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 8b1b3d7a2SStefano Zampini #undef __FUNCT__ 9b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 1008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 11b1b3d7a2SStefano Zampini { 12b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 1508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 1608122e43SStefano Zampini PetscScalar *work,lwork; 1708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 1808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 1908122e43SStefano Zampini PetscReal *eigs,thresh; 201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 21f6f667cfSStefano Zampini PetscBool allocated_S_St; 2208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2308122e43SStefano Zampini PetscReal *rwork; 2408122e43SStefano Zampini #endif 25b1b3d7a2SStefano Zampini PetscErrorCode ierr; 26b1b3d7a2SStefano Zampini 27b1b3d7a2SStefano Zampini PetscFunctionBegin; 2808122e43SStefano Zampini if (!sub_schurs->use_mumps) { 2908122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3008122e43SStefano Zampini } 3108122e43SStefano Zampini 32fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 33fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 34fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 35fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 36fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 5108122e43SStefano Zampini PetscInt subset_size; 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini } 5608122e43SStefano Zampini 5708122e43SStefano Zampini /* min/max and threshold */ 5808122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 59f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 6008122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 61f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 62f6f667cfSStefano Zampini if (nmin) { 63f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 64f6f667cfSStefano Zampini } 6508122e43SStefano Zampini 6608122e43SStefano Zampini /* allocate lapack workspace */ 6708122e43SStefano Zampini cum = cum2 = 0; 6808122e43SStefano Zampini maxneigs = 0; 6908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 7108122e43SStefano Zampini PetscInt n,subset_size; 72f6f667cfSStefano Zampini 7308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 759162d606SStefano Zampini cum += subset_size; 769162d606SStefano Zampini cum2 += subset_size*n; 7708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7808122e43SStefano Zampini } 7908122e43SStefano Zampini } 8008122e43SStefano Zampini if (mss) { 819ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 8208122e43SStefano Zampini PetscBLASInt B_itype = 1; 8308122e43SStefano Zampini PetscBLASInt B_N = mss; 844c6709b3SStefano Zampini PetscReal zero = 0.0; 854c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8608122e43SStefano Zampini 8708122e43SStefano Zampini B_lwork = -1; 8808122e43SStefano Zampini S = NULL; 8908122e43SStefano Zampini St = NULL; 90a58a30b4SStefano Zampini eigs = NULL; 91a58a30b4SStefano Zampini eigv = NULL; 92a58a30b4SStefano Zampini B_iwork = NULL; 93a58a30b4SStefano Zampini B_ifail = NULL; 948bec7fa6SStefano Zampini thresh = 1.0; 9508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9708122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 9808122e43SStefano Zampini #else 9908122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 10008122e43SStefano Zampini #endif 10108122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 10208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10308122e43SStefano Zampini } else { 10408122e43SStefano Zampini /* TODO */ 10508122e43SStefano Zampini } 10608122e43SStefano Zampini } else { 10708122e43SStefano Zampini lwork = 0; 10808122e43SStefano Zampini } 10908122e43SStefano Zampini 11008122e43SStefano Zampini nv = 0; 111d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 112d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 11308122e43SStefano Zampini } 1144c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 115f6f667cfSStefano Zampini if (allocated_S_St) { 116f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 117f6f667cfSStefano Zampini } 118f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12008122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 12108122e43SStefano Zampini #endif 1229162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1239162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1249162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1269162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12708122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12808122e43SStefano Zampini 12908122e43SStefano Zampini maxneigs = 0; 13008122e43SStefano Zampini cum = cum2 = cumarray = 0; 1319162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1329162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 133d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13408122e43SStefano Zampini const PetscInt *idxs; 13508122e43SStefano Zampini 136d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13708122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13808122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 14008122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1419162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1429162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 14308122e43SStefano Zampini } 14408122e43SStefano Zampini cum2 = cum; 145d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14608122e43SStefano Zampini } 14708122e43SStefano Zampini 14808122e43SStefano Zampini if (mss) { /* multilevel */ 14908122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 15008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 15108122e43SStefano Zampini } 15208122e43SStefano Zampini 15308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15408122e43SStefano Zampini PetscInt j,subset_size; 15508122e43SStefano Zampini 15608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 15708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 15808122e43SStefano Zampini const PetscInt *idxs; 159f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 160f6f667cfSStefano Zampini PetscInt eigs_start = 0; 16108122e43SStefano Zampini PetscBLASInt B_N; 162aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 16308122e43SStefano Zampini 164f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 165f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1669ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 167aff50787SStefano Zampini PetscInt j,k; 168aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 169aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 170aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17108122e43SStefano Zampini } 17208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 173aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 174aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 175aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 176aff50787SStefano Zampini } 17708122e43SStefano Zampini } 17808122e43SStefano Zampini } else { 17908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 18008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 18108122e43SStefano Zampini } 1828bec7fa6SStefano Zampini } else { 183f6f667cfSStefano Zampini S = Sarray + cumarray; 184f6f667cfSStefano Zampini St = Starray + cumarray; 1858bec7fa6SStefano Zampini } 18608122e43SStefano Zampini 187f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 188aff50787SStefano Zampini /* see if we can save some work */ 189aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 190aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 191aff50787SStefano Zampini } 192aff50787SStefano Zampini 193aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 194aff50787SStefano Zampini B_neigs = 0; 195aff50787SStefano Zampini } else { 196aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 197f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 198f6f667cfSStefano Zampini 1999ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 20008122e43SStefano Zampini PetscBLASInt B_itype = 1; 201f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2024c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2039552c7c7SStefano Zampini PetscInt nmin_s; 20408122e43SStefano Zampini 205f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 206fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2078bec7fa6SStefano 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]]); 208fd14bc51SStefano Zampini } 20908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 211f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 21208122e43SStefano Zampini #else 213f6f667cfSStefano 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)); 21408122e43SStefano Zampini #endif 21508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 21608122e43SStefano Zampini if (B_ierr) { 21708122e43SStefano Zampini if (B_ierr < 0 ) { 21808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 21908122e43SStefano Zampini } else if (B_ierr <= B_N) { 22008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 22108122e43SStefano Zampini } else { 2229552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 22308122e43SStefano Zampini } 22408122e43SStefano Zampini } 22508122e43SStefano Zampini 22608122e43SStefano Zampini if (B_neigs > nmax) { 227fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 228fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 229fd14bc51SStefano Zampini } 230f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 23108122e43SStefano Zampini B_neigs = nmax; 23208122e43SStefano Zampini } 23308122e43SStefano Zampini 2349552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2359552c7c7SStefano Zampini if (B_neigs < nmin_s) { 23608122e43SStefano Zampini PetscBLASInt B_neigs2; 23708122e43SStefano Zampini 238f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 239f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 240fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 241fd14bc51SStefano 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); 242fd14bc51SStefano Zampini } 2439ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 24408122e43SStefano Zampini PetscInt j; 24508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24608122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24708122e43SStefano Zampini } 24808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24908122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 25008122e43SStefano Zampini } 25108122e43SStefano Zampini } else { 25208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25408122e43SStefano Zampini } 25508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 257f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 25808122e43SStefano Zampini #else 259f6f667cfSStefano 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)); 26008122e43SStefano Zampini #endif 26108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 26208122e43SStefano Zampini B_neigs += B_neigs2; 26308122e43SStefano Zampini } 26408122e43SStefano Zampini if (B_ierr) { 26508122e43SStefano Zampini if (B_ierr < 0 ) { 26608122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 26708122e43SStefano Zampini } else if (B_ierr <= B_N) { 26808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 26908122e43SStefano Zampini } else { 2709552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 27108122e43SStefano Zampini } 27208122e43SStefano Zampini } 273fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 274fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 27508122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 27608122e43SStefano Zampini if (eigs[j] == 0.0) { 277fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 27808122e43SStefano Zampini } else { 279f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 280fd14bc51SStefano Zampini } 28108122e43SStefano Zampini } 28208122e43SStefano Zampini } 28308122e43SStefano Zampini } else { 28408122e43SStefano Zampini /* TODO */ 28508122e43SStefano Zampini } 286aff50787SStefano Zampini } 2878bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2888bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2899162d606SStefano Zampini if (B_neigs) { 2909162d606SStefano 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); 291fd14bc51SStefano Zampini 292fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2939552c7c7SStefano Zampini PetscInt ii; 2949552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 295fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2969552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2979162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2989552c7c7SStefano Zampini } 2999552c7c7SStefano Zampini } 300fd14bc51SStefano Zampini } 30108122e43SStefano Zampini #if 0 3029162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 30308122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 30408122e43SStefano Zampini PetscScalar norm; 30508122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3069162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3079162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 30808122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 30908122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 31008122e43SStefano Zampini } else { 31108122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 31208122e43SStefano Zampini } 3139162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 314b1b3d7a2SStefano Zampini } 315b1b3d7a2SStefano Zampini #endif 3169162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3179162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3189162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3199162d606SStefano Zampini cum++; 32008122e43SStefano Zampini } 32108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 32208122e43SStefano Zampini } 32308122e43SStefano Zampini /* shift for next computation */ 32408122e43SStefano Zampini cumarray += subset_size*subset_size; 32508122e43SStefano Zampini } 326fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 327fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 328fd14bc51SStefano Zampini } 32908122e43SStefano Zampini 33008122e43SStefano Zampini if (mss) { 33108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 33208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 333f6f667cfSStefano Zampini /* destroy matrices (junk) */ 334f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 335f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 33608122e43SStefano Zampini } 337f6f667cfSStefano Zampini if (allocated_S_St) { 338f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 339f6f667cfSStefano Zampini } 340f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 34108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 34208122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 34308122e43SStefano Zampini #endif 34408122e43SStefano Zampini if (pcbddc->dbg_flag) { 3451b968477SStefano Zampini PetscInt maxneigs_r; 34608122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3479b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 34808122e43SStefano Zampini } 34908122e43SStefano Zampini PetscFunctionReturn(0); 35008122e43SStefano Zampini } 351b1b3d7a2SStefano Zampini 352674ae819SStefano Zampini #undef __FUNCT__ 353c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 354c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 355c8587f34SStefano Zampini { 356c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3578629588bSStefano Zampini PetscScalar *coarse_submat_vals; 358c8587f34SStefano Zampini PetscErrorCode ierr; 359c8587f34SStefano Zampini 360c8587f34SStefano Zampini PetscFunctionBegin; 361f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3625e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 363c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 364c8587f34SStefano Zampini 365684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3660fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 367684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 368c8587f34SStefano Zampini 369c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 370b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 371c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 372c8587f34SStefano Zampini } 373c8587f34SStefano Zampini 3748629588bSStefano Zampini /* 3758629588bSStefano Zampini Setup local correction and local part of coarse basis. 3768629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3778629588bSStefano Zampini */ 37847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3798629588bSStefano Zampini 3808629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3818629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3828629588bSStefano Zampini 3838629588bSStefano Zampini /* free */ 3848629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 385c8587f34SStefano Zampini PetscFunctionReturn(0); 386c8587f34SStefano Zampini } 387c8587f34SStefano Zampini 388c8587f34SStefano Zampini #undef __FUNCT__ 389674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 390674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 391674ae819SStefano Zampini { 392674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 393674ae819SStefano Zampini PetscErrorCode ierr; 394674ae819SStefano Zampini 395674ae819SStefano Zampini PetscFunctionBegin; 396674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 397674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 398674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 399674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 400785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 401674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 402f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 403f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 404785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 40563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 40663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 407674ae819SStefano Zampini PetscFunctionReturn(0); 408674ae819SStefano Zampini } 409674ae819SStefano Zampini 410674ae819SStefano Zampini #undef __FUNCT__ 411674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 412674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 413674ae819SStefano Zampini { 414674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 415674ae819SStefano Zampini PetscErrorCode ierr; 416674ae819SStefano Zampini 417674ae819SStefano Zampini PetscFunctionBegin; 418b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 419674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 420674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 421674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 422b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 423674ae819SStefano Zampini PetscFunctionReturn(0); 424674ae819SStefano Zampini } 425674ae819SStefano Zampini 426674ae819SStefano Zampini #undef __FUNCT__ 427674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 428674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 429674ae819SStefano Zampini { 430674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 43106656605SStefano Zampini PetscScalar *array; 432674ae819SStefano Zampini PetscErrorCode ierr; 433674ae819SStefano Zampini 434674ae819SStefano Zampini PetscFunctionBegin; 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43658da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 43706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 43806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43958da7f69SStefano Zampini } 440674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 441674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 44215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 44315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 444674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 445674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 446674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 44706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 448674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 449674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4508ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 451674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 452674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 453674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 454f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 455f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 456f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 457f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 458727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4590e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 460f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 46170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4626e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 463674ae819SStefano Zampini PetscFunctionReturn(0); 464674ae819SStefano Zampini } 465674ae819SStefano Zampini 466674ae819SStefano Zampini #undef __FUNCT__ 467f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 468f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4696bfb1811SStefano Zampini { 4706bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4716bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4726bfb1811SStefano Zampini VecType impVecType; 473e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4746bfb1811SStefano Zampini PetscErrorCode ierr; 4756bfb1811SStefano Zampini 4766bfb1811SStefano Zampini PetscFunctionBegin; 477f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 478f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 479f4ddd8eeSStefano Zampini } 480e7b262bdSStefano Zampini /* get sizes */ 481b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 482b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4836bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 484e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 485e7b262bdSStefano Zampini /* R nodes */ 486e7b262bdSStefano Zampini old_size = -1; 487e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e7b262bdSStefano Zampini if (n_R != old_size) { 491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 492e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4936bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4946bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4956bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4966bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 497e7b262bdSStefano Zampini } 498e7b262bdSStefano Zampini /* local primal dofs */ 499e7b262bdSStefano Zampini old_size = -1; 500e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 501e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 504e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 506e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5076bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 508e7b262bdSStefano Zampini } 509e7b262bdSStefano Zampini /* local explicit constraints */ 510e7b262bdSStefano Zampini old_size = -1; 511e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 512e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 513e7b262bdSStefano Zampini } 514e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 515e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 51683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 51783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 51883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 51983b7ccabSStefano Zampini } 5206bfb1811SStefano Zampini PetscFunctionReturn(0); 5216bfb1811SStefano Zampini } 5226bfb1811SStefano Zampini 5236bfb1811SStefano Zampini #undef __FUNCT__ 52447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 52547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 52688ebb749SStefano Zampini { 52725084f0cSStefano Zampini PetscErrorCode ierr; 52825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 52988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 53088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 531d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 53225084f0cSStefano Zampini /* submatrices of local problem */ 53380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 53406656605SStefano Zampini /* submatrices of local coarse problem */ 53506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 53625084f0cSStefano Zampini /* working matrices */ 53706656605SStefano Zampini Mat C_CR; 53825084f0cSStefano Zampini /* additional working stuff */ 53906656605SStefano Zampini PC pc_R; 54006656605SStefano Zampini Mat F; 54106656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 54206656605SStefano Zampini 54325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 54406656605SStefano Zampini PetscScalar *work; 54506656605SStefano Zampini PetscInt *idx_V_B; 54606656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 54706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 548b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 54945a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 55088ebb749SStefano Zampini MatType impMatType; 55125084f0cSStefano Zampini /* some shortcuts to scalars */ 55206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 55388ebb749SStefano Zampini 55488ebb749SStefano Zampini PetscFunctionBegin; 555b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 55688ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 55788ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 558b371cd4fSStefano Zampini n_B = pcis->n_B; 559b371cd4fSStefano Zampini n_D = pcis->n - n_B; 56088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 56188ebb749SStefano Zampini 56288ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 56388ebb749SStefano Zampini impMatType = MATSEQDENSE; 56488ebb749SStefano Zampini 56588ebb749SStefano Zampini /* vertices in boundary numbering */ 566785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5670e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 56888ebb749SStefano Zampini if (i != n_vertices) { 56922d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 57088ebb749SStefano Zampini } 57188ebb749SStefano Zampini 57206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 57306656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 57406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 57506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 57606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 57706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 57806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 57906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 58206656605SStefano Zampini 58306656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 58406656605SStefano Zampini /* allocate workspace */ 58506656605SStefano Zampini n = 0; 58606656605SStefano Zampini if (n_constraints) { 58706656605SStefano Zampini n += n_R*n_constraints; 58806656605SStefano Zampini } 58906656605SStefano Zampini if (n_vertices) { 59006656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 59180677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 59206656605SStefano Zampini } 5933301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 59406656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 59506656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 59606656605SStefano Zampini } 59706656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 59806656605SStefano Zampini 59906656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 60006656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 60106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 60206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 60306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 60406656605SStefano Zampini if (isLU || isILU || isCHOL) { 60506656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 606d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 607d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 608d62866d3SStefano Zampini MatFactorType type; 609d62866d3SStefano Zampini 6106816873aSStefano Zampini F = reuse_mumps->F; 6116816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 612d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 61306656605SStefano Zampini } else { 61406656605SStefano Zampini F = NULL; 61506656605SStefano Zampini } 61606656605SStefano Zampini 61788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 61888ebb749SStefano Zampini if (n_constraints) { 61906656605SStefano Zampini Mat M1,M2,M3; 62080677318SStefano Zampini Mat auxmat; 62106656605SStefano Zampini IS is_aux; 62280677318SStefano Zampini PetscScalar *array,*array2; 62306656605SStefano Zampini 624f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 62580677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 62688ebb749SStefano Zampini 62725084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62825084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6298ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 63080677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 63188ebb749SStefano Zampini 63280677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 63380677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 63406656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 63588ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63606656605SStefano Zampini const PetscScalar *row_cmat_values; 63706656605SStefano Zampini const PetscInt *row_cmat_indices; 63806656605SStefano Zampini PetscInt size_of_constraint,j; 63988ebb749SStefano Zampini 64006656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64106656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 64206656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64306656605SStefano Zampini } 64406656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64506656605SStefano Zampini } 64680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 64706656605SStefano Zampini if (F) { 64806656605SStefano Zampini Mat B; 64906656605SStefano Zampini 65006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 65180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 65206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 65306656605SStefano Zampini } else { 65480677318SStefano Zampini PetscScalar *marr; 65580677318SStefano Zampini 65680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 65806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 66006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 66106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 66306656605SStefano Zampini } 66480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66506656605SStefano Zampini } 66680677318SStefano Zampini if (!pcbddc->switch_static) { 66780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 66880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 67180677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 67280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 67380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 67680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 67780677318SStefano Zampini } 67880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 68080677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68180677318SStefano Zampini } else { 68280677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 68380677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 68425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68580677318SStefano Zampini } 68680677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 68780677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 68880677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 68906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 69006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 69180677318SStefano Zampini if (isCHOL) { 69280677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 69380677318SStefano Zampini } else { 69425084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 69580677318SStefano Zampini } 69680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 69706656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 69825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 69925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 70025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 70180677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 70280677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 70380677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 70406656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 70506656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 70688ebb749SStefano Zampini } 70788ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 70888ebb749SStefano Zampini if (n_vertices) { 70906656605SStefano Zampini Mat newmat; 71006656605SStefano Zampini IS is_aux; 7113a50541eSStefano Zampini PetscInt ibs,mbs; 7123a50541eSStefano Zampini PetscBool issbaij; 7133a50541eSStefano Zampini 7146816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7156816873aSStefano Zampini IS tis; 7166816873aSStefano Zampini 7176816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7186816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7196816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7206816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7216816873aSStefano Zampini } else { 7223a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7236816873aSStefano Zampini } 7249577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 7259577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 7263a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 7279577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7289577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7299577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7309577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 73180677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7329577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7339577ea80SStefano Zampini } else { 7343a50541eSStefano Zampini /* this is safe */ 7359577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 7363a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7373a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 738b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 739b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7403a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7413a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 74206656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7433a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7443a50541eSStefano Zampini } else { 745b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7469577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7479577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7489577ea80SStefano Zampini } 74980677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 75080677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7513a50541eSStefano Zampini } 75225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 75388ebb749SStefano Zampini } 75488ebb749SStefano Zampini 75588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 756f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 75706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 75806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 75906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 76006656605SStefano Zampini } 761f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 76206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 76306656605SStefano Zampini PetscScalar *marray; 76406656605SStefano Zampini 76506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 76606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 767f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 768f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 769f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 770f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 771f4ddd8eeSStefano Zampini } 772f4ddd8eeSStefano Zampini } 77306656605SStefano Zampini 774f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 77506656605SStefano Zampini PetscScalar *marray; 77688ebb749SStefano Zampini 77706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7788eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 78088ebb749SStefano Zampini } 7813301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 78206656605SStefano Zampini n *= 2; 783f4ddd8eeSStefano Zampini } 78406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 78506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 78606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 78706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 78806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 78906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 79088ebb749SStefano Zampini } 7913301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 79206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7931b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 79406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 79506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 79688ebb749SStefano Zampini } 79788ebb749SStefano Zampini } else { 798c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 799c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8001b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 801c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 802c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 803c0553b1fSStefano Zampini } 80488ebb749SStefano Zampini } 80506656605SStefano Zampini } 80606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 80706656605SStefano Zampini /* vertices */ 80806656605SStefano Zampini if (n_vertices) { 80916f15bc4SStefano Zampini 81016f15bc4SStefano Zampini if (n_R) { 81106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 81206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 81316f15bc4SStefano Zampini PetscScalar *x,*y; 81406656605SStefano Zampini 81506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 81621eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 81706656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 8186816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 81906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 82006656605SStefano Zampini } else { 82106656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 82206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 82306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 82406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 82506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 82606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 82806656605SStefano Zampini } 82906656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 83006656605SStefano Zampini } 83180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 83206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 83306656605SStefano Zampini if (n_constraints) { 83406656605SStefano Zampini Mat B; 83580677318SStefano Zampini 836b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 83780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 83880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 83980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 84080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 84380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 84480677318SStefano Zampini } 84580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 84680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 84780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 84806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 84980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 85206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 85306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 85406656605SStefano Zampini } 85506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 85680677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 86006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 86106656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 86516f15bc4SStefano Zampini } else { 86616f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 86716f15bc4SStefano Zampini } 86821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 86906656605SStefano Zampini /* coarse basis functions */ 87006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 87116f15bc4SStefano Zampini PetscScalar *y; 87216f15bc4SStefano Zampini 87306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 87406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 87506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 87606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 87906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 88106656605SStefano Zampini 88206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 88306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 88406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 88906656605SStefano Zampini } 89006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 89106656605SStefano Zampini } 89206656605SStefano Zampini } 89306656605SStefano Zampini 89406656605SStefano Zampini if (n_constraints) { 89506656605SStefano Zampini Mat B; 89606656605SStefano Zampini 89706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 89806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 89980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 90006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 90106656605SStefano Zampini if (n_vertices) { 90280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 90380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 90480677318SStefano Zampini } else { 90580677318SStefano Zampini Mat S_VCt; 90680677318SStefano Zampini 90780677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 90880677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 90980677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 91080677318SStefano Zampini } 91106656605SStefano Zampini } 91206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 91306656605SStefano Zampini /* coarse basis functions */ 91406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 91506656605SStefano Zampini PetscScalar *y; 91606656605SStefano Zampini 91706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 92006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 92406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 92506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 92606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 92706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 93006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 93106656605SStefano Zampini } 93206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 93306656605SStefano Zampini } 93406656605SStefano Zampini } 93580677318SStefano Zampini if (n_constraints) { 93680677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 93780677318SStefano Zampini } 93806656605SStefano Zampini 93906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9403301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 94106656605SStefano Zampini 94206656605SStefano Zampini if (n_constraints) { 94316f15bc4SStefano Zampini Mat S_CCT,B_C; 94406656605SStefano Zampini 94580677318SStefano Zampini /* this is a lazy thing */ 94680677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 94706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 94806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 95016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 95106656605SStefano Zampini if (n_vertices) { 95216f15bc4SStefano Zampini Mat B_V,S_VCT; 95306656605SStefano Zampini 95406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 95506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 95706656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 95816f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 95906656605SStefano Zampini } 96006656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 96180677318SStefano Zampini } else { /* if there are no constraints, reset work */ 96280677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 96306656605SStefano Zampini } 96416f15bc4SStefano Zampini if (n_vertices && n_R) { 96506656605SStefano Zampini Mat A_VRT; 96680677318SStefano Zampini PetscScalar *marray; 96706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 96806656605SStefano Zampini 96980677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 97080677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 97180677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 97380677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 97480677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 97516f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 97606656605SStefano Zampini } 97706656605SStefano Zampini 97806656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 97906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 98506656605SStefano Zampini } 98606656605SStefano Zampini } else { 98706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 99306656605SStefano Zampini } 99406656605SStefano Zampini } 99506656605SStefano Zampini /* coarse basis functions */ 99606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 99706656605SStefano Zampini PetscScalar *y; 99806656605SStefano Zampini 99906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 100006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100406656605SStefano Zampini if (i<n_vertices) { 100506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 100606656605SStefano Zampini } 100706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 100906656605SStefano Zampini 101006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 101106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 101606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 101706656605SStefano Zampini } 101806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 101906656605SStefano Zampini } 102006656605SStefano Zampini } 1021d62866d3SStefano Zampini /* free memory */ 102288ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 102306656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 102406656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 102506656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 102606656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1027d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1028d62866d3SStefano Zampini if (n_vertices) { 1029d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1030d62866d3SStefano Zampini } 1031d62866d3SStefano Zampini if (n_constraints) { 1032d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1033d62866d3SStefano Zampini } 103488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 103588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 103688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 103725084f0cSStefano Zampini if (pcbddc->dbg_flag) { 103888ebb749SStefano Zampini Mat coarse_sub_mat; 103925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 104088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 104188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 104288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10438bec7fa6SStefano Zampini Mat C_B,CPHI; 10448bec7fa6SStefano Zampini IS is_dummy; 10458bec7fa6SStefano Zampini Vec mones; 104688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 104788ebb749SStefano Zampini PetscReal real_value; 104888ebb749SStefano Zampini 104988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1055c0553b1fSStefano Zampini if (unsymmetric_check) { 105688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 105888ebb749SStefano Zampini } 105988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 106088ebb749SStefano Zampini 106125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10623301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 106325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1064c0553b1fSStefano Zampini if (unsymmetric_check) { 106588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 107388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107788ebb749SStefano Zampini } else { 107888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 107988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 108088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 108188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 108288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 108388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 108488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 108588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 108688ebb749SStefano Zampini } 108788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 108888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 108988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 109181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10928bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10930fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 109406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10958bec7fa6SStefano Zampini 10968bec7fa6SStefano Zampini /* check constraints */ 10978bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10988bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10998bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 11008bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 11018bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 11028bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 11038bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1104*bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1105*bdae7319SStefano Zampini if (unsymmetric_check) { 1106*bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1107*bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1108*bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1109*bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1110*bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1111*bdae7319SStefano Zampini } 11128bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 11138bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11148bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11158bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11168bec7fa6SStefano Zampini 111725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 111888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 111988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 112088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 112188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 112288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 112388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 112488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 112588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 112688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 112788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1128c0553b1fSStefano Zampini if (unsymmetric_check) { 112988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 113088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 113188ebb749SStefano Zampini } 113288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 113388ebb749SStefano Zampini } 11348629588bSStefano Zampini /* get back data */ 11358629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 113688ebb749SStefano Zampini PetscFunctionReturn(0); 113788ebb749SStefano Zampini } 113888ebb749SStefano Zampini 113988ebb749SStefano Zampini #undef __FUNCT__ 1140d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1141d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1142d65f70fdSStefano Zampini { 1143d65f70fdSStefano Zampini Mat *work_mat; 1144d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1145d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1146d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1147d65f70fdSStefano Zampini PetscErrorCode ierr; 1148d65f70fdSStefano Zampini 1149d65f70fdSStefano Zampini PetscFunctionBegin; 1150d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini 1155d65f70fdSStefano Zampini if (!rsorted) { 1156d65f70fdSStefano Zampini const PetscInt *idxs; 1157d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1158d65f70fdSStefano Zampini 1159d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1161d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1162d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1163d65f70fdSStefano Zampini } 1164d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1167d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini } else { 1172d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1173d65f70fdSStefano Zampini isrow_s = isrow; 1174d65f70fdSStefano Zampini } 1175d65f70fdSStefano Zampini 1176d65f70fdSStefano Zampini if (!csorted) { 1177d65f70fdSStefano Zampini if (isrow == iscol) { 1178d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini iscol_s = isrow_s; 1180d65f70fdSStefano Zampini } else { 1181d65f70fdSStefano Zampini const PetscInt *idxs; 1182d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1183d65f70fdSStefano Zampini 1184d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1185d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1187d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1188d65f70fdSStefano Zampini } 1189d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1190d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1192d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini } 1197d65f70fdSStefano Zampini } else { 1198d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini iscol_s = iscol; 1200d65f70fdSStefano Zampini } 1201d65f70fdSStefano Zampini 1202d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini 1204d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1205d65f70fdSStefano Zampini Mat new_mat; 1206d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1207d65f70fdSStefano Zampini 1208d65f70fdSStefano Zampini if (!rsorted) { 1209d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1210d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1212d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1213d65f70fdSStefano Zampini } 1214d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1215d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini } else { 1217d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini } 1219d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini 1221d65f70fdSStefano Zampini if (!csorted) { 1222d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1223d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1225d65f70fdSStefano Zampini } else { 1226d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1227d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1228d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1229d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1230d65f70fdSStefano Zampini } 1231d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1232d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1233d65f70fdSStefano Zampini } 1234d65f70fdSStefano Zampini } else { 1235d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1236d65f70fdSStefano Zampini } 1237d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1238d65f70fdSStefano Zampini 1239d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1240d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1241d65f70fdSStefano Zampini work_mat[0] = new_mat; 1242d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1243d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1244d65f70fdSStefano Zampini } 1245d65f70fdSStefano Zampini 1246d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1247d65f70fdSStefano Zampini *B = work_mat[0]; 1248d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1249d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1250d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1251d65f70fdSStefano Zampini PetscFunctionReturn(0); 1252d65f70fdSStefano Zampini } 1253d65f70fdSStefano Zampini 1254d65f70fdSStefano Zampini #undef __FUNCT__ 12555e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12565e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1257aa0d41d4SStefano Zampini { 1258aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12595e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1260d65f70fdSStefano Zampini Mat new_mat; 12615e8657edSStefano Zampini IS is_local,is_global; 1262d65f70fdSStefano Zampini PetscInt local_size; 1263d65f70fdSStefano Zampini PetscBool isseqaij; 1264aa0d41d4SStefano Zampini PetscErrorCode ierr; 1265aa0d41d4SStefano Zampini 1266aa0d41d4SStefano Zampini PetscFunctionBegin; 1267aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12685e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12695e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1270906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1271906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1272d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1273906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1274906d46d4SStefano Zampini 1275906d46d4SStefano Zampini /* check */ 1276906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1277906d46d4SStefano Zampini Vec x,x_change; 1278906d46d4SStefano Zampini PetscReal error; 1279906d46d4SStefano Zampini 12805e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1281906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12825e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12835e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12845e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1285d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12865e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12875e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1288906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1289906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1290906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1291906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1292906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1293906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1294906d46d4SStefano Zampini } 1295906d46d4SStefano Zampini 129622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12979b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 129822d5777bSStefano Zampini if (isseqaij) { 1299d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1300aa0d41d4SStefano Zampini } else { 1301aa0d41d4SStefano Zampini Mat work_mat; 1302aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1303d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1304aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1305aa0d41d4SStefano Zampini } 13063301b35fSStefano Zampini if (matis->A->symmetric_set) { 13073301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1308e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 13093301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1310e496cd5dSStefano Zampini #endif 13113301b35fSStefano Zampini } 131245a1bb75SStefano Zampini /* 131345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1314d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 131545a1bb75SStefano Zampini */ 1316d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1317aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1318aa0d41d4SStefano Zampini } 1319aa0d41d4SStefano Zampini 1320aa0d41d4SStefano Zampini #undef __FUNCT__ 1321a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13228ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1323a64d13efSStefano Zampini { 1324a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1325a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1326d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 132753892102SStefano Zampini PetscInt *idx_R_local=NULL; 13283a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13293a50541eSStefano Zampini PetscInt vbs,bs; 13306816873aSStefano Zampini PetscBT bitmask=NULL; 1331a64d13efSStefano Zampini PetscErrorCode ierr; 1332a64d13efSStefano Zampini 1333a64d13efSStefano Zampini PetscFunctionBegin; 1334b23d619eSStefano Zampini /* 1335b23d619eSStefano Zampini No need to setup local scatters if 1336b23d619eSStefano Zampini - primal space is unchanged 1337b23d619eSStefano Zampini AND 1338b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1339b23d619eSStefano Zampini AND 1340b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1341b23d619eSStefano Zampini */ 1342b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1343f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1344f4ddd8eeSStefano Zampini } 1345f4ddd8eeSStefano Zampini /* destroy old objects */ 1346f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1347f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1348f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1349a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1350b371cd4fSStefano Zampini n_B = pcis->n_B; 1351b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1352b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13536816873aSStefano Zampini 13546816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13556816873aSStefano Zampini 135653892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13576816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 135853892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13594641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13604641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13610e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13624641a718SStefano Zampini } 13633a50541eSStefano Zampini 1364a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13654641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13666816873aSStefano Zampini idx_R_local[n_R++] = i; 1367a64d13efSStefano Zampini } 1368a64d13efSStefano Zampini } 136953892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13706816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13716816873aSStefano Zampini 137253892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 137353892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13746816873aSStefano Zampini } 13753a50541eSStefano Zampini 13763a50541eSStefano Zampini /* Block code */ 13773a50541eSStefano Zampini vbs = 1; 13783a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13793a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13803a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13813a50541eSStefano Zampini PetscInt *vary; 13823a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1383785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13843a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13850e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1386964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13873a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13883a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13893a50541eSStefano Zampini break; 13903a50541eSStefano Zampini } 13913a50541eSStefano Zampini } 13923a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13933a50541eSStefano Zampini vbs = bs; 13943a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13953a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13963a50541eSStefano Zampini } 13973a50541eSStefano Zampini } 13983a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13993a50541eSStefano Zampini } 14003a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 140153892102SStefano Zampini if (sub_schurs->reuse_mumps) { 140253892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 140353892102SStefano Zampini 140453892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140553892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140753892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140853892102SStefano Zampini } else { 14093a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 141053892102SStefano Zampini } 1411a64d13efSStefano Zampini 1412a64d13efSStefano Zampini /* print some info if requested */ 1413a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1414a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1415a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14160fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1417a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1418a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14193a50541eSStefano 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); 1420a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1421a64d13efSStefano Zampini } 1422a64d13efSStefano Zampini 1423a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14246816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14256816873aSStefano Zampini IS is_aux1,is_aux2; 14266816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14276816873aSStefano Zampini 14283a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1429854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1430854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1431a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14324641a718SStefano Zampini for (i=0; i<n_D; i++) { 14334641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14344641a718SStefano Zampini } 1435a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1436a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14374641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14384641a718SStefano Zampini aux_array1[j++] = i; 1439a64d13efSStefano Zampini } 1440a64d13efSStefano Zampini } 1441a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1442a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1443a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14444641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14454641a718SStefano Zampini aux_array2[j++] = i; 1446a64d13efSStefano Zampini } 1447a64d13efSStefano Zampini } 1448a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1449a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1450a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1451a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1452a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1453a64d13efSStefano Zampini 14548eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1455785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1456a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14574641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14584641a718SStefano Zampini aux_array1[j++] = i; 1459a64d13efSStefano Zampini } 1460a64d13efSStefano Zampini } 1461a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1462a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1463a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1464a64d13efSStefano Zampini } 14654641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14663a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1467d62866d3SStefano Zampini } else { 146853892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14696816873aSStefano Zampini IS tis; 14706816873aSStefano Zampini PetscInt schur_size; 14716816873aSStefano Zampini 147253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14736816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 147453892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14756816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14766816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14776816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14786816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14796816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1480d62866d3SStefano Zampini } 1481d62866d3SStefano Zampini } 1482a64d13efSStefano Zampini PetscFunctionReturn(0); 1483a64d13efSStefano Zampini } 1484a64d13efSStefano Zampini 1485304d26faSStefano Zampini 1486304d26faSStefano Zampini #undef __FUNCT__ 1487304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1488684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1489304d26faSStefano Zampini { 1490304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1491304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1492304d26faSStefano Zampini PC pc_temp; 1493304d26faSStefano Zampini Mat A_RR; 1494f4ddd8eeSStefano Zampini MatReuse reuse; 1495304d26faSStefano Zampini PetscScalar m_one = -1.0; 1496304d26faSStefano Zampini PetscReal value; 1497af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14989577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1499304d26faSStefano Zampini PetscErrorCode ierr; 1500e604994aSStefano Zampini /* prefixes stuff */ 1501312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1502e604994aSStefano Zampini size_t len; 1503304d26faSStefano Zampini 1504304d26faSStefano Zampini PetscFunctionBegin; 1505304d26faSStefano Zampini 1506e604994aSStefano Zampini /* compute prefixes */ 1507e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1508e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1509e604994aSStefano Zampini if (!pcbddc->current_level) { 1510e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1511e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1512e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1513e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1514e604994aSStefano Zampini } else { 1515e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1516312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1517e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1518e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1519312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1520312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 152134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 152234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1523e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1524e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1525e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1526e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1527e604994aSStefano Zampini } 1528e604994aSStefano Zampini 1529304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1530684f6988SStefano Zampini if (dirichlet) { 1531d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15323301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15333301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1534964fefecSStefano Zampini } 1535ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1536964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1537304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1538304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1539304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1540304d26faSStefano Zampini /* default */ 1541304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1542e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15439577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1544304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15459577ea80SStefano Zampini if (issbaij) { 15469577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15479577ea80SStefano Zampini } else { 1548304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15499577ea80SStefano Zampini } 1550304d26faSStefano Zampini /* Allow user's customization */ 1551304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1552304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1553304d26faSStefano Zampini } 1554d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1555d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1556d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1557d62866d3SStefano Zampini 1558d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1559d5574798SStefano Zampini } 1560304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1561304d26faSStefano Zampini if (!n_D) { 1562304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1563304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1564304d26faSStefano Zampini } 1565304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1566304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1567304d26faSStefano Zampini /* set ksp_D into pcis data */ 1568304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1569304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1570304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1571684f6988SStefano Zampini } 1572304d26faSStefano Zampini 1573304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1574684f6988SStefano Zampini A_RR = 0; 1575684f6988SStefano Zampini if (neumann) { 1576d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1577f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15788ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1579f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1580f4ddd8eeSStefano Zampini PetscInt nn_R; 158181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1582f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1583f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1585f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1586f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1587f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1588f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1589727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1590f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1591f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1592f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1593f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1594f4ddd8eeSStefano Zampini } 1595f4ddd8eeSStefano Zampini } 1596f4ddd8eeSStefano Zampini /* last check */ 1597d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1598f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1599f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1600f4ddd8eeSStefano Zampini } 1601f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1602f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1603f4ddd8eeSStefano Zampini } 16046816873aSStefano Zampini /* extract A_RR */ 16056816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16066816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16076816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 16086816873aSStefano Zampini if (ibs != mbs) { 16096816873aSStefano Zampini Mat newmat; 16106816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 16116816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16126816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 16136816873aSStefano Zampini } else { 16146816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16156816873aSStefano Zampini } 16163301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16173301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16186816873aSStefano Zampini } 16196816873aSStefano Zampini } else { 16206816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16216816873aSStefano Zampini 16226816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16236816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16246816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16256816873aSStefano Zampini } 1626f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1627304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1628304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1629304d26faSStefano Zampini /* default */ 1630304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1631e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1632304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16339577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16349577ea80SStefano Zampini if (issbaij) { 16359577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16369577ea80SStefano Zampini } else { 1637304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16389577ea80SStefano Zampini } 1639304d26faSStefano Zampini /* Allow user's customization */ 1640304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1641304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1642304d26faSStefano Zampini } 1643d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1644304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1645304d26faSStefano Zampini if (!n_R) { 1646304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1647304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1648304d26faSStefano Zampini } 1649d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1650d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1651d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1652d62866d3SStefano Zampini 1653d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1654d62866d3SStefano Zampini } 1655304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1656304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1657684f6988SStefano Zampini } 16586816873aSStefano Zampini /* free Neumann problem's matrix */ 16596816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1660304d26faSStefano Zampini 1661304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16620fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1663684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1664684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1665684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1666684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1667684f6988SStefano Zampini } 1668684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16690fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16700fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16710fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16720fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16730fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1674304d26faSStefano Zampini /* need to be adapted? */ 1675b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1676b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1677b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1678304d26faSStefano Zampini /* print info */ 1679304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1680e604994aSStefano 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); 1681304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1682304d26faSStefano Zampini } 1683b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16848ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1685304d26faSStefano Zampini } 1686684f6988SStefano Zampini } 1687684f6988SStefano Zampini if (neumann) { /* Neumann */ 16886816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16890fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16900fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16910fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16920fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16930fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1694304d26faSStefano Zampini /* need to be adapted? */ 1695b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1696b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1697304d26faSStefano Zampini /* print info */ 1698304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1699e604994aSStefano 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); 1700304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1701304d26faSStefano Zampini } 1702b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 17038ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1704304d26faSStefano Zampini } 17050fccc4e9SStefano Zampini } 1706684f6988SStefano Zampini } 1707304d26faSStefano Zampini PetscFunctionReturn(0); 1708304d26faSStefano Zampini } 1709304d26faSStefano Zampini 1710304d26faSStefano Zampini #undef __FUNCT__ 1711ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 171280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1713674ae819SStefano Zampini { 1714674ae819SStefano Zampini PetscErrorCode ierr; 1715674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1716be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1717674ae819SStefano Zampini 1718674ae819SStefano Zampini PetscFunctionBegin; 1719be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1721be83ff47SStefano Zampini } 172280677318SStefano Zampini if (!pcbddc->switch_static) { 172380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 172480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 172580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172620c7b377SStefano Zampini } 1727be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173020c7b377SStefano Zampini } else { 1731be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1732be83ff47SStefano Zampini 173353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 173453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1735be83ff47SStefano Zampini } 1736be83ff47SStefano Zampini } else { 173780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 174280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 174380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 174480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174620c7b377SStefano Zampini } 1747674ae819SStefano Zampini } 1748be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174980677318SStefano Zampini if (applytranspose) { 175080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175180677318SStefano Zampini } else { 175280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175380677318SStefano Zampini } 175453892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1755be83ff47SStefano Zampini } else { 1756be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1757be83ff47SStefano Zampini 1758be83ff47SStefano Zampini if (applytranspose) { 175953892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1760be83ff47SStefano Zampini } else { 176153892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1762be83ff47SStefano Zampini } 176353892102SStefano Zampini #endif 1764be83ff47SStefano Zampini } 176580677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 176680677318SStefano Zampini if (!pcbddc->switch_static) { 1767be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1770be83ff47SStefano Zampini } else { 1771be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1772be83ff47SStefano Zampini 177353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 177453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1775be83ff47SStefano Zampini } 177680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 177980677318SStefano Zampini } 178080677318SStefano Zampini } else { 178180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 178880677318SStefano Zampini } 178980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179380677318SStefano Zampini } 1794674ae819SStefano Zampini PetscFunctionReturn(0); 1795674ae819SStefano Zampini } 1796674ae819SStefano Zampini 1797dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1798674ae819SStefano Zampini #undef __FUNCT__ 1799674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1800dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1801674ae819SStefano Zampini { 1802674ae819SStefano Zampini PetscErrorCode ierr; 1803674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1804674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1805674ae819SStefano Zampini const PetscScalar zero = 0.0; 1806674ae819SStefano Zampini 1807674ae819SStefano Zampini PetscFunctionBegin; 1808dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1809dc359a40SStefano Zampini if (applytranspose) { 1810674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18118eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1812dc359a40SStefano Zampini } else { 1813674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1814674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 181515aaf578SStefano Zampini } 181612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 181712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 181812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182012edc857SStefano Zampini 18219f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 182212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 182312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1824964fefecSStefano Zampini Vec rhs,sol; 1825964fefecSStefano Zampini 1826964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1827964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 182812edc857SStefano Zampini if (applytranspose) { 1829964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183012edc857SStefano Zampini } else { 1831964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183212edc857SStefano Zampini } 183312edc857SStefano Zampini } 1834674ae819SStefano Zampini 1835674ae819SStefano Zampini /* Local solution on R nodes */ 183680677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 183780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18389f00e9b4SStefano Zampini } 1839674ae819SStefano Zampini 18409f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18419f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 184212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1843674ae819SStefano Zampini 1844674ae819SStefano Zampini /* Sum contributions from two levels */ 1845dc359a40SStefano Zampini if (applytranspose) { 1846dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1847dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1848dc359a40SStefano Zampini } else { 1849674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18508eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1851dc359a40SStefano Zampini } 1852674ae819SStefano Zampini PetscFunctionReturn(0); 1853674ae819SStefano Zampini } 1854674ae819SStefano Zampini 1855674ae819SStefano Zampini #undef __FUNCT__ 1856674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 185712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1858674ae819SStefano Zampini { 1859674ae819SStefano Zampini PetscErrorCode ierr; 1860674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 186158da7f69SStefano Zampini PetscScalar *array; 186212edc857SStefano Zampini Vec from,to; 1863674ae819SStefano Zampini 1864674ae819SStefano Zampini PetscFunctionBegin; 186512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 186612edc857SStefano Zampini from = pcbddc->coarse_vec; 186712edc857SStefano Zampini to = pcbddc->vec1_P; 186812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 186912edc857SStefano Zampini Vec tvec; 187058da7f69SStefano Zampini 187158da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 187258da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 187312edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 187458da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 187558da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 187658da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 187712edc857SStefano Zampini } 187812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187912edc857SStefano Zampini from = pcbddc->vec1_P; 188012edc857SStefano Zampini to = pcbddc->coarse_vec; 188112edc857SStefano Zampini } 188212edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1883674ae819SStefano Zampini PetscFunctionReturn(0); 1884674ae819SStefano Zampini } 1885674ae819SStefano Zampini 1886674ae819SStefano Zampini #undef __FUNCT__ 1887674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 188812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1889674ae819SStefano Zampini { 1890674ae819SStefano Zampini PetscErrorCode ierr; 1891674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 189258da7f69SStefano Zampini PetscScalar *array; 189312edc857SStefano Zampini Vec from,to; 1894674ae819SStefano Zampini 1895674ae819SStefano Zampini PetscFunctionBegin; 189612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 189712edc857SStefano Zampini from = pcbddc->coarse_vec; 189812edc857SStefano Zampini to = pcbddc->vec1_P; 189912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 190012edc857SStefano Zampini from = pcbddc->vec1_P; 190112edc857SStefano Zampini to = pcbddc->coarse_vec; 190212edc857SStefano Zampini } 190312edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 190412edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 190512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 190612edc857SStefano Zampini Vec tvec; 190758da7f69SStefano Zampini 190812edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 190958da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 191058da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 191158da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 191258da7f69SStefano Zampini } 191358da7f69SStefano Zampini } else { 191458da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 191558da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 191612edc857SStefano Zampini } 191712edc857SStefano Zampini } 1918674ae819SStefano Zampini PetscFunctionReturn(0); 1919674ae819SStefano Zampini } 1920674ae819SStefano Zampini 1921984c4197SStefano Zampini /* uncomment for testing purposes */ 1922984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1923674ae819SStefano Zampini #undef __FUNCT__ 1924674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1925674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1926674ae819SStefano Zampini { 1927674ae819SStefano Zampini PetscErrorCode ierr; 1928674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1929674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1930674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1931984c4197SStefano Zampini /* one and zero */ 1932984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1933984c4197SStefano Zampini /* space to store constraints and their local indices */ 19349162d606SStefano Zampini PetscScalar *constraints_data; 19359162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19369162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19379162d606SStefano Zampini PetscInt *constraints_n; 1938984c4197SStefano Zampini /* iterators */ 1939b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1940984c4197SStefano Zampini /* BLAS integers */ 1941e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1942e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1943c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1944727cdba6SStefano Zampini /* reuse */ 19450e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19460e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1947984c4197SStefano Zampini /* change of basis */ 1948b3d85658SStefano Zampini PetscBool qr_needed; 19499162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1950984c4197SStefano Zampini /* auxiliary stuff */ 195164efe560SStefano Zampini PetscInt *nnz,*is_indices; 19528a0068c3SStefano Zampini PetscInt ncc; 1953984c4197SStefano Zampini /* some quantities */ 195445a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1955a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1956984c4197SStefano Zampini 1957674ae819SStefano Zampini PetscFunctionBegin; 19588e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19598e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19608e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1961088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1962088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19630e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19640e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19650e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19660e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19670e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1968088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1969cf5a6209SStefano Zampini 1970cf5a6209SStefano Zampini /* print some info */ 1971cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1972cf5a6209SStefano Zampini IS vertices; 1973cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1974cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1975cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1976cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1977cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1978cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1979cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1980fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1981fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1982cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1983cf5a6209SStefano Zampini } 1984cf5a6209SStefano Zampini 1985cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19869162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1987cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1988cf5a6209SStefano Zampini const Vec *nearnullvecs; 1989cf5a6209SStefano Zampini Vec *localnearnullsp; 1990cf5a6209SStefano Zampini PetscScalar *array; 1991cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1992cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1993cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1994b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1995cf5a6209SStefano Zampini PetscScalar *work; 1996cf5a6209SStefano Zampini PetscReal *singular_vals; 1997cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1998cf5a6209SStefano Zampini PetscReal *rwork; 1999cf5a6209SStefano Zampini #endif 2000cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2001cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2002cf5a6209SStefano Zampini #else 2003964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2004964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2005cf5a6209SStefano Zampini #endif 2006cf5a6209SStefano Zampini 2007674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2008d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2009d06fc5fdSStefano Zampini /* free unneeded index sets */ 2010d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2011d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2012674ae819SStefano Zampini } 2013d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2014d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2015d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2016d06fc5fdSStefano Zampini } 2017d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2018d06fc5fdSStefano Zampini n_ISForEdges = 0; 2019d06fc5fdSStefano Zampini } 2020d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2021d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2022d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2023d06fc5fdSStefano Zampini } 2024d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2025d06fc5fdSStefano Zampini n_ISForFaces = 0; 2026d06fc5fdSStefano Zampini } 202770022509SStefano Zampini 202870022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 202970022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 203070022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 203170022509SStefano Zampini if (pcbddc->NullSpace) { 203270022509SStefano Zampini PetscBool tbool[2],gbool[2]; 203370022509SStefano Zampini 203470022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2035b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2036d06fc5fdSStefano Zampini if (!ISForEdges) { 2037d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2038d06fc5fdSStefano Zampini } 2039b8ffe317SStefano Zampini } 2040d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2041d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2042d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2043d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2044d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 204598a51de6SStefano Zampini } 204670022509SStefano Zampini #endif 204708122e43SStefano Zampini 2048674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2049674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2050674ae819SStefano Zampini if (nearnullsp) { 2051674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2052f4ddd8eeSStefano Zampini /* remove any stored info */ 2053f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2054f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2055f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2056f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2057f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2058473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2059f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2060f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2061f4ddd8eeSStefano Zampini } 2062984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2063984c4197SStefano Zampini nnsp_size = 0; 2064674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2065674ae819SStefano Zampini } 2066984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2067984c4197SStefano Zampini max_constraints = nnsp_size; 2068984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2069984c4197SStefano Zampini 2070674ae819SStefano Zampini /* 2071674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20729162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20739162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20749162d606SStefano Zampini There can be multiple constraints per connected component 2075674ae819SStefano Zampini */ 2076674ae819SStefano Zampini n_vertices = 0; 2077674ae819SStefano Zampini if (ISForVertices) { 2078674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2079674ae819SStefano Zampini } 20809162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20819162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20829162d606SStefano Zampini 20839162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20849162d606SStefano Zampini total_counts *= max_constraints; 2085674ae819SStefano Zampini total_counts += n_vertices; 20864641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20879162d606SStefano Zampini 2088674ae819SStefano Zampini total_counts = 0; 2089674ae819SStefano Zampini max_size_of_constraint = 0; 2090674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20919162d606SStefano Zampini IS used_is; 2092674ae819SStefano Zampini if (i<n_ISForEdges) { 20939162d606SStefano Zampini used_is = ISForEdges[i]; 2094674ae819SStefano Zampini } else { 20959162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2096674ae819SStefano Zampini } 20979162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2098674ae819SStefano Zampini total_counts += j; 2099674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2100674ae819SStefano Zampini } 21019162d606SStefano 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); 21029162d606SStefano Zampini 2103984c4197SStefano Zampini /* get local part of global near null space vectors */ 2104785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2105984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2106984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2107984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2108984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2109984c4197SStefano Zampini } 2110674ae819SStefano Zampini 2111242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2112242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2113a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2114242a89d7SStefano Zampini 2115984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2116a773dcb8SStefano Zampini if (!skip_lapack) { 2117674ae819SStefano Zampini PetscScalar temp_work; 2118911cabfeSStefano Zampini 2119674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2120984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2121785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2122785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2123785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2124674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2125785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2126674ae819SStefano Zampini #endif 2127674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2128c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2129c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2130674ae819SStefano Zampini lwork = -1; 2131674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2132674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2133c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2134674ae819SStefano Zampini #else 2135c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2136674ae819SStefano Zampini #endif 2137674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2138984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2139674ae819SStefano Zampini #else /* on missing GESVD */ 2140674ae819SStefano Zampini /* SVD */ 2141674ae819SStefano Zampini PetscInt max_n,min_n; 2142674ae819SStefano Zampini max_n = max_size_of_constraint; 2143984c4197SStefano Zampini min_n = max_constraints; 2144984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2145674ae819SStefano Zampini min_n = max_size_of_constraint; 2146984c4197SStefano Zampini max_n = max_constraints; 2147674ae819SStefano Zampini } 2148785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2149674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2150785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2151674ae819SStefano Zampini #endif 2152674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2153674ae819SStefano Zampini lwork = -1; 2154e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2155e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2156b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2157674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2158674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21599162d606SStefano 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)); 2160674ae819SStefano Zampini #else 21619162d606SStefano 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)); 2162674ae819SStefano Zampini #endif 2163674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2164984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2165984c4197SStefano Zampini #endif /* on missing GESVD */ 2166674ae819SStefano Zampini /* Allocate optimal workspace */ 2167674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2168854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2169674ae819SStefano Zampini } 2170674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2171674ae819SStefano Zampini total_counts = 0; 21729162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21739162d606SStefano Zampini constraints_data_ptr[0] = 0; 2174674ae819SStefano Zampini /* vertices */ 21759162d606SStefano Zampini if (n_vertices) { 2176674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21779162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21789162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2179674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21809162d606SStefano Zampini constraints_n[total_counts] = 1; 21819162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21829162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21839162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2184674ae819SStefano Zampini total_counts++; 2185674ae819SStefano Zampini } 2186674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2187984c4197SStefano Zampini PetscBool used_vertex; 2188674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2189674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2190674ae819SStefano Zampini k = 0; 2191674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2192984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2193984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21949162d606SStefano Zampini constraints_n[total_counts] = 1; 21959162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21969162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21979162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21989162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2199674ae819SStefano Zampini total_counts++; 2200674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2201674ae819SStefano Zampini } 2202984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2203674ae819SStefano Zampini k++; 2204674ae819SStefano Zampini } 2205674ae819SStefano Zampini } 2206674ae819SStefano Zampini } 2207674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2208674ae819SStefano Zampini n_vertices = total_counts; 2209674ae819SStefano Zampini } 2210984c4197SStefano Zampini 2211674ae819SStefano Zampini /* edges and faces */ 22129162d606SStefano Zampini total_counts_cc = total_counts; 2213911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22149162d606SStefano Zampini IS used_is; 22159162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22169162d606SStefano Zampini 2217911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22189162d606SStefano Zampini used_is = ISForEdges[ncc]; 2219984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2220674ae819SStefano Zampini } else { 22219162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2222984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2223674ae819SStefano Zampini } 2224674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22259162d606SStefano Zampini 22269162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22279162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2228984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2229984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2230674ae819SStefano Zampini if (nnsp_has_cnst) { 22315b08dc53SStefano Zampini PetscScalar quad_value; 22329162d606SStefano Zampini 22339162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22349162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22359162d606SStefano Zampini 2236a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2237674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2238a773dcb8SStefano Zampini } else { 2239a773dcb8SStefano Zampini quad_value = 1.0; 2240a773dcb8SStefano Zampini } 2241674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22429162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2243674ae819SStefano Zampini } 22449162d606SStefano Zampini temp_constraints++; 2245674ae819SStefano Zampini total_counts++; 2246674ae819SStefano Zampini } 2247674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2248984c4197SStefano Zampini PetscReal real_value; 22499162d606SStefano Zampini PetscScalar *ptr_to_data; 22509162d606SStefano Zampini 2251984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22529162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2253674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22549162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2255674ae819SStefano Zampini } 2256984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2257984c4197SStefano Zampini /* check if array is null on the connected component */ 2258e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22599162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22605b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2261674ae819SStefano Zampini temp_constraints++; 2262674ae819SStefano Zampini total_counts++; 22639162d606SStefano Zampini if (!idxs_copied) { 22649162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22659162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2266674ae819SStefano Zampini } 2267674ae819SStefano Zampini } 22689162d606SStefano Zampini } 22699162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 227045a1bb75SStefano Zampini valid_constraints = temp_constraints; 2271eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2272a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22739162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22749162d606SStefano Zampini 22759162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2276a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22779162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2278a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22799162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2280a773dcb8SStefano Zampini } else { /* perform SVD */ 2281984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22829162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2283674ae819SStefano Zampini 2284674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2285984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2286984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2287984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2288984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2289984c4197SStefano Zampini from that computed using LAPACKgesvd 2290984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2291984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2292984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2293674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2294e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2295984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2296674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2297674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22989162d606SStefano 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)); 2299674ae819SStefano Zampini } 2300674ae819SStefano Zampini } 2301e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2302e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2303e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2304674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2305c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2306674ae819SStefano Zampini #else 2307c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2308674ae819SStefano Zampini #endif 2309674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2310984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2311984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2312674ae819SStefano Zampini j = 0; 2313984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2314674ae819SStefano Zampini total_counts = total_counts-j; 231545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2316e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2317c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2318c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2319c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2320c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2321c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2322c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2323674ae819SStefano Zampini if (j<temp_constraints) { 2324984c4197SStefano Zampini PetscInt ii; 2325984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2326674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23279162d606SStefano 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)); 2328674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2329984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2330674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23319162d606SStefano 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]; 2332674ae819SStefano Zampini } 2333674ae819SStefano Zampini } 2334674ae819SStefano Zampini } 2335674ae819SStefano Zampini #else /* on missing GESVD */ 2336e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2337e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2338b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2339674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2340674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23419162d606SStefano 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)); 2342674ae819SStefano Zampini #else 23439162d606SStefano 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)); 2344674ae819SStefano Zampini #endif 2345984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2346674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2347984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2348e310c8b4SStefano Zampini k = temp_constraints; 2349e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2350674ae819SStefano Zampini j = 0; 2351e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 235245a1bb75SStefano Zampini valid_constraints = k-j; 2353911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2354984c4197SStefano Zampini #endif /* on missing GESVD */ 2355674ae819SStefano Zampini } 2356a773dcb8SStefano Zampini } 23579162d606SStefano Zampini /* update pointers information */ 23589162d606SStefano Zampini if (valid_constraints) { 23599162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23609162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23619162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23629162d606SStefano Zampini /* set change_of_basis flag */ 236345a1bb75SStefano Zampini if (boolforchange) { 2364b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23659162d606SStefano Zampini } 2366b3d85658SStefano Zampini total_counts_cc++; 236745a1bb75SStefano Zampini } 2368674ae819SStefano Zampini } 2369984c4197SStefano Zampini /* free workspace */ 23708f1c130eSStefano Zampini if (!skip_lapack) { 2371984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2372984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2373984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2374984c4197SStefano Zampini #endif 2375984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2376984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2377984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2378984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2379984c4197SStefano Zampini #endif 2380984c4197SStefano Zampini } 2381984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2382984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2383984c4197SStefano Zampini } 2384984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2385cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2386cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2387cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2388cf5a6209SStefano Zampini } 2389cf5a6209SStefano Zampini if (n_ISForFaces) { 2390cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2391cf5a6209SStefano Zampini } 2392cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2393cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2394cf5a6209SStefano Zampini } 2395cf5a6209SStefano Zampini if (n_ISForEdges) { 2396cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2397cf5a6209SStefano Zampini } 2398cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 239908122e43SStefano Zampini } else { 240008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 240108122e43SStefano Zampini 240208122e43SStefano Zampini total_counts = 0; 240308122e43SStefano Zampini n_vertices = 0; 2404d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2405d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 240608122e43SStefano Zampini } 240708122e43SStefano Zampini max_constraints = 0; 24089162d606SStefano Zampini total_counts_cc = 0; 240908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 241008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24119162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 241208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 241308122e43SStefano Zampini } 24149162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24159162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24169162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24179162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 241874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24199162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24209162d606SStefano Zampini total_counts_cc = 0; 24219162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24229162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24239162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 242408122e43SStefano Zampini } 242508122e43SStefano Zampini } 24269162d606SStefano Zampini #if 0 24279162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24289162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24299162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24309162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24319162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24329162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24339162d606SStefano Zampini } 24349162d606SStefano Zampini printf("\n"); 24359162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24369162d606SStefano Zampini } 24371b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24388bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24391b968477SStefano Zampini } 24401b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24418bec7fa6SStefano 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]); 24421b968477SStefano Zampini } 244308122e43SStefano Zampini #endif 244408122e43SStefano Zampini 24458bec7fa6SStefano Zampini max_size_of_constraint = 0; 24469162d606SStefano 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]); 24479162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 244808122e43SStefano Zampini /* Change of basis */ 2449b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 245008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 245108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 245208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2453b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 245408122e43SStefano Zampini } 245508122e43SStefano Zampini } 245608122e43SStefano Zampini } 245708122e43SStefano Zampini } 2458b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24590e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 246008122e43SStefano Zampini 24619162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24629162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24639162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24649162d606SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",constraints_idxs_ptr[total_counts_cc],i); 246508122e43SStefano Zampini } 2466984c4197SStefano Zampini 2467674ae819SStefano Zampini /* Create constraint matrix */ 2468674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 246916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2470984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2471984c4197SStefano Zampini 2472984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 247374d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 247474d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 247574d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2476984c4197SStefano Zampini total_primal_vertices=0; 2477b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24789162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24799162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2480984c4197SStefano Zampini if (size_of_constraint == 1) { 24819162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2482b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 248364efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24849162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24859162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24869162d606SStefano Zampini } 2487b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 248874d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 248974d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 249074d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 249174d5cdf7SStefano Zampini } 2492b3d85658SStefano Zampini } else { 2493b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2494984c4197SStefano Zampini } 2495984c4197SStefano Zampini } 2496b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2497b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 249870022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 249970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2500b3d85658SStefano Zampini 2501b3d85658SStefano 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); 25020e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25030e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 250470022509SStefano Zampini 2505984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 250674d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2507785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2508984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 250974d5cdf7SStefano Zampini 2510984c4197SStefano Zampini j = total_primal_vertices; 251174d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2512b3d85658SStefano Zampini cum = total_primal_vertices; 25139162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25144641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2515b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2516b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2517b3d85658SStefano Zampini cum++; 25189162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 251974d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 252074d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 252174d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 252274d5cdf7SStefano Zampini } 25239162d606SStefano Zampini j += constraints_n[i]; 2524674ae819SStefano Zampini } 2525674ae819SStefano Zampini } 2526674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2527674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2528088faed8SStefano Zampini 2529674ae819SStefano Zampini /* set values in constraint matrix */ 2530984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25310e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2532674ae819SStefano Zampini } 2533984c4197SStefano Zampini total_counts = total_primal_vertices; 25349162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25354641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25369162d606SStefano Zampini PetscInt *cols; 25379162d606SStefano Zampini 25389162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25399162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25409162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25419162d606SStefano Zampini PetscInt row = total_counts+k; 25429162d606SStefano Zampini PetscScalar *vals; 25439162d606SStefano Zampini 25449162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25459162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25469162d606SStefano Zampini } 25479162d606SStefano Zampini total_counts += constraints_n[i]; 2548674ae819SStefano Zampini } 2549674ae819SStefano Zampini } 2550674ae819SStefano Zampini /* assembling */ 2551674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2552674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2553088faed8SStefano Zampini 2554984c4197SStefano Zampini /* 255545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2556984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2557984c4197SStefano Zampini */ 2558674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2559674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2560026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2561984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2562984c4197SStefano Zampini /* working stuff for GEQRF */ 256381d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2564984c4197SStefano Zampini PetscBLASInt lqr_work; 2565984c4197SStefano Zampini /* working stuff for UNGQR */ 2566984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2567984c4197SStefano Zampini PetscBLASInt lgqr_work; 2568984c4197SStefano Zampini /* working stuff for TRTRS */ 2569984c4197SStefano Zampini PetscScalar *trs_rhs; 25703f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2571984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2572984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2573984c4197SStefano Zampini PetscScalar *start_vals; 2574984c4197SStefano Zampini /* working stuff for values insertion */ 25754641a718SStefano Zampini PetscBT is_primal; 257664efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2577906d46d4SStefano Zampini /* matrix sizes */ 2578906d46d4SStefano Zampini PetscInt global_size,local_size; 2579906d46d4SStefano Zampini /* temporary change of basis */ 2580906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2581cf5a6209SStefano Zampini /* extra space for debugging */ 2582cf5a6209SStefano Zampini PetscScalar *dbg_work; 2583984c4197SStefano Zampini 2584906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2585906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 258616f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2587bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2588906d46d4SStefano Zampini /* nonzeros for local mat */ 2589bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2590bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25919162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2592a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25939162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2594a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25959162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2596a717540cSStefano Zampini } else { 25979162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25989162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2599a717540cSStefano Zampini } 2600a717540cSStefano Zampini } 2601a717540cSStefano Zampini } 2602906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2603bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2604a717540cSStefano Zampini /* Set initial identity in the matrix */ 2605bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2606906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2607a717540cSStefano Zampini } 2608a717540cSStefano Zampini 2609a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2610a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2611a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2612a717540cSStefano Zampini } 2613a717540cSStefano Zampini 2614a717540cSStefano Zampini 2615a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2616a717540cSStefano Zampini /* 2617a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2618a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2619a717540cSStefano Zampini 2620a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2621a717540cSStefano Zampini 2622a6b551f4SStefano 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) 2623a6b551f4SStefano Zampini 2624a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2625a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2626a717540cSStefano Zampini | ... | 2627a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2628a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2629a717540cSStefano Zampini 2630a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2631a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2632a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2633a6b551f4SStefano Zampini 2634a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2635a717540cSStefano Zampini */ 2636a717540cSStefano Zampini if (qr_needed) { 2637984c4197SStefano Zampini /* space to store Q */ 2638854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2639984c4197SStefano Zampini /* first we issue queries for optimal work */ 26403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2643984c4197SStefano Zampini lqr_work = -1; 26443f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2645984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2646984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2647785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2648984c4197SStefano Zampini lgqr_work = -1; 26493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26523f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26533f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26543f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2655984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2656984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2657785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2658984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2659785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2660984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2661785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2662a717540cSStefano Zampini /* allocating workspace for check */ 2663a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2664cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2665a717540cSStefano Zampini } 2666a717540cSStefano Zampini } 2667984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26684641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2669473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26700e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 267139e2fb2aSStefano Zampini if (i != total_primal_vertices) { 267239e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26734641a718SStefano Zampini } 267439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 267539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 267639e2fb2aSStefano Zampini } 267739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2678984c4197SStefano Zampini 2679a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26809162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26819162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26824641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2683984c4197SStefano Zampini /* get constraint info */ 26849162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2685984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2686984c4197SStefano Zampini 2687984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26889162d606SStefano 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); 2689674ae819SStefano Zampini } 2690984c4197SStefano Zampini 2691fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2692a717540cSStefano Zampini 2693a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2694a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26959162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2696a717540cSStefano Zampini } 2697984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26989162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2699984c4197SStefano Zampini 2700984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2704674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2706984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2707674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2708984c4197SStefano Zampini 2709984c4197SStefano Zampini /* explictly compute R^-T */ 2710984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2711984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2716984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27173f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2718984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2719984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2720984c4197SStefano Zampini 2721a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27233f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27243f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2726984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2728984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2729984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2730984c4197SStefano Zampini 2731984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2732984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2733984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27353f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27383f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27393f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2740984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27419162d606SStefano 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)); 2742984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27439162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2744984c4197SStefano Zampini 2745984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27469162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2747984c4197SStefano Zampini /* insert cols for primal dofs */ 2748984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2749984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27509162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2751906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2752984c4197SStefano Zampini } 2753984c4197SStefano Zampini /* insert cols for dual dofs */ 2754984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27559162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2756984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27579162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2758906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2759984c4197SStefano Zampini j++; 2760674ae819SStefano Zampini } 2761674ae819SStefano Zampini } 2762984c4197SStefano Zampini 2763984c4197SStefano Zampini /* check change of basis */ 2764984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2765984c4197SStefano Zampini PetscInt ii,jj; 2766984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2767c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2768c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2769c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2770c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2771c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2772c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2773984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2774cf5a6209SStefano 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)); 2775984c4197SStefano Zampini ierr = PetscFPTrapPop();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) valid_qr = PETSC_FALSE; 2779cf5a6209SStefano 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; 2780674ae819SStefano Zampini } 2781674ae819SStefano Zampini } 2782984c4197SStefano Zampini if (!valid_qr) { 278322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2784984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2785984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2786cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2787cf5a6209SStefano 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])); 2788674ae819SStefano Zampini } 2789cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2790cf5a6209SStefano 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])); 2791984c4197SStefano Zampini } 2792984c4197SStefano Zampini } 2793984c4197SStefano Zampini } 2794674ae819SStefano Zampini } else { 279522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2796674ae819SStefano Zampini } 2797674ae819SStefano Zampini } 2798a717540cSStefano Zampini } else { /* simple transformation block */ 2799a717540cSStefano Zampini PetscInt row,col; 2800a6b551f4SStefano Zampini PetscScalar val,norm; 2801a6b551f4SStefano Zampini 2802a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28039162d606SStefano 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)); 2804a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28059162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28069162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2807bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28089162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2809906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28109162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2811a717540cSStefano Zampini } else { 2812a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28139162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2814a717540cSStefano Zampini if (row != col) { 28159162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2816a717540cSStefano Zampini } else { 28179162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2818a717540cSStefano Zampini } 2819906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2820a717540cSStefano Zampini } 2821a717540cSStefano Zampini } 2822a717540cSStefano Zampini } 282398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 282422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2825a717540cSStefano Zampini } 2826674ae819SStefano Zampini } 2827984c4197SStefano Zampini } else { 2828984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28299162d606SStefano 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); 2830674ae819SStefano Zampini } 2831674ae819SStefano Zampini } 2832674ae819SStefano Zampini } 2833a717540cSStefano Zampini 2834a717540cSStefano Zampini /* free workspace */ 2835a717540cSStefano Zampini if (qr_needed) { 2836984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2837cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2838984c4197SStefano Zampini } 2839984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2840984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2841984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2842984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2843984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2844674ae819SStefano Zampini } 2845a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2846906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2847906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2848906d46d4SStefano Zampini 2849906d46d4SStefano Zampini /* assembling of global change of variable */ 2850bbb9e6c6SStefano Zampini { 2851bbb9e6c6SStefano Zampini Mat tmat; 285216f15bc4SStefano Zampini PetscInt bs; 285316f15bc4SStefano Zampini 2854906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2855906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 286016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 286116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2862906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2866bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2867bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2868bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2870bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2872906d46d4SStefano Zampini } 2873906d46d4SStefano Zampini /* check */ 2874906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2875906d46d4SStefano Zampini PetscReal error; 2876906d46d4SStefano Zampini Vec x,x_change; 2877906d46d4SStefano Zampini 2878906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2879906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2880906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2882bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2883bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2884bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2885bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2886bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2887906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2888906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2889906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2890906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2891bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2892906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2893906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2894906d46d4SStefano Zampini } 2895b96c3477SStefano Zampini 2896b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2897b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2898b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2899b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2900ac632422SStefano Zampini Mat S_new,tmat; 2901bbb9e6c6SStefano Zampini IS is_all_N; 2902bbb9e6c6SStefano Zampini 2903bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29046816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2905bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2906ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2907b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2908ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2909ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2910ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2911ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2912ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2913b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2914ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2915ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2916ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2917ac632422SStefano Zampini } 2918b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2919b96c3477SStefano Zampini } 2920b96c3477SStefano Zampini } 2921906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2922906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2923b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2924b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2925b9b85e73SStefano Zampini } 2926906d46d4SStefano Zampini 2927906d46d4SStefano Zampini /* set up change of basis context */ 2928906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2929906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2930906d46d4SStefano Zampini 2931906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2932906d46d4SStefano Zampini PetscInt global_size,local_size; 2933906d46d4SStefano Zampini 2934906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2935906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2936906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2937906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2938906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2939906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2940906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2941906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2942906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2943906d46d4SStefano Zampini } else { 2944906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2945906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2946906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2947906d46d4SStefano Zampini } 2948906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2949906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2950906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2951906d46d4SStefano Zampini } else { 2952906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2953906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2954906d46d4SStefano Zampini } 2955906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2956906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2957906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2958906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2959b9b85e73SStefano Zampini } 2960a717540cSStefano Zampini 2961727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2962727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2963727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2964aff50787SStefano 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); 29650e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29660e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2967aff50787SStefano 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); 2968c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2969727cdba6SStefano Zampini } 29700e6343abSStefano Zampini } 29710e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2972727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2973727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2974727cdba6SStefano Zampini 2975a717540cSStefano Zampini /* flush dbg viewer */ 2976b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2977b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2978b8ffe317SStefano Zampini } 2979a717540cSStefano Zampini 2980e310c8b4SStefano Zampini /* free workspace */ 2981a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29824641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 298308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29849162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29859162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 298608122e43SStefano Zampini } else { 29879162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29889162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29899162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 299008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 299108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29929162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29939162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 299408122e43SStefano Zampini } 2995674ae819SStefano Zampini PetscFunctionReturn(0); 2996674ae819SStefano Zampini } 2997674ae819SStefano Zampini 2998674ae819SStefano Zampini #undef __FUNCT__ 2999674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3000674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3001674ae819SStefano Zampini { 3002674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3003674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3004674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30057fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3006674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3007674ae819SStefano Zampini 3008674ae819SStefano Zampini PetscFunctionBegin; 30098e61c736SStefano Zampini /* Reset previously computed graph */ 30108e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3011674ae819SStefano Zampini /* Init local Graph struct */ 30127fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30137fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3014674ae819SStefano Zampini 3015575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3016575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3017575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3018575ad6abSStefano Zampini } 30199577ea80SStefano Zampini 3020674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30214d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30224d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30234d379d7bSStefano Zampini PetscInt nvtxs; 3024e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3025674ae819SStefano Zampini 30264d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30272fffb893SStefano Zampini 30282fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30292fffb893SStefano Zampini if (flg_row) { 30304d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3031b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30322fffb893SStefano Zampini } 30332fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30349b28b941SStefano 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 */ 30354d379d7bSStefano Zampini IS is_dummy; 30364d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30374d379d7bSStefano Zampini PetscInt j,sum; 30384d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30394d379d7bSStefano Zampini const PetscInt *idxs; 30404d379d7bSStefano Zampini PCBDDCGraph graph; 30414d379d7bSStefano Zampini PetscBT is_on_boundary; 30424d379d7bSStefano Zampini 30434d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30444d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30454d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30464d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30477fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30484d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3049e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3050e496cd5dSStefano Zampini if (flg_row) { 30514d379d7bSStefano Zampini graph->xadj = xadj; 30524d379d7bSStefano Zampini graph->adjncy = adjncy; 3053e496cd5dSStefano Zampini } 30544d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30554d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3056e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30574d379d7bSStefano Zampini 30584d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30599b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30604d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30614d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30624d379d7bSStefano Zampini } 30634d379d7bSStefano Zampini } 30644d379d7bSStefano Zampini 3065e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30664d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30674d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30684d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30694d379d7bSStefano Zampini } 30704d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30714d379d7bSStefano Zampini 3072e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30734d379d7bSStefano Zampini sum = 0; 30744d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30754d379d7bSStefano Zampini PetscInt sizecc = 0; 30764d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30774d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30784d379d7bSStefano Zampini sizecc++; 30794d379d7bSStefano Zampini } 30804d379d7bSStefano Zampini } 30814d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30824d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30834d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30844d379d7bSStefano Zampini } 30854d379d7bSStefano Zampini } 30864d379d7bSStefano Zampini sum += sizecc*sizecc; 30874d379d7bSStefano Zampini } 30884d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30894d379d7bSStefano Zampini sum = 0; 3090e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30914d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30924d379d7bSStefano Zampini cxadj[i] = sum; 30934d379d7bSStefano Zampini sum += temp; 30944d379d7bSStefano Zampini } 3095e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30964d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30974d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30984d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30994d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31004d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31014d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31024d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31034d379d7bSStefano Zampini sizecc++; 31044d379d7bSStefano Zampini } 31054d379d7bSStefano Zampini } 31064d379d7bSStefano Zampini } 31074d379d7bSStefano Zampini } 31084d379d7bSStefano Zampini } 31099b28b941SStefano Zampini if (sum) { 3110e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31114d379d7bSStefano Zampini } else { 31124d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31134d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31144d379d7bSStefano Zampini } 31154d379d7bSStefano Zampini graph->xadj = 0; 31164d379d7bSStefano Zampini graph->adjncy = 0; 31174d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31184d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31194d379d7bSStefano Zampini } 3120674ae819SStefano Zampini } 31219b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31229b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31239b28b941SStefano Zampini } 3124674ae819SStefano Zampini 312563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3126674ae819SStefano Zampini vertex_size = 1; 312763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 312863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 312995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3131a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3133674ae819SStefano Zampini } 313463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 313563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 313663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3137674ae819SStefano Zampini } 313863602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3139674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 314063602bcaSStefano Zampini } else { 314163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 314263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3143854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 314463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 314563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314663602bcaSStefano Zampini } 314763602bcaSStefano Zampini } 3148674ae819SStefano Zampini } 3149674ae819SStefano Zampini 3150674ae819SStefano Zampini /* Setup of Graph */ 3151785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3152a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3153785d1243SStefano Zampini } 3154785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3155a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3156785d1243SStefano Zampini } 3157302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3158674ae819SStefano Zampini 3159674ae819SStefano Zampini /* Graph's connected components analysis */ 3160674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3161674ae819SStefano Zampini 3162674ae819SStefano Zampini /* print some info to stdout */ 3163674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3164302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3165674ae819SStefano Zampini } 3166fb180af4SStefano Zampini 3167fb180af4SStefano Zampini /* mark topography has done */ 3168fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3169674ae819SStefano Zampini PetscFunctionReturn(0); 3170674ae819SStefano Zampini } 3171674ae819SStefano Zampini 3172674ae819SStefano Zampini #undef __FUNCT__ 3173674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3174674ae819SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(MPI_Comm comm,ISLocalToGlobalMapping l2gmap, PetscInt n_local_dofs, PetscInt local_dofs[], PetscInt local_dofs_mult[], PetscInt* n_global_subset, PetscInt* global_numbering_subset[]) 3175674ae819SStefano Zampini { 3176674ae819SStefano Zampini Vec local_vec,global_vec; 3177674ae819SStefano Zampini IS seqis,paris; 3178674ae819SStefano Zampini VecScatter scatter_ctx; 3179674ae819SStefano Zampini PetscScalar *array; 3180674ae819SStefano Zampini PetscInt *temp_global_dofs; 3181674ae819SStefano Zampini PetscScalar globalsum; 3182674ae819SStefano Zampini PetscInt i,j,s; 318317ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 318417ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 318517ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3186674ae819SStefano Zampini PetscBool first_found; 3187674ae819SStefano Zampini PetscErrorCode ierr; 3188674ae819SStefano Zampini 3189674ae819SStefano Zampini PetscFunctionBegin; 3190674ae819SStefano Zampini /* mpi buffers */ 3191b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3192b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3193674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 319417ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3195674ae819SStefano Zampini /* get maximum size of subset */ 3196785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3197674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3198674ae819SStefano Zampini max_local = 0; 3199b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3200674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3201674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3202674ae819SStefano Zampini } 3203674ae819SStefano Zampini } 3204b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3205674ae819SStefano Zampini max_global++; 3206674ae819SStefano Zampini max_local = 0; 3207b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3208674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3209674ae819SStefano Zampini max_local = local_dofs[i]; 3210674ae819SStefano Zampini } 3211674ae819SStefano Zampini } 3212674ae819SStefano Zampini max_local++; 3213674ae819SStefano Zampini /* allocate workspace */ 3214674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3215674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3216674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3218674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3219674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3220674ae819SStefano Zampini /* create scatter */ 3221674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3222674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3223674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3224674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3225674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3226674ae819SStefano Zampini /* init array */ 3227674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3228674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3229674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3230674ae819SStefano Zampini if (local_dofs_mult) { 3231674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3232674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3233674ae819SStefano Zampini } 3234674ae819SStefano Zampini } else { 3235674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3236674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3237674ae819SStefano Zampini } 3238674ae819SStefano Zampini } 3239674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3240674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3241674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3242674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3243674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 32445b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3245674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3246674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3247674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3248674ae819SStefano Zampini nlocals = 0; 3249674ae819SStefano Zampini first_index = -1; 3250674ae819SStefano Zampini first_found = PETSC_FALSE; 3251674ae819SStefano Zampini for (i=0;i<s;i++) { 3252b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3253674ae819SStefano Zampini first_found = PETSC_TRUE; 3254674ae819SStefano Zampini first_index = i; 3255674ae819SStefano Zampini } 32565b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3257674ae819SStefano Zampini } 3258674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3259674ae819SStefano Zampini if (!rank_prec_comm) { 3260674ae819SStefano Zampini dof_displs[0]=0; 3261674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3262674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3263674ae819SStefano Zampini } 3264674ae819SStefano Zampini } 3265674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3266674ae819SStefano Zampini if (first_found) { 3267674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3268674ae819SStefano Zampini old_index = first_index; 3269674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3270b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3271674ae819SStefano Zampini array[i] += array[old_index]; 3272674ae819SStefano Zampini old_index = i; 3273674ae819SStefano Zampini } 3274674ae819SStefano Zampini } 3275674ae819SStefano Zampini } 3276674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3277674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3278674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3279674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3280674ae819SStefano Zampini /* get global ordering of local dofs */ 32812b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3282674ae819SStefano Zampini if (local_dofs_mult) { 3283674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32845b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3285674ae819SStefano Zampini } 3286674ae819SStefano Zampini } else { 3287674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32885b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3289674ae819SStefano Zampini } 3290674ae819SStefano Zampini } 32912b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3292674ae819SStefano Zampini /* free workspace */ 3293674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3294674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3295674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 329617ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3297674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3298674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3299674ae819SStefano Zampini PetscFunctionReturn(0); 3300674ae819SStefano Zampini } 33019a7d3425SStefano Zampini 33029a7d3425SStefano Zampini #undef __FUNCT__ 33039a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33049a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33059a7d3425SStefano Zampini { 33069a7d3425SStefano Zampini PetscInt i,j; 33079a7d3425SStefano Zampini PetscScalar *alphas; 33089a7d3425SStefano Zampini PetscErrorCode ierr; 33099a7d3425SStefano Zampini 33109a7d3425SStefano Zampini PetscFunctionBegin; 33119a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3312785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33139a7d3425SStefano Zampini for (i=0;i<n;i++) { 33149a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33159a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33169a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33179a7d3425SStefano Zampini } 33189a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33199a7d3425SStefano Zampini PetscFunctionReturn(0); 33209a7d3425SStefano Zampini } 33219a7d3425SStefano Zampini 3322e7931f94SStefano Zampini #undef __FUNCT__ 332370cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 332428143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3325e7931f94SStefano Zampini { 3326e7931f94SStefano Zampini Mat subdomain_adj; 3327e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3328e7931f94SStefano Zampini MatPartitioning partitioner; 3329e7931f94SStefano Zampini Mat_IS *matis; 3330e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3331e7931f94SStefano Zampini PetscInt prank; 3332e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3333e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3334e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33353837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3336e7931f94SStefano Zampini PetscErrorCode ierr; 33372b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3338e7931f94SStefano Zampini PetscSubcomm subcomm; 3339a57a6d2fSStefano Zampini 3340e7931f94SStefano Zampini PetscFunctionBegin; 33412b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33422b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33432b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3344e7931f94SStefano Zampini 3345e7931f94SStefano Zampini /* Get info on mapping */ 3346e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3347e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3348e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3349e7931f94SStefano Zampini 3350e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3351785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3352e7931f94SStefano Zampini xadj[0] = 0; 3353e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3354785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3355785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3356e7931f94SStefano Zampini 33572b510759SStefano Zampini if (threshold) { 3358d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33592b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3360d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3361d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3362d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3363d023bfaeSStefano Zampini xadj_count++; 3364e7931f94SStefano Zampini } 3365e7931f94SStefano Zampini } 3366d023bfaeSStefano Zampini xadj[1] = xadj_count; 3367c8587f34SStefano Zampini } else { 3368e7931f94SStefano Zampini if (xadj[1]) { 3369e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3370e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3371c8587f34SStefano Zampini } 3372e7931f94SStefano Zampini } 3373e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3374e7931f94SStefano Zampini if (use_square) { 3375e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3376e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3377e7931f94SStefano Zampini } 3378e7931f94SStefano Zampini } 3379e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3380e7931f94SStefano Zampini 33813837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3382e7931f94SStefano Zampini 3383e7931f94SStefano Zampini /* 3384e7931f94SStefano Zampini Restrict work on active processes only. 3385e7931f94SStefano Zampini */ 3386e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3387e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3388e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33892b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3390d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3391e7931f94SStefano Zampini if (color) { 3392e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3393e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3394e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3395c8587f34SStefano Zampini } else { 339628143c3dSStefano Zampini PetscInt coarsening_ratio; 3397306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3398785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3399e7931f94SStefano Zampini prank = rank; 3400306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34018002ef2cSStefano Zampini /* 3402e7931f94SStefano Zampini for (i=0;i<size;i++) { 3403e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3404c8587f34SStefano Zampini } 34058002ef2cSStefano Zampini */ 3406e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3407e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3408c8587f34SStefano Zampini } 3409e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3410306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 341122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3412e7931f94SStefano Zampini 3413e7931f94SStefano Zampini /* Partition */ 3414306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3415e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3416e7931f94SStefano Zampini if (use_vwgt) { 34173837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3418e7931f94SStefano Zampini v_wgt[0] = local_size; 3419e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3420c8587f34SStefano Zampini } 342128143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 342228143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 342328143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3424e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3425e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 342622b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3427e7931f94SStefano Zampini 3428e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 342928143c3dSStefano Zampini if (contiguous) { 343028143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 343128143c3dSStefano Zampini } else { 343284ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 343328143c3dSStefano Zampini } 3434e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3435e7931f94SStefano Zampini /* clean up */ 3436e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3437e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3438e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3439e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3440e7931f94SStefano Zampini } 3441e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3442e7931f94SStefano Zampini 3443e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3444e7931f94SStefano Zampini i = 1; 3445e7931f94SStefano Zampini if (color) i=0; 3446e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3447e7931f94SStefano Zampini 3448e7931f94SStefano Zampini /* get back IS */ 3449e7931f94SStefano Zampini *is_sends = ranks_send_to; 3450e7931f94SStefano Zampini PetscFunctionReturn(0); 3451e7931f94SStefano Zampini } 3452e7931f94SStefano Zampini 3453e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3454e7931f94SStefano Zampini 3455e7931f94SStefano Zampini #undef __FUNCT__ 3456e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 345728143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3458e7931f94SStefano Zampini { 345970cf5478SStefano Zampini Mat local_mat; 3460e7931f94SStefano Zampini Mat_IS *matis; 3461e7931f94SStefano Zampini IS is_sends_internal; 34629d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 346328143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34649d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3465e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3466e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3467e7931f94SStefano Zampini const PetscInt* is_indices; 3468e7931f94SStefano Zampini MatType new_local_type; 3469e7931f94SStefano Zampini /* buffers */ 3470e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 347128143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34729d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3473e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3474e7931f94SStefano Zampini /* MPI */ 347528143c3dSStefano Zampini MPI_Comm comm,comm_n; 347628143c3dSStefano Zampini PetscSubcomm subcomm; 3477e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 347828143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 347928143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 348028143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 348128143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 348228143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3483e7931f94SStefano Zampini PetscErrorCode ierr; 3484e7931f94SStefano Zampini 3485e7931f94SStefano Zampini PetscFunctionBegin; 348628143c3dSStefano Zampini /* TODO: add missing checks */ 348728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 348828143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 348928143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 349028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3491e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 349228143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3493e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3494e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3495e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3496e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3497e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 349828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 349970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 350070cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 350128143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 350270cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 350370cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 350470cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 350570cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 350670cf5478SStefano Zampini } 3507e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3508e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3509e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3510e7931f94SStefano Zampini if (!is_sends) { 351128143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 351228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 351328143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3514c8587f34SStefano Zampini } else { 3515e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3516e7931f94SStefano Zampini is_sends_internal = is_sends; 3517c8587f34SStefano Zampini } 3518e7931f94SStefano Zampini 3519e7931f94SStefano Zampini /* get pointer of MATIS data */ 3520e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3521e7931f94SStefano Zampini 3522e7931f94SStefano Zampini /* get comm */ 3523a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3524e7931f94SStefano Zampini 3525e7931f94SStefano Zampini /* compute number of sends */ 3526e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3527e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3528e7931f94SStefano Zampini 3529e7931f94SStefano Zampini /* compute number of receives */ 3530e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3531785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3532e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3533e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3534e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3535e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3536e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3537e7931f94SStefano Zampini 353828143c3dSStefano Zampini /* restrict comm if requested */ 353928143c3dSStefano Zampini subcomm = 0; 354028143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 354128143c3dSStefano Zampini if (restrict_comm) { 3542779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3543779c1cceSStefano Zampini 354428143c3dSStefano Zampini color = 0; 35456366ad86SStefano Zampini if (!n_recvs && n_sends) color = 1; /* processes sending and not receiving anything will not partecipate in new comm */ 354628143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 354728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 354828143c3dSStefano Zampini /* check if reuse has been requested */ 354928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 355028143c3dSStefano Zampini if (*mat_n) { 355128143c3dSStefano Zampini PetscMPIInt subcommsize2; 355228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 355328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 355428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 355528143c3dSStefano Zampini } else { 355628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 355728143c3dSStefano Zampini } 355828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3559779c1cceSStefano Zampini PetscMPIInt rank; 3560779c1cceSStefano Zampini 3561779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 356228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 356328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 356428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3565306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 356628143c3dSStefano Zampini } 356728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 356828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 356928143c3dSStefano Zampini } else { 357028143c3dSStefano Zampini comm_n = comm; 357128143c3dSStefano Zampini } 357228143c3dSStefano Zampini 3573e7931f94SStefano Zampini /* prepare send/receive buffers */ 3574785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3575e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3576785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3577e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 357828143c3dSStefano Zampini if (nis) { 3579854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 358028143c3dSStefano Zampini } 3581e7931f94SStefano Zampini 358228143c3dSStefano Zampini /* Get data from local matrices */ 3583e7931f94SStefano Zampini if (!isdense) { 3584a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3585e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3586e7931f94SStefano Zampini /* 3587e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3588e7931f94SStefano Zampini send_buffer_idxs should contain: 3589e7931f94SStefano Zampini - MatType_PRIVATE type 3590e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3591e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3592e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3593e7931f94SStefano Zampini */ 3594e7931f94SStefano Zampini } else { 3595e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3596e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3597854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3598e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3599e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3600e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3601e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3602e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3603e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3604e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3605e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3606e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3607c8587f34SStefano Zampini } 3608c8587f34SStefano Zampini } 3609e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 361028143c3dSStefano Zampini /* additional is (if any) */ 361128143c3dSStefano Zampini if (nis) { 361228143c3dSStefano Zampini PetscMPIInt psum; 361328143c3dSStefano Zampini PetscInt j; 361428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 361528143c3dSStefano Zampini PetscInt plen; 361628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 361828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 361928143c3dSStefano Zampini } 3620854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 362128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 362228143c3dSStefano Zampini PetscInt plen; 362328143c3dSStefano Zampini const PetscInt *is_array_idxs; 362428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 362528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 362628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 362828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 363028143c3dSStefano Zampini } 363128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 363228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 363328143c3dSStefano Zampini } 363428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 363528143c3dSStefano Zampini } 363628143c3dSStefano Zampini 3637e7931f94SStefano Zampini buf_size_idxs = 0; 3638e7931f94SStefano Zampini buf_size_vals = 0; 363928143c3dSStefano Zampini buf_size_idxs_is = 0; 3640e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3641e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3642e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 364328143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3644e7931f94SStefano Zampini } 3645785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3646785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 364795ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3648e7931f94SStefano Zampini 3649e7931f94SStefano Zampini /* get new tags for clean communications */ 3650e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3651e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 365228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3653e7931f94SStefano Zampini 3654e7931f94SStefano Zampini /* allocate for requests */ 3655785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3656785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 365795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3658785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3659785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 366095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3661e7931f94SStefano Zampini 3662e7931f94SStefano Zampini /* communications */ 3663e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3664e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 366528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3666e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3667e7931f94SStefano Zampini source_dest = onodes[i]; 3668e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3669e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3670e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3671e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 367228143c3dSStefano Zampini if (nis) { 367328143c3dSStefano 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); 367428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 367528143c3dSStefano Zampini } 3676e7931f94SStefano Zampini } 3677e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3678e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3679e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3680e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 368128143c3dSStefano Zampini if (nis) { 368228143c3dSStefano 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); 368328143c3dSStefano Zampini } 3684e7931f94SStefano Zampini } 3685e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3686e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3687e7931f94SStefano Zampini 3688e7931f94SStefano Zampini /* assemble new l2g map */ 3689e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3690e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36919d30be91SStefano Zampini new_local_rows = 0; 3692e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36939d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3694e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3695e7931f94SStefano Zampini } 36969d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3697e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36989d30be91SStefano Zampini new_local_rows = 0; 3699e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37009d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37019d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3702e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3703e7931f94SStefano Zampini } 37049d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37059d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3706e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3707e7931f94SStefano Zampini 3708e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3709e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3710e7931f94SStefano 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) */ 3711e7931f94SStefano Zampini if (n_recvs) { 371228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3713e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3714e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3715e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3716e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3717e7931f94SStefano Zampini break; 3718e7931f94SStefano Zampini } 3719e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3720e7931f94SStefano Zampini } 3721e7931f94SStefano Zampini switch (new_local_type_private) { 372228143c3dSStefano Zampini case MATDENSE_PRIVATE: 372328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3724e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3725e7931f94SStefano Zampini bs = 1; 372628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 372728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 372828143c3dSStefano Zampini bs = 1; 372928143c3dSStefano Zampini } 3730e7931f94SStefano Zampini break; 3731e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3732e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3733e7931f94SStefano Zampini bs = 1; 3734e7931f94SStefano Zampini break; 3735e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3736e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3737e7931f94SStefano Zampini break; 3738e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3739e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3740e7931f94SStefano Zampini break; 3741e7931f94SStefano Zampini default: 37429d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3743e7931f94SStefano Zampini break; 3744e7931f94SStefano Zampini } 374528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 374628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 374728143c3dSStefano Zampini bs = 1; 3748e7931f94SStefano Zampini } 3749e7931f94SStefano Zampini 375070cf5478SStefano Zampini /* create MATIS object if needed */ 375170cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3752e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 375328143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 375470cf5478SStefano Zampini } else { 375570cf5478SStefano Zampini /* it also destroys the local matrices */ 375670cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 375770cf5478SStefano Zampini } 375870cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3759e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37609d30be91SStefano Zampini 37619d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37629d30be91SStefano Zampini 37639d30be91SStefano Zampini /* Global to local map of received indices */ 37649d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37659d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37669d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37679d30be91SStefano Zampini 37689d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37699d30be91SStefano Zampini buf_size_idxs = 0; 37709d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37719d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37729d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37739d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37749d30be91SStefano Zampini } 37759d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37769d30be91SStefano Zampini 37779d30be91SStefano Zampini /* set preallocation */ 37789d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37799d30be91SStefano Zampini if (!newisdense) { 37809d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37819d30be91SStefano Zampini 37829d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37839d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37849d30be91SStefano Zampini if (n_recvs) { 37859d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37869d30be91SStefano Zampini } 37879d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37889d30be91SStefano Zampini PetscInt j; 37899d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37909d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37919d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37929d30be91SStefano Zampini } 37939d30be91SStefano Zampini } else { 37949d30be91SStefano Zampini /* TODO */ 37959d30be91SStefano Zampini } 37969d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37979d30be91SStefano Zampini } 37989d30be91SStefano Zampini if (new_local_nnz) { 37999d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38009d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38019d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38029d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38039d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38049d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38059d30be91SStefano Zampini } else { 38069d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38079d30be91SStefano Zampini } 38089d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38099d30be91SStefano Zampini } else { 38109d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38119d30be91SStefano Zampini } 3812e7931f94SStefano Zampini 3813e7931f94SStefano Zampini /* set values */ 3814e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38159d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3816e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3817e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3818e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38199d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3820e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3821e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3822e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 382328143c3dSStefano Zampini } else { 382428143c3dSStefano Zampini /* TODO */ 3825e7931f94SStefano Zampini } 3826e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3827e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3828e7931f94SStefano Zampini } 3829e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3830e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38339d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38349d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3835e7931f94SStefano Zampini 3836dfd14d43SStefano Zampini #if 0 383728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3838e7931f94SStefano Zampini Vec lvec,rvec; 3839e7931f94SStefano Zampini PetscReal infty_error; 3840e7931f94SStefano Zampini 38412a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3842e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3843e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3844e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 384570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3846e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3847e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3848e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3849e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3850e7931f94SStefano Zampini } 385128143c3dSStefano Zampini #endif 3852e7931f94SStefano Zampini 385328143c3dSStefano Zampini /* assemble new additional is (if any) */ 385428143c3dSStefano Zampini if (nis) { 385528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 385628143c3dSStefano Zampini 385728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3858854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 385928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 386028143c3dSStefano Zampini psum = 0; 386128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 386228143c3dSStefano Zampini for (j=0;j<nis;j++) { 386328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 386428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 386528143c3dSStefano Zampini psum += plen; 386628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 386728143c3dSStefano Zampini } 386828143c3dSStefano Zampini } 3869854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3870854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 387128143c3dSStefano Zampini for (i=1;i<nis;i++) { 387228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 387328143c3dSStefano Zampini } 387428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 387528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 387628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 387728143c3dSStefano Zampini for (j=0;j<nis;j++) { 387828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 388028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 388128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 388228143c3dSStefano Zampini } 388328143c3dSStefano Zampini } 388428143c3dSStefano Zampini for (i=0;i<nis;i++) { 388528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 388628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 388728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 388828143c3dSStefano Zampini } 388928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 389028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 389128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 389228143c3dSStefano Zampini } 3893e7931f94SStefano Zampini /* free workspace */ 389428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3895e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3896e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3897e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3898e7931f94SStefano Zampini if (isdense) { 3899e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3900e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3901e7931f94SStefano Zampini } else { 3902e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3903e7931f94SStefano Zampini } 390428143c3dSStefano Zampini if (nis) { 390528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 390628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 390728143c3dSStefano Zampini } 3908e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 391028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3911e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3912e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 391328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3914e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3915e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3916e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3917e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3918e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 391928143c3dSStefano Zampini if (nis) { 392028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 392128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 392228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 392328143c3dSStefano Zampini } 392428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 392528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 392628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 392728143c3dSStefano Zampini for (i=0;i<nis;i++) { 392828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 392928143c3dSStefano Zampini } 393028143c3dSStefano Zampini } 3931e7931f94SStefano Zampini PetscFunctionReturn(0); 3932e7931f94SStefano Zampini } 3933a57a6d2fSStefano Zampini 393412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3935af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 393612edc857SStefano Zampini 3937c8587f34SStefano Zampini #undef __FUNCT__ 3938c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3939c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3940c8587f34SStefano Zampini { 3941c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3942c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 394320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39449881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 394520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39466e683305SStefano Zampini IS coarse_is,*isarray; 39476e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39486e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3949f9eb5b7dSStefano Zampini PC pc_temp; 3950c8587f34SStefano Zampini PCType coarse_pc_type; 3951c8587f34SStefano Zampini KSPType coarse_ksp_type; 3952f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39534f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39546e683305SStefano Zampini Mat t_coarse_mat_is; 39556e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39566e683305SStefano Zampini PetscMPIInt all_procs; 395774e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 395868457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 395922bc73bbSStefano Zampini PetscScalar *array; 39609881197aSStefano Zampini PetscErrorCode ierr; 3961fdc09c96SStefano Zampini 3962c8587f34SStefano Zampini PetscFunctionBegin; 3963c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 396468457ee5SStefano 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 */ 3965fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39665a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3967fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3968f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3969f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3970f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3971fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3972fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3973727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3974fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3975fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3976fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3977f4ddd8eeSStefano Zampini } 3978fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3979fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3980f4ddd8eeSStefano Zampini } 398170cf5478SStefano Zampini /* reset any subassembling information */ 398270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39836e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39846e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3985fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3986f4ddd8eeSStefano Zampini } 3987c8587f34SStefano Zampini 39886e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39892b510759SStefano Zampini im_active = !!(pcis->n); 39902b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39916e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39926e683305SStefano Zampini void_procs = all_procs-active_procs; 39936e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 399474e2c79eSStefano Zampini redist = PETSC_FALSE; 399522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39966e683305SStefano Zampini csin_ml = PETSC_TRUE; 39976e683305SStefano Zampini ncoarse_ml = void_procs; 3998779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3999779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 400018a45a71SStefano Zampini csin_ds = PETSC_TRUE; 400118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 400218a45a71SStefano Zampini redist = PETSC_TRUE; 400318a45a71SStefano Zampini } else { 40046e683305SStefano Zampini csin_ds = PETSC_TRUE; 4005779c1cceSStefano Zampini ncoarse_ds = active_procs; 4006779c1cceSStefano Zampini redist = PETSC_TRUE; 400718a45a71SStefano Zampini } 40086e683305SStefano Zampini } else { 40096e683305SStefano Zampini csin_ml = PETSC_FALSE; 40106e683305SStefano Zampini ncoarse_ml = all_procs; 40116e683305SStefano Zampini if (void_procs) { 40126e683305SStefano Zampini csin_ds = PETSC_TRUE; 40136e683305SStefano Zampini ncoarse_ds = void_procs; 40146e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40156e683305SStefano Zampini } else { 4016779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 401774e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 401874e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401974e2c79eSStefano Zampini redist = PETSC_TRUE; 402074e2c79eSStefano Zampini } else { 40216e683305SStefano Zampini csin_ds = PETSC_FALSE; 40226e683305SStefano Zampini ncoarse_ds = all_procs; 40236e683305SStefano Zampini } 40246e683305SStefano Zampini } 402574e2c79eSStefano Zampini } 40266e683305SStefano Zampini 40276e683305SStefano Zampini /* 40286e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40296e683305SStefano Zampini - we have not exceeded the number of levels requested 40306e683305SStefano Zampini - we can actually subassemble the active processes 40316e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40326e683305SStefano Zampini */ 40336e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40346e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40356e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40366e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40376e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4038f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40392b510759SStefano Zampini } else { 4040f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4041c8587f34SStefano Zampini } 4042c8587f34SStefano Zampini } 40436e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40446e683305SStefano Zampini if (multilevel_allowed) { 40456e683305SStefano Zampini ncoarse = ncoarse_ml; 40466e683305SStefano Zampini csin = csin_ml; 404758da7f69SStefano Zampini redist = PETSC_FALSE; 40486e683305SStefano Zampini } else { 40496e683305SStefano Zampini ncoarse = ncoarse_ds; 40506e683305SStefano Zampini csin = csin_ds; 40516e683305SStefano Zampini } 4052e7931f94SStefano Zampini 4053abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4054abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4055abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4056abbbba34SStefano Zampini 4057abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 405822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 405922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 406022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 406122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4062b9b85e73SStefano Zampini #if 0 4063b9b85e73SStefano Zampini { 4064b9b85e73SStefano Zampini PetscViewer viewer; 4065b9b85e73SStefano Zampini char filename[256]; 4066b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4067b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4068b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4069b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4070b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4071b9b85e73SStefano Zampini } 4072b9b85e73SStefano Zampini #endif 40736e683305SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,&t_coarse_mat_is);CHKERRQ(ierr); 40746e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40756e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40766e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4077abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4078abbbba34SStefano Zampini 40796e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40806e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40816e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40826e683305SStefano Zampini const PetscInt *idxs; 40836e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40846e683305SStefano Zampini 40856e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40860be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40876e683305SStefano Zampini /* allocate space for temporary storage */ 4088854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4089854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40906e683305SStefano Zampini /* allocate for IS array */ 40916e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40926e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40936e683305SStefano Zampini nis = nisdofs + nisneu; 4094854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40956e683305SStefano Zampini /* dofs splitting */ 40966e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40976e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40986e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40996e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41006e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41016e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41026e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41036e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41046e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41056e683305SStefano Zampini } 41066e683305SStefano Zampini /* neumann boundaries */ 41076e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41086e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41096e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41106e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41116e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41126e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41156e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41166e683305SStefano Zampini } 41176e683305SStefano Zampini /* free memory */ 41186e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41196e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41206e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41216e683305SStefano Zampini } else { 41226e683305SStefano Zampini nis = 0; 41236e683305SStefano Zampini nisdofs = 0; 41246e683305SStefano Zampini nisneu = 0; 41256e683305SStefano Zampini isarray = NULL; 41266e683305SStefano Zampini } 41276e683305SStefano Zampini /* destroy no longer needed map */ 41286e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41296e683305SStefano Zampini 41306e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41316e683305SStefano Zampini coarse_mat_is = NULL; 41326e683305SStefano Zampini if (csin) { 41336e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 413474e2c79eSStefano Zampini if (redist) { 413574e2c79eSStefano Zampini PetscMPIInt rank; 4136779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 413774e2c79eSStefano Zampini 413874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 413958da7f69SStefano Zampini spc = active_procs/ncoarse; 414058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4141779c1cceSStefano Zampini if (im_active) { 4142779c1cceSStefano Zampini destsize = 1; 414374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 414474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 414574e2c79eSStefano Zampini } else { 414674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 414774e2c79eSStefano Zampini } 414874e2c79eSStefano Zampini } else { 4149779c1cceSStefano Zampini destsize = 0; 41506e683305SStefano Zampini } 4151779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4152779c1cceSStefano Zampini } else if (csin_type_simple) { 41536e683305SStefano Zampini PetscMPIInt rank; 41546e683305SStefano Zampini PetscInt issize,isidx; 4155779c1cceSStefano Zampini 41566e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41576e683305SStefano Zampini if (im_active) { 41586e683305SStefano Zampini issize = 1; 41596e683305SStefano Zampini isidx = (PetscInt)rank; 41606e683305SStefano Zampini } else { 41616e683305SStefano Zampini issize = 0; 41626e683305SStefano Zampini isidx = -1; 41636e683305SStefano Zampini } 41646e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4165779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41666e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41676e683305SStefano Zampini } 4168779c1cceSStefano Zampini 4169779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4170779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4171779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4172779c1cceSStefano Zampini PetscInt *coarse_candidates; 4173779c1cceSStefano Zampini const PetscInt* tisindices; 4174779c1cceSStefano Zampini 4175779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4176779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4177779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4178779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4179779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4180779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4181779c1cceSStefano Zampini } 4182779c1cceSStefano Zampini } 4183779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4184779c1cceSStefano Zampini 4185779c1cceSStefano Zampini 41866e683305SStefano Zampini if (pcbddc->dbg_flag) { 41876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41896e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4191779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41936e683305SStefano Zampini } 41946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41956e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41966e683305SStefano Zampini } 41976e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41986e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41996e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4200854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42016e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42026e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42036e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42046e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42056e683305SStefano Zampini } 42066e683305SStefano Zampini if (pcbddc->dbg_flag) { 42076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42096e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42106e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42116e683305SStefano Zampini } 4212779c1cceSStefano Zampini } 42136e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 42146e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 42156e683305SStefano Zampini } else { 42166e683305SStefano Zampini if (pcbddc->dbg_flag) { 42176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42196e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42206e683305SStefano Zampini } 42216e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42226e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42236e683305SStefano Zampini } 42246e683305SStefano Zampini 42256e683305SStefano Zampini /* create local to global scatters for coarse problem */ 422668457ee5SStefano Zampini if (compute_vecs) { 42276e683305SStefano Zampini PetscInt lrows; 42286e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42296e683305SStefano Zampini if (coarse_mat_is) { 42306e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42316e683305SStefano Zampini } else { 42326e683305SStefano Zampini lrows = 0; 42336e683305SStefano Zampini } 42346e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42356e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42366e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42376e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42386e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42396e683305SStefano Zampini } 42406e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42416e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4242c8587f34SStefano Zampini 4243f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4244f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4245f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4246f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4247f9eb5b7dSStefano Zampini } else { 4248f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4249f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4250c8587f34SStefano Zampini } 4251c8587f34SStefano Zampini 42526e683305SStefano Zampini /* print some info if requested */ 42536e683305SStefano Zampini if (pcbddc->dbg_flag) { 42546e683305SStefano Zampini if (!multilevel_allowed) { 42556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42566e683305SStefano Zampini if (multilevel_requested) { 42576e683305SStefano 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); 42586e683305SStefano Zampini } else if (pcbddc->max_levels) { 42596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42606e683305SStefano Zampini } 42616e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42626e683305SStefano Zampini } 42636e683305SStefano Zampini } 42646e683305SStefano Zampini 4265f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42666e683305SStefano Zampini if (coarse_mat_is) { 42676e683305SStefano Zampini MatReuse coarse_mat_reuse; 42686a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42696e683305SStefano Zampini if (pcbddc->dbg_flag) { 42706e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42716e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42726e683305SStefano Zampini } 4273f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4274312be037SStefano Zampini char prefix[256],str_level[16]; 4275e604994aSStefano Zampini size_t len; 42766e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4277c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4278f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42795f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4280c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42816e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4282c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4283c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42843e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4285e604994aSStefano Zampini /* prefix */ 4286e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4287e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4288e604994aSStefano Zampini if (!pcbddc->current_level) { 4289e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4290e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4291c8587f34SStefano Zampini } else { 4292e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4293312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4294312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 429534d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4296312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4297e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4298e604994aSStefano Zampini } 4299e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43003e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43013e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43023e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43033e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4304f9eb5b7dSStefano Zampini /* allow user customization */ 4305f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43063e3c6dadSStefano Zampini } 43073e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43083e3c6dadSStefano Zampini if (nisdofs) { 43093e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43103e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43113e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43123e3c6dadSStefano Zampini } 43133e3c6dadSStefano Zampini } 43143e3c6dadSStefano Zampini if (nisneu) { 43153e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43163e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4317312be037SStefano Zampini } 4318f9eb5b7dSStefano Zampini 4319f9eb5b7dSStefano Zampini /* get some info after set from options */ 4320f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4321f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4322f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43234f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43246e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4325f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4326f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4327f9eb5b7dSStefano Zampini } 43284f3a063dSStefano Zampini if (isredundant) { 43294f3a063dSStefano Zampini KSP inner_ksp; 43304f3a063dSStefano Zampini PC inner_pc; 43314f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43324f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43334f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43344f3a063dSStefano Zampini } 4335f9eb5b7dSStefano Zampini 4336f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4337fa7f1dd8SStefano Zampini if (coarse_reuse) { 433881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4339fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43406e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4341fa7f1dd8SStefano Zampini } else { 43426e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4343fa7f1dd8SStefano Zampini } 4344c8587f34SStefano Zampini if (isbddc || isnn) { 434522bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 434670cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 434728143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 434822b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43516e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43526e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 435322b6e8a2SStefano Zampini } 435470cf5478SStefano Zampini } 43556e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 435670cf5478SStefano Zampini } else { 435722bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 435822bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 435922bc73bbSStefano Zampini } 436022bc73bbSStefano Zampini } else { 43612e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4362c8587f34SStefano Zampini } 4363c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4364c8587f34SStefano Zampini 43653301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43665a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 43673301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 43683301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 43693301b35fSStefano Zampini } 43703301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 43713301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 43723301b35fSStefano Zampini } 43733301b35fSStefano Zampini if (pc->pmat->spd_set) { 43743301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 43753301b35fSStefano Zampini } 43766e683305SStefano Zampini /* set operators */ 43775f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43786e683305SStefano Zampini if (pcbddc->dbg_flag) { 43796e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43806e683305SStefano Zampini } 43816e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43826e683305SStefano Zampini coarse_mat = 0; 43836e683305SStefano Zampini } 43846e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4385b9b85e73SStefano Zampini #if 0 4386b9b85e73SStefano Zampini { 4387b9b85e73SStefano Zampini PetscViewer viewer; 4388b9b85e73SStefano Zampini char filename[256]; 4389b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4390b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4391b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4392b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4393b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4394b9b85e73SStefano Zampini } 4395b9b85e73SStefano Zampini #endif 4396c8587f34SStefano Zampini 4397c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4398c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4399c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 440098a51de6SStefano Zampini } 440198a51de6SStefano Zampini 440298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 440398a51de6SStefano Zampini Vec crhs,csol; 440498a51de6SStefano Zampini PetscBool ispreonly; 440598a51de6SStefano Zampini if (CoarseNullSpace) { 4406c8587f34SStefano Zampini if (isbddc) { 4407c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4408c8587f34SStefano Zampini } else { 4409c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4410c8587f34SStefano Zampini } 4411c8587f34SStefano Zampini } 4412f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4413f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4414f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4415f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44166e683305SStefano Zampini /* hack */ 4417f347579bSStefano Zampini if (!csol) { 44182a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4419f9eb5b7dSStefano Zampini } 4420f347579bSStefano Zampini if (!crhs) { 44212a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4422f347579bSStefano Zampini } 4423cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4424cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44256e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4426c8587f34SStefano Zampini KSP check_ksp; 44272b510759SStefano Zampini KSPType check_ksp_type; 4428c8587f34SStefano Zampini PC check_pc; 44296e683305SStefano Zampini Vec check_vec,coarse_vec; 44306a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44312b510759SStefano Zampini PetscInt its; 44326e683305SStefano Zampini PetscBool compute_eigs; 44336e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44346e683305SStefano Zampini PetscInt neigs; 44358e185a42SStefano Zampini const char *prefix; 4436c8587f34SStefano Zampini 44372b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44386e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 443923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4440f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44412b510759SStefano Zampini if (ispreonly) { 44422b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44436e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44442b510759SStefano Zampini } else { 4445cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44466e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4447c8587f34SStefano Zampini } 4448c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44496e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44506e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44516e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4452a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4453a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4454a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4455a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4456c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4457c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4458c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4459c8587f34SStefano Zampini /* create random vec */ 44606e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44616e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4462c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4463c8587f34SStefano Zampini if (CoarseNullSpace) { 4464c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4465c8587f34SStefano Zampini } 44666e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4467c8587f34SStefano Zampini /* solve coarse problem */ 44686e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4469c8587f34SStefano Zampini if (CoarseNullSpace) { 44706e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4471c8587f34SStefano Zampini } 4472cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44736e683305SStefano Zampini if (compute_eigs) { 4474854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4475854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44776e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44786e683305SStefano Zampini lambda_min = eigs_r[0]; 44796e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44806e683305SStefano Zampini if (lambda_max>lambda_min) { 4481cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4482cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4483cbcc2c2aSStefano Zampini } 4484c8587f34SStefano Zampini } 4485c8587f34SStefano Zampini } 4486cbcc2c2aSStefano Zampini 4487c8587f34SStefano Zampini /* check coarse problem residual error */ 44886e683305SStefano Zampini if (pcbddc->dbg_flag) { 44896e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44906e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44916e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4492c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44936e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44946e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4495c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4496779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44976e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44986e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45006e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45016e683305SStefano Zampini if (compute_eigs) { 45026e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4503deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4504c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45056e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45066e683305SStefano 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); 45076e683305SStefano Zampini for (i=0;i<neigs;i++) { 45086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4509c8587f34SStefano Zampini } 45106e683305SStefano Zampini } 45116e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45126e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45136e683305SStefano Zampini } 4514c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45156e683305SStefano Zampini if (compute_eigs) { 45166e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45176e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4518c8587f34SStefano Zampini } 45196e683305SStefano Zampini } 45206e683305SStefano Zampini } 4521cbcc2c2aSStefano Zampini /* print additional info */ 4522cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45236e683305SStefano Zampini /* waits until all processes reaches this point */ 45246e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4525cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4526cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4527cbcc2c2aSStefano Zampini } 4528cbcc2c2aSStefano Zampini 45292b510759SStefano Zampini /* free memory */ 4530c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4531fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4532c8587f34SStefano Zampini PetscFunctionReturn(0); 4533c8587f34SStefano Zampini } 4534674ae819SStefano Zampini 4535f34684f1SStefano Zampini #undef __FUNCT__ 4536f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4537f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4538f34684f1SStefano Zampini { 4539f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4540f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4541f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 45420e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 454373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 45440e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4545f34684f1SStefano Zampini PetscErrorCode ierr; 4546f34684f1SStefano Zampini 4547f34684f1SStefano Zampini PetscFunctionBegin; 4548f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45490e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45500e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4551727cdba6SStefano Zampini } 45520e6343abSStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult,&coarse_size,&t_local_primal_indices);CHKERRQ(ierr); 45530e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 45540e6343abSStefano Zampini local_size = 0; 45550e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 45560e6343abSStefano Zampini PetscInt j; 45570e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 45580e6343abSStefano Zampini } 45590e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 45600e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45610e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45620e6343abSStefano Zampini } 4563f34684f1SStefano Zampini 4564f34684f1SStefano Zampini /* check numbering */ 4565f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4566f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4567b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4568f34684f1SStefano Zampini 4569f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4570f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4571f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45720fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4573f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4574f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4575727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4576f34684f1SStefano Zampini } 4577f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4578f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4579f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4580f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4581f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4582f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4583f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4584f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4585f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4586f34684f1SStefano Zampini if (array[i] == 1.0) { 4587b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4588f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4589f34684f1SStefano Zampini } 4590f34684f1SStefano Zampini } 4591b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4592f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4593f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4594f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4595f34684f1SStefano Zampini } 4596f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4597f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4598f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4599f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4600f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4601f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4602b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4603ca8b9ea9SStefano Zampini PetscInt *gidxs; 4604ca8b9ea9SStefano Zampini 4605ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4606ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4607f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4608f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4609f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4610f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46114bc2dc4bSStefano 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); 4612f34684f1SStefano Zampini } 4613f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4614ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4615f34684f1SStefano Zampini } 4616f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4617302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4618f34684f1SStefano Zampini } 46198bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4620f34684f1SStefano Zampini /* get back data */ 4621f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4622f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4623674ae819SStefano Zampini PetscFunctionReturn(0); 4624674ae819SStefano Zampini } 4625674ae819SStefano Zampini 4626e456f2a8SStefano Zampini #undef __FUNCT__ 4627e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4628a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4629e456f2a8SStefano Zampini { 4630e456f2a8SStefano Zampini IS localis_t; 4631a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4632e456f2a8SStefano Zampini PetscScalar *vals; 4633e456f2a8SStefano Zampini PetscErrorCode ierr; 4634e456f2a8SStefano Zampini 4635e456f2a8SStefano Zampini PetscFunctionBegin; 4636a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4637e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4638854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4639e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4640e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4641a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4642a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46431035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4644a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46451035eff8SStefano Zampini } 4646a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4647e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4648e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4649a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4650a7dc3881SStefano Zampini /* now compute set in local ordering */ 4651a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4652a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4653a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4654a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4655a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4656ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4657e456f2a8SStefano Zampini lsize++; 4658e456f2a8SStefano Zampini } 4659e456f2a8SStefano Zampini } 4660854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4661a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4662ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4663e456f2a8SStefano Zampini idxs[lsize++] = i; 4664e456f2a8SStefano Zampini } 4665e456f2a8SStefano Zampini } 4666a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4667a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4668e456f2a8SStefano Zampini *localis = localis_t; 4669e456f2a8SStefano Zampini PetscFunctionReturn(0); 4670e456f2a8SStefano Zampini } 4671906d46d4SStefano Zampini 4672906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4673906d46d4SStefano Zampini #undef __FUNCT__ 4674906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4675906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4676906d46d4SStefano Zampini { 4677906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4678906d46d4SStefano Zampini PetscErrorCode ierr; 4679906d46d4SStefano Zampini 4680906d46d4SStefano Zampini PetscFunctionBegin; 4681906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4682906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4683906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4684906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4685906d46d4SStefano Zampini PetscFunctionReturn(0); 4686906d46d4SStefano Zampini } 4687906d46d4SStefano Zampini 4688906d46d4SStefano Zampini #undef __FUNCT__ 4689906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4690906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4691906d46d4SStefano Zampini { 4692906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4693906d46d4SStefano Zampini PetscErrorCode ierr; 4694906d46d4SStefano Zampini 4695906d46d4SStefano Zampini PetscFunctionBegin; 4696906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4697906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4698906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4699906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4700906d46d4SStefano Zampini PetscFunctionReturn(0); 4701906d46d4SStefano Zampini } 4702b96c3477SStefano Zampini 4703b96c3477SStefano Zampini #undef __FUNCT__ 4704b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 470508122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4706b96c3477SStefano Zampini { 4707a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4708b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4709b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4710a64f4aa4SStefano Zampini Mat S_j; 4711b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4712b96c3477SStefano Zampini PetscBool free_used_adj; 4713b96c3477SStefano Zampini PetscErrorCode ierr; 4714b96c3477SStefano Zampini 4715b96c3477SStefano Zampini PetscFunctionBegin; 4716b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4717b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 471808122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4719b96c3477SStefano Zampini used_xadj = NULL; 4720b96c3477SStefano Zampini used_adjncy = NULL; 4721b96c3477SStefano Zampini } else { 472208122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 472308122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 472408122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 472508122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4726b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4727b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4728b96c3477SStefano Zampini } else { 47292fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4730b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4731b96c3477SStefano Zampini PetscInt nvtxs; 4732b96c3477SStefano Zampini 47332fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47342fffb893SStefano Zampini if (flg_row) { 4735b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4736b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4737b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4738b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47392fffb893SStefano Zampini } else { 47402fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47412fffb893SStefano Zampini used_xadj = NULL; 47422fffb893SStefano Zampini used_adjncy = NULL; 47432fffb893SStefano Zampini } 47442fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4745b96c3477SStefano Zampini } 4746b96c3477SStefano Zampini } 4747d5574798SStefano Zampini 4748d5574798SStefano Zampini /* setup sub_schurs data */ 4749a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4750a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4751a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4752a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 47536816873aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4754a64f4aa4SStefano Zampini } else { 47556816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 47566816873aSStefano 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,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4757a64f4aa4SStefano Zampini } 4758a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4759b96c3477SStefano Zampini 4760b96c3477SStefano Zampini /* free adjacency */ 4761b96c3477SStefano Zampini if (free_used_adj) { 4762b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4763b96c3477SStefano Zampini } 4764b96c3477SStefano Zampini PetscFunctionReturn(0); 4765b96c3477SStefano Zampini } 4766b96c3477SStefano Zampini 4767b96c3477SStefano Zampini #undef __FUNCT__ 4768b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 476908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4770b96c3477SStefano Zampini { 4771b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4772b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4773b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4774b96c3477SStefano Zampini PCBDDCGraph graph; 4775b96c3477SStefano Zampini PetscErrorCode ierr; 4776b96c3477SStefano Zampini 4777b96c3477SStefano Zampini PetscFunctionBegin; 4778b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 477908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 47803301b35fSStefano Zampini IS verticesIS,verticescomm; 47813301b35fSStefano Zampini PetscInt vsize,*idxs; 4782b96c3477SStefano Zampini 4783b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 47843301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 47853301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 47863301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 47873301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 47883301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4789b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47907fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 47913301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 47923301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4793b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4794b96c3477SStefano Zampini /* 4795b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4796b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4797b96c3477SStefano Zampini } 4798b96c3477SStefano Zampini */ 4799b96c3477SStefano Zampini } else { 4800b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4801b96c3477SStefano Zampini } 4802b96c3477SStefano Zampini 4803b96c3477SStefano Zampini /* sub_schurs init */ 4804a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4805a64f4aa4SStefano Zampini 4806b96c3477SStefano Zampini /* free graph struct */ 480708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4808b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4809b96c3477SStefano Zampini } 4810b96c3477SStefano Zampini PetscFunctionReturn(0); 4811b96c3477SStefano Zampini } 4812