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; 12608122e43SStefano Zampini cum = cum2 = 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 } 14008122e43SStefano Zampini cum2 = cum; 141d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14208122e43SStefano Zampini } 14308122e43SStefano Zampini 14408122e43SStefano Zampini if (mss) { /* multilevel */ 14508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 14608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 14708122e43SStefano Zampini } 14808122e43SStefano Zampini 14908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15008122e43SStefano Zampini 15108122e43SStefano Zampini const PetscInt *idxs; 152f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 153862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15408122e43SStefano Zampini PetscBLASInt B_N; 155aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 15608122e43SStefano Zampini 157862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 158f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 159f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1609ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 161aff50787SStefano Zampini PetscInt j,k; 162aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 163aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 164aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 16508122e43SStefano Zampini } 16608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 167aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 168aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 169aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 170aff50787SStefano Zampini } 17108122e43SStefano Zampini } 17208122e43SStefano Zampini } else { 17308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17508122e43SStefano Zampini } 1768bec7fa6SStefano Zampini } else { 177f6f667cfSStefano Zampini S = Sarray + cumarray; 178f6f667cfSStefano Zampini St = Starray + cumarray; 1798bec7fa6SStefano Zampini } 18008122e43SStefano Zampini 181f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 182aff50787SStefano Zampini /* see if we can save some work */ 183aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 184aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 185aff50787SStefano Zampini } 186aff50787SStefano Zampini 187aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 188aff50787SStefano Zampini B_neigs = 0; 189aff50787SStefano Zampini } else { 190aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 191f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 192f6f667cfSStefano Zampini 1939ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 19408122e43SStefano Zampini PetscBLASInt B_itype = 1; 195f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1964c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1979552c7c7SStefano Zampini PetscInt nmin_s; 19808122e43SStefano Zampini 199f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 200fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2018bec7fa6SStefano 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]]); 202fd14bc51SStefano Zampini } 20308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 20408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 205f6f667cfSStefano 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)); 20608122e43SStefano Zampini #else 207f6f667cfSStefano 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)); 20808122e43SStefano Zampini #endif 20908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 21008122e43SStefano Zampini if (B_ierr) { 2116c4ed002SBarry 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); 2126c4ed002SBarry 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); 2136c4ed002SBarry 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); 21408122e43SStefano Zampini } 21508122e43SStefano Zampini 21608122e43SStefano Zampini if (B_neigs > nmax) { 217fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 218fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 219fd14bc51SStefano Zampini } 220f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 22108122e43SStefano Zampini B_neigs = nmax; 22208122e43SStefano Zampini } 22308122e43SStefano Zampini 2249552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2259552c7c7SStefano Zampini if (B_neigs < nmin_s) { 22608122e43SStefano Zampini PetscBLASInt B_neigs2; 22708122e43SStefano Zampini 228f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 229f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 230fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 231fd14bc51SStefano 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); 232fd14bc51SStefano Zampini } 2339ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23408122e43SStefano Zampini PetscInt j; 23508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23608122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23708122e43SStefano Zampini } 23808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23908122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24008122e43SStefano Zampini } 24108122e43SStefano Zampini } else { 24208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24408122e43SStefano Zampini } 24508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 247f6f667cfSStefano 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)); 24808122e43SStefano Zampini #else 249f6f667cfSStefano 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)); 25008122e43SStefano Zampini #endif 25108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 25208122e43SStefano Zampini B_neigs += B_neigs2; 25308122e43SStefano Zampini } 25408122e43SStefano Zampini if (B_ierr) { 2556c4ed002SBarry 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); 2566c4ed002SBarry 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); 2576c4ed002SBarry 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); 25808122e43SStefano Zampini } 259fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 260ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 26108122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 26208122e43SStefano Zampini if (eigs[j] == 0.0) { 263ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 26408122e43SStefano Zampini } else { 265ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 266fd14bc51SStefano Zampini } 26708122e43SStefano Zampini } 26808122e43SStefano Zampini } 26908122e43SStefano Zampini } else { 27008122e43SStefano Zampini /* TODO */ 27108122e43SStefano Zampini } 272aff50787SStefano Zampini } 2738bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2748bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2759162d606SStefano Zampini if (B_neigs) { 2769162d606SStefano 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); 277fd14bc51SStefano Zampini 278fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2799552c7c7SStefano Zampini PetscInt ii; 2809552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 281ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 2829552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 283ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 284ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 285ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 286ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 287ac47001eSStefano Zampini #else 288ac47001eSStefano 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); 289ac47001eSStefano Zampini #endif 2909552c7c7SStefano Zampini } 2919552c7c7SStefano Zampini } 292fd14bc51SStefano Zampini } 29308122e43SStefano Zampini #if 0 2949162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29508122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29608122e43SStefano Zampini PetscScalar norm; 29708122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 2989162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 2999162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 30008122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 30108122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30208122e43SStefano Zampini } else { 30308122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 30408122e43SStefano Zampini } 3059162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 306b1b3d7a2SStefano Zampini } 307b1b3d7a2SStefano Zampini #endif 3089162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3099162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3109162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3119162d606SStefano Zampini cum++; 31208122e43SStefano Zampini } 31308122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 31408122e43SStefano Zampini /* shift for next computation */ 31508122e43SStefano Zampini cumarray += subset_size*subset_size; 31608122e43SStefano Zampini } 317fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 318fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 319fd14bc51SStefano Zampini } 32008122e43SStefano Zampini 32108122e43SStefano Zampini if (mss) { 32208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 32308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 324f6f667cfSStefano Zampini /* destroy matrices (junk) */ 325f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 326f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 32708122e43SStefano Zampini } 328f6f667cfSStefano Zampini if (allocated_S_St) { 329f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 330f6f667cfSStefano Zampini } 331f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 33208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 33308122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33408122e43SStefano Zampini #endif 33508122e43SStefano Zampini if (pcbddc->dbg_flag) { 3361b968477SStefano Zampini PetscInt maxneigs_r; 337*b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3389b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33908122e43SStefano Zampini } 34008122e43SStefano Zampini PetscFunctionReturn(0); 34108122e43SStefano Zampini } 342b1b3d7a2SStefano Zampini 343674ae819SStefano Zampini #undef __FUNCT__ 344c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 345c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 346c8587f34SStefano Zampini { 347c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3488629588bSStefano Zampini PetscScalar *coarse_submat_vals; 349c8587f34SStefano Zampini PetscErrorCode ierr; 350c8587f34SStefano Zampini 351c8587f34SStefano Zampini PetscFunctionBegin; 352f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3535e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 354c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 355c8587f34SStefano Zampini 356684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3570fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 358684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 359c8587f34SStefano Zampini 360c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 361b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 362c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 363c8587f34SStefano Zampini } 364c8587f34SStefano Zampini 3658629588bSStefano Zampini /* 3668629588bSStefano Zampini Setup local correction and local part of coarse basis. 3678629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3688629588bSStefano Zampini */ 36947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3708629588bSStefano Zampini 3718629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3728629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3738629588bSStefano Zampini 3748629588bSStefano Zampini /* free */ 3758629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 376c8587f34SStefano Zampini PetscFunctionReturn(0); 377c8587f34SStefano Zampini } 378c8587f34SStefano Zampini 379c8587f34SStefano Zampini #undef __FUNCT__ 380674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 381674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 382674ae819SStefano Zampini { 383674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 384674ae819SStefano Zampini PetscErrorCode ierr; 385674ae819SStefano Zampini 386674ae819SStefano Zampini PetscFunctionBegin; 387674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 388674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 389674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 390674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 391785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 392674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 393f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 394f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 395785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 39763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 398674ae819SStefano Zampini PetscFunctionReturn(0); 399674ae819SStefano Zampini } 400674ae819SStefano Zampini 401674ae819SStefano Zampini #undef __FUNCT__ 402674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 403674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 404674ae819SStefano Zampini { 405674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 406674ae819SStefano Zampini PetscErrorCode ierr; 407674ae819SStefano Zampini 408674ae819SStefano Zampini PetscFunctionBegin; 409b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 410674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 411674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 412674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 413b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 414674ae819SStefano Zampini PetscFunctionReturn(0); 415674ae819SStefano Zampini } 416674ae819SStefano Zampini 417674ae819SStefano Zampini #undef __FUNCT__ 418674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 419674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 420674ae819SStefano Zampini { 421674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 42206656605SStefano Zampini PetscScalar *array; 423674ae819SStefano Zampini PetscErrorCode ierr; 424674ae819SStefano Zampini 425674ae819SStefano Zampini PetscFunctionBegin; 426674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42906656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43058da7f69SStefano Zampini } 431674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 432674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 43315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 437674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43806656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4418ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 442674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 443674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 444674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 445f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 446f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 448f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 449727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4500e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 451f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 45270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4536e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 454674ae819SStefano Zampini PetscFunctionReturn(0); 455674ae819SStefano Zampini } 456674ae819SStefano Zampini 457674ae819SStefano Zampini #undef __FUNCT__ 458f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 459f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4606bfb1811SStefano Zampini { 4616bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4626bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4636bfb1811SStefano Zampini VecType impVecType; 464e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4656bfb1811SStefano Zampini PetscErrorCode ierr; 4666bfb1811SStefano Zampini 4676bfb1811SStefano Zampini PetscFunctionBegin; 4686c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 469e7b262bdSStefano Zampini /* get sizes */ 470b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 471b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4726bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 473e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 474e7b262bdSStefano Zampini /* R nodes */ 475e7b262bdSStefano Zampini old_size = -1; 476e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 477e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 478e7b262bdSStefano Zampini } 479e7b262bdSStefano Zampini if (n_R != old_size) { 480e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 481e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4826bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4846bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4856bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 486e7b262bdSStefano Zampini } 487e7b262bdSStefano Zampini /* local primal dofs */ 488e7b262bdSStefano Zampini old_size = -1; 489e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 490e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 491e7b262bdSStefano Zampini } 492e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 493e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 495e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4966bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 497e7b262bdSStefano Zampini } 498e7b262bdSStefano Zampini /* local explicit constraints */ 499e7b262bdSStefano Zampini old_size = -1; 500e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 501e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 504e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50683b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50783b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50883b7ccabSStefano Zampini } 5096bfb1811SStefano Zampini PetscFunctionReturn(0); 5106bfb1811SStefano Zampini } 5116bfb1811SStefano Zampini 5126bfb1811SStefano Zampini #undef __FUNCT__ 51347f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51447f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51588ebb749SStefano Zampini { 51625084f0cSStefano Zampini PetscErrorCode ierr; 51725084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51888ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51988ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 520d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 52125084f0cSStefano Zampini /* submatrices of local problem */ 52280677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52306656605SStefano Zampini /* submatrices of local coarse problem */ 52406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52525084f0cSStefano Zampini /* working matrices */ 52606656605SStefano Zampini Mat C_CR; 52725084f0cSStefano Zampini /* additional working stuff */ 52806656605SStefano Zampini PC pc_R; 52906656605SStefano Zampini Mat F; 53006656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 53106656605SStefano Zampini 53225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53306656605SStefano Zampini PetscScalar *work; 53406656605SStefano Zampini PetscInt *idx_V_B; 53506656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 537b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53845a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53988ebb749SStefano Zampini MatType impMatType; 54025084f0cSStefano Zampini /* some shortcuts to scalars */ 54106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54288ebb749SStefano Zampini 54388ebb749SStefano Zampini PetscFunctionBegin; 544b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54588ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54688ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 547b371cd4fSStefano Zampini n_B = pcis->n_B; 548b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54988ebb749SStefano Zampini n_R = pcis->n - n_vertices; 55088ebb749SStefano Zampini 55188ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55288ebb749SStefano Zampini impMatType = MATSEQDENSE; 55388ebb749SStefano Zampini 55488ebb749SStefano Zampini /* vertices in boundary numbering */ 555785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5560e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 5576c4ed002SBarry 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); 55888ebb749SStefano Zampini 55906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56006656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56906656605SStefano Zampini 57006656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57106656605SStefano Zampini /* allocate workspace */ 57206656605SStefano Zampini n = 0; 57306656605SStefano Zampini if (n_constraints) { 57406656605SStefano Zampini n += n_R*n_constraints; 57506656605SStefano Zampini } 57606656605SStefano Zampini if (n_vertices) { 57706656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57880677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57906656605SStefano Zampini } 5803301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 58106656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58206656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58306656605SStefano Zampini } 58406656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58506656605SStefano Zampini 58606656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58706656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59106656605SStefano Zampini if (isLU || isILU || isCHOL) { 59206656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 593d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 594d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 595d62866d3SStefano Zampini MatFactorType type; 596d62866d3SStefano Zampini 5976816873aSStefano Zampini F = reuse_mumps->F; 5986816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 599d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 60006656605SStefano Zampini } else { 60106656605SStefano Zampini F = NULL; 60206656605SStefano Zampini } 60306656605SStefano Zampini 60488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 60588ebb749SStefano Zampini if (n_constraints) { 60606656605SStefano Zampini Mat M1,M2,M3; 60780677318SStefano Zampini Mat auxmat; 60806656605SStefano Zampini IS is_aux; 60980677318SStefano Zampini PetscScalar *array,*array2; 61006656605SStefano Zampini 611f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 61388ebb749SStefano Zampini 61425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 61525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6168ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 61780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61888ebb749SStefano Zampini 61980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 62080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62106656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 62306656605SStefano Zampini const PetscScalar *row_cmat_values; 62406656605SStefano Zampini const PetscInt *row_cmat_indices; 62506656605SStefano Zampini PetscInt size_of_constraint,j; 62688ebb749SStefano Zampini 62706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62906656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63006656605SStefano Zampini } 63106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63206656605SStefano Zampini } 63380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 63406656605SStefano Zampini if (F) { 63506656605SStefano Zampini Mat B; 63606656605SStefano Zampini 63706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 63880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64006656605SStefano Zampini } else { 64180677318SStefano Zampini PetscScalar *marr; 64280677318SStefano Zampini 64380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 64506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 64680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 64706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65006656605SStefano Zampini } 65180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65206656605SStefano Zampini } 65380677318SStefano Zampini if (!pcbddc->switch_static) { 65480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 66080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 66380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66480677318SStefano Zampini } 66580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66780677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66880677318SStefano Zampini } else { 66980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 67080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67125084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67280677318SStefano Zampini } 67380677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 67480677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 67580677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 67706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 67880677318SStefano Zampini if (isCHOL) { 67980677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 68080677318SStefano Zampini } else { 68125084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68280677318SStefano Zampini } 68380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 68406656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 68525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 68625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 68725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 68880677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68980677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 69080677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 693f4ddd8eeSStefano Zampini } 69488ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 69588ebb749SStefano Zampini if (n_vertices) { 69606656605SStefano Zampini IS is_aux; 6973a50541eSStefano Zampini 6986816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 6996816873aSStefano Zampini IS tis; 7006816873aSStefano Zampini 7016816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7026816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7036816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7046816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7056816873aSStefano Zampini } else { 7063a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7076816873aSStefano Zampini } 7089577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7099577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 71004708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 71125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 71288ebb749SStefano Zampini } 71388ebb749SStefano Zampini 71488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 715f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 71606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 71706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 71806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 71906656605SStefano Zampini } 720f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 72106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 72206656605SStefano Zampini PetscScalar *marray; 72306656605SStefano Zampini 72406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 72506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 726f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 727f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 728f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 729f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 730f4ddd8eeSStefano Zampini } 731f4ddd8eeSStefano Zampini } 73206656605SStefano Zampini 733f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 73406656605SStefano Zampini PetscScalar *marray; 73588ebb749SStefano Zampini 73606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 73988ebb749SStefano Zampini } 7403301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 74106656605SStefano Zampini n *= 2; 74288ebb749SStefano Zampini } 74306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 74406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 74506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7468eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 74806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74988ebb749SStefano Zampini } 7503301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7528eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 75588ebb749SStefano Zampini } 75688ebb749SStefano Zampini } else { 757c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 758c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7591b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 760c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 761c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 762c0553b1fSStefano Zampini } 76388ebb749SStefano Zampini } 76406656605SStefano Zampini } 76506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 76606656605SStefano Zampini /* vertices */ 76706656605SStefano Zampini if (n_vertices) { 76816f15bc4SStefano Zampini 76904708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 77004708bb6SStefano Zampini 77116f15bc4SStefano Zampini if (n_R) { 77206656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 77306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 77416f15bc4SStefano Zampini PetscScalar *x,*y; 77504708bb6SStefano Zampini PetscBool isseqaij; 77606656605SStefano Zampini 77721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 77806656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 77904708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7806816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 78106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 78206656605SStefano Zampini } else { 78306656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 78406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 78506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 78606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 78706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 78806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 78906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 79006656605SStefano Zampini } 79106656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 79206656605SStefano Zampini } 79380677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 79406656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 79506656605SStefano Zampini if (n_constraints) { 79606656605SStefano Zampini Mat B; 79780677318SStefano Zampini 798b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 79980677318SStefano Zampini for (i=0;i<n_vertices;i++) { 80080677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 80180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 80280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 80380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 80480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 80580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80680677318SStefano Zampini } 80780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 80880677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 80980677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81180677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 81206656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 81306656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 81406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 81506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81606656605SStefano Zampini } 81704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 81804708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 81904708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 82004708bb6SStefano Zampini } 82106656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 82280677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 82306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 82406656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 82506656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 82606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 82706656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83006656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 83116f15bc4SStefano Zampini } else { 83216f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83316f15bc4SStefano Zampini } 83421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 83506656605SStefano Zampini /* coarse basis functions */ 83606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 83716f15bc4SStefano Zampini PetscScalar *y; 83816f15bc4SStefano Zampini 83906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 84506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 84706656605SStefano Zampini 84806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 84906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85506656605SStefano Zampini } 85606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 85706656605SStefano Zampini } 85804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 85904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 86006656605SStefano Zampini } 86106656605SStefano Zampini 86206656605SStefano Zampini if (n_constraints) { 86306656605SStefano Zampini Mat B; 86406656605SStefano Zampini 86506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86906656605SStefano Zampini if (n_vertices) { 87080677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 87180677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 87280677318SStefano Zampini } else { 87380677318SStefano Zampini Mat S_VCt; 87480677318SStefano Zampini 87580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 87680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 87780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 87880677318SStefano Zampini } 87906656605SStefano Zampini } 88006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88106656605SStefano Zampini /* coarse basis functions */ 88206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 88306656605SStefano Zampini PetscScalar *y; 88406656605SStefano Zampini 88506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 88606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 88806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 89206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 89306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 89506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 89806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89906656605SStefano Zampini } 90006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90106656605SStefano Zampini } 90206656605SStefano Zampini } 90380677318SStefano Zampini if (n_constraints) { 90480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 90580677318SStefano Zampini } 90606656605SStefano Zampini 90706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9083301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 90906656605SStefano Zampini 91006656605SStefano Zampini if (n_constraints) { 91116f15bc4SStefano Zampini Mat S_CCT,B_C; 91206656605SStefano Zampini 91380677318SStefano Zampini /* this is a lazy thing */ 91480677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 91816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 91906656605SStefano Zampini if (n_vertices) { 92016f15bc4SStefano Zampini Mat B_V,S_VCT; 92106656605SStefano Zampini 92206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 92616f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 92706656605SStefano Zampini } 92806656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 92980677318SStefano Zampini } else { /* if there are no constraints, reset work */ 93080677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 93106656605SStefano Zampini } 93216f15bc4SStefano Zampini if (n_vertices && n_R) { 93306656605SStefano Zampini Mat A_VRT; 93480677318SStefano Zampini PetscScalar *marray; 93506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 93606656605SStefano Zampini 93780677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 93880677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 93980677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 94006656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 94180677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 94280677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 94316f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 94406656605SStefano Zampini } 94506656605SStefano Zampini 94606656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 94706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 94806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95306656605SStefano Zampini } 95406656605SStefano Zampini } else { 95506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96106656605SStefano Zampini } 96206656605SStefano Zampini } 96306656605SStefano Zampini /* coarse basis functions */ 96406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96506656605SStefano Zampini PetscScalar *y; 96606656605SStefano Zampini 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 96806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97206656605SStefano Zampini if (i<n_vertices) { 97306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 97406656605SStefano Zampini } 97506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 97706656605SStefano Zampini 97806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 97906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98506656605SStefano Zampini } 98606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98706656605SStefano Zampini } 98806656605SStefano Zampini } 989d62866d3SStefano Zampini /* free memory */ 99088ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 995d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 996d62866d3SStefano Zampini if (n_vertices) { 997d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 998d62866d3SStefano Zampini } 999d62866d3SStefano Zampini if (n_constraints) { 1000d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1001d62866d3SStefano Zampini } 100288ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 100388ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 100488ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 100525084f0cSStefano Zampini if (pcbddc->dbg_flag) { 100688ebb749SStefano Zampini Mat coarse_sub_mat; 100725084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 100888ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 100988ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 101088ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10118bec7fa6SStefano Zampini Mat C_B,CPHI; 10128bec7fa6SStefano Zampini IS is_dummy; 10138bec7fa6SStefano Zampini Vec mones; 101488ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 101588ebb749SStefano Zampini PetscReal real_value; 101688ebb749SStefano Zampini 101788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 101888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 101988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 102088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 102188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 102288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1023c0553b1fSStefano Zampini if (unsymmetric_check) { 102488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 102588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 102688ebb749SStefano Zampini } 102788ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 102888ebb749SStefano Zampini 102925084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10303301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 103125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1032c0553b1fSStefano Zampini if (unsymmetric_check) { 103388ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104588ebb749SStefano Zampini } else { 104688ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105488ebb749SStefano Zampini } 105588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 105981d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10608bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10611575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 106206656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10638bec7fa6SStefano Zampini 10648bec7fa6SStefano Zampini /* check constraints */ 10658bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10668bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10678bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10688bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10698bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10718bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1072bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1073c0553b1fSStefano Zampini if (unsymmetric_check) { 1074bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1075bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1076bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1077bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1078bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 107988ebb749SStefano Zampini } 10808bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10818bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10838bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10848bec7fa6SStefano Zampini 108525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 108688ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 108788ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 108888ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 108988ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 109188ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 109288ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 109388ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 109488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 109588ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1096c0553b1fSStefano Zampini if (unsymmetric_check) { 109788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 109888ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 109988ebb749SStefano Zampini } 110088ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 110188ebb749SStefano Zampini } 11028629588bSStefano Zampini /* get back data */ 11038629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 110488ebb749SStefano Zampini PetscFunctionReturn(0); 110588ebb749SStefano Zampini } 110688ebb749SStefano Zampini 110788ebb749SStefano Zampini #undef __FUNCT__ 1108d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1109d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1110aa0d41d4SStefano Zampini { 1111d65f70fdSStefano Zampini Mat *work_mat; 1112d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1113d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1114d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1115aa0d41d4SStefano Zampini PetscErrorCode ierr; 1116aa0d41d4SStefano Zampini 1117aa0d41d4SStefano Zampini PetscFunctionBegin; 1118d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1119d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1120d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1121d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1122aa0d41d4SStefano Zampini 1123d65f70fdSStefano Zampini if (!rsorted) { 1124906d46d4SStefano Zampini const PetscInt *idxs; 1125906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1126aa0d41d4SStefano Zampini 1127d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1128d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1129d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1130d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1131aa0d41d4SStefano Zampini } 1132d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1134d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1135d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1136aa0d41d4SStefano Zampini } 1137d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1138d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini } else { 1140d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini isrow_s = isrow; 1142aa0d41d4SStefano Zampini } 1143906d46d4SStefano Zampini 1144d65f70fdSStefano Zampini if (!csorted) { 1145d65f70fdSStefano Zampini if (isrow == iscol) { 1146d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini iscol_s = isrow_s; 1148d65f70fdSStefano Zampini } else { 1149d65f70fdSStefano Zampini const PetscInt *idxs; 1150d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1151906d46d4SStefano Zampini 1152d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1155d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1156d65f70fdSStefano Zampini } 1157d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1160d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1161d65f70fdSStefano Zampini } 1162d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1163d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1164d65f70fdSStefano Zampini } 1165d65f70fdSStefano Zampini } else { 1166d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1167d65f70fdSStefano Zampini iscol_s = iscol; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini 1170d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini 1172d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1173906d46d4SStefano Zampini Mat new_mat; 1174d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1175906d46d4SStefano Zampini 1176d65f70fdSStefano Zampini if (!rsorted) { 1177d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1178d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1180d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1181906d46d4SStefano Zampini } 1182d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini } else { 1185d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1186906d46d4SStefano Zampini } 1187d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1188d65f70fdSStefano Zampini 1189d65f70fdSStefano Zampini if (!csorted) { 1190d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1191d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1192d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1193d65f70fdSStefano Zampini } else { 1194d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1195d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1197d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1198d65f70fdSStefano Zampini } 1199d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1201d65f70fdSStefano Zampini } 1202d65f70fdSStefano Zampini } else { 1203d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini } 1205d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1206d65f70fdSStefano Zampini 1207d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini work_mat[0] = new_mat; 1210d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini } 1213d65f70fdSStefano Zampini 1214d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1215d65f70fdSStefano Zampini *B = work_mat[0]; 1216d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1217d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini PetscFunctionReturn(0); 1220d65f70fdSStefano Zampini } 1221d65f70fdSStefano Zampini 1222d65f70fdSStefano Zampini #undef __FUNCT__ 12235e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12245e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1225aa0d41d4SStefano Zampini { 1226aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12273bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 12285e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1229d65f70fdSStefano Zampini Mat new_mat; 12305e8657edSStefano Zampini IS is_local,is_global; 1231d65f70fdSStefano Zampini PetscInt local_size; 1232d65f70fdSStefano Zampini PetscBool isseqaij; 1233aa0d41d4SStefano Zampini PetscErrorCode ierr; 1234aa0d41d4SStefano Zampini 1235aa0d41d4SStefano Zampini PetscFunctionBegin; 1236aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12375e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12385e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 12393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1240aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1241d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1242aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1243906d46d4SStefano Zampini 1244906d46d4SStefano Zampini /* check */ 1245906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1246906d46d4SStefano Zampini Vec x,x_change; 1247906d46d4SStefano Zampini PetscReal error; 1248906d46d4SStefano Zampini 12495e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1250906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12515e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1252e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1253e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1254d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1255e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1256e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1257906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1258906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1259906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1260906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1261906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1262906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1263906d46d4SStefano Zampini } 1264906d46d4SStefano Zampini 126522d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12669b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 126722d5777bSStefano Zampini if (isseqaij) { 1268d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1269aa0d41d4SStefano Zampini } else { 1270aa0d41d4SStefano Zampini Mat work_mat; 1271aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1272d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1273aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1274aa0d41d4SStefano Zampini } 12753301b35fSStefano Zampini if (matis->A->symmetric_set) { 12763301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1277e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12783301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1279e496cd5dSStefano Zampini #endif 12803301b35fSStefano Zampini } 128145a1bb75SStefano Zampini /* 128245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1283d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 128445a1bb75SStefano Zampini */ 1285d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1286aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1287aa0d41d4SStefano Zampini } 1288aa0d41d4SStefano Zampini 1289aa0d41d4SStefano Zampini #undef __FUNCT__ 1290a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12918ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1292a64d13efSStefano Zampini { 1293a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1294a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1295d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 129653892102SStefano Zampini PetscInt *idx_R_local=NULL; 12973a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12983a50541eSStefano Zampini PetscInt vbs,bs; 12996816873aSStefano Zampini PetscBT bitmask=NULL; 1300a64d13efSStefano Zampini PetscErrorCode ierr; 1301a64d13efSStefano Zampini 1302a64d13efSStefano Zampini PetscFunctionBegin; 1303b23d619eSStefano Zampini /* 1304b23d619eSStefano Zampini No need to setup local scatters if 1305b23d619eSStefano Zampini - primal space is unchanged 1306b23d619eSStefano Zampini AND 1307b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1308b23d619eSStefano Zampini AND 1309b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1310b23d619eSStefano Zampini */ 1311b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1312f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1313f4ddd8eeSStefano Zampini } 1314f4ddd8eeSStefano Zampini /* destroy old objects */ 1315f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1316f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1317f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1318a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1319b371cd4fSStefano Zampini n_B = pcis->n_B; 1320b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1321b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13223a50541eSStefano Zampini 1323a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13246816873aSStefano Zampini 132553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13266816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1327854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1328a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1329a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13300e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1331a64d13efSStefano Zampini } 1332a64d13efSStefano Zampini 1333a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13344641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13356816873aSStefano Zampini idx_R_local[n_R++] = i; 1336a64d13efSStefano Zampini } 1337a64d13efSStefano Zampini } 133853892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13396816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13406816873aSStefano Zampini 134153892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 134253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13436816873aSStefano Zampini } 13443a50541eSStefano Zampini 13453a50541eSStefano Zampini /* Block code */ 13463a50541eSStefano Zampini vbs = 1; 13473a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13483a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13493a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13503a50541eSStefano Zampini PetscInt *vary; 1351d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1352785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13533a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1354d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1355d3df7717SStefano 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 */ 13560e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1357d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13583a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13593a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13603a50541eSStefano Zampini break; 13613a50541eSStefano Zampini } 13623a50541eSStefano Zampini } 1363d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1364d3df7717SStefano Zampini } else { 1365d3df7717SStefano Zampini /* Verify directly the R set */ 1366d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1367d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1368d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1369d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1370d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1371d3df7717SStefano Zampini break; 1372d3df7717SStefano Zampini } 1373d3df7717SStefano Zampini } 1374d3df7717SStefano Zampini } 1375d3df7717SStefano Zampini } 13763a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13773a50541eSStefano Zampini vbs = bs; 13783a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13793a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13803a50541eSStefano Zampini } 13813a50541eSStefano Zampini } 13823a50541eSStefano Zampini } 13833a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 138453892102SStefano Zampini if (sub_schurs->reuse_mumps) { 138553892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 138653892102SStefano Zampini 138753892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 138853892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 138953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 139053892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 139153892102SStefano Zampini } else { 13923a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 139353892102SStefano Zampini } 1394a64d13efSStefano Zampini 1395a64d13efSStefano Zampini /* print some info if requested */ 1396a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1397a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1398a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13991575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1400a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1401a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14023a50541eSStefano 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); 1403a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1404a64d13efSStefano Zampini } 1405a64d13efSStefano Zampini 1406a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14076816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14086816873aSStefano Zampini IS is_aux1,is_aux2; 14096816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14106816873aSStefano Zampini 14113a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1412854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1413854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1414a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14154641a718SStefano Zampini for (i=0; i<n_D; i++) { 14164641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14174641a718SStefano Zampini } 1418a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1419a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14204641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14214641a718SStefano Zampini aux_array1[j++] = i; 1422a64d13efSStefano Zampini } 1423a64d13efSStefano Zampini } 1424a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1425a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1426a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14274641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14284641a718SStefano Zampini aux_array2[j++] = i; 1429a64d13efSStefano Zampini } 1430a64d13efSStefano Zampini } 1431a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1432a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1433a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1434a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1435a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1436a64d13efSStefano Zampini 14378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1438785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1439a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14404641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14414641a718SStefano Zampini aux_array1[j++] = i; 1442a64d13efSStefano Zampini } 1443a64d13efSStefano Zampini } 1444a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1445a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1446a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1447a64d13efSStefano Zampini } 14484641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14493a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1450d62866d3SStefano Zampini } else { 145153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14526816873aSStefano Zampini IS tis; 14536816873aSStefano Zampini PetscInt schur_size; 14546816873aSStefano Zampini 145553892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14566816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 145753892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14586816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14596816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14606816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14616816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14626816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1463d62866d3SStefano Zampini } 1464d62866d3SStefano Zampini } 1465a64d13efSStefano Zampini PetscFunctionReturn(0); 1466a64d13efSStefano Zampini } 1467a64d13efSStefano Zampini 1468304d26faSStefano Zampini 1469304d26faSStefano Zampini #undef __FUNCT__ 1470304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1471684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1472304d26faSStefano Zampini { 1473304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1474304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1475304d26faSStefano Zampini PC pc_temp; 1476304d26faSStefano Zampini Mat A_RR; 1477f4ddd8eeSStefano Zampini MatReuse reuse; 1478304d26faSStefano Zampini PetscScalar m_one = -1.0; 1479304d26faSStefano Zampini PetscReal value; 148004708bb6SStefano Zampini PetscInt n_D,n_R; 14819577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1482304d26faSStefano Zampini PetscErrorCode ierr; 1483e604994aSStefano Zampini /* prefixes stuff */ 1484312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1485e604994aSStefano Zampini size_t len; 1486304d26faSStefano Zampini 1487304d26faSStefano Zampini PetscFunctionBegin; 1488304d26faSStefano Zampini 1489e604994aSStefano Zampini /* compute prefixes */ 1490e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1491e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1492e604994aSStefano Zampini if (!pcbddc->current_level) { 1493e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1494e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1495e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1496e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1497e604994aSStefano Zampini } else { 1498e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1499312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1500e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1501e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1502312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1503312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 150434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 150534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1506e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1507e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1508e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1509e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1510e604994aSStefano Zampini } 1511e604994aSStefano Zampini 1512304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1513684f6988SStefano Zampini if (dirichlet) { 1514d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15153301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15163301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1517964fefecSStefano Zampini } 1518ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1519964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1520304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1521304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1522304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1523304d26faSStefano Zampini /* default */ 1524304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1525e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15269577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1527304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15289577ea80SStefano Zampini if (issbaij) { 15299577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15309577ea80SStefano Zampini } else { 1531304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15329577ea80SStefano Zampini } 1533304d26faSStefano Zampini /* Allow user's customization */ 1534304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1535304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1536304d26faSStefano Zampini } 1537d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1538d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1539d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1540d62866d3SStefano Zampini 1541d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1542d5574798SStefano Zampini } 1543304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1544304d26faSStefano Zampini if (!n_D) { 1545304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1546304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1547304d26faSStefano Zampini } 1548304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1549304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1550304d26faSStefano Zampini /* set ksp_D into pcis data */ 1551304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1552304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1553304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1554684f6988SStefano Zampini } 1555304d26faSStefano Zampini 1556304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1557684f6988SStefano Zampini A_RR = 0; 1558684f6988SStefano Zampini if (neumann) { 1559d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 156004708bb6SStefano Zampini PetscInt ibs,mbs; 156104708bb6SStefano Zampini PetscBool issbaij; 156204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1563f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15648ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1565f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1566f4ddd8eeSStefano Zampini PetscInt nn_R; 156781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1568f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1569f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1570f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1571f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1572f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1573f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1574f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1575727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1576f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1578f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1579f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1580f4ddd8eeSStefano Zampini } 1581f4ddd8eeSStefano Zampini } 1582f4ddd8eeSStefano Zampini /* last check */ 1583d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1584f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1585f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1586f4ddd8eeSStefano Zampini } 1587f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1588f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1589f4ddd8eeSStefano Zampini } 1590f4ddd8eeSStefano Zampini /* extract A_RR */ 1591af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1592af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 159304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 159404708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 159504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 159604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 159704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1598af732b37SStefano Zampini } else { 159904708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16006816873aSStefano Zampini } 160104708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 160204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 160304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 160404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 160504708bb6SStefano Zampini } else { 160604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 160704708bb6SStefano Zampini } 160804708bb6SStefano Zampini } 160904708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1610f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16113301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16123301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16136816873aSStefano Zampini } 16146816873aSStefano Zampini } else { 16156816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16166816873aSStefano Zampini 16176816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16186816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16196816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1620af732b37SStefano Zampini } 1621f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1622304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1623304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1624304d26faSStefano Zampini /* default */ 1625304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1626e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1627304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16289577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16299577ea80SStefano Zampini if (issbaij) { 16309577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16319577ea80SStefano Zampini } else { 1632304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16339577ea80SStefano Zampini } 1634304d26faSStefano Zampini /* Allow user's customization */ 1635304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1636304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1637304d26faSStefano Zampini } 1638d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1639304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1640304d26faSStefano Zampini if (!n_R) { 1641304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1642304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1643304d26faSStefano Zampini } 1644d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1645d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1646d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1647d62866d3SStefano Zampini 1648d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1649d62866d3SStefano Zampini } 1650304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1651304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1652684f6988SStefano Zampini } 16536816873aSStefano Zampini /* free Neumann problem's matrix */ 16546816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1655304d26faSStefano Zampini 1656304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16570fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1658684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1659684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16601575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1661684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1662684f6988SStefano Zampini } 1663684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16640fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16650fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16660fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16670fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16680fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1669304d26faSStefano Zampini /* need to be adapted? */ 1670b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1671*b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1672b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1673304d26faSStefano Zampini /* print info */ 1674304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1675e604994aSStefano 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); 1676304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1677304d26faSStefano Zampini } 1678b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1679298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1680304d26faSStefano Zampini } 1681684f6988SStefano Zampini } 1682684f6988SStefano Zampini if (neumann) { /* Neumann */ 16836816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16870fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16880fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1689304d26faSStefano Zampini /* need to be adapted? */ 1690b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1691*b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1692304d26faSStefano Zampini /* print info */ 1693304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1694e604994aSStefano 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); 1695304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1696304d26faSStefano Zampini } 1697b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1698298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1699304d26faSStefano Zampini } 17000fccc4e9SStefano Zampini } 1701684f6988SStefano Zampini } 1702304d26faSStefano Zampini PetscFunctionReturn(0); 1703304d26faSStefano Zampini } 1704304d26faSStefano Zampini 1705304d26faSStefano Zampini #undef __FUNCT__ 1706ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 170780677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1708674ae819SStefano Zampini { 1709674ae819SStefano Zampini PetscErrorCode ierr; 1710674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1711be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1712674ae819SStefano Zampini 1713674ae819SStefano Zampini PetscFunctionBegin; 1714be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 171620c7b377SStefano Zampini } 171780677318SStefano Zampini if (!pcbddc->switch_static) { 171880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 171980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 172080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172120c7b377SStefano Zampini } 1722be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172520c7b377SStefano Zampini } else { 1726be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1727be83ff47SStefano Zampini 172853892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172953892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 173020c7b377SStefano Zampini } 1731be83ff47SStefano Zampini } else { 173280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 173880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1741674ae819SStefano Zampini } 1742674ae819SStefano Zampini } 1743be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174480677318SStefano Zampini if (applytranspose) { 174580677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174680677318SStefano Zampini } else { 174780677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174880677318SStefano Zampini } 174953892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1750be83ff47SStefano Zampini } else { 1751be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1752be83ff47SStefano Zampini 1753be83ff47SStefano Zampini if (applytranspose) { 175453892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1755be83ff47SStefano Zampini } else { 175653892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1757be83ff47SStefano Zampini } 175853892102SStefano Zampini #endif 1759be83ff47SStefano Zampini } 176080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 176180677318SStefano Zampini if (!pcbddc->switch_static) { 1762be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1765be83ff47SStefano Zampini } else { 1766be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1767be83ff47SStefano Zampini 176853892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 176953892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1770be83ff47SStefano Zampini } 177180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 177480677318SStefano Zampini } 177580677318SStefano Zampini } else { 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); 178080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 178380677318SStefano Zampini } 178480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1788674ae819SStefano Zampini } 1789674ae819SStefano Zampini PetscFunctionReturn(0); 1790674ae819SStefano Zampini } 1791674ae819SStefano Zampini 1792dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1793674ae819SStefano Zampini #undef __FUNCT__ 1794674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1795dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1796674ae819SStefano Zampini { 1797674ae819SStefano Zampini PetscErrorCode ierr; 1798674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1799674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1800674ae819SStefano Zampini const PetscScalar zero = 0.0; 1801674ae819SStefano Zampini 1802674ae819SStefano Zampini PetscFunctionBegin; 1803dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1804dc359a40SStefano Zampini if (applytranspose) { 1805674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18068eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1807dc359a40SStefano Zampini } else { 1808674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1809674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 181015aaf578SStefano Zampini } 181112edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 181212edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 181312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181512edc857SStefano Zampini 18169f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 181712edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 181812edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1819964fefecSStefano Zampini Vec rhs,sol; 1820964fefecSStefano Zampini 1821964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1822964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 182312edc857SStefano Zampini if (applytranspose) { 1824964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182512edc857SStefano Zampini } else { 1826964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182712edc857SStefano Zampini } 182812edc857SStefano Zampini } 1829674ae819SStefano Zampini 1830674ae819SStefano Zampini /* Local solution on R nodes */ 183180677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 183280677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18339f00e9b4SStefano Zampini } 1834674ae819SStefano Zampini 18359f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18369f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 183712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1838674ae819SStefano Zampini 1839674ae819SStefano Zampini /* Sum contributions from two levels */ 1840dc359a40SStefano Zampini if (applytranspose) { 1841dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1842dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1843dc359a40SStefano Zampini } else { 1844674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18458eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1846dc359a40SStefano Zampini } 1847674ae819SStefano Zampini PetscFunctionReturn(0); 1848674ae819SStefano Zampini } 1849674ae819SStefano Zampini 1850674ae819SStefano Zampini #undef __FUNCT__ 1851674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 185212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1853674ae819SStefano Zampini { 1854674ae819SStefano Zampini PetscErrorCode ierr; 1855674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 185658da7f69SStefano Zampini PetscScalar *array; 185712edc857SStefano Zampini Vec from,to; 1858674ae819SStefano Zampini 1859674ae819SStefano Zampini PetscFunctionBegin; 186012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 186112edc857SStefano Zampini from = pcbddc->coarse_vec; 186212edc857SStefano Zampini to = pcbddc->vec1_P; 186312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 186412edc857SStefano Zampini Vec tvec; 186558da7f69SStefano Zampini 186658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 186812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 187058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 187158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 187212edc857SStefano Zampini } 187312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187412edc857SStefano Zampini from = pcbddc->vec1_P; 187512edc857SStefano Zampini to = pcbddc->coarse_vec; 187612edc857SStefano Zampini } 187712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1878674ae819SStefano Zampini PetscFunctionReturn(0); 1879674ae819SStefano Zampini } 1880674ae819SStefano Zampini 1881674ae819SStefano Zampini #undef __FUNCT__ 1882674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 188312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1884674ae819SStefano Zampini { 1885674ae819SStefano Zampini PetscErrorCode ierr; 1886674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 188758da7f69SStefano Zampini PetscScalar *array; 188812edc857SStefano Zampini Vec from,to; 1889674ae819SStefano Zampini 1890674ae819SStefano Zampini PetscFunctionBegin; 189112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 189212edc857SStefano Zampini from = pcbddc->coarse_vec; 189312edc857SStefano Zampini to = pcbddc->vec1_P; 189412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189512edc857SStefano Zampini from = pcbddc->vec1_P; 189612edc857SStefano Zampini to = pcbddc->coarse_vec; 189712edc857SStefano Zampini } 189812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 189912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 190012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 190112edc857SStefano Zampini Vec tvec; 190258da7f69SStefano Zampini 190312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 190458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 190558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 190658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 190758da7f69SStefano Zampini } 190858da7f69SStefano Zampini } else { 190958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 191058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 191112edc857SStefano Zampini } 191212edc857SStefano Zampini } 1913674ae819SStefano Zampini PetscFunctionReturn(0); 1914674ae819SStefano Zampini } 1915674ae819SStefano Zampini 1916984c4197SStefano Zampini /* uncomment for testing purposes */ 1917984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1918674ae819SStefano Zampini #undef __FUNCT__ 1919674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1920674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1921674ae819SStefano Zampini { 1922674ae819SStefano Zampini PetscErrorCode ierr; 1923674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1924674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1925674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1926984c4197SStefano Zampini /* one and zero */ 1927984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1928984c4197SStefano Zampini /* space to store constraints and their local indices */ 19299162d606SStefano Zampini PetscScalar *constraints_data; 19309162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19319162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19329162d606SStefano Zampini PetscInt *constraints_n; 1933984c4197SStefano Zampini /* iterators */ 1934b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1935984c4197SStefano Zampini /* BLAS integers */ 1936e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1937e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1938c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1939727cdba6SStefano Zampini /* reuse */ 19400e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19410e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1942984c4197SStefano Zampini /* change of basis */ 1943b3d85658SStefano Zampini PetscBool qr_needed; 19449162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1945984c4197SStefano Zampini /* auxiliary stuff */ 194664efe560SStefano Zampini PetscInt *nnz,*is_indices; 19478a0068c3SStefano Zampini PetscInt ncc; 1948984c4197SStefano Zampini /* some quantities */ 194945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1950a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1951984c4197SStefano Zampini 1952674ae819SStefano Zampini PetscFunctionBegin; 19538e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19548e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19558e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1956088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1957088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19580e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19590e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19600e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19610e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19620e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1963088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1964cf5a6209SStefano Zampini 1965cf5a6209SStefano Zampini /* print some info */ 1966cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1967cf5a6209SStefano Zampini IS vertices; 1968cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1969cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1970cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1971cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 19721575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1973cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1974cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1975fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1976fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1977cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 19781575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1979cf5a6209SStefano Zampini } 1980cf5a6209SStefano Zampini 1981cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19829162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1983cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1984cf5a6209SStefano Zampini const Vec *nearnullvecs; 1985cf5a6209SStefano Zampini Vec *localnearnullsp; 1986cf5a6209SStefano Zampini PetscScalar *array; 1987cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1988cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1989674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1990b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1991674ae819SStefano Zampini PetscScalar *work; 1992674ae819SStefano Zampini PetscReal *singular_vals; 1993674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1994674ae819SStefano Zampini PetscReal *rwork; 1995674ae819SStefano Zampini #endif 1996674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1997674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1998674ae819SStefano Zampini #else 1999964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2000964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2001674ae819SStefano Zampini #endif 2002674ae819SStefano Zampini 2003674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2004d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2005d06fc5fdSStefano Zampini /* free unneeded index sets */ 2006d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2007d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2008674ae819SStefano Zampini } 2009d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2010d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2011d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2012d06fc5fdSStefano Zampini } 2013d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2014d06fc5fdSStefano Zampini n_ISForEdges = 0; 2015d06fc5fdSStefano Zampini } 2016d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2017d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2018d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2019d06fc5fdSStefano Zampini } 2020d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2021d06fc5fdSStefano Zampini n_ISForFaces = 0; 2022d06fc5fdSStefano Zampini } 202370022509SStefano Zampini 202470022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 202570022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 202670022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 202770022509SStefano Zampini if (pcbddc->NullSpace) { 202870022509SStefano Zampini PetscBool tbool[2],gbool[2]; 202970022509SStefano Zampini 203070022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2031b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2032d06fc5fdSStefano Zampini if (!ISForEdges) { 2033d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2034d06fc5fdSStefano Zampini } 2035b8ffe317SStefano Zampini } 2036d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2037d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2038*b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2039d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2040d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 204198a51de6SStefano Zampini } 204270022509SStefano Zampini #endif 204308122e43SStefano Zampini 2044674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2045674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2046674ae819SStefano Zampini if (nearnullsp) { 2047674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini /* remove any stored info */ 2049f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2050f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2051f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2052f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2053f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2054473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2055f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2056f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2057f4ddd8eeSStefano Zampini } 2058984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2059984c4197SStefano Zampini nnsp_size = 0; 2060674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2061674ae819SStefano Zampini } 2062984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2063984c4197SStefano Zampini max_constraints = nnsp_size; 2064984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2065984c4197SStefano Zampini 2066674ae819SStefano Zampini /* 2067674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20689162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20699162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20709162d606SStefano Zampini There can be multiple constraints per connected component 2071674ae819SStefano Zampini */ 2072674ae819SStefano Zampini n_vertices = 0; 2073674ae819SStefano Zampini if (ISForVertices) { 2074674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2075674ae819SStefano Zampini } 20769162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20779162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20789162d606SStefano Zampini 20799162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20809162d606SStefano Zampini total_counts *= max_constraints; 2081674ae819SStefano Zampini total_counts += n_vertices; 20824641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20839162d606SStefano Zampini 2084674ae819SStefano Zampini total_counts = 0; 2085674ae819SStefano Zampini max_size_of_constraint = 0; 2086674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20879162d606SStefano Zampini IS used_is; 2088674ae819SStefano Zampini if (i<n_ISForEdges) { 20899162d606SStefano Zampini used_is = ISForEdges[i]; 2090674ae819SStefano Zampini } else { 20919162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2092674ae819SStefano Zampini } 20939162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2094674ae819SStefano Zampini total_counts += j; 2095674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2096674ae819SStefano Zampini } 20979162d606SStefano 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); 20989162d606SStefano Zampini 2099984c4197SStefano Zampini /* get local part of global near null space vectors */ 2100785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2101984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2102984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2103e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2104e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2105984c4197SStefano Zampini } 2106674ae819SStefano Zampini 2107242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2108242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2109a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2110242a89d7SStefano Zampini 2111984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2112a773dcb8SStefano Zampini if (!skip_lapack) { 2113674ae819SStefano Zampini PetscScalar temp_work; 2114911cabfeSStefano Zampini 2115674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2116984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2117785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2118785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2119785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2120674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2121785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2122674ae819SStefano Zampini #endif 2123674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2124c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2125c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2126674ae819SStefano Zampini lwork = -1; 2127674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2128674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2129c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2130674ae819SStefano Zampini #else 2131c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2132674ae819SStefano Zampini #endif 2133674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2134984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2135674ae819SStefano Zampini #else /* on missing GESVD */ 2136674ae819SStefano Zampini /* SVD */ 2137674ae819SStefano Zampini PetscInt max_n,min_n; 2138674ae819SStefano Zampini max_n = max_size_of_constraint; 2139984c4197SStefano Zampini min_n = max_constraints; 2140984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2141674ae819SStefano Zampini min_n = max_size_of_constraint; 2142984c4197SStefano Zampini max_n = max_constraints; 2143674ae819SStefano Zampini } 2144785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2145674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2146785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2147674ae819SStefano Zampini #endif 2148674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2149674ae819SStefano Zampini lwork = -1; 2150e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2151e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2152b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2153674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2154674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21559162d606SStefano 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)); 2156674ae819SStefano Zampini #else 21579162d606SStefano 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)); 2158674ae819SStefano Zampini #endif 2159674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2160984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2161984c4197SStefano Zampini #endif /* on missing GESVD */ 2162674ae819SStefano Zampini /* Allocate optimal workspace */ 2163674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2164854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2165674ae819SStefano Zampini } 2166674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2167674ae819SStefano Zampini total_counts = 0; 21689162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21699162d606SStefano Zampini constraints_data_ptr[0] = 0; 2170674ae819SStefano Zampini /* vertices */ 21719162d606SStefano Zampini if (n_vertices) { 2172674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21739162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21749162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2175674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21769162d606SStefano Zampini constraints_n[total_counts] = 1; 21779162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21789162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21799162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2180674ae819SStefano Zampini total_counts++; 2181674ae819SStefano Zampini } 2182674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2183984c4197SStefano Zampini PetscBool used_vertex; 2184674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2185674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2186674ae819SStefano Zampini k = 0; 2187674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2188984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2189984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21909162d606SStefano Zampini constraints_n[total_counts] = 1; 21919162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21929162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21939162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21949162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2195674ae819SStefano Zampini total_counts++; 2196674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2197674ae819SStefano Zampini } 2198984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2199674ae819SStefano Zampini k++; 2200674ae819SStefano Zampini } 2201674ae819SStefano Zampini } 2202674ae819SStefano Zampini } 2203674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2204674ae819SStefano Zampini n_vertices = total_counts; 2205674ae819SStefano Zampini } 2206984c4197SStefano Zampini 2207674ae819SStefano Zampini /* edges and faces */ 22089162d606SStefano Zampini total_counts_cc = total_counts; 2209911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22109162d606SStefano Zampini IS used_is; 22119162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22129162d606SStefano Zampini 2213911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22149162d606SStefano Zampini used_is = ISForEdges[ncc]; 2215984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2216674ae819SStefano Zampini } else { 22179162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2218984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2219674ae819SStefano Zampini } 2220674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22219162d606SStefano Zampini 22229162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22239162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2224984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2225984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2226674ae819SStefano Zampini if (nnsp_has_cnst) { 22275b08dc53SStefano Zampini PetscScalar quad_value; 22289162d606SStefano Zampini 22299162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22309162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22319162d606SStefano Zampini 2232a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2233674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2234a773dcb8SStefano Zampini } else { 2235a773dcb8SStefano Zampini quad_value = 1.0; 2236a773dcb8SStefano Zampini } 2237674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22389162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2239674ae819SStefano Zampini } 22409162d606SStefano Zampini temp_constraints++; 2241674ae819SStefano Zampini total_counts++; 2242674ae819SStefano Zampini } 2243674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2244984c4197SStefano Zampini PetscReal real_value; 22459162d606SStefano Zampini PetscScalar *ptr_to_data; 22469162d606SStefano Zampini 2247984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22489162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2249674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22509162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2251674ae819SStefano Zampini } 2252984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2253984c4197SStefano Zampini /* check if array is null on the connected component */ 2254e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22559162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22565b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2257674ae819SStefano Zampini temp_constraints++; 2258674ae819SStefano Zampini total_counts++; 22599162d606SStefano Zampini if (!idxs_copied) { 22609162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22619162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2262674ae819SStefano Zampini } 2263674ae819SStefano Zampini } 22649162d606SStefano Zampini } 22659162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 226645a1bb75SStefano Zampini valid_constraints = temp_constraints; 2267eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2268a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22699162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22709162d606SStefano Zampini 22719162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2272a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22739162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2274a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22759162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2276a773dcb8SStefano Zampini } else { /* perform SVD */ 2277984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22789162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2279674ae819SStefano Zampini 2280674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2281984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2282984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2283984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2284984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2285984c4197SStefano Zampini from that computed using LAPACKgesvd 2286984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2287984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2288984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2289674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2290e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2291984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2292674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2293674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22949162d606SStefano 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)); 2295674ae819SStefano Zampini } 2296674ae819SStefano Zampini } 2297e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2298e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2299e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2300674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2301c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2302674ae819SStefano Zampini #else 2303c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2304674ae819SStefano Zampini #endif 2305674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2306984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2307984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2308674ae819SStefano Zampini j = 0; 2309984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2310674ae819SStefano Zampini total_counts = total_counts-j; 231145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2312e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2313c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2314c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2315c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2316c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2317c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2318c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2319674ae819SStefano Zampini if (j<temp_constraints) { 2320984c4197SStefano Zampini PetscInt ii; 2321984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2322674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23239162d606SStefano 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)); 2324674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2325984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2326674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23279162d606SStefano 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]; 2328674ae819SStefano Zampini } 2329674ae819SStefano Zampini } 2330674ae819SStefano Zampini } 2331674ae819SStefano Zampini #else /* on missing GESVD */ 2332e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2333e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2334b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2335674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2336674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23379162d606SStefano 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)); 2338674ae819SStefano Zampini #else 23399162d606SStefano 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)); 2340674ae819SStefano Zampini #endif 2341984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2342674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2343984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2344e310c8b4SStefano Zampini k = temp_constraints; 2345e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2346674ae819SStefano Zampini j = 0; 2347e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 234845a1bb75SStefano Zampini valid_constraints = k-j; 2349911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2350984c4197SStefano Zampini #endif /* on missing GESVD */ 2351674ae819SStefano Zampini } 2352a773dcb8SStefano Zampini } 23539162d606SStefano Zampini /* update pointers information */ 23549162d606SStefano Zampini if (valid_constraints) { 23559162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23569162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23579162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23589162d606SStefano Zampini /* set change_of_basis flag */ 235945a1bb75SStefano Zampini if (boolforchange) { 2360b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23619162d606SStefano Zampini } 2362b3d85658SStefano Zampini total_counts_cc++; 236345a1bb75SStefano Zampini } 236445a1bb75SStefano Zampini } 2365984c4197SStefano Zampini /* free workspace */ 23668f1c130eSStefano Zampini if (!skip_lapack) { 2367984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2368984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2369984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2370984c4197SStefano Zampini #endif 2371984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2372984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2373984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2374984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2375984c4197SStefano Zampini #endif 2376984c4197SStefano Zampini } 2377984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2378984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2379984c4197SStefano Zampini } 2380984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2381cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2382cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2383cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2384cf5a6209SStefano Zampini } 2385cf5a6209SStefano Zampini if (n_ISForFaces) { 2386cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2387cf5a6209SStefano Zampini } 2388cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2389cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2390cf5a6209SStefano Zampini } 2391cf5a6209SStefano Zampini if (n_ISForEdges) { 2392cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2393cf5a6209SStefano Zampini } 2394cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 239508122e43SStefano Zampini } else { 239608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2397984c4197SStefano Zampini 239808122e43SStefano Zampini total_counts = 0; 239908122e43SStefano Zampini n_vertices = 0; 2400d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2401d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 240208122e43SStefano Zampini } 240308122e43SStefano Zampini max_constraints = 0; 24049162d606SStefano Zampini total_counts_cc = 0; 240508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 240608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24079162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 240808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 240908122e43SStefano Zampini } 24109162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24119162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24129162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24139162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 241474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24159162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24169162d606SStefano Zampini total_counts_cc = 0; 24179162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24189162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24199162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 242008122e43SStefano Zampini } 242108122e43SStefano Zampini } 24229162d606SStefano Zampini #if 0 24239162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24249162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24259162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24269162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24279162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24289162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24299162d606SStefano Zampini } 24309162d606SStefano Zampini printf("\n"); 24319162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24329162d606SStefano Zampini } 24331b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24348bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24351b968477SStefano Zampini } 24361b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24378bec7fa6SStefano 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]); 24381b968477SStefano Zampini } 243908122e43SStefano Zampini #endif 244008122e43SStefano Zampini 24418bec7fa6SStefano Zampini max_size_of_constraint = 0; 24429162d606SStefano 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]); 24439162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 244408122e43SStefano Zampini /* Change of basis */ 2445b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 244608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 244708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 244808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2449b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 245008122e43SStefano Zampini } 245108122e43SStefano Zampini } 245208122e43SStefano Zampini } 245308122e43SStefano Zampini } 2454984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 24550e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 245608122e43SStefano Zampini 24579162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24589162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24596c4ed002SBarry 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); 2460674ae819SStefano Zampini 2461674ae819SStefano Zampini /* Create constraint matrix */ 2462674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 246316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2464984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2465984c4197SStefano Zampini 2466984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2467a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2468a717540cSStefano Zampini qr_needed = PETSC_FALSE; 246974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2470984c4197SStefano Zampini total_primal_vertices=0; 2471b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24729162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24739162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2474984c4197SStefano Zampini if (size_of_constraint == 1) { 24759162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2476b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 247764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24789162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24799162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2480a717540cSStefano Zampini } 2481b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 248274d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2483a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2484a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2485a717540cSStefano Zampini } 2486fa434743SStefano Zampini } else { 2487b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2488fa434743SStefano Zampini } 2489a717540cSStefano Zampini } 2490b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2491b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2492674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 249370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2494b3d85658SStefano Zampini 2495b3d85658SStefano 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); 24960e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24970e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2498984c4197SStefano Zampini 2499984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 250074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2501785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2502984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 250374d5cdf7SStefano Zampini 2504984c4197SStefano Zampini j = total_primal_vertices; 250574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2506b3d85658SStefano Zampini cum = total_primal_vertices; 25079162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25084641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2509b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2510b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2511b3d85658SStefano Zampini cum++; 25129162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 251374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 251474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 251574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 251674d5cdf7SStefano Zampini } 25179162d606SStefano Zampini j += constraints_n[i]; 2518674ae819SStefano Zampini } 2519674ae819SStefano Zampini } 2520674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2521674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2522088faed8SStefano Zampini 2523674ae819SStefano Zampini /* set values in constraint matrix */ 2524984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25250e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2526674ae819SStefano Zampini } 2527984c4197SStefano Zampini total_counts = total_primal_vertices; 25289162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25294641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25309162d606SStefano Zampini PetscInt *cols; 25319162d606SStefano Zampini 25329162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25339162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25349162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25359162d606SStefano Zampini PetscInt row = total_counts+k; 25369162d606SStefano Zampini PetscScalar *vals; 25379162d606SStefano Zampini 25389162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25399162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25409162d606SStefano Zampini } 25419162d606SStefano Zampini total_counts += constraints_n[i]; 2542674ae819SStefano Zampini } 2543674ae819SStefano Zampini } 2544674ae819SStefano Zampini /* assembling */ 2545674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2546674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2547088faed8SStefano Zampini 2548984c4197SStefano Zampini /* 254945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2550984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2551984c4197SStefano Zampini */ 2552674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2553674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2554026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2555984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2556984c4197SStefano Zampini /* working stuff for GEQRF */ 255781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2558984c4197SStefano Zampini PetscBLASInt lqr_work; 2559984c4197SStefano Zampini /* working stuff for UNGQR */ 2560984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2561984c4197SStefano Zampini PetscBLASInt lgqr_work; 2562984c4197SStefano Zampini /* working stuff for TRTRS */ 2563984c4197SStefano Zampini PetscScalar *trs_rhs; 25643f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2565984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2566984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2567984c4197SStefano Zampini PetscScalar *start_vals; 2568984c4197SStefano Zampini /* working stuff for values insertion */ 25694641a718SStefano Zampini PetscBT is_primal; 257064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2571906d46d4SStefano Zampini /* matrix sizes */ 2572906d46d4SStefano Zampini PetscInt global_size,local_size; 2573906d46d4SStefano Zampini /* temporary change of basis */ 2574906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2575cf5a6209SStefano Zampini /* extra space for debugging */ 2576cf5a6209SStefano Zampini PetscScalar *dbg_work; 2577984c4197SStefano Zampini 2578906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2579906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 258016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2581bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2582906d46d4SStefano Zampini /* nonzeros for local mat */ 2583bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2584bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25859162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2586a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25879162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2588a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25899162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2590a717540cSStefano Zampini } else { 25919162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25929162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2593a717540cSStefano Zampini } 2594a717540cSStefano Zampini } 2595a717540cSStefano Zampini } 2596906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2597bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2598a717540cSStefano Zampini /* Set initial identity in the matrix */ 2599bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2600906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2601a717540cSStefano Zampini } 2602a717540cSStefano Zampini 2603a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2604a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2605a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2606a717540cSStefano Zampini } 2607a717540cSStefano Zampini 2608a717540cSStefano Zampini 2609a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2610a717540cSStefano Zampini /* 2611a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2612a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2613a717540cSStefano Zampini 2614a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2615a717540cSStefano Zampini 2616a6b551f4SStefano 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) 2617a6b551f4SStefano Zampini 2618a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2619a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2620a717540cSStefano Zampini | ... | 2621a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2622a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2623a717540cSStefano Zampini 2624a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2625a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2626a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2627a6b551f4SStefano Zampini 2628a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2629a717540cSStefano Zampini */ 2630a717540cSStefano Zampini if (qr_needed) { 2631984c4197SStefano Zampini /* space to store Q */ 2632854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2633984c4197SStefano Zampini /* first we issue queries for optimal work */ 26343f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26363f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2637984c4197SStefano Zampini lqr_work = -1; 26383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2639984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2640984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2641785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2642984c4197SStefano Zampini lgqr_work = -1; 26433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26473f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2649984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2650984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2651785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2652984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2653785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2654984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2655785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2656a717540cSStefano Zampini /* allocating workspace for check */ 2657a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2658cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2659a717540cSStefano Zampini } 2660a717540cSStefano Zampini } 2661984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26624641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2663473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26640e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 26656c4ed002SBarry 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); 266639e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 266739e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 266839e2fb2aSStefano Zampini } 266939e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2670984c4197SStefano Zampini 2671a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26729162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26739162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26744641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2675984c4197SStefano Zampini /* get constraint info */ 26769162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2677984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2678984c4197SStefano Zampini 2679984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26809162d606SStefano 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); 2681674ae819SStefano Zampini } 2682984c4197SStefano Zampini 2683fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2684a717540cSStefano Zampini 2685a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2686a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26879162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2688a717540cSStefano Zampini } 2689984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26909162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2691984c4197SStefano Zampini 2692984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26943f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26953f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2696674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26973f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2698984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2699674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2700984c4197SStefano Zampini 2701984c4197SStefano Zampini /* explictly compute R^-T */ 2702984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2703984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27053f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27063f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27073f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2708984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27093f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2710984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2711984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2712984c4197SStefano Zampini 2713a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2718984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27193f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2720984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2721984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2722984c4197SStefano Zampini 2723984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2724984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2725984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27263f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2732984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27339162d606SStefano 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)); 2734984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27359162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2736984c4197SStefano Zampini 2737984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27389162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2739984c4197SStefano Zampini /* insert cols for primal dofs */ 2740984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2741984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27429162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2743906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2744984c4197SStefano Zampini } 2745984c4197SStefano Zampini /* insert cols for dual dofs */ 2746984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27479162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2748984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27499162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2750906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2751984c4197SStefano Zampini j++; 2752674ae819SStefano Zampini } 2753674ae819SStefano Zampini } 2754984c4197SStefano Zampini 2755984c4197SStefano Zampini /* check change of basis */ 2756984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2757984c4197SStefano Zampini PetscInt ii,jj; 2758984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2759c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2760c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2761c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2762c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2763c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2764c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2765984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2766cf5a6209SStefano 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)); 2767984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2768984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2769984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2770cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2771cf5a6209SStefano 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; 2772674ae819SStefano Zampini } 2773674ae819SStefano Zampini } 2774984c4197SStefano Zampini if (!valid_qr) { 277522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2776984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2777984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2778cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2779cf5a6209SStefano 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])); 2780674ae819SStefano Zampini } 2781cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2782cf5a6209SStefano 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])); 2783984c4197SStefano Zampini } 2784984c4197SStefano Zampini } 2785984c4197SStefano Zampini } 2786674ae819SStefano Zampini } else { 278722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2788674ae819SStefano Zampini } 2789674ae819SStefano Zampini } 2790a717540cSStefano Zampini } else { /* simple transformation block */ 2791a717540cSStefano Zampini PetscInt row,col; 2792a6b551f4SStefano Zampini PetscScalar val,norm; 2793a6b551f4SStefano Zampini 2794a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27959162d606SStefano 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)); 2796a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27979162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27989162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2799bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28009162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2801906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28029162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2803a717540cSStefano Zampini } else { 2804a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28059162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2806a717540cSStefano Zampini if (row != col) { 28079162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2808a717540cSStefano Zampini } else { 28099162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2810a717540cSStefano Zampini } 2811906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2812a717540cSStefano Zampini } 2813a717540cSStefano Zampini } 2814a717540cSStefano Zampini } 281598a51de6SStefano Zampini if (pcbddc->dbg_flag) { 281622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2817a717540cSStefano Zampini } 2818674ae819SStefano Zampini } 2819984c4197SStefano Zampini } else { 2820984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28219162d606SStefano 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); 2822674ae819SStefano Zampini } 2823674ae819SStefano Zampini } 2824674ae819SStefano Zampini } 2825a717540cSStefano Zampini 2826a717540cSStefano Zampini /* free workspace */ 2827a717540cSStefano Zampini if (qr_needed) { 2828984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2829cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2830984c4197SStefano Zampini } 2831984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2832984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2833984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2834984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2835984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2836674ae819SStefano Zampini } 2837a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2838906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2839906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2840906d46d4SStefano Zampini 2841906d46d4SStefano Zampini /* assembling of global change of variable */ 2842bbb9e6c6SStefano Zampini { 2843bbb9e6c6SStefano Zampini Mat tmat; 284416f15bc4SStefano Zampini PetscInt bs; 284516f15bc4SStefano Zampini 2846906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2847906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2848bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2850bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2851bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 285216f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 285316f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2854906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2855bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2860e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2861e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2862bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2864906d46d4SStefano Zampini } 2865906d46d4SStefano Zampini /* check */ 2866906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2867906d46d4SStefano Zampini PetscReal error; 2868906d46d4SStefano Zampini Vec x,x_change; 2869906d46d4SStefano Zampini 2870906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2871906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2874e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2875e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2877e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2878e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2879906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2880906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2882906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2883bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2884906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2885906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2886906d46d4SStefano Zampini } 2887b96c3477SStefano Zampini 2888b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2889b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2890b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2891b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2892ac632422SStefano Zampini Mat S_new,tmat; 2893bbb9e6c6SStefano Zampini IS is_all_N; 2894bbb9e6c6SStefano Zampini 2895bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28966816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2897bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2898ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2899b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2900ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2901ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2902ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2903ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2904ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2905b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2906ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2907ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2908ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2909ac632422SStefano Zampini } 2910b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2911b96c3477SStefano Zampini } 2912b96c3477SStefano Zampini } 2913906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2914906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2915b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2916b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2917b9b85e73SStefano Zampini } 2918906d46d4SStefano Zampini 2919906d46d4SStefano Zampini /* set up change of basis context */ 2920906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2921906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2922906d46d4SStefano Zampini 2923906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2924906d46d4SStefano Zampini PetscInt global_size,local_size; 2925906d46d4SStefano Zampini 2926906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2927906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2928906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2929906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2930906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2931906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2932906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2933906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2934906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2935906d46d4SStefano Zampini } else { 2936906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2937906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2938906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2939906d46d4SStefano Zampini } 2940906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2941906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2942906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2943906d46d4SStefano Zampini } else { 2944906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2945906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2946906d46d4SStefano Zampini } 2947906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2951b9b85e73SStefano Zampini } 2952a717540cSStefano Zampini 2953727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2954727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2955727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2956aff50787SStefano 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); 2957c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29580e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2959aff50787SStefano 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); 2960727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2961727cdba6SStefano Zampini } 29620e6343abSStefano Zampini } 29630e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2964727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2965*b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2966727cdba6SStefano Zampini 2967a717540cSStefano Zampini /* flush dbg viewer */ 2968b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2969b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2970b8ffe317SStefano Zampini } 2971a717540cSStefano Zampini 2972e310c8b4SStefano Zampini /* free workspace */ 2973a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29744641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 297508122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29769162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29779162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 297808122e43SStefano Zampini } else { 29799162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29809162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29819162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 298208122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 298308122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29849162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29859162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 298608122e43SStefano Zampini } 2987674ae819SStefano Zampini PetscFunctionReturn(0); 2988674ae819SStefano Zampini } 2989674ae819SStefano Zampini 2990674ae819SStefano Zampini #undef __FUNCT__ 2991674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2992674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2993674ae819SStefano Zampini { 2994674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2995674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2996674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29977fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2998674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2999674ae819SStefano Zampini 3000674ae819SStefano Zampini PetscFunctionBegin; 30018e61c736SStefano Zampini /* Reset previously computed graph */ 30028e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3003674ae819SStefano Zampini /* Init local Graph struct */ 30047fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30053bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3006674ae819SStefano Zampini 3007575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3008575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3009575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3010575ad6abSStefano Zampini } 30119577ea80SStefano Zampini 3012674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30134d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30144d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30154d379d7bSStefano Zampini PetscInt nvtxs; 3016e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3017674ae819SStefano Zampini 30184d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30192fffb893SStefano Zampini 30202fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30212fffb893SStefano Zampini if (flg_row) { 30224d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3023b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30242fffb893SStefano Zampini } 30252fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30269b28b941SStefano 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 */ 30274d379d7bSStefano Zampini IS is_dummy; 30284d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30294d379d7bSStefano Zampini PetscInt j,sum; 30304d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30314d379d7bSStefano Zampini const PetscInt *idxs; 30324d379d7bSStefano Zampini PCBDDCGraph graph; 30334d379d7bSStefano Zampini PetscBT is_on_boundary; 30344d379d7bSStefano Zampini 30354d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30364d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30374d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30384d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30397fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30404d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3041e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3042e496cd5dSStefano Zampini if (flg_row) { 30434d379d7bSStefano Zampini graph->xadj = xadj; 30444d379d7bSStefano Zampini graph->adjncy = adjncy; 3045e496cd5dSStefano Zampini } 30464d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30474d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3048e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30494d379d7bSStefano Zampini 30504d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30519b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30524d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30534d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30544d379d7bSStefano Zampini } 30554d379d7bSStefano Zampini } 30564d379d7bSStefano Zampini 3057e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30584d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30594d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30604d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30614d379d7bSStefano Zampini } 30624d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30634d379d7bSStefano Zampini 3064e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30654d379d7bSStefano Zampini sum = 0; 30664d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30674d379d7bSStefano Zampini PetscInt sizecc = 0; 30684d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30694d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30704d379d7bSStefano Zampini sizecc++; 30714d379d7bSStefano Zampini } 30724d379d7bSStefano Zampini } 30734d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30744d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30754d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30764d379d7bSStefano Zampini } 30774d379d7bSStefano Zampini } 30784d379d7bSStefano Zampini sum += sizecc*sizecc; 30794d379d7bSStefano Zampini } 30804d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30814d379d7bSStefano Zampini sum = 0; 3082e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30834d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30844d379d7bSStefano Zampini cxadj[i] = sum; 30854d379d7bSStefano Zampini sum += temp; 30864d379d7bSStefano Zampini } 3087e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30884d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30894d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30904d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30914d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30924d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30934d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30944d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30954d379d7bSStefano Zampini sizecc++; 30964d379d7bSStefano Zampini } 30974d379d7bSStefano Zampini } 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini } 31004d379d7bSStefano Zampini } 31019b28b941SStefano Zampini if (sum) { 3102e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31034d379d7bSStefano Zampini } else { 31044d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31054d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31064d379d7bSStefano Zampini } 31074d379d7bSStefano Zampini graph->xadj = 0; 31084d379d7bSStefano Zampini graph->adjncy = 0; 31094d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31104d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31114d379d7bSStefano Zampini } 3112674ae819SStefano Zampini } 31139b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31149b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3115674ae819SStefano Zampini } 3116674ae819SStefano Zampini 311763602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3118674ae819SStefano Zampini vertex_size = 1; 311963602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 312063602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 312195ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3123e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 312463602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3125674ae819SStefano Zampini } 312663602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 312763602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 312863602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3129674ae819SStefano Zampini } 313063602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3131674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 313263602bcaSStefano Zampini } else { 313363602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 313463602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3135854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313663602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 313763602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313863602bcaSStefano Zampini } 313963602bcaSStefano Zampini } 3140674ae819SStefano Zampini } 3141674ae819SStefano Zampini 3142674ae819SStefano Zampini /* Setup of Graph */ 3143785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3144e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3145785d1243SStefano Zampini } 3146785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3147e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3148785d1243SStefano Zampini } 3149302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3150674ae819SStefano Zampini 3151674ae819SStefano Zampini /* Graph's connected components analysis */ 3152674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3153674ae819SStefano Zampini 3154674ae819SStefano Zampini /* print some info to stdout */ 3155674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3156302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3157674ae819SStefano Zampini } 3158fb180af4SStefano Zampini 3159fb180af4SStefano Zampini /* mark topography has done */ 3160fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3161674ae819SStefano Zampini PetscFunctionReturn(0); 3162674ae819SStefano Zampini } 3163674ae819SStefano Zampini 3164dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3165674ae819SStefano Zampini #undef __FUNCT__ 3166674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3167dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3168674ae819SStefano Zampini { 3169dc456d91SStefano Zampini PetscSF sf; 3170dc456d91SStefano Zampini PetscLayout map; 3171dc456d91SStefano Zampini const PetscInt *idxs; 3172dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3173dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3174dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3175dc456d91SStefano Zampini PetscMPIInt commsize; 3176674ae819SStefano Zampini PetscBool first_found; 3177674ae819SStefano Zampini PetscErrorCode ierr; 3178674ae819SStefano Zampini 3179674ae819SStefano Zampini PetscFunctionBegin; 3180dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3181dc456d91SStefano Zampini if (subset_mult) { 3182dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3183dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3184dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3185674ae819SStefano Zampini } 3186dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3187dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3188dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3189dc456d91SStefano Zampini for (i=0;i<n;i++) { 3190dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3191dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3192674ae819SStefano Zampini } 3193dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3194*b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3195dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3196dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3197dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3198dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3199dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3200dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3201dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3202dc456d91SStefano Zampini 3203dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3204dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3205dc456d91SStefano Zampini if (subset_mult) { 3206dc456d91SStefano Zampini const PetscInt* idxs_mult; 3207dc456d91SStefano Zampini 3208dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3209dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3210dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3211674ae819SStefano Zampini } else { 3212dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3213674ae819SStefano Zampini } 3214dc456d91SStefano Zampini /* local size of new subset */ 3215dc456d91SStefano Zampini n_n = 0; 3216dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3217dc456d91SStefano Zampini 3218dc456d91SStefano Zampini /* global indexes in layout */ 3219dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3220dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3221dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3222dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3223dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3224dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3225dc456d91SStefano Zampini 3226dc456d91SStefano Zampini /* reduce from leaves to roots */ 3227dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 322864a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 322964a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3230dc456d91SStefano Zampini 3231dc456d91SStefano Zampini /* count indexes in local part of layout */ 3232674ae819SStefano Zampini nlocals = 0; 3233674ae819SStefano Zampini first_index = -1; 3234674ae819SStefano Zampini first_found = PETSC_FALSE; 3235dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3236dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3237674ae819SStefano Zampini first_found = PETSC_TRUE; 3238674ae819SStefano Zampini first_index = i; 3239674ae819SStefano Zampini } 3240dc456d91SStefano Zampini nlocals += root_data[i]; 3241674ae819SStefano Zampini } 3242dc456d91SStefano Zampini 3243dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 32445fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3245dc456d91SStefano Zampini start = 0; 324664a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32475fa240b1SStefano Zampini #else 324864a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 32495fa240b1SStefano Zampini start = start-nlocals; 32505fa240b1SStefano Zampini #endif 32515fa240b1SStefano Zampini 3252dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3253dc456d91SStefano Zampini *N_n = start + nlocals; 3254dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3255dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3256674ae819SStefano Zampini } 32575fa240b1SStefano Zampini 32585fa240b1SStefano Zampini /* adapt root data with cumulative */ 3259674ae819SStefano Zampini if (first_found) { 3260dc456d91SStefano Zampini PetscInt old_index; 3261dc456d91SStefano Zampini 3262dc456d91SStefano Zampini root_data[first_index] += start; 3263674ae819SStefano Zampini old_index = first_index; 3264dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3265dc456d91SStefano Zampini if (root_data[i]) { 3266dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3267674ae819SStefano Zampini old_index = i; 3268674ae819SStefano Zampini } 3269674ae819SStefano Zampini } 3270674ae819SStefano Zampini } 3271dc456d91SStefano Zampini 3272dc456d91SStefano Zampini /* from roots to leaves */ 3273dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3274dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3275dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3276dc456d91SStefano Zampini 3277dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3278dc456d91SStefano Zampini if (subset_mult) { 3279dc456d91SStefano Zampini const PetscInt* idxs_mult; 3280dc456d91SStefano Zampini PetscInt cum; 3281dc456d91SStefano Zampini 3282dc456d91SStefano Zampini cum = 0; 3283dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3284dc456d91SStefano Zampini for (i=0;i<n;i++) { 3285dc456d91SStefano Zampini PetscInt j; 3286dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3287674ae819SStefano Zampini } 3288dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3289674ae819SStefano Zampini } else { 3290dc456d91SStefano Zampini for (i=0;i<n;i++) { 3291dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3292674ae819SStefano Zampini } 3293674ae819SStefano Zampini } 3294dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3295dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3296674ae819SStefano Zampini PetscFunctionReturn(0); 3297674ae819SStefano Zampini } 32989a7d3425SStefano Zampini 32999a7d3425SStefano Zampini #undef __FUNCT__ 33009a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33019a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33029a7d3425SStefano Zampini { 33039a7d3425SStefano Zampini PetscInt i,j; 33049a7d3425SStefano Zampini PetscScalar *alphas; 33059a7d3425SStefano Zampini PetscErrorCode ierr; 33069a7d3425SStefano Zampini 33079a7d3425SStefano Zampini PetscFunctionBegin; 33089a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3309785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33109a7d3425SStefano Zampini for (i=0;i<n;i++) { 33119a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33129a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33139a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33149a7d3425SStefano Zampini } 33159a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33169a7d3425SStefano Zampini PetscFunctionReturn(0); 33179a7d3425SStefano Zampini } 33189a7d3425SStefano Zampini 3319e7931f94SStefano Zampini #undef __FUNCT__ 332070cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3321b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3322e7931f94SStefano Zampini { 332352e5ac9dSStefano Zampini IS ranks_send_to; 3324e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3325e7931f94SStefano Zampini PetscMPIInt size,rank,color; 332652e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 332752e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33283837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33292b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3330e7931f94SStefano Zampini PetscSubcomm subcomm; 333152e5ac9dSStefano Zampini PetscErrorCode ierr; 3332a57a6d2fSStefano Zampini 3333e7931f94SStefano Zampini PetscFunctionBegin; 33342b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33352b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33362b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3337e7931f94SStefano Zampini 3338e7931f94SStefano Zampini /* Get info on mapping */ 33393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 33403bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3341e7931f94SStefano Zampini 3342e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3343785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3344e7931f94SStefano Zampini xadj[0] = 0; 3345e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3346785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3347785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3348e7931f94SStefano Zampini 33492b510759SStefano Zampini if (threshold) { 3350d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33512b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3352d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3353d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3354d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3355d023bfaeSStefano Zampini xadj_count++; 3356e7931f94SStefano Zampini } 3357e7931f94SStefano Zampini } 3358d023bfaeSStefano Zampini xadj[1] = xadj_count; 3359c8587f34SStefano Zampini } else { 3360e7931f94SStefano Zampini if (xadj[1]) { 3361e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3362e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3363c8587f34SStefano Zampini } 3364e7931f94SStefano Zampini } 33653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3366e7931f94SStefano Zampini if (use_square) { 3367e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3368e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3369e7931f94SStefano Zampini } 3370e7931f94SStefano Zampini } 3371e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3372e7931f94SStefano Zampini 33733837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3374e7931f94SStefano Zampini 3375e7931f94SStefano Zampini /* 3376e7931f94SStefano Zampini Restrict work on active processes only. 3377e7931f94SStefano Zampini */ 3378e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3379e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3380e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33812b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3382d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3383e7931f94SStefano Zampini if (color) { 3384e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3385e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3386e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3387c8587f34SStefano Zampini } else { 338852e5ac9dSStefano Zampini Mat subdomain_adj; 338952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 339052e5ac9dSStefano Zampini MatPartitioning partitioner; 339152e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 339252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3393b0c7d250SStefano Zampini PetscBool aggregate; 3394b0c7d250SStefano Zampini 3395306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3396785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3397e7931f94SStefano Zampini prank = rank; 3398306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33998002ef2cSStefano Zampini /* 3400e7931f94SStefano Zampini for (i=0;i<size;i++) { 3401e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3402c8587f34SStefano Zampini } 34038002ef2cSStefano Zampini */ 3404e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3405e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3406c8587f34SStefano Zampini } 3407e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3408b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3409b0c7d250SStefano Zampini if (aggregate) { 3410b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3411b0c7d250SStefano Zampini PetscMPIInt nrank; 3412b0c7d250SStefano Zampini PetscScalar *vals; 3413b0c7d250SStefano Zampini 3414b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3415b0c7d250SStefano Zampini lrows = 0; 3416b0c7d250SStefano Zampini if (nrank<redprocs) { 3417b0c7d250SStefano Zampini lrows = size/redprocs; 3418b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3419b0c7d250SStefano Zampini } 34205fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3421b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3422b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3423b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3424b0c7d250SStefano Zampini row = nrank; 3425b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3426b0c7d250SStefano Zampini cols = adjncy; 3427b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3428b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3429b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3430b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3431b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 343252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 343352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 343452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3435b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3436b0c7d250SStefano Zampini } else { 3437306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3438b0c7d250SStefano Zampini } 343922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3440e7931f94SStefano Zampini 3441e7931f94SStefano Zampini /* Partition */ 3442306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3443e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3444e7931f94SStefano Zampini if (use_vwgt) { 34453837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3446e7931f94SStefano Zampini v_wgt[0] = local_size; 3447e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3448c8587f34SStefano Zampini } 344928143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 345028143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3451e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3452e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 345322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3454e7931f94SStefano Zampini 345552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 345652e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 345752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 345852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3459b0c7d250SStefano Zampini if (!redprocs) { 3460b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 346128143c3dSStefano Zampini } else { 3462b0c7d250SStefano Zampini PetscInt idxs[1]; 3463b0c7d250SStefano Zampini PetscMPIInt tag; 3464b0c7d250SStefano Zampini MPI_Request *reqs; 3465b0c7d250SStefano Zampini 3466b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3467b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3468b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3469b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 347028143c3dSStefano Zampini } 3471b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3472b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3473b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3474b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3475e7931f94SStefano Zampini } 347652e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3477e7931f94SStefano Zampini /* clean up */ 3478e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 347952e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3480e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3481e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3482e7931f94SStefano Zampini } 3483e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3484e7931f94SStefano Zampini 3485e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3486e7931f94SStefano Zampini i = 1; 3487e7931f94SStefano Zampini if (color) i=0; 3488e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3489e7931f94SStefano Zampini /* get back IS */ 3490e7931f94SStefano Zampini *is_sends = ranks_send_to; 3491e7931f94SStefano Zampini PetscFunctionReturn(0); 3492e7931f94SStefano Zampini } 3493e7931f94SStefano Zampini 3494e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3495e7931f94SStefano Zampini 3496e7931f94SStefano Zampini #undef __FUNCT__ 3497e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 349853a05cb3SStefano 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[]) 3499e7931f94SStefano Zampini { 350070cf5478SStefano Zampini Mat local_mat; 3501e7931f94SStefano Zampini IS is_sends_internal; 35029d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 350328143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35049d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3505e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3506e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3507e7931f94SStefano Zampini const PetscInt* is_indices; 3508e7931f94SStefano Zampini MatType new_local_type; 3509e7931f94SStefano Zampini /* buffers */ 3510e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 351128143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35129d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3513e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3514e7931f94SStefano Zampini /* MPI */ 351528143c3dSStefano Zampini MPI_Comm comm,comm_n; 351628143c3dSStefano Zampini PetscSubcomm subcomm; 3517e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 351828143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 351928143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 352028143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 352128143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 352228143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3523e7931f94SStefano Zampini PetscErrorCode ierr; 3524e7931f94SStefano Zampini 3525e7931f94SStefano Zampini PetscFunctionBegin; 352628143c3dSStefano Zampini /* TODO: add missing checks */ 352728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 352828143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 352928143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 353028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3531e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 353228143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3533e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3534e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3535e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3536e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3537e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 353828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 353970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 354070cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 354128143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 354270cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 354370cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 354470cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 354570cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 354670cf5478SStefano Zampini } 3547e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3548e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3549e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3550e7931f94SStefano Zampini if (!is_sends) { 355128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3552b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3553c8587f34SStefano Zampini } else { 3554e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3555e7931f94SStefano Zampini is_sends_internal = is_sends; 3556c8587f34SStefano Zampini } 3557e7931f94SStefano Zampini 3558e7931f94SStefano Zampini /* get comm */ 3559a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3560e7931f94SStefano Zampini 3561e7931f94SStefano Zampini /* compute number of sends */ 3562e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3563e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3564e7931f94SStefano Zampini 3565e7931f94SStefano Zampini /* compute number of receives */ 3566e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3567785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3568e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3569e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3570e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3571e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3572e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3573e7931f94SStefano Zampini 357428143c3dSStefano Zampini /* restrict comm if requested */ 357528143c3dSStefano Zampini subcomm = 0; 357628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 357728143c3dSStefano Zampini if (restrict_comm) { 3578779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3579779c1cceSStefano Zampini 358028143c3dSStefano Zampini color = 0; 358153a05cb3SStefano Zampini if (restrict_full) { 358253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 358353a05cb3SStefano Zampini } else { 358453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 358553a05cb3SStefano Zampini } 3586*b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 358728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 358828143c3dSStefano Zampini /* check if reuse has been requested */ 358928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 359028143c3dSStefano Zampini if (*mat_n) { 359128143c3dSStefano Zampini PetscMPIInt subcommsize2; 359228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 359328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 359428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 359528143c3dSStefano Zampini } else { 359628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 359728143c3dSStefano Zampini } 359828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3599779c1cceSStefano Zampini PetscMPIInt rank; 3600779c1cceSStefano Zampini 3601779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 360228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 360328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 360428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3605306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 360628143c3dSStefano Zampini } 360728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 360828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 360928143c3dSStefano Zampini } else { 361028143c3dSStefano Zampini comm_n = comm; 361128143c3dSStefano Zampini } 361228143c3dSStefano Zampini 3613e7931f94SStefano Zampini /* prepare send/receive buffers */ 3614785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3615e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3616785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3617e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 361828143c3dSStefano Zampini if (nis) { 3619854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 362028143c3dSStefano Zampini } 3621e7931f94SStefano Zampini 362228143c3dSStefano Zampini /* Get data from local matrices */ 36236c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3624e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3625e7931f94SStefano Zampini /* 3626e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3627e7931f94SStefano Zampini send_buffer_idxs should contain: 3628e7931f94SStefano Zampini - MatType_PRIVATE type 3629e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3630e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3631e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3632e7931f94SStefano Zampini */ 36336c4ed002SBarry Smith else { 3634e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 36353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3636854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3637e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3638e7931f94SStefano Zampini send_buffer_idxs[1] = i; 36393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3640e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 36413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3642e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3643e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3644e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3645e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3646c8587f34SStefano Zampini } 3647c8587f34SStefano Zampini } 3648e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 364928143c3dSStefano Zampini /* additional is (if any) */ 365028143c3dSStefano Zampini if (nis) { 365128143c3dSStefano Zampini PetscMPIInt psum; 365228143c3dSStefano Zampini PetscInt j; 365328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 365428143c3dSStefano Zampini PetscInt plen; 365528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 365628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 365728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 365828143c3dSStefano Zampini } 3659854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 366028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 366128143c3dSStefano Zampini PetscInt plen; 366228143c3dSStefano Zampini const PetscInt *is_array_idxs; 366328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 366428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 366528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 366628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 366728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 366828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 366928143c3dSStefano Zampini } 367028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 367128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 367228143c3dSStefano Zampini } 367328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 367428143c3dSStefano Zampini } 367528143c3dSStefano Zampini 3676e7931f94SStefano Zampini buf_size_idxs = 0; 3677e7931f94SStefano Zampini buf_size_vals = 0; 367828143c3dSStefano Zampini buf_size_idxs_is = 0; 3679e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3680e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3681e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 368228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3683e7931f94SStefano Zampini } 3684785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3685785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 368695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3687e7931f94SStefano Zampini 3688e7931f94SStefano Zampini /* get new tags for clean communications */ 3689e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3690e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 369128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3692e7931f94SStefano Zampini 3693e7931f94SStefano Zampini /* allocate for requests */ 3694785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3695785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 369695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3697785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3698785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 369995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3700e7931f94SStefano Zampini 3701e7931f94SStefano Zampini /* communications */ 3702e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3703e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 370428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3705e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3706e7931f94SStefano Zampini source_dest = onodes[i]; 3707e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3708e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3709e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3710e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 371128143c3dSStefano Zampini if (nis) { 371228143c3dSStefano 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); 371328143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 371428143c3dSStefano Zampini } 3715e7931f94SStefano Zampini } 3716e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3717e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3718e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3719e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 372028143c3dSStefano Zampini if (nis) { 372128143c3dSStefano 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); 372228143c3dSStefano Zampini } 3723e7931f94SStefano Zampini } 3724e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3725e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3726e7931f94SStefano Zampini 3727e7931f94SStefano Zampini /* assemble new l2g map */ 3728e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);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 new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3733e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3734e7931f94SStefano Zampini } 37359d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3736e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37379d30be91SStefano Zampini new_local_rows = 0; 3738e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37399d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37409d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3741e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3742e7931f94SStefano Zampini } 37439d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37449d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3745e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3746e7931f94SStefano Zampini 3747e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3748e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3749e7931f94SStefano 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) */ 3750e7931f94SStefano Zampini if (n_recvs) { 375128143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3752e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3753e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3754e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3755e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3756e7931f94SStefano Zampini break; 3757e7931f94SStefano Zampini } 3758e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3759e7931f94SStefano Zampini } 3760e7931f94SStefano Zampini switch (new_local_type_private) { 376128143c3dSStefano Zampini case MATDENSE_PRIVATE: 376228143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3763e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3764e7931f94SStefano Zampini bs = 1; 376528143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 376628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 376728143c3dSStefano Zampini bs = 1; 376828143c3dSStefano Zampini } 3769e7931f94SStefano Zampini break; 3770e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3771e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3772e7931f94SStefano Zampini bs = 1; 3773e7931f94SStefano Zampini break; 3774e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3775e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3776e7931f94SStefano Zampini break; 3777e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3778e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3779e7931f94SStefano Zampini break; 3780e7931f94SStefano Zampini default: 37819d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3782e7931f94SStefano Zampini break; 3783e7931f94SStefano Zampini } 378428143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 378528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 378628143c3dSStefano Zampini bs = 1; 3787e7931f94SStefano Zampini } 3788e7931f94SStefano Zampini 378970cf5478SStefano Zampini /* create MATIS object if needed */ 379070cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3791e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3792e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 379370cf5478SStefano Zampini } else { 379470cf5478SStefano Zampini /* it also destroys the local matrices */ 379570cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 379670cf5478SStefano Zampini } 379770cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3798e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37999d30be91SStefano Zampini 38009d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38019d30be91SStefano Zampini 38029d30be91SStefano Zampini /* Global to local map of received indices */ 38039d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38049d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38059d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38069d30be91SStefano Zampini 38079d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38089d30be91SStefano Zampini buf_size_idxs = 0; 38099d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38109d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38119d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38129d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38139d30be91SStefano Zampini } 38149d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38159d30be91SStefano Zampini 38169d30be91SStefano Zampini /* set preallocation */ 38179d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38189d30be91SStefano Zampini if (!newisdense) { 38199d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38209d30be91SStefano Zampini 38219d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38229d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38239d30be91SStefano Zampini if (n_recvs) { 38249d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38259d30be91SStefano Zampini } 38269d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38279d30be91SStefano Zampini PetscInt j; 38289d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38299d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38309d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38319d30be91SStefano Zampini } 38329d30be91SStefano Zampini } else { 38339d30be91SStefano Zampini /* TODO */ 38349d30be91SStefano Zampini } 38359d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38369d30be91SStefano Zampini } 38379d30be91SStefano Zampini if (new_local_nnz) { 38389d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38399d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38409d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38419d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38429d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38439d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38449d30be91SStefano Zampini } else { 38459d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38469d30be91SStefano Zampini } 38479d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38489d30be91SStefano Zampini } else { 38499d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38509d30be91SStefano Zampini } 3851e7931f94SStefano Zampini 3852e7931f94SStefano Zampini /* set values */ 3853e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38549d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3855e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3856e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3857e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38589d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3859e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3860e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3861e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 386228143c3dSStefano Zampini } else { 386328143c3dSStefano Zampini /* TODO */ 3864e7931f94SStefano Zampini } 3865e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3866e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3867e7931f94SStefano Zampini } 3868e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3869e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 387070cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 387170cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38729d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38739d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3874e7931f94SStefano Zampini 3875dfd14d43SStefano Zampini #if 0 387628143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3877e7931f94SStefano Zampini Vec lvec,rvec; 3878e7931f94SStefano Zampini PetscReal infty_error; 3879e7931f94SStefano Zampini 38802a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3881e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3882e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3883e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 388470cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3885e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3886e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3887e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3888e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3889e7931f94SStefano Zampini } 389028143c3dSStefano Zampini #endif 3891e7931f94SStefano Zampini 389228143c3dSStefano Zampini /* assemble new additional is (if any) */ 389328143c3dSStefano Zampini if (nis) { 389428143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 389528143c3dSStefano Zampini 389628143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3897854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 389828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 389928143c3dSStefano Zampini psum = 0; 390028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 390128143c3dSStefano Zampini for (j=0;j<nis;j++) { 390228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 390328143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 390428143c3dSStefano Zampini psum += plen; 390528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 390628143c3dSStefano Zampini } 390728143c3dSStefano Zampini } 3908854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3909854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 391028143c3dSStefano Zampini for (i=1;i<nis;i++) { 391128143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 391228143c3dSStefano Zampini } 391328143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 391428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 391528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 391628143c3dSStefano Zampini for (j=0;j<nis;j++) { 391728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 391828143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 391928143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 392028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 392128143c3dSStefano Zampini } 392228143c3dSStefano Zampini } 392328143c3dSStefano Zampini for (i=0;i<nis;i++) { 392428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 392528143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 392628143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 392728143c3dSStefano Zampini } 392828143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 392928143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 393028143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 393128143c3dSStefano Zampini } 3932e7931f94SStefano Zampini /* free workspace */ 393328143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3934e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3935e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3936e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3937e7931f94SStefano Zampini if (isdense) { 3938e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3939e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3940e7931f94SStefano Zampini } else { 3941e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3942e7931f94SStefano Zampini } 394328143c3dSStefano Zampini if (nis) { 394428143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 394528143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 394628143c3dSStefano Zampini } 3947e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3948e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 394928143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3950e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3951e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 395228143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3953e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3954e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3955e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3956e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 395828143c3dSStefano Zampini if (nis) { 395928143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 396028143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 396128143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 396228143c3dSStefano Zampini } 396328143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 396428143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 396528143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 396628143c3dSStefano Zampini for (i=0;i<nis;i++) { 396728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 396828143c3dSStefano Zampini } 396953a05cb3SStefano Zampini *mat_n = NULL; 397028143c3dSStefano Zampini } 3971e7931f94SStefano Zampini PetscFunctionReturn(0); 3972e7931f94SStefano Zampini } 3973a57a6d2fSStefano Zampini 397412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3975af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 397612edc857SStefano Zampini 3977c8587f34SStefano Zampini #undef __FUNCT__ 3978c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3979c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3980c8587f34SStefano Zampini { 3981c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3982c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 398320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39849881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 398520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39866e683305SStefano Zampini IS coarse_is,*isarray; 39876e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39886e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3989f9eb5b7dSStefano Zampini PC pc_temp; 3990c8587f34SStefano Zampini PCType coarse_pc_type; 3991c8587f34SStefano Zampini KSPType coarse_ksp_type; 3992f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39934f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39946e683305SStefano Zampini Mat t_coarse_mat_is; 39956e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39966e683305SStefano Zampini PetscMPIInt all_procs; 399774e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 399868457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 399922bc73bbSStefano Zampini PetscScalar *array; 40009881197aSStefano Zampini PetscErrorCode ierr; 4001fdc09c96SStefano Zampini 4002c8587f34SStefano Zampini PetscFunctionBegin; 4003c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 400468457ee5SStefano 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 */ 4005fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40065a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4007fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4008f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4009f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4010f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4011fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 401251bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 401351bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4014dc4bcba2SStefano Zampini PC pc; 4015dc4bcba2SStefano Zampini PetscBool isbddc; 4016dc4bcba2SStefano Zampini 4017dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4018dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4019dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4020dc4bcba2SStefano Zampini if (isbddc) { 4021dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4022dc4bcba2SStefano Zampini } 4023727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4024fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4025fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4026fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4027f4ddd8eeSStefano Zampini } 4028fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4029fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4030f4ddd8eeSStefano Zampini } 403170cf5478SStefano Zampini /* reset any subassembling information */ 403270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40336e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40346e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4035fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4036f4ddd8eeSStefano Zampini } 4037c8587f34SStefano Zampini 40386e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40392b510759SStefano Zampini im_active = !!(pcis->n); 4040*b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40416e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40426e683305SStefano Zampini void_procs = all_procs-active_procs; 40436e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 404474e2c79eSStefano Zampini redist = PETSC_FALSE; 404522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40466e683305SStefano Zampini csin_ml = PETSC_TRUE; 40476e683305SStefano Zampini ncoarse_ml = void_procs; 4048779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4049779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 40506e683305SStefano Zampini csin_ds = PETSC_TRUE; 405118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 405218a45a71SStefano Zampini redist = PETSC_TRUE; 405318a45a71SStefano Zampini } else { 40546e683305SStefano Zampini csin_ds = PETSC_TRUE; 4055779c1cceSStefano Zampini ncoarse_ds = active_procs; 4056779c1cceSStefano Zampini redist = PETSC_TRUE; 405718a45a71SStefano Zampini } 40586e683305SStefano Zampini } else { 40596e683305SStefano Zampini csin_ml = PETSC_FALSE; 40606e683305SStefano Zampini ncoarse_ml = all_procs; 40616e683305SStefano Zampini if (void_procs) { 40626e683305SStefano Zampini csin_ds = PETSC_TRUE; 40636e683305SStefano Zampini ncoarse_ds = void_procs; 40646e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40656e683305SStefano Zampini } else { 4066779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 406774e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 406874e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 406974e2c79eSStefano Zampini redist = PETSC_TRUE; 407074e2c79eSStefano Zampini } else { 40716e683305SStefano Zampini csin_ds = PETSC_FALSE; 40726e683305SStefano Zampini ncoarse_ds = all_procs; 40736e683305SStefano Zampini } 40746e683305SStefano Zampini } 407574e2c79eSStefano Zampini } 40766e683305SStefano Zampini 40776e683305SStefano Zampini /* 40786e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40796e683305SStefano Zampini - we have not exceeded the number of levels requested 40806e683305SStefano Zampini - we can actually subassemble the active processes 40816e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40826e683305SStefano Zampini */ 40836e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40846e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40856e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40866e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40876e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4088f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40892b510759SStefano Zampini } else { 4090f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4091c8587f34SStefano Zampini } 4092c8587f34SStefano Zampini } 40936e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40946e683305SStefano Zampini if (multilevel_allowed) { 40956e683305SStefano Zampini ncoarse = ncoarse_ml; 40966e683305SStefano Zampini csin = csin_ml; 409758da7f69SStefano Zampini redist = PETSC_FALSE; 40986e683305SStefano Zampini } else { 40996e683305SStefano Zampini ncoarse = ncoarse_ds; 41006e683305SStefano Zampini csin = csin_ds; 41016e683305SStefano Zampini } 4102e7931f94SStefano Zampini 4103abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4104abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4105abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4106abbbba34SStefano Zampini 4107abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 410822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 410922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 411022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 411122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4112b9b85e73SStefano Zampini #if 0 4113b9b85e73SStefano Zampini { 4114b9b85e73SStefano Zampini PetscViewer viewer; 4115b9b85e73SStefano Zampini char filename[256]; 4116b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4117b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4118b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4119b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4120b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4121b9b85e73SStefano Zampini } 4122b9b85e73SStefano Zampini #endif 4123e176bc59SStefano 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); 41246e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41256e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41266e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4127abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4128abbbba34SStefano Zampini 41296e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41306e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41316e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41326e683305SStefano Zampini const PetscInt *idxs; 41336e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41346e683305SStefano Zampini 41356e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41360be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41376e683305SStefano Zampini /* allocate space for temporary storage */ 4138854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4139854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41406e683305SStefano Zampini /* allocate for IS array */ 41416e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41426e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41436e683305SStefano Zampini nis = nisdofs + nisneu; 4144854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41456e683305SStefano Zampini /* dofs splitting */ 41466e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41476e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41486e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41496e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41506e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41516e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41526e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41536e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41546e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41556e683305SStefano Zampini } 41566e683305SStefano Zampini /* neumann boundaries */ 41576e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41586e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41596e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41606e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41616e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41626e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41636e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41646e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41656e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41666e683305SStefano Zampini } 41676e683305SStefano Zampini /* free memory */ 41686e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41696e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41706e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41716e683305SStefano Zampini } else { 41726e683305SStefano Zampini nis = 0; 41736e683305SStefano Zampini nisdofs = 0; 41746e683305SStefano Zampini nisneu = 0; 41756e683305SStefano Zampini isarray = NULL; 41766e683305SStefano Zampini } 41776e683305SStefano Zampini /* destroy no longer needed map */ 41786e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41796e683305SStefano Zampini 41806e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41816e683305SStefano Zampini coarse_mat_is = NULL; 41826e683305SStefano Zampini if (csin) { 41836e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 418474e2c79eSStefano Zampini if (redist) { 418574e2c79eSStefano Zampini PetscMPIInt rank; 4186779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 418774e2c79eSStefano Zampini 418874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 418958da7f69SStefano Zampini spc = active_procs/ncoarse; 419058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4191779c1cceSStefano Zampini if (im_active) { 4192779c1cceSStefano Zampini destsize = 1; 419374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 419474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 419574e2c79eSStefano Zampini } else { 419674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 419774e2c79eSStefano Zampini } 419874e2c79eSStefano Zampini } else { 4199779c1cceSStefano Zampini destsize = 0; 42006e683305SStefano Zampini } 4201779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4202779c1cceSStefano Zampini } else if (csin_type_simple) { 42036e683305SStefano Zampini PetscMPIInt rank; 42046e683305SStefano Zampini PetscInt issize,isidx; 4205779c1cceSStefano Zampini 42066e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42076e683305SStefano Zampini if (im_active) { 42086e683305SStefano Zampini issize = 1; 42096e683305SStefano Zampini isidx = (PetscInt)rank; 42106e683305SStefano Zampini } else { 42116e683305SStefano Zampini issize = 0; 42126e683305SStefano Zampini isidx = -1; 42136e683305SStefano Zampini } 42146e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4215779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4216b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42176e683305SStefano Zampini } 4218779c1cceSStefano Zampini 4219779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4220779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4221779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4222779c1cceSStefano Zampini PetscInt *coarse_candidates; 4223779c1cceSStefano Zampini const PetscInt* tisindices; 4224779c1cceSStefano Zampini 4225779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4226779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4227779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4228779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4229779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4230779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4231779c1cceSStefano Zampini } 4232779c1cceSStefano Zampini } 4233779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4234779c1cceSStefano Zampini 4235779c1cceSStefano Zampini 42366e683305SStefano Zampini if (pcbddc->dbg_flag) { 42376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42386e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42396e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4241779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42436e683305SStefano Zampini } 42446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42456e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42466e683305SStefano Zampini } 42476e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42486e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42496e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4250854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42516e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42526e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42536e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42546e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42556e683305SStefano Zampini } 42566e683305SStefano Zampini if (pcbddc->dbg_flag) { 42576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42586e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42596e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42606e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42616e683305SStefano Zampini } 4262779c1cceSStefano Zampini } 42636e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 426453a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 426553a05cb3SStefano 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); 426653a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 426753a05cb3SStefano 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); 426853a05cb3SStefano Zampini } 42696e683305SStefano Zampini } else { 42706e683305SStefano Zampini if (pcbddc->dbg_flag) { 42716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42736e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42746e683305SStefano Zampini } 42756e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42766e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42776e683305SStefano Zampini } 42786e683305SStefano Zampini 42796e683305SStefano Zampini /* create local to global scatters for coarse problem */ 428068457ee5SStefano Zampini if (compute_vecs) { 42816e683305SStefano Zampini PetscInt lrows; 42826e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42836e683305SStefano Zampini if (coarse_mat_is) { 42846e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42856e683305SStefano Zampini } else { 42866e683305SStefano Zampini lrows = 0; 42876e683305SStefano Zampini } 42886e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42896e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42906e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42916e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42926e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42936e683305SStefano Zampini } 42946e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42956e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4296c8587f34SStefano Zampini 4297f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4298f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4299f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4300f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4301f9eb5b7dSStefano Zampini } else { 4302f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4303f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4304c8587f34SStefano Zampini } 4305c8587f34SStefano Zampini 43066e683305SStefano Zampini /* print some info if requested */ 43076e683305SStefano Zampini if (pcbddc->dbg_flag) { 43086e683305SStefano Zampini if (!multilevel_allowed) { 43096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43106e683305SStefano Zampini if (multilevel_requested) { 43116e683305SStefano 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); 43126e683305SStefano Zampini } else if (pcbddc->max_levels) { 43136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43146e683305SStefano Zampini } 43156e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43166e683305SStefano Zampini } 43176e683305SStefano Zampini } 43186e683305SStefano Zampini 4319f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43206e683305SStefano Zampini if (coarse_mat_is) { 43216e683305SStefano Zampini MatReuse coarse_mat_reuse; 43226a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43236e683305SStefano Zampini if (pcbddc->dbg_flag) { 43246e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43256e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43266e683305SStefano Zampini } 4327f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4328312be037SStefano Zampini char prefix[256],str_level[16]; 4329e604994aSStefano Zampini size_t len; 43306e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4331422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4332c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4333f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43345f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4335c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43366e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4337c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4338c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4339e604994aSStefano Zampini /* prefix */ 4340e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4341e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4342e604994aSStefano Zampini if (!pcbddc->current_level) { 4343e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4344e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4345c8587f34SStefano Zampini } else { 4346e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4347312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4348312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 434934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4350312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4351e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4352e604994aSStefano Zampini } 4353e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43543e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43553e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43563e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43573e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4358f9eb5b7dSStefano Zampini /* allow user customization */ 4359f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43603e3c6dadSStefano Zampini } 43613e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 436251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 43633e3c6dadSStefano Zampini if (nisdofs) { 43643e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43653e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43663e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43673e3c6dadSStefano Zampini } 43683e3c6dadSStefano Zampini } 43693e3c6dadSStefano Zampini if (nisneu) { 43703e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43713e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4372312be037SStefano Zampini } 4373f9eb5b7dSStefano Zampini 4374f9eb5b7dSStefano Zampini /* get some info after set from options */ 4375f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4376f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43774f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43786e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4379f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4380f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4381f9eb5b7dSStefano Zampini } 438239f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 43834f3a063dSStefano Zampini if (isredundant) { 43844f3a063dSStefano Zampini KSP inner_ksp; 43854f3a063dSStefano Zampini PC inner_pc; 43864f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43874f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43884f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43894f3a063dSStefano Zampini } 4390f9eb5b7dSStefano Zampini 4391f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4392fa7f1dd8SStefano Zampini if (coarse_reuse) { 439381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4394fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43956e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4396fa7f1dd8SStefano Zampini } else { 43976e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4398fa7f1dd8SStefano Zampini } 4399c8587f34SStefano Zampini if (isbddc || isnn) { 440022bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 440170cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4402b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 440322b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44066e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44076e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 440822b6e8a2SStefano Zampini } 440970cf5478SStefano Zampini } 441053a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 441170cf5478SStefano Zampini } else { 441222bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 441322bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 441422bc73bbSStefano Zampini } 441522bc73bbSStefano Zampini } else { 44162e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4417c8587f34SStefano Zampini } 4418c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4419c8587f34SStefano Zampini 44203301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44215a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44223301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44233301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44243301b35fSStefano Zampini } 44253301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44263301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44273301b35fSStefano Zampini } 44283301b35fSStefano Zampini if (pc->pmat->spd_set) { 44293301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44303301b35fSStefano Zampini } 44316e683305SStefano Zampini /* set operators */ 44325f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44336e683305SStefano Zampini if (pcbddc->dbg_flag) { 44346e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44356e683305SStefano Zampini } 44366e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44376e683305SStefano Zampini coarse_mat = 0; 44386e683305SStefano Zampini } 44396e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4440b9b85e73SStefano Zampini #if 0 4441b9b85e73SStefano Zampini { 4442b9b85e73SStefano Zampini PetscViewer viewer; 4443b9b85e73SStefano Zampini char filename[256]; 4444b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4445b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4446b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4447b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4448b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4449b9b85e73SStefano Zampini } 4450b9b85e73SStefano Zampini #endif 4451c8587f34SStefano Zampini 4452c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4453298c0119SStefano Zampini #if 0 4454c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4455c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 445698a51de6SStefano Zampini } 4457298c0119SStefano Zampini #endif 445898a51de6SStefano Zampini 445998a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 446098a51de6SStefano Zampini Vec crhs,csol; 446198a51de6SStefano Zampini PetscBool ispreonly; 446204708bb6SStefano Zampini 446398a51de6SStefano Zampini if (CoarseNullSpace) { 4464c8587f34SStefano Zampini if (isbddc) { 4465c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4466c8587f34SStefano Zampini } else { 44675fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4468c8587f34SStefano Zampini } 4469c8587f34SStefano Zampini } 4470f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4471f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4472f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4473f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44746e683305SStefano Zampini /* hack */ 4475f347579bSStefano Zampini if (!csol) { 44762a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4477f9eb5b7dSStefano Zampini } 4478f347579bSStefano Zampini if (!crhs) { 44792a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4480f347579bSStefano Zampini } 4481cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4482cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44836e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4484c8587f34SStefano Zampini KSP check_ksp; 44852b510759SStefano Zampini KSPType check_ksp_type; 4486c8587f34SStefano Zampini PC check_pc; 44876e683305SStefano Zampini Vec check_vec,coarse_vec; 44886a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44892b510759SStefano Zampini PetscInt its; 44906e683305SStefano Zampini PetscBool compute_eigs; 44916e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44926e683305SStefano Zampini PetscInt neigs; 44938e185a42SStefano Zampini const char *prefix; 4494c8587f34SStefano Zampini 44952b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44966e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4497422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 449823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4499f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 45002b510759SStefano Zampini if (ispreonly) { 45012b510759SStefano Zampini check_ksp_type = KSPPREONLY; 45026e683305SStefano Zampini compute_eigs = PETSC_FALSE; 45032b510759SStefano Zampini } else { 4504cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45056e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4506c8587f34SStefano Zampini } 4507c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45086e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45096e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45106e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4511a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4512a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4513a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4514a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4515c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4516c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4517c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4518c8587f34SStefano Zampini /* create random vec */ 45196e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45206e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4521c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4522c8587f34SStefano Zampini if (CoarseNullSpace) { 4523c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4524c8587f34SStefano Zampini } 45256e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4526c8587f34SStefano Zampini /* solve coarse problem */ 45276e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4528c8587f34SStefano Zampini if (CoarseNullSpace) { 45296e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4530c8587f34SStefano Zampini } 4531cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45326e683305SStefano Zampini if (compute_eigs) { 4533854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4534854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45356e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45366e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45376e683305SStefano Zampini lambda_min = eigs_r[0]; 45386e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45396e683305SStefano Zampini if (lambda_max>lambda_min) { 4540cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4541cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4542cbcc2c2aSStefano Zampini } 4543c8587f34SStefano Zampini } 4544c8587f34SStefano Zampini } 4545cbcc2c2aSStefano Zampini 4546c8587f34SStefano Zampini /* check coarse problem residual error */ 45476e683305SStefano Zampini if (pcbddc->dbg_flag) { 45486e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45496e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45506e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4551c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45526e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45536e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4554c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4555779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45566e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45576e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45586e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45606e683305SStefano Zampini if (compute_eigs) { 45616e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4562deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4563c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45646e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45656e683305SStefano 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); 45666e683305SStefano Zampini for (i=0;i<neigs;i++) { 45676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4568c8587f34SStefano Zampini } 45696e683305SStefano Zampini } 45706e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45716e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45726e683305SStefano Zampini } 4573c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45746e683305SStefano Zampini if (compute_eigs) { 45756e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45766e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4577c8587f34SStefano Zampini } 45786e683305SStefano Zampini } 45796e683305SStefano Zampini } 4580cbcc2c2aSStefano Zampini /* print additional info */ 4581cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45826e683305SStefano Zampini /* waits until all processes reaches this point */ 45836e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4584cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4585cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4586cbcc2c2aSStefano Zampini } 4587cbcc2c2aSStefano Zampini 45882b510759SStefano Zampini /* free memory */ 4589c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4590fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4591c8587f34SStefano Zampini PetscFunctionReturn(0); 4592c8587f34SStefano Zampini } 4593674ae819SStefano Zampini 4594f34684f1SStefano Zampini #undef __FUNCT__ 4595f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4596f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4597f34684f1SStefano Zampini { 4598f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4599f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4600f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4601dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4602dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 460373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4604dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4605f34684f1SStefano Zampini PetscErrorCode ierr; 4606f34684f1SStefano Zampini 4607f34684f1SStefano Zampini PetscFunctionBegin; 4608f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46096c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4610dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46113bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4612dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4613dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4614dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4615dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4616dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4617dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46186c4ed002SBarry 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); 4619dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4620dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4621dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4622dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4623dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4624f34684f1SStefano Zampini 4625f34684f1SStefano Zampini /* check numbering */ 4626f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4627f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4628dc456d91SStefano Zampini PetscInt i; 4629b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4630f34684f1SStefano Zampini 4631f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4632f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4633f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46341575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4635f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4636f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4637727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4638f34684f1SStefano Zampini } 4639f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4640f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4641f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4642e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4643e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4644e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4645e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4646f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4647f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4648f34684f1SStefano Zampini if (array[i] == 1.0) { 4649b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4650f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4651f34684f1SStefano Zampini } 4652f34684f1SStefano Zampini } 4653*b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4654f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4655f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4656f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4657f34684f1SStefano Zampini } 4658f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4659f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4660e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4661e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4662f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4663f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4664b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4665ca8b9ea9SStefano Zampini PetscInt *gidxs; 4666ca8b9ea9SStefano Zampini 4667ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 46683bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4669f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4670f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4671f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4672f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46734bc2dc4bSStefano 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); 4674f34684f1SStefano Zampini } 4675f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4676ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4677f34684f1SStefano Zampini } 4678f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46791575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4680302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4681f34684f1SStefano Zampini } 46828bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4683f34684f1SStefano Zampini /* get back data */ 4684f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4685f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4686674ae819SStefano Zampini PetscFunctionReturn(0); 4687674ae819SStefano Zampini } 4688674ae819SStefano Zampini 4689e456f2a8SStefano Zampini #undef __FUNCT__ 4690e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4691a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4692e456f2a8SStefano Zampini { 4693e456f2a8SStefano Zampini IS localis_t; 4694a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4695e456f2a8SStefano Zampini PetscScalar *vals; 4696e456f2a8SStefano Zampini PetscErrorCode ierr; 4697e456f2a8SStefano Zampini 4698e456f2a8SStefano Zampini PetscFunctionBegin; 4699a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4700e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4701854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4702e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4703e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4704a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4705a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47061035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4707a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47081035eff8SStefano Zampini } 4709a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4710e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4711e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4712a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4713a7dc3881SStefano Zampini /* now compute set in local ordering */ 4714a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4715a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4716a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4717a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4718a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4719ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4720e456f2a8SStefano Zampini lsize++; 4721e456f2a8SStefano Zampini } 4722e456f2a8SStefano Zampini } 4723854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4724a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4725ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4726e456f2a8SStefano Zampini idxs[lsize++] = i; 4727e456f2a8SStefano Zampini } 4728e456f2a8SStefano Zampini } 4729a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4730a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4731e456f2a8SStefano Zampini *localis = localis_t; 4732e456f2a8SStefano Zampini PetscFunctionReturn(0); 4733e456f2a8SStefano Zampini } 4734906d46d4SStefano Zampini 4735906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4736906d46d4SStefano Zampini #undef __FUNCT__ 4737906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4738906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4739906d46d4SStefano Zampini { 4740906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4741906d46d4SStefano Zampini PetscErrorCode ierr; 4742906d46d4SStefano Zampini 4743906d46d4SStefano Zampini PetscFunctionBegin; 4744906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4745906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4746906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4747906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4748906d46d4SStefano Zampini PetscFunctionReturn(0); 4749906d46d4SStefano Zampini } 4750906d46d4SStefano Zampini 4751906d46d4SStefano Zampini #undef __FUNCT__ 4752906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4753906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4754906d46d4SStefano Zampini { 4755906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4756906d46d4SStefano Zampini PetscErrorCode ierr; 4757906d46d4SStefano Zampini 4758906d46d4SStefano Zampini PetscFunctionBegin; 4759906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4760906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4761906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4762906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4763906d46d4SStefano Zampini PetscFunctionReturn(0); 4764906d46d4SStefano Zampini } 4765b96c3477SStefano Zampini 4766b96c3477SStefano Zampini #undef __FUNCT__ 4767b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 476808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4769b96c3477SStefano Zampini { 4770a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4771b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4772b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4773a64f4aa4SStefano Zampini Mat S_j; 4774b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4775b96c3477SStefano Zampini PetscBool free_used_adj; 4776b96c3477SStefano Zampini PetscErrorCode ierr; 4777b96c3477SStefano Zampini 4778b96c3477SStefano Zampini PetscFunctionBegin; 4779b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4780b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 478108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4782b96c3477SStefano Zampini used_xadj = NULL; 4783b96c3477SStefano Zampini used_adjncy = NULL; 4784b96c3477SStefano Zampini } else { 478508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 478608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 478708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 478808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4789b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4790b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4791b96c3477SStefano Zampini } else { 47922fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4793b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4794b96c3477SStefano Zampini PetscInt nvtxs; 4795b96c3477SStefano Zampini 47962fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47972fffb893SStefano Zampini if (flg_row) { 4798b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4799b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4800b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4801b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48022fffb893SStefano Zampini } else { 48032fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48042fffb893SStefano Zampini used_xadj = NULL; 48052fffb893SStefano Zampini used_adjncy = NULL; 48062fffb893SStefano Zampini } 48072fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4808b96c3477SStefano Zampini } 4809b96c3477SStefano Zampini } 4810d5574798SStefano Zampini 4811d5574798SStefano Zampini /* setup sub_schurs data */ 4812a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4813a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4814a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4815a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4816862806e4SStefano 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); 4817a64f4aa4SStefano Zampini } else { 48186816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 481904708bb6SStefano Zampini PetscBool isseqaij; 48205feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48215feab87aSStefano Zampini PetscInt n_vertices; 48225feab87aSStefano Zampini 48235feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48242034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48255feab87aSStefano Zampini } 482604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 482704708bb6SStefano Zampini if (!isseqaij) { 482804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 482904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 483004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 483104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483204708bb6SStefano Zampini } else { 483304708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483404708bb6SStefano Zampini } 483504708bb6SStefano Zampini } 4836862806e4SStefano 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); 4837a64f4aa4SStefano Zampini } 4838a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4839b96c3477SStefano Zampini 4840b96c3477SStefano Zampini /* free adjacency */ 4841b96c3477SStefano Zampini if (free_used_adj) { 4842b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4843b96c3477SStefano Zampini } 4844b96c3477SStefano Zampini PetscFunctionReturn(0); 4845b96c3477SStefano Zampini } 4846b96c3477SStefano Zampini 4847b96c3477SStefano Zampini #undef __FUNCT__ 4848b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 484908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4850b96c3477SStefano Zampini { 4851b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4852b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4853b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4854b96c3477SStefano Zampini PCBDDCGraph graph; 4855b96c3477SStefano Zampini PetscErrorCode ierr; 4856b96c3477SStefano Zampini 4857b96c3477SStefano Zampini PetscFunctionBegin; 4858b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 485908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48603301b35fSStefano Zampini IS verticesIS,verticescomm; 48613301b35fSStefano Zampini PetscInt vsize,*idxs; 4862b96c3477SStefano Zampini 4863b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48643301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48653301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48663301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48673301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48683301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4869b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48707fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48713301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48723301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4873b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4874b96c3477SStefano Zampini /* 4875b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4876b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4877b96c3477SStefano Zampini } 4878b96c3477SStefano Zampini */ 4879b96c3477SStefano Zampini } else { 4880b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4881b96c3477SStefano Zampini } 4882b96c3477SStefano Zampini 4883b96c3477SStefano Zampini /* sub_schurs init */ 4884a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4885a64f4aa4SStefano Zampini 4886b96c3477SStefano Zampini /* free graph struct */ 488708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4888b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4889b96c3477SStefano Zampini } 4890b96c3477SStefano Zampini PetscFunctionReturn(0); 4891b96c3477SStefano Zampini } 4892