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 } 593*3301b35fSStefano 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 } 781*3301b35fSStefano 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 } 791*3301b35fSStefano 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 */ 940*3301b35fSStefano 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); 1062*3301b35fSStefano 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); 11048bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 11058bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 11068bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11078bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11088bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11098bec7fa6SStefano Zampini 111025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 111188ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 111288ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 111388ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 111488ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 111588ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 111688ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 111788ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 111888ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 111988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 112088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1121c0553b1fSStefano Zampini if (unsymmetric_check) { 112288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 112388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 112488ebb749SStefano Zampini } 112588ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 112688ebb749SStefano Zampini } 11278629588bSStefano Zampini /* get back data */ 11288629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 112988ebb749SStefano Zampini PetscFunctionReturn(0); 113088ebb749SStefano Zampini } 113188ebb749SStefano Zampini 113288ebb749SStefano Zampini #undef __FUNCT__ 1133d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1134d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1135d65f70fdSStefano Zampini { 1136d65f70fdSStefano Zampini Mat *work_mat; 1137d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1138d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1139d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1140d65f70fdSStefano Zampini PetscErrorCode ierr; 1141d65f70fdSStefano Zampini 1142d65f70fdSStefano Zampini PetscFunctionBegin; 1143d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1144d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini 1148d65f70fdSStefano Zampini if (!rsorted) { 1149d65f70fdSStefano Zampini const PetscInt *idxs; 1150d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1151d65f70fdSStefano Zampini 1152d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1155d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1156d65f70fdSStefano Zampini } 1157d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1160d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1161d65f70fdSStefano Zampini } 1162d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1163d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1164d65f70fdSStefano Zampini } else { 1165d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini isrow_s = isrow; 1167d65f70fdSStefano Zampini } 1168d65f70fdSStefano Zampini 1169d65f70fdSStefano Zampini if (!csorted) { 1170d65f70fdSStefano Zampini if (isrow == iscol) { 1171d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1172d65f70fdSStefano Zampini iscol_s = isrow_s; 1173d65f70fdSStefano Zampini } else { 1174d65f70fdSStefano Zampini const PetscInt *idxs; 1175d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1176d65f70fdSStefano Zampini 1177d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1180d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1181d65f70fdSStefano Zampini } 1182d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1185d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1186d65f70fdSStefano Zampini } 1187d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1188d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1189d65f70fdSStefano Zampini } 1190d65f70fdSStefano Zampini } else { 1191d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1192d65f70fdSStefano Zampini iscol_s = iscol; 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini 1195d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini 1197d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1198d65f70fdSStefano Zampini Mat new_mat; 1199d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1200d65f70fdSStefano Zampini 1201d65f70fdSStefano Zampini if (!rsorted) { 1202d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1203d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1205d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1206d65f70fdSStefano Zampini } 1207d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini } else { 1210d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini } 1212d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini 1214d65f70fdSStefano Zampini if (!csorted) { 1215d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1216d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1217d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1218d65f70fdSStefano Zampini } else { 1219d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1220d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1222d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1223d65f70fdSStefano Zampini } 1224d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1225d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1226d65f70fdSStefano Zampini } 1227d65f70fdSStefano Zampini } else { 1228d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1229d65f70fdSStefano Zampini } 1230d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini 1232d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1233d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1234d65f70fdSStefano Zampini work_mat[0] = new_mat; 1235d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1236d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1237d65f70fdSStefano Zampini } 1238d65f70fdSStefano Zampini 1239d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1240d65f70fdSStefano Zampini *B = work_mat[0]; 1241d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1242d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1243d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1244d65f70fdSStefano Zampini PetscFunctionReturn(0); 1245d65f70fdSStefano Zampini } 1246d65f70fdSStefano Zampini 1247d65f70fdSStefano Zampini #undef __FUNCT__ 12485e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12495e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1250aa0d41d4SStefano Zampini { 1251aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12525e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1253d65f70fdSStefano Zampini Mat new_mat; 12545e8657edSStefano Zampini IS is_local,is_global; 1255d65f70fdSStefano Zampini PetscInt local_size; 1256d65f70fdSStefano Zampini PetscBool isseqaij; 1257aa0d41d4SStefano Zampini PetscErrorCode ierr; 1258aa0d41d4SStefano Zampini 1259aa0d41d4SStefano Zampini PetscFunctionBegin; 1260aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12615e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12625e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1263906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1264906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1265d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1266906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1267906d46d4SStefano Zampini 1268906d46d4SStefano Zampini /* check */ 1269906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1270906d46d4SStefano Zampini Vec x,x_change; 1271906d46d4SStefano Zampini PetscReal error; 1272906d46d4SStefano Zampini 12735e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1274906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12755e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12765e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12775e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1278d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12795e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12805e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1281906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1282906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1283906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1284906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1285906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1286906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1287906d46d4SStefano Zampini } 1288906d46d4SStefano Zampini 128922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12909b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 129122d5777bSStefano Zampini if (isseqaij) { 1292d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1293aa0d41d4SStefano Zampini } else { 1294aa0d41d4SStefano Zampini Mat work_mat; 1295aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1296d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1297aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1298aa0d41d4SStefano Zampini } 1299*3301b35fSStefano Zampini if (matis->A->symmetric_set) { 1300*3301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1301e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1302*3301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1303e496cd5dSStefano Zampini #endif 1304*3301b35fSStefano Zampini } 130545a1bb75SStefano Zampini /* 130645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1307d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 130845a1bb75SStefano Zampini */ 1309d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1310aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1311aa0d41d4SStefano Zampini } 1312aa0d41d4SStefano Zampini 1313aa0d41d4SStefano Zampini #undef __FUNCT__ 1314a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13158ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1316a64d13efSStefano Zampini { 1317a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1318a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1319d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 132053892102SStefano Zampini PetscInt *idx_R_local=NULL; 13213a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13223a50541eSStefano Zampini PetscInt vbs,bs; 13236816873aSStefano Zampini PetscBT bitmask=NULL; 1324a64d13efSStefano Zampini PetscErrorCode ierr; 1325a64d13efSStefano Zampini 1326a64d13efSStefano Zampini PetscFunctionBegin; 1327b23d619eSStefano Zampini /* 1328b23d619eSStefano Zampini No need to setup local scatters if 1329b23d619eSStefano Zampini - primal space is unchanged 1330b23d619eSStefano Zampini AND 1331b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1332b23d619eSStefano Zampini AND 1333b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1334b23d619eSStefano Zampini */ 1335b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1336f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1337f4ddd8eeSStefano Zampini } 1338f4ddd8eeSStefano Zampini /* destroy old objects */ 1339f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1340f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1341f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1342a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1343b371cd4fSStefano Zampini n_B = pcis->n_B; 1344b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1345b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13466816873aSStefano Zampini 13476816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13486816873aSStefano Zampini 134953892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13506816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 135153892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13524641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13534641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13540e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13554641a718SStefano Zampini } 13563a50541eSStefano Zampini 1357a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13584641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13596816873aSStefano Zampini idx_R_local[n_R++] = i; 1360a64d13efSStefano Zampini } 1361a64d13efSStefano Zampini } 136253892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13636816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13646816873aSStefano Zampini 136553892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 136653892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13676816873aSStefano Zampini } 13683a50541eSStefano Zampini 13693a50541eSStefano Zampini /* Block code */ 13703a50541eSStefano Zampini vbs = 1; 13713a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13723a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13733a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13743a50541eSStefano Zampini PetscInt *vary; 13753a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1376785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13773a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13780e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1379964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13803a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13813a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13823a50541eSStefano Zampini break; 13833a50541eSStefano Zampini } 13843a50541eSStefano Zampini } 13853a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13863a50541eSStefano Zampini vbs = bs; 13873a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13883a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13893a50541eSStefano Zampini } 13903a50541eSStefano Zampini } 13913a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13923a50541eSStefano Zampini } 13933a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 139453892102SStefano Zampini if (sub_schurs->reuse_mumps) { 139553892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 139653892102SStefano Zampini 139753892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 139853892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 139953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140053892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140153892102SStefano Zampini } else { 14023a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 140353892102SStefano Zampini } 1404a64d13efSStefano Zampini 1405a64d13efSStefano Zampini /* print some info if requested */ 1406a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1407a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1408a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14090fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1410a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1411a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14123a50541eSStefano 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); 1413a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1414a64d13efSStefano Zampini } 1415a64d13efSStefano Zampini 1416a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14176816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14186816873aSStefano Zampini IS is_aux1,is_aux2; 14196816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14206816873aSStefano Zampini 14213a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1422854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1423854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1424a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14254641a718SStefano Zampini for (i=0; i<n_D; i++) { 14264641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14274641a718SStefano Zampini } 1428a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1429a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14304641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14314641a718SStefano Zampini aux_array1[j++] = i; 1432a64d13efSStefano Zampini } 1433a64d13efSStefano Zampini } 1434a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1435a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1436a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14374641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14384641a718SStefano Zampini aux_array2[j++] = i; 1439a64d13efSStefano Zampini } 1440a64d13efSStefano Zampini } 1441a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1442a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1443a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1444a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1445a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1446a64d13efSStefano Zampini 14478eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1448785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1449a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14504641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14514641a718SStefano Zampini aux_array1[j++] = i; 1452a64d13efSStefano Zampini } 1453a64d13efSStefano Zampini } 1454a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1455a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1456a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1457a64d13efSStefano Zampini } 14584641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14593a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1460d62866d3SStefano Zampini } else { 146153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14626816873aSStefano Zampini IS tis; 14636816873aSStefano Zampini PetscInt schur_size; 14646816873aSStefano Zampini 146553892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14666816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 146753892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14696816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14706816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14716816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14726816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1473d62866d3SStefano Zampini } 1474d62866d3SStefano Zampini } 1475a64d13efSStefano Zampini PetscFunctionReturn(0); 1476a64d13efSStefano Zampini } 1477a64d13efSStefano Zampini 1478304d26faSStefano Zampini 1479304d26faSStefano Zampini #undef __FUNCT__ 1480304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1481684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1482304d26faSStefano Zampini { 1483304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1484304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1485304d26faSStefano Zampini PC pc_temp; 1486304d26faSStefano Zampini Mat A_RR; 1487f4ddd8eeSStefano Zampini MatReuse reuse; 1488304d26faSStefano Zampini PetscScalar m_one = -1.0; 1489304d26faSStefano Zampini PetscReal value; 1490af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14919577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1492304d26faSStefano Zampini PetscErrorCode ierr; 1493e604994aSStefano Zampini /* prefixes stuff */ 1494312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1495e604994aSStefano Zampini size_t len; 1496304d26faSStefano Zampini 1497304d26faSStefano Zampini PetscFunctionBegin; 1498304d26faSStefano Zampini 1499e604994aSStefano Zampini /* compute prefixes */ 1500e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1501e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1502e604994aSStefano Zampini if (!pcbddc->current_level) { 1503e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1504e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1505e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1506e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1507e604994aSStefano Zampini } else { 1508e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1509312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1510e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1511e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1512312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1513312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 151434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 151534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1516e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1517e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1518e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1520e604994aSStefano Zampini } 1521e604994aSStefano Zampini 1522304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1523684f6988SStefano Zampini if (dirichlet) { 1524d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1525*3301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 1526*3301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1527964fefecSStefano Zampini } 1528ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1529964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1530304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1531304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1532304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1533304d26faSStefano Zampini /* default */ 1534304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1535e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15369577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1537304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15389577ea80SStefano Zampini if (issbaij) { 15399577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15409577ea80SStefano Zampini } else { 1541304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15429577ea80SStefano Zampini } 1543304d26faSStefano Zampini /* Allow user's customization */ 1544304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1545304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1546304d26faSStefano Zampini } 1547d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1548d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1549d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1550d62866d3SStefano Zampini 1551d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1552d5574798SStefano Zampini } 1553304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1554304d26faSStefano Zampini if (!n_D) { 1555304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1556304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1557304d26faSStefano Zampini } 1558304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1559304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1560304d26faSStefano Zampini /* set ksp_D into pcis data */ 1561304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1562304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1563304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1564684f6988SStefano Zampini } 1565304d26faSStefano Zampini 1566304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1567684f6988SStefano Zampini A_RR = 0; 1568684f6988SStefano Zampini if (neumann) { 1569d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1570f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15718ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1572f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1573f4ddd8eeSStefano Zampini PetscInt nn_R; 157481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1575f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1576f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1578f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1579f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1580f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1581f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1582727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1583f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1585f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1586f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1587f4ddd8eeSStefano Zampini } 1588f4ddd8eeSStefano Zampini } 1589f4ddd8eeSStefano Zampini /* last check */ 1590d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1591f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1592f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1593f4ddd8eeSStefano Zampini } 1594f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1595f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1596f4ddd8eeSStefano Zampini } 15976816873aSStefano Zampini /* extract A_RR */ 15986816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 15996816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16006816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 16016816873aSStefano Zampini if (ibs != mbs) { 16026816873aSStefano Zampini Mat newmat; 16036816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 16046816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16056816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 16066816873aSStefano Zampini } else { 16076816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16086816873aSStefano Zampini } 1609*3301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 1610*3301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16116816873aSStefano Zampini } 16126816873aSStefano Zampini } else { 16136816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16146816873aSStefano Zampini 16156816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16166816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16176816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16186816873aSStefano Zampini } 1619f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1620304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1621304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1622304d26faSStefano Zampini /* default */ 1623304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1624e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1625304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16269577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16279577ea80SStefano Zampini if (issbaij) { 16289577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16299577ea80SStefano Zampini } else { 1630304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16319577ea80SStefano Zampini } 1632304d26faSStefano Zampini /* Allow user's customization */ 1633304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1634304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1635304d26faSStefano Zampini } 1636d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1637304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1638304d26faSStefano Zampini if (!n_R) { 1639304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1640304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1641304d26faSStefano Zampini } 1642d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1643d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1644d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1645d62866d3SStefano Zampini 1646d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1647d62866d3SStefano Zampini } 1648304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1649304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1650684f6988SStefano Zampini } 16516816873aSStefano Zampini /* free Neumann problem's matrix */ 16526816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1653304d26faSStefano Zampini 1654304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16550fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1656684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1657684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1658684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1659684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1660684f6988SStefano Zampini } 1661684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16620fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16630fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16640fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16650fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16660fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1667304d26faSStefano Zampini /* need to be adapted? */ 1668b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1669b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1670b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1671304d26faSStefano Zampini /* print info */ 1672304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1673e604994aSStefano 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); 1674304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1675304d26faSStefano Zampini } 1676b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16778ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1678304d26faSStefano Zampini } 1679684f6988SStefano Zampini } 1680684f6988SStefano Zampini if (neumann) { /* Neumann */ 16816816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16820fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1687304d26faSStefano Zampini /* need to be adapted? */ 1688b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1689b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1690304d26faSStefano Zampini /* print info */ 1691304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1692e604994aSStefano 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); 1693304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1694304d26faSStefano Zampini } 1695b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16968ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1697304d26faSStefano Zampini } 16980fccc4e9SStefano Zampini } 1699684f6988SStefano Zampini } 1700304d26faSStefano Zampini PetscFunctionReturn(0); 1701304d26faSStefano Zampini } 1702304d26faSStefano Zampini 1703304d26faSStefano Zampini #undef __FUNCT__ 1704ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 170580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1706674ae819SStefano Zampini { 1707674ae819SStefano Zampini PetscErrorCode ierr; 1708674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1709be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1710674ae819SStefano Zampini 1711674ae819SStefano Zampini PetscFunctionBegin; 1712be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1714be83ff47SStefano Zampini } 171580677318SStefano Zampini if (!pcbddc->switch_static) { 171680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 171780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 171880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 171920c7b377SStefano Zampini } 1720be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172320c7b377SStefano Zampini } else { 1724be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1725be83ff47SStefano Zampini 172653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1728be83ff47SStefano Zampini } 1729be83ff47SStefano Zampini } else { 173080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 173680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 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); 173920c7b377SStefano Zampini } 1740674ae819SStefano Zampini } 1741be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174280677318SStefano Zampini if (applytranspose) { 174380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174480677318SStefano Zampini } else { 174580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174680677318SStefano Zampini } 174753892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1748be83ff47SStefano Zampini } else { 1749be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1750be83ff47SStefano Zampini 1751be83ff47SStefano Zampini if (applytranspose) { 175253892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1753be83ff47SStefano Zampini } else { 175453892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1755be83ff47SStefano Zampini } 175653892102SStefano Zampini #endif 1757be83ff47SStefano Zampini } 175880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 175980677318SStefano Zampini if (!pcbddc->switch_static) { 1760be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1763be83ff47SStefano Zampini } else { 1764be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1765be83ff47SStefano Zampini 176653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 176753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1768be83ff47SStefano Zampini } 176980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 177280677318SStefano Zampini } 177380677318SStefano Zampini } else { 177480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 178180677318SStefano Zampini } 178280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178680677318SStefano Zampini } 1787674ae819SStefano Zampini PetscFunctionReturn(0); 1788674ae819SStefano Zampini } 1789674ae819SStefano Zampini 1790dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1791674ae819SStefano Zampini #undef __FUNCT__ 1792674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1793dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1794674ae819SStefano Zampini { 1795674ae819SStefano Zampini PetscErrorCode ierr; 1796674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1797674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1798674ae819SStefano Zampini const PetscScalar zero = 0.0; 1799674ae819SStefano Zampini 1800674ae819SStefano Zampini PetscFunctionBegin; 1801dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1802dc359a40SStefano Zampini if (applytranspose) { 1803674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18048eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1805dc359a40SStefano Zampini } else { 1806674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1807674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 180815aaf578SStefano Zampini } 180912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 181012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 181112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181312edc857SStefano Zampini 18149f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 181512edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 181612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1817964fefecSStefano Zampini Vec rhs,sol; 1818964fefecSStefano Zampini 1819964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1820964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 182112edc857SStefano Zampini if (applytranspose) { 1822964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182312edc857SStefano Zampini } else { 1824964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182512edc857SStefano Zampini } 182612edc857SStefano Zampini } 1827674ae819SStefano Zampini 1828674ae819SStefano Zampini /* Local solution on R nodes */ 182980677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 183080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18319f00e9b4SStefano Zampini } 1832674ae819SStefano Zampini 18339f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18349f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 183512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1836674ae819SStefano Zampini 1837674ae819SStefano Zampini /* Sum contributions from two levels */ 1838dc359a40SStefano Zampini if (applytranspose) { 1839dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1840dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1841dc359a40SStefano Zampini } else { 1842674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18438eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1844dc359a40SStefano Zampini } 1845674ae819SStefano Zampini PetscFunctionReturn(0); 1846674ae819SStefano Zampini } 1847674ae819SStefano Zampini 1848674ae819SStefano Zampini #undef __FUNCT__ 1849674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 185012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1851674ae819SStefano Zampini { 1852674ae819SStefano Zampini PetscErrorCode ierr; 1853674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 185458da7f69SStefano Zampini PetscScalar *array; 185512edc857SStefano Zampini Vec from,to; 1856674ae819SStefano Zampini 1857674ae819SStefano Zampini PetscFunctionBegin; 185812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 185912edc857SStefano Zampini from = pcbddc->coarse_vec; 186012edc857SStefano Zampini to = pcbddc->vec1_P; 186112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 186212edc857SStefano Zampini Vec tvec; 186358da7f69SStefano Zampini 186458da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186558da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 186612edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186758da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 186858da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 186958da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 187012edc857SStefano Zampini } 187112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187212edc857SStefano Zampini from = pcbddc->vec1_P; 187312edc857SStefano Zampini to = pcbddc->coarse_vec; 187412edc857SStefano Zampini } 187512edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1876674ae819SStefano Zampini PetscFunctionReturn(0); 1877674ae819SStefano Zampini } 1878674ae819SStefano Zampini 1879674ae819SStefano Zampini #undef __FUNCT__ 1880674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 188112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1882674ae819SStefano Zampini { 1883674ae819SStefano Zampini PetscErrorCode ierr; 1884674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 188558da7f69SStefano Zampini PetscScalar *array; 188612edc857SStefano Zampini Vec from,to; 1887674ae819SStefano Zampini 1888674ae819SStefano Zampini PetscFunctionBegin; 188912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 189012edc857SStefano Zampini from = pcbddc->coarse_vec; 189112edc857SStefano Zampini to = pcbddc->vec1_P; 189212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189312edc857SStefano Zampini from = pcbddc->vec1_P; 189412edc857SStefano Zampini to = pcbddc->coarse_vec; 189512edc857SStefano Zampini } 189612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 189712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 189812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 189912edc857SStefano Zampini Vec tvec; 190058da7f69SStefano Zampini 190112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 190258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 190358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 190458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 190558da7f69SStefano Zampini } 190658da7f69SStefano Zampini } else { 190758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 190858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 190912edc857SStefano Zampini } 191012edc857SStefano Zampini } 1911674ae819SStefano Zampini PetscFunctionReturn(0); 1912674ae819SStefano Zampini } 1913674ae819SStefano Zampini 1914984c4197SStefano Zampini /* uncomment for testing purposes */ 1915984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1916674ae819SStefano Zampini #undef __FUNCT__ 1917674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1918674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1919674ae819SStefano Zampini { 1920674ae819SStefano Zampini PetscErrorCode ierr; 1921674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1922674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1923674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1924984c4197SStefano Zampini /* one and zero */ 1925984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1926984c4197SStefano Zampini /* space to store constraints and their local indices */ 19279162d606SStefano Zampini PetscScalar *constraints_data; 19289162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19299162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19309162d606SStefano Zampini PetscInt *constraints_n; 1931984c4197SStefano Zampini /* iterators */ 1932b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1933984c4197SStefano Zampini /* BLAS integers */ 1934e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1935e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1936c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1937727cdba6SStefano Zampini /* reuse */ 19380e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19390e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1940984c4197SStefano Zampini /* change of basis */ 1941b3d85658SStefano Zampini PetscBool qr_needed; 19429162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1943984c4197SStefano Zampini /* auxiliary stuff */ 194464efe560SStefano Zampini PetscInt *nnz,*is_indices; 19458a0068c3SStefano Zampini PetscInt ncc; 1946984c4197SStefano Zampini /* some quantities */ 194745a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1948a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1949984c4197SStefano Zampini 1950674ae819SStefano Zampini PetscFunctionBegin; 19518e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19528e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19538e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1954088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1955088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19560e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19570e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19580e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19590e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19600e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1961088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1962cf5a6209SStefano Zampini 1963cf5a6209SStefano Zampini /* print some info */ 1964cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1965cf5a6209SStefano Zampini IS vertices; 1966cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1967cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1968cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1969cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1970cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1971cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1972cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1973fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1974fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1975cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1976cf5a6209SStefano Zampini } 1977cf5a6209SStefano Zampini 1978cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19799162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1980cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1981cf5a6209SStefano Zampini const Vec *nearnullvecs; 1982cf5a6209SStefano Zampini Vec *localnearnullsp; 1983cf5a6209SStefano Zampini PetscScalar *array; 1984cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1985cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1986cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1987b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1988cf5a6209SStefano Zampini PetscScalar *work; 1989cf5a6209SStefano Zampini PetscReal *singular_vals; 1990cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1991cf5a6209SStefano Zampini PetscReal *rwork; 1992cf5a6209SStefano Zampini #endif 1993cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1994cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1995cf5a6209SStefano Zampini #else 1996964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1997964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1998cf5a6209SStefano Zampini #endif 1999cf5a6209SStefano Zampini 2000674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2001d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2002d06fc5fdSStefano Zampini /* free unneeded index sets */ 2003d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2004d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2005674ae819SStefano Zampini } 2006d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2007d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2008d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2009d06fc5fdSStefano Zampini } 2010d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2011d06fc5fdSStefano Zampini n_ISForEdges = 0; 2012d06fc5fdSStefano Zampini } 2013d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2014d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2015d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2016d06fc5fdSStefano Zampini } 2017d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2018d06fc5fdSStefano Zampini n_ISForFaces = 0; 2019d06fc5fdSStefano Zampini } 202070022509SStefano Zampini 202170022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 202270022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 202370022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 202470022509SStefano Zampini if (pcbddc->NullSpace) { 202570022509SStefano Zampini PetscBool tbool[2],gbool[2]; 202670022509SStefano Zampini 202770022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2028b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2029d06fc5fdSStefano Zampini if (!ISForEdges) { 2030d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2031d06fc5fdSStefano Zampini } 2032b8ffe317SStefano Zampini } 2033d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2034d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2035d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2036d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2037d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 203898a51de6SStefano Zampini } 203970022509SStefano Zampini #endif 204008122e43SStefano Zampini 2041674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2042674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2043674ae819SStefano Zampini if (nearnullsp) { 2044674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2045f4ddd8eeSStefano Zampini /* remove any stored info */ 2046f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2047f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2049f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2050f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2051473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2052f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2053f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2054f4ddd8eeSStefano Zampini } 2055984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2056984c4197SStefano Zampini nnsp_size = 0; 2057674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2058674ae819SStefano Zampini } 2059984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2060984c4197SStefano Zampini max_constraints = nnsp_size; 2061984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2062984c4197SStefano Zampini 2063674ae819SStefano Zampini /* 2064674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20659162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20669162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20679162d606SStefano Zampini There can be multiple constraints per connected component 2068674ae819SStefano Zampini */ 2069674ae819SStefano Zampini n_vertices = 0; 2070674ae819SStefano Zampini if (ISForVertices) { 2071674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2072674ae819SStefano Zampini } 20739162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20749162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20759162d606SStefano Zampini 20769162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20779162d606SStefano Zampini total_counts *= max_constraints; 2078674ae819SStefano Zampini total_counts += n_vertices; 20794641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20809162d606SStefano Zampini 2081674ae819SStefano Zampini total_counts = 0; 2082674ae819SStefano Zampini max_size_of_constraint = 0; 2083674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20849162d606SStefano Zampini IS used_is; 2085674ae819SStefano Zampini if (i<n_ISForEdges) { 20869162d606SStefano Zampini used_is = ISForEdges[i]; 2087674ae819SStefano Zampini } else { 20889162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2089674ae819SStefano Zampini } 20909162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2091674ae819SStefano Zampini total_counts += j; 2092674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2093674ae819SStefano Zampini } 20949162d606SStefano 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); 20959162d606SStefano Zampini 2096984c4197SStefano Zampini /* get local part of global near null space vectors */ 2097785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2098984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2099984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2100984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2101984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2102984c4197SStefano Zampini } 2103674ae819SStefano Zampini 2104242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2105242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2106a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2107242a89d7SStefano Zampini 2108984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2109a773dcb8SStefano Zampini if (!skip_lapack) { 2110674ae819SStefano Zampini PetscScalar temp_work; 2111911cabfeSStefano Zampini 2112674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2113984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2114785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2115785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2116785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2117674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2118785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2119674ae819SStefano Zampini #endif 2120674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2121c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2122c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2123674ae819SStefano Zampini lwork = -1; 2124674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2125674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2126c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2127674ae819SStefano Zampini #else 2128c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2129674ae819SStefano Zampini #endif 2130674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2131984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2132674ae819SStefano Zampini #else /* on missing GESVD */ 2133674ae819SStefano Zampini /* SVD */ 2134674ae819SStefano Zampini PetscInt max_n,min_n; 2135674ae819SStefano Zampini max_n = max_size_of_constraint; 2136984c4197SStefano Zampini min_n = max_constraints; 2137984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2138674ae819SStefano Zampini min_n = max_size_of_constraint; 2139984c4197SStefano Zampini max_n = max_constraints; 2140674ae819SStefano Zampini } 2141785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2142674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2143785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2144674ae819SStefano Zampini #endif 2145674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2146674ae819SStefano Zampini lwork = -1; 2147e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2148e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2149b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2150674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2151674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21529162d606SStefano 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)); 2153674ae819SStefano Zampini #else 21549162d606SStefano 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)); 2155674ae819SStefano Zampini #endif 2156674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2157984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2158984c4197SStefano Zampini #endif /* on missing GESVD */ 2159674ae819SStefano Zampini /* Allocate optimal workspace */ 2160674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2161854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2162674ae819SStefano Zampini } 2163674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2164674ae819SStefano Zampini total_counts = 0; 21659162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21669162d606SStefano Zampini constraints_data_ptr[0] = 0; 2167674ae819SStefano Zampini /* vertices */ 21689162d606SStefano Zampini if (n_vertices) { 2169674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21709162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21719162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2172674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21739162d606SStefano Zampini constraints_n[total_counts] = 1; 21749162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21759162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21769162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2177674ae819SStefano Zampini total_counts++; 2178674ae819SStefano Zampini } 2179674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2180984c4197SStefano Zampini PetscBool used_vertex; 2181674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2182674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2183674ae819SStefano Zampini k = 0; 2184674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2185984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2186984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21879162d606SStefano Zampini constraints_n[total_counts] = 1; 21889162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21899162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21909162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21919162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2192674ae819SStefano Zampini total_counts++; 2193674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2194674ae819SStefano Zampini } 2195984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2196674ae819SStefano Zampini k++; 2197674ae819SStefano Zampini } 2198674ae819SStefano Zampini } 2199674ae819SStefano Zampini } 2200674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2201674ae819SStefano Zampini n_vertices = total_counts; 2202674ae819SStefano Zampini } 2203984c4197SStefano Zampini 2204674ae819SStefano Zampini /* edges and faces */ 22059162d606SStefano Zampini total_counts_cc = total_counts; 2206911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22079162d606SStefano Zampini IS used_is; 22089162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22099162d606SStefano Zampini 2210911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22119162d606SStefano Zampini used_is = ISForEdges[ncc]; 2212984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2213674ae819SStefano Zampini } else { 22149162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2215984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2216674ae819SStefano Zampini } 2217674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22189162d606SStefano Zampini 22199162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22209162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2221984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2222984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2223674ae819SStefano Zampini if (nnsp_has_cnst) { 22245b08dc53SStefano Zampini PetscScalar quad_value; 22259162d606SStefano Zampini 22269162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22279162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22289162d606SStefano Zampini 2229a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2230674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2231a773dcb8SStefano Zampini } else { 2232a773dcb8SStefano Zampini quad_value = 1.0; 2233a773dcb8SStefano Zampini } 2234674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22359162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2236674ae819SStefano Zampini } 22379162d606SStefano Zampini temp_constraints++; 2238674ae819SStefano Zampini total_counts++; 2239674ae819SStefano Zampini } 2240674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2241984c4197SStefano Zampini PetscReal real_value; 22429162d606SStefano Zampini PetscScalar *ptr_to_data; 22439162d606SStefano Zampini 2244984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22459162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2246674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22479162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2248674ae819SStefano Zampini } 2249984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2250984c4197SStefano Zampini /* check if array is null on the connected component */ 2251e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22529162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22535b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2254674ae819SStefano Zampini temp_constraints++; 2255674ae819SStefano Zampini total_counts++; 22569162d606SStefano Zampini if (!idxs_copied) { 22579162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22589162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2259674ae819SStefano Zampini } 2260674ae819SStefano Zampini } 22619162d606SStefano Zampini } 22629162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 226345a1bb75SStefano Zampini valid_constraints = temp_constraints; 2264eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2265a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22669162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22679162d606SStefano Zampini 22689162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2269a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22709162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2271a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22729162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2273a773dcb8SStefano Zampini } else { /* perform SVD */ 2274984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22759162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2276674ae819SStefano Zampini 2277674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2278984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2279984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2280984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2281984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2282984c4197SStefano Zampini from that computed using LAPACKgesvd 2283984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2284984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2285984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2286674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2287e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2288984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2289674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2290674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22919162d606SStefano 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)); 2292674ae819SStefano Zampini } 2293674ae819SStefano Zampini } 2294e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2295e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2296e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2297674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2298c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2299674ae819SStefano Zampini #else 2300c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2301674ae819SStefano Zampini #endif 2302674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2303984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2304984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2305674ae819SStefano Zampini j = 0; 2306984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2307674ae819SStefano Zampini total_counts = total_counts-j; 230845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2309e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2310c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2311c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2312c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2313c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2314c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2315c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2316674ae819SStefano Zampini if (j<temp_constraints) { 2317984c4197SStefano Zampini PetscInt ii; 2318984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2319674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23209162d606SStefano 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)); 2321674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2322984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2323674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23249162d606SStefano 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]; 2325674ae819SStefano Zampini } 2326674ae819SStefano Zampini } 2327674ae819SStefano Zampini } 2328674ae819SStefano Zampini #else /* on missing GESVD */ 2329e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2330e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2331b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2332674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2333674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23349162d606SStefano 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)); 2335674ae819SStefano Zampini #else 23369162d606SStefano 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)); 2337674ae819SStefano Zampini #endif 2338984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2339674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2340984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2341e310c8b4SStefano Zampini k = temp_constraints; 2342e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2343674ae819SStefano Zampini j = 0; 2344e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 234545a1bb75SStefano Zampini valid_constraints = k-j; 2346911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2347984c4197SStefano Zampini #endif /* on missing GESVD */ 2348674ae819SStefano Zampini } 2349a773dcb8SStefano Zampini } 23509162d606SStefano Zampini /* update pointers information */ 23519162d606SStefano Zampini if (valid_constraints) { 23529162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23539162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23549162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23559162d606SStefano Zampini /* set change_of_basis flag */ 235645a1bb75SStefano Zampini if (boolforchange) { 2357b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23589162d606SStefano Zampini } 2359b3d85658SStefano Zampini total_counts_cc++; 236045a1bb75SStefano Zampini } 2361674ae819SStefano Zampini } 2362984c4197SStefano Zampini /* free workspace */ 23638f1c130eSStefano Zampini if (!skip_lapack) { 2364984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2365984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2366984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2367984c4197SStefano Zampini #endif 2368984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2369984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2370984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2371984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2372984c4197SStefano Zampini #endif 2373984c4197SStefano Zampini } 2374984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2375984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2376984c4197SStefano Zampini } 2377984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2378cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2379cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2380cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2381cf5a6209SStefano Zampini } 2382cf5a6209SStefano Zampini if (n_ISForFaces) { 2383cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2384cf5a6209SStefano Zampini } 2385cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2386cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2387cf5a6209SStefano Zampini } 2388cf5a6209SStefano Zampini if (n_ISForEdges) { 2389cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2390cf5a6209SStefano Zampini } 2391cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 239208122e43SStefano Zampini } else { 239308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 239408122e43SStefano Zampini 239508122e43SStefano Zampini total_counts = 0; 239608122e43SStefano Zampini n_vertices = 0; 2397d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2398d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 239908122e43SStefano Zampini } 240008122e43SStefano Zampini max_constraints = 0; 24019162d606SStefano Zampini total_counts_cc = 0; 240208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 240308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24049162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 240508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 240608122e43SStefano Zampini } 24079162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24089162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24099162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24109162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 241174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24129162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24139162d606SStefano Zampini total_counts_cc = 0; 24149162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24159162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24169162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 241708122e43SStefano Zampini } 241808122e43SStefano Zampini } 24199162d606SStefano Zampini #if 0 24209162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24219162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24229162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24239162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24249162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24259162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24269162d606SStefano Zampini } 24279162d606SStefano Zampini printf("\n"); 24289162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24299162d606SStefano Zampini } 24301b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24318bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24321b968477SStefano Zampini } 24331b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24348bec7fa6SStefano 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]); 24351b968477SStefano Zampini } 243608122e43SStefano Zampini #endif 243708122e43SStefano Zampini 24388bec7fa6SStefano Zampini max_size_of_constraint = 0; 24399162d606SStefano 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]); 24409162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 244108122e43SStefano Zampini /* Change of basis */ 2442b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 244308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 244408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 244508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2446b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 244708122e43SStefano Zampini } 244808122e43SStefano Zampini } 244908122e43SStefano Zampini } 245008122e43SStefano Zampini } 2451b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24520e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 245308122e43SStefano Zampini 24549162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24559162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24569162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24579162d606SStefano 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); 245808122e43SStefano Zampini } 2459984c4197SStefano Zampini 2460674ae819SStefano Zampini /* Create constraint matrix */ 2461674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 246216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2463984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2464984c4197SStefano Zampini 2465984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 246674d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 246774d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 246874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2469984c4197SStefano Zampini total_primal_vertices=0; 2470b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24719162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2473984c4197SStefano Zampini if (size_of_constraint == 1) { 24749162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2475b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 247664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24779162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24789162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24799162d606SStefano Zampini } 2480b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 248174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 248274d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 248374d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 248474d5cdf7SStefano Zampini } 2485b3d85658SStefano Zampini } else { 2486b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2487984c4197SStefano Zampini } 2488984c4197SStefano Zampini } 2489b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2490b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 249170022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 249270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2493b3d85658SStefano Zampini 2494b3d85658SStefano 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); 24950e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24960e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 249770022509SStefano Zampini 2498984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 249974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2500785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2501984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 250274d5cdf7SStefano Zampini 2503984c4197SStefano Zampini j = total_primal_vertices; 250474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2505b3d85658SStefano Zampini cum = total_primal_vertices; 25069162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25074641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2508b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2509b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2510b3d85658SStefano Zampini cum++; 25119162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 251274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 251374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 251474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 251574d5cdf7SStefano Zampini } 25169162d606SStefano Zampini j += constraints_n[i]; 2517674ae819SStefano Zampini } 2518674ae819SStefano Zampini } 2519674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2520674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2521088faed8SStefano Zampini 2522674ae819SStefano Zampini /* set values in constraint matrix */ 2523984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25240e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2525674ae819SStefano Zampini } 2526984c4197SStefano Zampini total_counts = total_primal_vertices; 25279162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25284641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25299162d606SStefano Zampini PetscInt *cols; 25309162d606SStefano Zampini 25319162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25329162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25339162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25349162d606SStefano Zampini PetscInt row = total_counts+k; 25359162d606SStefano Zampini PetscScalar *vals; 25369162d606SStefano Zampini 25379162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25389162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25399162d606SStefano Zampini } 25409162d606SStefano Zampini total_counts += constraints_n[i]; 2541674ae819SStefano Zampini } 2542674ae819SStefano Zampini } 2543674ae819SStefano Zampini /* assembling */ 2544674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2545674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2546088faed8SStefano Zampini 2547984c4197SStefano Zampini /* 254845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2549984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2550984c4197SStefano Zampini */ 2551674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2552674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2553026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2554984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2555984c4197SStefano Zampini /* working stuff for GEQRF */ 255681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2557984c4197SStefano Zampini PetscBLASInt lqr_work; 2558984c4197SStefano Zampini /* working stuff for UNGQR */ 2559984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2560984c4197SStefano Zampini PetscBLASInt lgqr_work; 2561984c4197SStefano Zampini /* working stuff for TRTRS */ 2562984c4197SStefano Zampini PetscScalar *trs_rhs; 25633f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2564984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2565984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2566984c4197SStefano Zampini PetscScalar *start_vals; 2567984c4197SStefano Zampini /* working stuff for values insertion */ 25684641a718SStefano Zampini PetscBT is_primal; 256964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2570906d46d4SStefano Zampini /* matrix sizes */ 2571906d46d4SStefano Zampini PetscInt global_size,local_size; 2572906d46d4SStefano Zampini /* temporary change of basis */ 2573906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2574cf5a6209SStefano Zampini /* extra space for debugging */ 2575cf5a6209SStefano Zampini PetscScalar *dbg_work; 2576984c4197SStefano Zampini 2577906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2578906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 257916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2580bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2581906d46d4SStefano Zampini /* nonzeros for local mat */ 2582bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2583bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25849162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2585a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25869162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2587a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25889162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2589a717540cSStefano Zampini } else { 25909162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25919162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2592a717540cSStefano Zampini } 2593a717540cSStefano Zampini } 2594a717540cSStefano Zampini } 2595906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2596bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2597a717540cSStefano Zampini /* Set initial identity in the matrix */ 2598bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2599906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2600a717540cSStefano Zampini } 2601a717540cSStefano Zampini 2602a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2603a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2604a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2605a717540cSStefano Zampini } 2606a717540cSStefano Zampini 2607a717540cSStefano Zampini 2608a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2609a717540cSStefano Zampini /* 2610a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2611a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2612a717540cSStefano Zampini 2613a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2614a717540cSStefano Zampini 2615a6b551f4SStefano 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) 2616a6b551f4SStefano Zampini 2617a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2618a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2619a717540cSStefano Zampini | ... | 2620a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2621a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2622a717540cSStefano Zampini 2623a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2624a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2625a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2626a6b551f4SStefano Zampini 2627a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2628a717540cSStefano Zampini */ 2629a717540cSStefano Zampini if (qr_needed) { 2630984c4197SStefano Zampini /* space to store Q */ 2631854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2632984c4197SStefano Zampini /* first we issue queries for optimal work */ 26333f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26343f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2636984c4197SStefano Zampini lqr_work = -1; 26373f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2638984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2639984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2640785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2641984c4197SStefano Zampini lgqr_work = -1; 26423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26463f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26473f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2648984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2649984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2650785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2651984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2652785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2653984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2654785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2655a717540cSStefano Zampini /* allocating workspace for check */ 2656a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2657cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2658a717540cSStefano Zampini } 2659a717540cSStefano Zampini } 2660984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26614641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2662473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26630e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 266439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 266539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26664641a718SStefano Zampini } 266739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 266839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 266939e2fb2aSStefano Zampini } 267039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2671984c4197SStefano Zampini 2672a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26739162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26749162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26754641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2676984c4197SStefano Zampini /* get constraint info */ 26779162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2678984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2679984c4197SStefano Zampini 2680984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26819162d606SStefano 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); 2682674ae819SStefano Zampini } 2683984c4197SStefano Zampini 2684fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2685a717540cSStefano Zampini 2686a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2687a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26889162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2689a717540cSStefano Zampini } 2690984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26919162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2692984c4197SStefano Zampini 2693984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2697674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26983f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2699984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2700674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2701984c4197SStefano Zampini 2702984c4197SStefano Zampini /* explictly compute R^-T */ 2703984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2704984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27053f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27063f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27073f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27083f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2709984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27103f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2711984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2712984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2713984c4197SStefano Zampini 2714a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27173f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27183f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2719984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2721984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2722984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2723984c4197SStefano Zampini 2724984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2725984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2726984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27273f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27313f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2733984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27349162d606SStefano 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)); 2735984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27369162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2737984c4197SStefano Zampini 2738984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27399162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2740984c4197SStefano Zampini /* insert cols for primal dofs */ 2741984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2742984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27439162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2744906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2745984c4197SStefano Zampini } 2746984c4197SStefano Zampini /* insert cols for dual dofs */ 2747984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27489162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2749984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27509162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2751906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2752984c4197SStefano Zampini j++; 2753674ae819SStefano Zampini } 2754674ae819SStefano Zampini } 2755984c4197SStefano Zampini 2756984c4197SStefano Zampini /* check change of basis */ 2757984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2758984c4197SStefano Zampini PetscInt ii,jj; 2759984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2760c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2761c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2762c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2763c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2764c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2765c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2766984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2767cf5a6209SStefano 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)); 2768984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2769984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2770984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2771cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2772cf5a6209SStefano 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; 2773674ae819SStefano Zampini } 2774674ae819SStefano Zampini } 2775984c4197SStefano Zampini if (!valid_qr) { 277622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2777984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2778984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2779cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2780cf5a6209SStefano 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])); 2781674ae819SStefano Zampini } 2782cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2783cf5a6209SStefano 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])); 2784984c4197SStefano Zampini } 2785984c4197SStefano Zampini } 2786984c4197SStefano Zampini } 2787674ae819SStefano Zampini } else { 278822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2789674ae819SStefano Zampini } 2790674ae819SStefano Zampini } 2791a717540cSStefano Zampini } else { /* simple transformation block */ 2792a717540cSStefano Zampini PetscInt row,col; 2793a6b551f4SStefano Zampini PetscScalar val,norm; 2794a6b551f4SStefano Zampini 2795a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27969162d606SStefano 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)); 2797a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27989162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27999162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2800bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28019162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2802906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28039162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2804a717540cSStefano Zampini } else { 2805a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28069162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2807a717540cSStefano Zampini if (row != col) { 28089162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2809a717540cSStefano Zampini } else { 28109162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2811a717540cSStefano Zampini } 2812906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2813a717540cSStefano Zampini } 2814a717540cSStefano Zampini } 2815a717540cSStefano Zampini } 281698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 281722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2818a717540cSStefano Zampini } 2819674ae819SStefano Zampini } 2820984c4197SStefano Zampini } else { 2821984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28229162d606SStefano 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); 2823674ae819SStefano Zampini } 2824674ae819SStefano Zampini } 2825674ae819SStefano Zampini } 2826a717540cSStefano Zampini 2827a717540cSStefano Zampini /* free workspace */ 2828a717540cSStefano Zampini if (qr_needed) { 2829984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2830cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2831984c4197SStefano Zampini } 2832984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2833984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2834984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2835984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2836984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2837674ae819SStefano Zampini } 2838a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2839906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2840906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2841906d46d4SStefano Zampini 2842906d46d4SStefano Zampini /* assembling of global change of variable */ 2843bbb9e6c6SStefano Zampini { 2844bbb9e6c6SStefano Zampini Mat tmat; 284516f15bc4SStefano Zampini PetscInt bs; 284616f15bc4SStefano Zampini 2847906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2848906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2850bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2851bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2852bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 285316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 285416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2855906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2860bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2861bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2862bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2865906d46d4SStefano Zampini } 2866906d46d4SStefano Zampini /* check */ 2867906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2868906d46d4SStefano Zampini PetscReal error; 2869906d46d4SStefano Zampini Vec x,x_change; 2870906d46d4SStefano Zampini 2871906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2874906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2875bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2879bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2880906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2882906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2883906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2884bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2885906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2886906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2887906d46d4SStefano Zampini } 2888b96c3477SStefano Zampini 2889b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2890b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2891b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2892b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2893ac632422SStefano Zampini Mat S_new,tmat; 2894bbb9e6c6SStefano Zampini IS is_all_N; 2895bbb9e6c6SStefano Zampini 2896bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28976816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2898bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2899ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2900b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2901ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2902ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2903ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2904ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2905ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2906b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2907ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2908ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2909ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2910ac632422SStefano Zampini } 2911b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2912b96c3477SStefano Zampini } 2913b96c3477SStefano Zampini } 2914906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2915906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2916b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2917b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2918b9b85e73SStefano Zampini } 2919906d46d4SStefano Zampini 2920906d46d4SStefano Zampini /* set up change of basis context */ 2921906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2922906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2923906d46d4SStefano Zampini 2924906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2925906d46d4SStefano Zampini PetscInt global_size,local_size; 2926906d46d4SStefano Zampini 2927906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2928906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2929906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2930906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2931906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2932906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2933906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2934906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2935906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2936906d46d4SStefano Zampini } else { 2937906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2938906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2939906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2940906d46d4SStefano Zampini } 2941906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2942906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2943906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2944906d46d4SStefano Zampini } else { 2945906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2946906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2947906d46d4SStefano Zampini } 2948906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2951906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2952b9b85e73SStefano Zampini } 2953a717540cSStefano Zampini 2954727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2955727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2956727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2957aff50787SStefano 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); 29580e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29590e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2960aff50787SStefano 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); 2961c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2962727cdba6SStefano Zampini } 29630e6343abSStefano Zampini } 29640e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2965727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2966727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2967727cdba6SStefano Zampini 2968a717540cSStefano Zampini /* flush dbg viewer */ 2969b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2970b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2971b8ffe317SStefano Zampini } 2972a717540cSStefano Zampini 2973e310c8b4SStefano Zampini /* free workspace */ 2974a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29754641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 297608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29779162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29789162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 297908122e43SStefano Zampini } else { 29809162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29819162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29829162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 298308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 298408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29859162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29869162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 298708122e43SStefano Zampini } 2988674ae819SStefano Zampini PetscFunctionReturn(0); 2989674ae819SStefano Zampini } 2990674ae819SStefano Zampini 2991674ae819SStefano Zampini #undef __FUNCT__ 2992674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2993674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2994674ae819SStefano Zampini { 2995674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2996674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2997674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29987fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2999674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3000674ae819SStefano Zampini 3001674ae819SStefano Zampini PetscFunctionBegin; 30028e61c736SStefano Zampini /* Reset previously computed graph */ 30038e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3004674ae819SStefano Zampini /* Init local Graph struct */ 30057fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30067fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3007674ae819SStefano Zampini 3008575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3009575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3010575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3011575ad6abSStefano Zampini } 30129577ea80SStefano Zampini 3013674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30144d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30154d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30164d379d7bSStefano Zampini PetscInt nvtxs; 3017e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3018674ae819SStefano Zampini 30194d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30202fffb893SStefano Zampini 30212fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30222fffb893SStefano Zampini if (flg_row) { 30234d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3024b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30252fffb893SStefano Zampini } 30262fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30279b28b941SStefano 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 */ 30284d379d7bSStefano Zampini IS is_dummy; 30294d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30304d379d7bSStefano Zampini PetscInt j,sum; 30314d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30324d379d7bSStefano Zampini const PetscInt *idxs; 30334d379d7bSStefano Zampini PCBDDCGraph graph; 30344d379d7bSStefano Zampini PetscBT is_on_boundary; 30354d379d7bSStefano Zampini 30364d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30374d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30384d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30394d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30407fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30414d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3042e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3043e496cd5dSStefano Zampini if (flg_row) { 30444d379d7bSStefano Zampini graph->xadj = xadj; 30454d379d7bSStefano Zampini graph->adjncy = adjncy; 3046e496cd5dSStefano Zampini } 30474d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30484d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3049e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30504d379d7bSStefano Zampini 30514d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30529b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30534d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30544d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30554d379d7bSStefano Zampini } 30564d379d7bSStefano Zampini } 30574d379d7bSStefano Zampini 3058e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30594d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30604d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30614d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30624d379d7bSStefano Zampini } 30634d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30644d379d7bSStefano Zampini 3065e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30664d379d7bSStefano Zampini sum = 0; 30674d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30684d379d7bSStefano Zampini PetscInt sizecc = 0; 30694d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30704d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30714d379d7bSStefano Zampini sizecc++; 30724d379d7bSStefano Zampini } 30734d379d7bSStefano Zampini } 30744d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30754d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30764d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30774d379d7bSStefano Zampini } 30784d379d7bSStefano Zampini } 30794d379d7bSStefano Zampini sum += sizecc*sizecc; 30804d379d7bSStefano Zampini } 30814d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30824d379d7bSStefano Zampini sum = 0; 3083e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30844d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30854d379d7bSStefano Zampini cxadj[i] = sum; 30864d379d7bSStefano Zampini sum += temp; 30874d379d7bSStefano Zampini } 3088e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30894d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30904d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30914d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30924d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30934d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30944d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30954d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30964d379d7bSStefano Zampini sizecc++; 30974d379d7bSStefano Zampini } 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini } 31004d379d7bSStefano Zampini } 31014d379d7bSStefano Zampini } 31029b28b941SStefano Zampini if (sum) { 3103e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31044d379d7bSStefano Zampini } else { 31054d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31064d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31074d379d7bSStefano Zampini } 31084d379d7bSStefano Zampini graph->xadj = 0; 31094d379d7bSStefano Zampini graph->adjncy = 0; 31104d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31114d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31124d379d7bSStefano Zampini } 3113674ae819SStefano Zampini } 31149b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31159b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31169b28b941SStefano Zampini } 3117674ae819SStefano Zampini 311863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3119674ae819SStefano Zampini vertex_size = 1; 312063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 312163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 312295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3124a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 312563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3126674ae819SStefano Zampini } 312763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 312863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 312963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3130674ae819SStefano Zampini } 313163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3132674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 313363602bcaSStefano Zampini } else { 313463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 313563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3136854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 313863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313963602bcaSStefano Zampini } 314063602bcaSStefano Zampini } 3141674ae819SStefano Zampini } 3142674ae819SStefano Zampini 3143674ae819SStefano Zampini /* Setup of Graph */ 3144785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3145a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3146785d1243SStefano Zampini } 3147785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3148a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3149785d1243SStefano Zampini } 3150302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3151674ae819SStefano Zampini 3152674ae819SStefano Zampini /* Graph's connected components analysis */ 3153674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3154674ae819SStefano Zampini 3155674ae819SStefano Zampini /* print some info to stdout */ 3156674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3157302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3158674ae819SStefano Zampini } 3159fb180af4SStefano Zampini 3160fb180af4SStefano Zampini /* mark topography has done */ 3161fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3162674ae819SStefano Zampini PetscFunctionReturn(0); 3163674ae819SStefano Zampini } 3164674ae819SStefano Zampini 3165674ae819SStefano Zampini #undef __FUNCT__ 3166674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3167674ae819SStefano 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[]) 3168674ae819SStefano Zampini { 3169674ae819SStefano Zampini Vec local_vec,global_vec; 3170674ae819SStefano Zampini IS seqis,paris; 3171674ae819SStefano Zampini VecScatter scatter_ctx; 3172674ae819SStefano Zampini PetscScalar *array; 3173674ae819SStefano Zampini PetscInt *temp_global_dofs; 3174674ae819SStefano Zampini PetscScalar globalsum; 3175674ae819SStefano Zampini PetscInt i,j,s; 317617ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 317717ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 317817ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3179674ae819SStefano Zampini PetscBool first_found; 3180674ae819SStefano Zampini PetscErrorCode ierr; 3181674ae819SStefano Zampini 3182674ae819SStefano Zampini PetscFunctionBegin; 3183674ae819SStefano Zampini /* mpi buffers */ 3184b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3185b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3186674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 318717ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3188674ae819SStefano Zampini /* get maximum size of subset */ 3189785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3190674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3191674ae819SStefano Zampini max_local = 0; 3192b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3193674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3194674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3195674ae819SStefano Zampini } 3196674ae819SStefano Zampini } 3197b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3198674ae819SStefano Zampini max_global++; 3199674ae819SStefano Zampini max_local = 0; 3200b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3201674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3202674ae819SStefano Zampini max_local = local_dofs[i]; 3203674ae819SStefano Zampini } 3204674ae819SStefano Zampini } 3205674ae819SStefano Zampini max_local++; 3206674ae819SStefano Zampini /* allocate workspace */ 3207674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3208674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3209674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3210674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3211674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3212674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3213674ae819SStefano Zampini /* create scatter */ 3214674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3215674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3216674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3218674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3219674ae819SStefano Zampini /* init array */ 3220674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3221674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3222674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3223674ae819SStefano Zampini if (local_dofs_mult) { 3224674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3225674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3226674ae819SStefano Zampini } 3227674ae819SStefano Zampini } else { 3228674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3229674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3230674ae819SStefano Zampini } 3231674ae819SStefano Zampini } 3232674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3233674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3234674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3235674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3236674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 32375b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3238674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3239674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3240674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3241674ae819SStefano Zampini nlocals = 0; 3242674ae819SStefano Zampini first_index = -1; 3243674ae819SStefano Zampini first_found = PETSC_FALSE; 3244674ae819SStefano Zampini for (i=0;i<s;i++) { 3245b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3246674ae819SStefano Zampini first_found = PETSC_TRUE; 3247674ae819SStefano Zampini first_index = i; 3248674ae819SStefano Zampini } 32495b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3250674ae819SStefano Zampini } 3251674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3252674ae819SStefano Zampini if (!rank_prec_comm) { 3253674ae819SStefano Zampini dof_displs[0]=0; 3254674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3255674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3256674ae819SStefano Zampini } 3257674ae819SStefano Zampini } 3258674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3259674ae819SStefano Zampini if (first_found) { 3260674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3261674ae819SStefano Zampini old_index = first_index; 3262674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3263b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3264674ae819SStefano Zampini array[i] += array[old_index]; 3265674ae819SStefano Zampini old_index = i; 3266674ae819SStefano Zampini } 3267674ae819SStefano Zampini } 3268674ae819SStefano Zampini } 3269674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3270674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3271674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3272674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3273674ae819SStefano Zampini /* get global ordering of local dofs */ 32742b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3275674ae819SStefano Zampini if (local_dofs_mult) { 3276674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32775b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3278674ae819SStefano Zampini } 3279674ae819SStefano Zampini } else { 3280674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32815b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3282674ae819SStefano Zampini } 3283674ae819SStefano Zampini } 32842b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3285674ae819SStefano Zampini /* free workspace */ 3286674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3287674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3288674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 328917ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3290674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3291674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3292674ae819SStefano Zampini PetscFunctionReturn(0); 3293674ae819SStefano Zampini } 32949a7d3425SStefano Zampini 32959a7d3425SStefano Zampini #undef __FUNCT__ 32969a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32979a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32989a7d3425SStefano Zampini { 32999a7d3425SStefano Zampini PetscInt i,j; 33009a7d3425SStefano Zampini PetscScalar *alphas; 33019a7d3425SStefano Zampini PetscErrorCode ierr; 33029a7d3425SStefano Zampini 33039a7d3425SStefano Zampini PetscFunctionBegin; 33049a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3305785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33069a7d3425SStefano Zampini for (i=0;i<n;i++) { 33079a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33089a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33099a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33109a7d3425SStefano Zampini } 33119a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33129a7d3425SStefano Zampini PetscFunctionReturn(0); 33139a7d3425SStefano Zampini } 33149a7d3425SStefano Zampini 3315e7931f94SStefano Zampini #undef __FUNCT__ 331670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 331728143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3318e7931f94SStefano Zampini { 3319e7931f94SStefano Zampini Mat subdomain_adj; 3320e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3321e7931f94SStefano Zampini MatPartitioning partitioner; 3322e7931f94SStefano Zampini Mat_IS *matis; 3323e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3324e7931f94SStefano Zampini PetscInt prank; 3325e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3326e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3327e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33283837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3329e7931f94SStefano Zampini PetscErrorCode ierr; 33302b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3331e7931f94SStefano Zampini PetscSubcomm subcomm; 3332a57a6d2fSStefano Zampini 3333e7931f94SStefano Zampini PetscFunctionBegin; 33342b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33352b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33362b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3337e7931f94SStefano Zampini 3338e7931f94SStefano Zampini /* Get info on mapping */ 3339e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3340e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3341e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3342e7931f94SStefano Zampini 3343e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3344785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3345e7931f94SStefano Zampini xadj[0] = 0; 3346e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3347785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3348785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3349e7931f94SStefano Zampini 33502b510759SStefano Zampini if (threshold) { 3351d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33522b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3353d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3354d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3355d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3356d023bfaeSStefano Zampini xadj_count++; 3357e7931f94SStefano Zampini } 3358e7931f94SStefano Zampini } 3359d023bfaeSStefano Zampini xadj[1] = xadj_count; 3360c8587f34SStefano Zampini } else { 3361e7931f94SStefano Zampini if (xadj[1]) { 3362e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3363e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3364c8587f34SStefano Zampini } 3365e7931f94SStefano Zampini } 3366e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3367e7931f94SStefano Zampini if (use_square) { 3368e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3369e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3370e7931f94SStefano Zampini } 3371e7931f94SStefano Zampini } 3372e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3373e7931f94SStefano Zampini 33743837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3375e7931f94SStefano Zampini 3376e7931f94SStefano Zampini /* 3377e7931f94SStefano Zampini Restrict work on active processes only. 3378e7931f94SStefano Zampini */ 3379e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3380e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3381e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33822b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3383d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3384e7931f94SStefano Zampini if (color) { 3385e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3386e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3387e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3388c8587f34SStefano Zampini } else { 338928143c3dSStefano Zampini PetscInt coarsening_ratio; 3390306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3391785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3392e7931f94SStefano Zampini prank = rank; 3393306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33948002ef2cSStefano Zampini /* 3395e7931f94SStefano Zampini for (i=0;i<size;i++) { 3396e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3397c8587f34SStefano Zampini } 33988002ef2cSStefano Zampini */ 3399e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3400e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3401c8587f34SStefano Zampini } 3402e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3403306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 340422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3405e7931f94SStefano Zampini 3406e7931f94SStefano Zampini /* Partition */ 3407306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3408e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3409e7931f94SStefano Zampini if (use_vwgt) { 34103837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3411e7931f94SStefano Zampini v_wgt[0] = local_size; 3412e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3413c8587f34SStefano Zampini } 341428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 341528143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 341628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3417e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3418e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 341922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3420e7931f94SStefano Zampini 3421e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 342228143c3dSStefano Zampini if (contiguous) { 342328143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 342428143c3dSStefano Zampini } else { 342584ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 342628143c3dSStefano Zampini } 3427e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3428e7931f94SStefano Zampini /* clean up */ 3429e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3431e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3432e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3433e7931f94SStefano Zampini } 3434e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3435e7931f94SStefano Zampini 3436e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3437e7931f94SStefano Zampini i = 1; 3438e7931f94SStefano Zampini if (color) i=0; 3439e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3440e7931f94SStefano Zampini 3441e7931f94SStefano Zampini /* get back IS */ 3442e7931f94SStefano Zampini *is_sends = ranks_send_to; 3443e7931f94SStefano Zampini PetscFunctionReturn(0); 3444e7931f94SStefano Zampini } 3445e7931f94SStefano Zampini 3446e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3447e7931f94SStefano Zampini 3448e7931f94SStefano Zampini #undef __FUNCT__ 3449e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 345028143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3451e7931f94SStefano Zampini { 345270cf5478SStefano Zampini Mat local_mat; 3453e7931f94SStefano Zampini Mat_IS *matis; 3454e7931f94SStefano Zampini IS is_sends_internal; 34559d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 345628143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34579d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3458e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3459e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3460e7931f94SStefano Zampini const PetscInt* is_indices; 3461e7931f94SStefano Zampini MatType new_local_type; 3462e7931f94SStefano Zampini /* buffers */ 3463e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 346428143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34659d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3466e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3467e7931f94SStefano Zampini /* MPI */ 346828143c3dSStefano Zampini MPI_Comm comm,comm_n; 346928143c3dSStefano Zampini PetscSubcomm subcomm; 3470e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 347128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 347228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 347328143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 347428143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 347528143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3476e7931f94SStefano Zampini PetscErrorCode ierr; 3477e7931f94SStefano Zampini 3478e7931f94SStefano Zampini PetscFunctionBegin; 347928143c3dSStefano Zampini /* TODO: add missing checks */ 348028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 348128143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 348228143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 348328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3484e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 348528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3486e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3487e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3488e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3489e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3490e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 349128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 349270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 349370cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 349428143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 349570cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 349670cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 349770cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 349870cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 349970cf5478SStefano Zampini } 3500e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3501e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3502e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3503e7931f94SStefano Zampini if (!is_sends) { 350428143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 350528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 350628143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3507c8587f34SStefano Zampini } else { 3508e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3509e7931f94SStefano Zampini is_sends_internal = is_sends; 3510c8587f34SStefano Zampini } 3511e7931f94SStefano Zampini 3512e7931f94SStefano Zampini /* get pointer of MATIS data */ 3513e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3514e7931f94SStefano Zampini 3515e7931f94SStefano Zampini /* get comm */ 3516a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3517e7931f94SStefano Zampini 3518e7931f94SStefano Zampini /* compute number of sends */ 3519e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3520e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3521e7931f94SStefano Zampini 3522e7931f94SStefano Zampini /* compute number of receives */ 3523e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3524785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3525e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3526e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3527e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3528e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3529e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3530e7931f94SStefano Zampini 353128143c3dSStefano Zampini /* restrict comm if requested */ 353228143c3dSStefano Zampini subcomm = 0; 353328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 353428143c3dSStefano Zampini if (restrict_comm) { 3535779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3536779c1cceSStefano Zampini 353728143c3dSStefano Zampini color = 0; 35386366ad86SStefano Zampini if (!n_recvs && n_sends) color = 1; /* processes sending and not receiving anything will not partecipate in new comm */ 353928143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 354028143c3dSStefano Zampini subcommsize = commsize - subcommsize; 354128143c3dSStefano Zampini /* check if reuse has been requested */ 354228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 354328143c3dSStefano Zampini if (*mat_n) { 354428143c3dSStefano Zampini PetscMPIInt subcommsize2; 354528143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 354628143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 354728143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 354828143c3dSStefano Zampini } else { 354928143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 355028143c3dSStefano Zampini } 355128143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3552779c1cceSStefano Zampini PetscMPIInt rank; 3553779c1cceSStefano Zampini 3554779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 355528143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 355628143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 355728143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3558306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 355928143c3dSStefano Zampini } 356028143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 356128143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 356228143c3dSStefano Zampini } else { 356328143c3dSStefano Zampini comm_n = comm; 356428143c3dSStefano Zampini } 356528143c3dSStefano Zampini 3566e7931f94SStefano Zampini /* prepare send/receive buffers */ 3567785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3568e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3569785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3570e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 357128143c3dSStefano Zampini if (nis) { 3572854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 357328143c3dSStefano Zampini } 3574e7931f94SStefano Zampini 357528143c3dSStefano Zampini /* Get data from local matrices */ 3576e7931f94SStefano Zampini if (!isdense) { 3577a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3578e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3579e7931f94SStefano Zampini /* 3580e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3581e7931f94SStefano Zampini send_buffer_idxs should contain: 3582e7931f94SStefano Zampini - MatType_PRIVATE type 3583e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3584e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3585e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3586e7931f94SStefano Zampini */ 3587e7931f94SStefano Zampini } else { 3588e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3589e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3590854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3591e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3592e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3593e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3594e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3595e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3596e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3597e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3598e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3599e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3600c8587f34SStefano Zampini } 3601c8587f34SStefano Zampini } 3602e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 360328143c3dSStefano Zampini /* additional is (if any) */ 360428143c3dSStefano Zampini if (nis) { 360528143c3dSStefano Zampini PetscMPIInt psum; 360628143c3dSStefano Zampini PetscInt j; 360728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 360828143c3dSStefano Zampini PetscInt plen; 360928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361028143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 361128143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 361228143c3dSStefano Zampini } 3613854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 361428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 361528143c3dSStefano Zampini PetscInt plen; 361628143c3dSStefano Zampini const PetscInt *is_array_idxs; 361728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361828143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 361928143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362028143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 362128143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362228143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 362328143c3dSStefano Zampini } 362428143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 362528143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 362628143c3dSStefano Zampini } 362728143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 362828143c3dSStefano Zampini } 362928143c3dSStefano Zampini 3630e7931f94SStefano Zampini buf_size_idxs = 0; 3631e7931f94SStefano Zampini buf_size_vals = 0; 363228143c3dSStefano Zampini buf_size_idxs_is = 0; 3633e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3634e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3635e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 363628143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3637e7931f94SStefano Zampini } 3638785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3639785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 364095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3641e7931f94SStefano Zampini 3642e7931f94SStefano Zampini /* get new tags for clean communications */ 3643e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3644e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 364528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3646e7931f94SStefano Zampini 3647e7931f94SStefano Zampini /* allocate for requests */ 3648785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3649785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 365095ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3651785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3652785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 365395ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3654e7931f94SStefano Zampini 3655e7931f94SStefano Zampini /* communications */ 3656e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3657e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 365828143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3659e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3660e7931f94SStefano Zampini source_dest = onodes[i]; 3661e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3662e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3663e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3664e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 366528143c3dSStefano Zampini if (nis) { 366628143c3dSStefano 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); 366728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 366828143c3dSStefano Zampini } 3669e7931f94SStefano Zampini } 3670e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3671e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3672e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3673e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 367428143c3dSStefano Zampini if (nis) { 367528143c3dSStefano 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); 367628143c3dSStefano Zampini } 3677e7931f94SStefano Zampini } 3678e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3679e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3680e7931f94SStefano Zampini 3681e7931f94SStefano Zampini /* assemble new l2g map */ 3682e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3683e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36849d30be91SStefano Zampini new_local_rows = 0; 3685e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36869d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3687e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3688e7931f94SStefano Zampini } 36899d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);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 ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36949d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3695e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3696e7931f94SStefano Zampini } 36979d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36989d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3699e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3700e7931f94SStefano Zampini 3701e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3702e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3703e7931f94SStefano 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) */ 3704e7931f94SStefano Zampini if (n_recvs) { 370528143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3706e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3707e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3708e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3709e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3710e7931f94SStefano Zampini break; 3711e7931f94SStefano Zampini } 3712e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3713e7931f94SStefano Zampini } 3714e7931f94SStefano Zampini switch (new_local_type_private) { 371528143c3dSStefano Zampini case MATDENSE_PRIVATE: 371628143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3717e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3718e7931f94SStefano Zampini bs = 1; 371928143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 372028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 372128143c3dSStefano Zampini bs = 1; 372228143c3dSStefano Zampini } 3723e7931f94SStefano Zampini break; 3724e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3725e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3726e7931f94SStefano Zampini bs = 1; 3727e7931f94SStefano Zampini break; 3728e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3729e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3730e7931f94SStefano Zampini break; 3731e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3732e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3733e7931f94SStefano Zampini break; 3734e7931f94SStefano Zampini default: 37359d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3736e7931f94SStefano Zampini break; 3737e7931f94SStefano Zampini } 373828143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 373928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 374028143c3dSStefano Zampini bs = 1; 3741e7931f94SStefano Zampini } 3742e7931f94SStefano Zampini 374370cf5478SStefano Zampini /* create MATIS object if needed */ 374470cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3745e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 374628143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 374770cf5478SStefano Zampini } else { 374870cf5478SStefano Zampini /* it also destroys the local matrices */ 374970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 375070cf5478SStefano Zampini } 375170cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3752e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37539d30be91SStefano Zampini 37549d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37559d30be91SStefano Zampini 37569d30be91SStefano Zampini /* Global to local map of received indices */ 37579d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37589d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37599d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37609d30be91SStefano Zampini 37619d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37629d30be91SStefano Zampini buf_size_idxs = 0; 37639d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37649d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37659d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37669d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37679d30be91SStefano Zampini } 37689d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37699d30be91SStefano Zampini 37709d30be91SStefano Zampini /* set preallocation */ 37719d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37729d30be91SStefano Zampini if (!newisdense) { 37739d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37749d30be91SStefano Zampini 37759d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37769d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37779d30be91SStefano Zampini if (n_recvs) { 37789d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37799d30be91SStefano Zampini } 37809d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37819d30be91SStefano Zampini PetscInt j; 37829d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37839d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37849d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37859d30be91SStefano Zampini } 37869d30be91SStefano Zampini } else { 37879d30be91SStefano Zampini /* TODO */ 37889d30be91SStefano Zampini } 37899d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37909d30be91SStefano Zampini } 37919d30be91SStefano Zampini if (new_local_nnz) { 37929d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37939d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37949d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37959d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37969d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37979d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37989d30be91SStefano Zampini } else { 37999d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38009d30be91SStefano Zampini } 38019d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38029d30be91SStefano Zampini } else { 38039d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38049d30be91SStefano Zampini } 3805e7931f94SStefano Zampini 3806e7931f94SStefano Zampini /* set values */ 3807e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38089d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3809e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3810e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3811e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38129d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3813e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3814e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3815e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 381628143c3dSStefano Zampini } else { 381728143c3dSStefano Zampini /* TODO */ 3818e7931f94SStefano Zampini } 3819e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3820e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3821e7931f94SStefano Zampini } 3822e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3823e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38269d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38279d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3828e7931f94SStefano Zampini 3829dfd14d43SStefano Zampini #if 0 383028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3831e7931f94SStefano Zampini Vec lvec,rvec; 3832e7931f94SStefano Zampini PetscReal infty_error; 3833e7931f94SStefano Zampini 38342a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3835e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3836e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3837e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 383870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3839e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3840e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3841e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3842e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3843e7931f94SStefano Zampini } 384428143c3dSStefano Zampini #endif 3845e7931f94SStefano Zampini 384628143c3dSStefano Zampini /* assemble new additional is (if any) */ 384728143c3dSStefano Zampini if (nis) { 384828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 384928143c3dSStefano Zampini 385028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3851854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 385228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 385328143c3dSStefano Zampini psum = 0; 385428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 385528143c3dSStefano Zampini for (j=0;j<nis;j++) { 385628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 385728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 385828143c3dSStefano Zampini psum += plen; 385928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 386028143c3dSStefano Zampini } 386128143c3dSStefano Zampini } 3862854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3863854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 386428143c3dSStefano Zampini for (i=1;i<nis;i++) { 386528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 386628143c3dSStefano Zampini } 386728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 386828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 386928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 387028143c3dSStefano Zampini for (j=0;j<nis;j++) { 387128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 387328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 387428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387528143c3dSStefano Zampini } 387628143c3dSStefano Zampini } 387728143c3dSStefano Zampini for (i=0;i<nis;i++) { 387828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 387928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 388028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 388128143c3dSStefano Zampini } 388228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 388328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 388428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 388528143c3dSStefano Zampini } 3886e7931f94SStefano Zampini /* free workspace */ 388728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3888e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3889e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3890e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3891e7931f94SStefano Zampini if (isdense) { 3892e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3893e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3894e7931f94SStefano Zampini } else { 3895e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3896e7931f94SStefano Zampini } 389728143c3dSStefano Zampini if (nis) { 389828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 389928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 390028143c3dSStefano Zampini } 3901e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3902e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 390328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3904e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 390628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3907e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3910e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3911e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 391228143c3dSStefano Zampini if (nis) { 391328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 391428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 391528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 391628143c3dSStefano Zampini } 391728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 391828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 391928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 392028143c3dSStefano Zampini for (i=0;i<nis;i++) { 392128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 392228143c3dSStefano Zampini } 392328143c3dSStefano Zampini } 3924e7931f94SStefano Zampini PetscFunctionReturn(0); 3925e7931f94SStefano Zampini } 3926a57a6d2fSStefano Zampini 392712edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3928af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 392912edc857SStefano Zampini 3930c8587f34SStefano Zampini #undef __FUNCT__ 3931c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3932c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3933c8587f34SStefano Zampini { 3934c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3935c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 393620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39379881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 393820a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39396e683305SStefano Zampini IS coarse_is,*isarray; 39406e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39416e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3942f9eb5b7dSStefano Zampini PC pc_temp; 3943c8587f34SStefano Zampini PCType coarse_pc_type; 3944c8587f34SStefano Zampini KSPType coarse_ksp_type; 3945f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39464f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39476e683305SStefano Zampini Mat t_coarse_mat_is; 39486e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39496e683305SStefano Zampini PetscMPIInt all_procs; 395074e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 395168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 395222bc73bbSStefano Zampini PetscScalar *array; 39539881197aSStefano Zampini PetscErrorCode ierr; 3954fdc09c96SStefano Zampini 3955c8587f34SStefano Zampini PetscFunctionBegin; 3956c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 395768457ee5SStefano 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 */ 3958fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39595a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3960fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3961f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3962f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3963f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3964fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3965fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3966727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3967fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3968fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3969fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3970f4ddd8eeSStefano Zampini } 3971fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3972fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3973f4ddd8eeSStefano Zampini } 397470cf5478SStefano Zampini /* reset any subassembling information */ 397570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39766e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39776e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3978fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3979f4ddd8eeSStefano Zampini } 3980c8587f34SStefano Zampini 39816e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39822b510759SStefano Zampini im_active = !!(pcis->n); 39832b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39846e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39856e683305SStefano Zampini void_procs = all_procs-active_procs; 39866e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 398774e2c79eSStefano Zampini redist = PETSC_FALSE; 398822bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39896e683305SStefano Zampini csin_ml = PETSC_TRUE; 39906e683305SStefano Zampini ncoarse_ml = void_procs; 3991779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3992779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 399318a45a71SStefano Zampini csin_ds = PETSC_TRUE; 399418a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 399518a45a71SStefano Zampini redist = PETSC_TRUE; 399618a45a71SStefano Zampini } else { 39976e683305SStefano Zampini csin_ds = PETSC_TRUE; 3998779c1cceSStefano Zampini ncoarse_ds = active_procs; 3999779c1cceSStefano Zampini redist = PETSC_TRUE; 400018a45a71SStefano Zampini } 40016e683305SStefano Zampini } else { 40026e683305SStefano Zampini csin_ml = PETSC_FALSE; 40036e683305SStefano Zampini ncoarse_ml = all_procs; 40046e683305SStefano Zampini if (void_procs) { 40056e683305SStefano Zampini csin_ds = PETSC_TRUE; 40066e683305SStefano Zampini ncoarse_ds = void_procs; 40076e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40086e683305SStefano Zampini } else { 4009779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 401074e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 401174e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401274e2c79eSStefano Zampini redist = PETSC_TRUE; 401374e2c79eSStefano Zampini } else { 40146e683305SStefano Zampini csin_ds = PETSC_FALSE; 40156e683305SStefano Zampini ncoarse_ds = all_procs; 40166e683305SStefano Zampini } 40176e683305SStefano Zampini } 401874e2c79eSStefano Zampini } 40196e683305SStefano Zampini 40206e683305SStefano Zampini /* 40216e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40226e683305SStefano Zampini - we have not exceeded the number of levels requested 40236e683305SStefano Zampini - we can actually subassemble the active processes 40246e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40256e683305SStefano Zampini */ 40266e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40276e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40286e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40296e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40306e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4031f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40322b510759SStefano Zampini } else { 4033f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4034c8587f34SStefano Zampini } 4035c8587f34SStefano Zampini } 40366e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40376e683305SStefano Zampini if (multilevel_allowed) { 40386e683305SStefano Zampini ncoarse = ncoarse_ml; 40396e683305SStefano Zampini csin = csin_ml; 404058da7f69SStefano Zampini redist = PETSC_FALSE; 40416e683305SStefano Zampini } else { 40426e683305SStefano Zampini ncoarse = ncoarse_ds; 40436e683305SStefano Zampini csin = csin_ds; 40446e683305SStefano Zampini } 4045e7931f94SStefano Zampini 4046abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4047abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4048abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4049abbbba34SStefano Zampini 4050abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 405122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 405222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 405322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 405422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4055b9b85e73SStefano Zampini #if 0 4056b9b85e73SStefano Zampini { 4057b9b85e73SStefano Zampini PetscViewer viewer; 4058b9b85e73SStefano Zampini char filename[256]; 4059b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4060b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4061b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4062b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4063b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4064b9b85e73SStefano Zampini } 4065b9b85e73SStefano Zampini #endif 40666e683305SStefano 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); 40676e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40686e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40696e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4070abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4071abbbba34SStefano Zampini 40726e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40736e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40746e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40756e683305SStefano Zampini const PetscInt *idxs; 40766e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40776e683305SStefano Zampini 40786e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40790be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40806e683305SStefano Zampini /* allocate space for temporary storage */ 4081854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4082854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40836e683305SStefano Zampini /* allocate for IS array */ 40846e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40856e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40866e683305SStefano Zampini nis = nisdofs + nisneu; 4087854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40886e683305SStefano Zampini /* dofs splitting */ 40896e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40906e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40916e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40926e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40936e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40946e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40956e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40966e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40976e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40986e683305SStefano Zampini } 40996e683305SStefano Zampini /* neumann boundaries */ 41006e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41016e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41026e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41036e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41046e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41056e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41066e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41076e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41086e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41096e683305SStefano Zampini } 41106e683305SStefano Zampini /* free memory */ 41116e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41126e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41146e683305SStefano Zampini } else { 41156e683305SStefano Zampini nis = 0; 41166e683305SStefano Zampini nisdofs = 0; 41176e683305SStefano Zampini nisneu = 0; 41186e683305SStefano Zampini isarray = NULL; 41196e683305SStefano Zampini } 41206e683305SStefano Zampini /* destroy no longer needed map */ 41216e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41226e683305SStefano Zampini 41236e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41246e683305SStefano Zampini coarse_mat_is = NULL; 41256e683305SStefano Zampini if (csin) { 41266e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 412774e2c79eSStefano Zampini if (redist) { 412874e2c79eSStefano Zampini PetscMPIInt rank; 4129779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 413074e2c79eSStefano Zampini 413174e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 413258da7f69SStefano Zampini spc = active_procs/ncoarse; 413358da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4134779c1cceSStefano Zampini if (im_active) { 4135779c1cceSStefano Zampini destsize = 1; 413674e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 413774e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 413874e2c79eSStefano Zampini } else { 413974e2c79eSStefano Zampini dest[0] = rank/(spc+1); 414074e2c79eSStefano Zampini } 414174e2c79eSStefano Zampini } else { 4142779c1cceSStefano Zampini destsize = 0; 41436e683305SStefano Zampini } 4144779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4145779c1cceSStefano Zampini } else if (csin_type_simple) { 41466e683305SStefano Zampini PetscMPIInt rank; 41476e683305SStefano Zampini PetscInt issize,isidx; 4148779c1cceSStefano Zampini 41496e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41506e683305SStefano Zampini if (im_active) { 41516e683305SStefano Zampini issize = 1; 41526e683305SStefano Zampini isidx = (PetscInt)rank; 41536e683305SStefano Zampini } else { 41546e683305SStefano Zampini issize = 0; 41556e683305SStefano Zampini isidx = -1; 41566e683305SStefano Zampini } 41576e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4158779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41596e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41606e683305SStefano Zampini } 4161779c1cceSStefano Zampini 4162779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4163779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4164779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4165779c1cceSStefano Zampini PetscInt *coarse_candidates; 4166779c1cceSStefano Zampini const PetscInt* tisindices; 4167779c1cceSStefano Zampini 4168779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4169779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4170779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4171779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4172779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4173779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4174779c1cceSStefano Zampini } 4175779c1cceSStefano Zampini } 4176779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4177779c1cceSStefano Zampini 4178779c1cceSStefano Zampini 41796e683305SStefano Zampini if (pcbddc->dbg_flag) { 41806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4184779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41866e683305SStefano Zampini } 41876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41886e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41896e683305SStefano Zampini } 41906e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41916e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41926e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4193854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41946e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41956e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41966e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41976e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41986e683305SStefano Zampini } 41996e683305SStefano Zampini if (pcbddc->dbg_flag) { 42006e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42026e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42036e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42046e683305SStefano Zampini } 4205779c1cceSStefano Zampini } 42066e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 42076e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 42086e683305SStefano Zampini } else { 42096e683305SStefano Zampini if (pcbddc->dbg_flag) { 42106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42126e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42136e683305SStefano Zampini } 42146e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42156e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42166e683305SStefano Zampini } 42176e683305SStefano Zampini 42186e683305SStefano Zampini /* create local to global scatters for coarse problem */ 421968457ee5SStefano Zampini if (compute_vecs) { 42206e683305SStefano Zampini PetscInt lrows; 42216e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42226e683305SStefano Zampini if (coarse_mat_is) { 42236e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42246e683305SStefano Zampini } else { 42256e683305SStefano Zampini lrows = 0; 42266e683305SStefano Zampini } 42276e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42286e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42296e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42306e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42316e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42326e683305SStefano Zampini } 42336e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42346e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4235c8587f34SStefano Zampini 4236f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4237f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4238f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4239f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4240f9eb5b7dSStefano Zampini } else { 4241f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4242f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4243c8587f34SStefano Zampini } 4244c8587f34SStefano Zampini 42456e683305SStefano Zampini /* print some info if requested */ 42466e683305SStefano Zampini if (pcbddc->dbg_flag) { 42476e683305SStefano Zampini if (!multilevel_allowed) { 42486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42496e683305SStefano Zampini if (multilevel_requested) { 42506e683305SStefano 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); 42516e683305SStefano Zampini } else if (pcbddc->max_levels) { 42526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42536e683305SStefano Zampini } 42546e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42556e683305SStefano Zampini } 42566e683305SStefano Zampini } 42576e683305SStefano Zampini 4258f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42596e683305SStefano Zampini if (coarse_mat_is) { 42606e683305SStefano Zampini MatReuse coarse_mat_reuse; 42616a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42626e683305SStefano Zampini if (pcbddc->dbg_flag) { 42636e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42646e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42656e683305SStefano Zampini } 4266f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4267312be037SStefano Zampini char prefix[256],str_level[16]; 4268e604994aSStefano Zampini size_t len; 42696e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4270c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4271f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42725f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4273c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42746e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4275c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4276c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42773e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4278e604994aSStefano Zampini /* prefix */ 4279e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4280e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4281e604994aSStefano Zampini if (!pcbddc->current_level) { 4282e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4283e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4284c8587f34SStefano Zampini } else { 4285e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4286312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4287312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 428834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4289312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4290e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4291e604994aSStefano Zampini } 4292e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 42933e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42943e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 42953e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 42963e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4297f9eb5b7dSStefano Zampini /* allow user customization */ 4298f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 42993e3c6dadSStefano Zampini } 43003e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43013e3c6dadSStefano Zampini if (nisdofs) { 43023e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43033e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43043e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43053e3c6dadSStefano Zampini } 43063e3c6dadSStefano Zampini } 43073e3c6dadSStefano Zampini if (nisneu) { 43083e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43093e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4310312be037SStefano Zampini } 4311f9eb5b7dSStefano Zampini 4312f9eb5b7dSStefano Zampini /* get some info after set from options */ 4313f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4314f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4315f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43164f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43176e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4318f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4319f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4320f9eb5b7dSStefano Zampini } 43214f3a063dSStefano Zampini if (isredundant) { 43224f3a063dSStefano Zampini KSP inner_ksp; 43234f3a063dSStefano Zampini PC inner_pc; 43244f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43254f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43264f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43274f3a063dSStefano Zampini } 4328f9eb5b7dSStefano Zampini 4329f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4330fa7f1dd8SStefano Zampini if (coarse_reuse) { 433181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4332fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43336e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4334fa7f1dd8SStefano Zampini } else { 43356e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4336fa7f1dd8SStefano Zampini } 4337c8587f34SStefano Zampini if (isbddc || isnn) { 433822bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 433970cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 434028143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 434122b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43446e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43456e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 434622b6e8a2SStefano Zampini } 434770cf5478SStefano Zampini } 43486e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 434970cf5478SStefano Zampini } else { 435022bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 435122bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 435222bc73bbSStefano Zampini } 435322bc73bbSStefano Zampini } else { 43542e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4355c8587f34SStefano Zampini } 4356c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4357c8587f34SStefano Zampini 4358*3301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43595a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4360*3301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 4361*3301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 4362*3301b35fSStefano Zampini } 4363*3301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 4364*3301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 4365*3301b35fSStefano Zampini } 4366*3301b35fSStefano Zampini if (pc->pmat->spd_set) { 4367*3301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 4368*3301b35fSStefano Zampini } 43696e683305SStefano Zampini /* set operators */ 43705f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43716e683305SStefano Zampini if (pcbddc->dbg_flag) { 43726e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43736e683305SStefano Zampini } 43746e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43756e683305SStefano Zampini coarse_mat = 0; 43766e683305SStefano Zampini } 43776e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4378b9b85e73SStefano Zampini #if 0 4379b9b85e73SStefano Zampini { 4380b9b85e73SStefano Zampini PetscViewer viewer; 4381b9b85e73SStefano Zampini char filename[256]; 4382b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4383b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4384b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4385b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4386b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4387b9b85e73SStefano Zampini } 4388b9b85e73SStefano Zampini #endif 4389c8587f34SStefano Zampini 4390c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4391c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4392c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 439398a51de6SStefano Zampini } 439498a51de6SStefano Zampini 439598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 439698a51de6SStefano Zampini Vec crhs,csol; 439798a51de6SStefano Zampini PetscBool ispreonly; 439898a51de6SStefano Zampini if (CoarseNullSpace) { 4399c8587f34SStefano Zampini if (isbddc) { 4400c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4401c8587f34SStefano Zampini } else { 4402c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4403c8587f34SStefano Zampini } 4404c8587f34SStefano Zampini } 4405f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4406f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4407f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4408f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44096e683305SStefano Zampini /* hack */ 4410f347579bSStefano Zampini if (!csol) { 44112a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4412f9eb5b7dSStefano Zampini } 4413f347579bSStefano Zampini if (!crhs) { 44142a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4415f347579bSStefano Zampini } 4416cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4417cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44186e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4419c8587f34SStefano Zampini KSP check_ksp; 44202b510759SStefano Zampini KSPType check_ksp_type; 4421c8587f34SStefano Zampini PC check_pc; 44226e683305SStefano Zampini Vec check_vec,coarse_vec; 44236a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44242b510759SStefano Zampini PetscInt its; 44256e683305SStefano Zampini PetscBool compute_eigs; 44266e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44276e683305SStefano Zampini PetscInt neigs; 44288e185a42SStefano Zampini const char *prefix; 4429c8587f34SStefano Zampini 44302b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44316e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 443223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4433f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44342b510759SStefano Zampini if (ispreonly) { 44352b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44366e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44372b510759SStefano Zampini } else { 4438cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44396e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4440c8587f34SStefano Zampini } 4441c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44426e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44436e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44446e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4445a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4446a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4447a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4448a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4449c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4450c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4451c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4452c8587f34SStefano Zampini /* create random vec */ 44536e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44546e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4455c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4456c8587f34SStefano Zampini if (CoarseNullSpace) { 4457c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4458c8587f34SStefano Zampini } 44596e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4460c8587f34SStefano Zampini /* solve coarse problem */ 44616e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4462c8587f34SStefano Zampini if (CoarseNullSpace) { 44636e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4464c8587f34SStefano Zampini } 4465cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44666e683305SStefano Zampini if (compute_eigs) { 4467854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4468854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44696e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44706e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44716e683305SStefano Zampini lambda_min = eigs_r[0]; 44726e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44736e683305SStefano Zampini if (lambda_max>lambda_min) { 4474cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4475cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4476cbcc2c2aSStefano Zampini } 4477c8587f34SStefano Zampini } 4478c8587f34SStefano Zampini } 4479cbcc2c2aSStefano Zampini 4480c8587f34SStefano Zampini /* check coarse problem residual error */ 44816e683305SStefano Zampini if (pcbddc->dbg_flag) { 44826e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44836e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44846e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4485c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44866e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44876e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4488c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4489779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44906e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44916e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44946e683305SStefano Zampini if (compute_eigs) { 44956e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4496deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4497c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44986e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 44996e683305SStefano 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); 45006e683305SStefano Zampini for (i=0;i<neigs;i++) { 45016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4502c8587f34SStefano Zampini } 45036e683305SStefano Zampini } 45046e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45056e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45066e683305SStefano Zampini } 4507c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45086e683305SStefano Zampini if (compute_eigs) { 45096e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45106e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4511c8587f34SStefano Zampini } 45126e683305SStefano Zampini } 45136e683305SStefano Zampini } 4514cbcc2c2aSStefano Zampini /* print additional info */ 4515cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45166e683305SStefano Zampini /* waits until all processes reaches this point */ 45176e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4518cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4519cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4520cbcc2c2aSStefano Zampini } 4521cbcc2c2aSStefano Zampini 45222b510759SStefano Zampini /* free memory */ 4523c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4524fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4525c8587f34SStefano Zampini PetscFunctionReturn(0); 4526c8587f34SStefano Zampini } 4527674ae819SStefano Zampini 4528f34684f1SStefano Zampini #undef __FUNCT__ 4529f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4530f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4531f34684f1SStefano Zampini { 4532f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4533f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4534f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 45350e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 453673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 45370e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4538f34684f1SStefano Zampini PetscErrorCode ierr; 4539f34684f1SStefano Zampini 4540f34684f1SStefano Zampini PetscFunctionBegin; 4541f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45420e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45430e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4544727cdba6SStefano Zampini } 45450e6343abSStefano 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); 45460e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 45470e6343abSStefano Zampini local_size = 0; 45480e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 45490e6343abSStefano Zampini PetscInt j; 45500e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 45510e6343abSStefano Zampini } 45520e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 45530e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45540e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45550e6343abSStefano Zampini } 4556f34684f1SStefano Zampini 4557f34684f1SStefano Zampini /* check numbering */ 4558f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4559f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4560b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4561f34684f1SStefano Zampini 4562f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4563f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4564f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45650fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4566f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4567f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4568727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4569f34684f1SStefano Zampini } 4570f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4571f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4572f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4573f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4574f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4575f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4576f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4577f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4578f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4579f34684f1SStefano Zampini if (array[i] == 1.0) { 4580b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4581f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4582f34684f1SStefano Zampini } 4583f34684f1SStefano Zampini } 4584b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4585f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4586f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4587f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4588f34684f1SStefano Zampini } 4589f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4590f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4591f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4592f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4593f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4594f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4595b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4596ca8b9ea9SStefano Zampini PetscInt *gidxs; 4597ca8b9ea9SStefano Zampini 4598ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4599ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4600f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4601f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4602f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4603f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46044bc2dc4bSStefano 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); 4605f34684f1SStefano Zampini } 4606f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4607ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4608f34684f1SStefano Zampini } 4609f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4610302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4611f34684f1SStefano Zampini } 46128bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4613f34684f1SStefano Zampini /* get back data */ 4614f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4615f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4616674ae819SStefano Zampini PetscFunctionReturn(0); 4617674ae819SStefano Zampini } 4618674ae819SStefano Zampini 4619e456f2a8SStefano Zampini #undef __FUNCT__ 4620e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4621a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4622e456f2a8SStefano Zampini { 4623e456f2a8SStefano Zampini IS localis_t; 4624a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4625e456f2a8SStefano Zampini PetscScalar *vals; 4626e456f2a8SStefano Zampini PetscErrorCode ierr; 4627e456f2a8SStefano Zampini 4628e456f2a8SStefano Zampini PetscFunctionBegin; 4629a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4630e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4631854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4632e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4633e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4634a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4635a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46361035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4637a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46381035eff8SStefano Zampini } 4639a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4640e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4641e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4642a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4643a7dc3881SStefano Zampini /* now compute set in local ordering */ 4644a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4645a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4646a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4647a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4648a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4649ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4650e456f2a8SStefano Zampini lsize++; 4651e456f2a8SStefano Zampini } 4652e456f2a8SStefano Zampini } 4653854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4654a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4655ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4656e456f2a8SStefano Zampini idxs[lsize++] = i; 4657e456f2a8SStefano Zampini } 4658e456f2a8SStefano Zampini } 4659a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4660a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4661e456f2a8SStefano Zampini *localis = localis_t; 4662e456f2a8SStefano Zampini PetscFunctionReturn(0); 4663e456f2a8SStefano Zampini } 4664906d46d4SStefano Zampini 4665906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4666906d46d4SStefano Zampini #undef __FUNCT__ 4667906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4668906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4669906d46d4SStefano Zampini { 4670906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4671906d46d4SStefano Zampini PetscErrorCode ierr; 4672906d46d4SStefano Zampini 4673906d46d4SStefano Zampini PetscFunctionBegin; 4674906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4675906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4676906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4677906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4678906d46d4SStefano Zampini PetscFunctionReturn(0); 4679906d46d4SStefano Zampini } 4680906d46d4SStefano Zampini 4681906d46d4SStefano Zampini #undef __FUNCT__ 4682906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4683906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4684906d46d4SStefano Zampini { 4685906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4686906d46d4SStefano Zampini PetscErrorCode ierr; 4687906d46d4SStefano Zampini 4688906d46d4SStefano Zampini PetscFunctionBegin; 4689906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4690906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4691906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4692906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4693906d46d4SStefano Zampini PetscFunctionReturn(0); 4694906d46d4SStefano Zampini } 4695b96c3477SStefano Zampini 4696b96c3477SStefano Zampini #undef __FUNCT__ 4697b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 469808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4699b96c3477SStefano Zampini { 4700a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4701b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4702b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4703a64f4aa4SStefano Zampini Mat S_j; 4704b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4705b96c3477SStefano Zampini PetscBool free_used_adj; 4706b96c3477SStefano Zampini PetscErrorCode ierr; 4707b96c3477SStefano Zampini 4708b96c3477SStefano Zampini PetscFunctionBegin; 4709b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4710b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 471108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4712b96c3477SStefano Zampini used_xadj = NULL; 4713b96c3477SStefano Zampini used_adjncy = NULL; 4714b96c3477SStefano Zampini } else { 471508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 471608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 471708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 471808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4719b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4720b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4721b96c3477SStefano Zampini } else { 47222fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4723b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4724b96c3477SStefano Zampini PetscInt nvtxs; 4725b96c3477SStefano Zampini 47262fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47272fffb893SStefano Zampini if (flg_row) { 4728b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4729b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4730b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4731b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47322fffb893SStefano Zampini } else { 47332fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47342fffb893SStefano Zampini used_xadj = NULL; 47352fffb893SStefano Zampini used_adjncy = NULL; 47362fffb893SStefano Zampini } 47372fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4738b96c3477SStefano Zampini } 4739b96c3477SStefano Zampini } 4740d5574798SStefano Zampini 4741d5574798SStefano Zampini /* setup sub_schurs data */ 4742a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4743a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4744a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4745a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 47466816873aSStefano 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); 4747a64f4aa4SStefano Zampini } else { 47486816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 47496816873aSStefano 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); 4750a64f4aa4SStefano Zampini } 4751a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4752b96c3477SStefano Zampini 4753b96c3477SStefano Zampini /* free adjacency */ 4754b96c3477SStefano Zampini if (free_used_adj) { 4755b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4756b96c3477SStefano Zampini } 4757b96c3477SStefano Zampini PetscFunctionReturn(0); 4758b96c3477SStefano Zampini } 4759b96c3477SStefano Zampini 4760b96c3477SStefano Zampini #undef __FUNCT__ 4761b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 476208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4763b96c3477SStefano Zampini { 4764b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4765b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4766b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4767b96c3477SStefano Zampini PCBDDCGraph graph; 4768b96c3477SStefano Zampini PetscErrorCode ierr; 4769b96c3477SStefano Zampini 4770b96c3477SStefano Zampini PetscFunctionBegin; 4771b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 477208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4773*3301b35fSStefano Zampini IS verticesIS,verticescomm; 4774*3301b35fSStefano Zampini PetscInt vsize,*idxs; 4775b96c3477SStefano Zampini 4776b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4777*3301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 4778*3301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 4779*3301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 4780*3301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 4781*3301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4782b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47837fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4784*3301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 4785*3301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4786b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4787b96c3477SStefano Zampini /* 4788b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4789b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4790b96c3477SStefano Zampini } 4791b96c3477SStefano Zampini */ 4792b96c3477SStefano Zampini } else { 4793b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4794b96c3477SStefano Zampini } 4795b96c3477SStefano Zampini 4796b96c3477SStefano Zampini /* sub_schurs init */ 4797a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4798a64f4aa4SStefano Zampini 4799b96c3477SStefano Zampini /* free graph struct */ 480008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4801b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4802b96c3477SStefano Zampini } 4803b96c3477SStefano Zampini PetscFunctionReturn(0); 4804b96c3477SStefano Zampini } 4805