStefano Zampini /* 4869b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4870b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4871b96c3477SStefano Zampini } 4872b96c3477SStefano Zampini */ 4873b96c3477SStefano Zampini } else { 4874b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4875b96c3477SStefano Zampini } 4876b96c3477SStefano Zampini 4877b96c3477SStefano Zampini /* sub_schurs init */ 4878a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4879a64f4aa4SStefano Zampini 4880b96c3477SStefano Zampini /* free graph struct */ 488108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4882b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4883b96c3477SStefano Zampini } 4884b96c3477SStefano Zampini PetscFunctionReturn(0); 4885b96c3477SStefano Zampini } 4886