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; 1118bec7fa6SStefano Zampini if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { /* complement of subsets, each entry is a vertex */ 11208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&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; 1338bec7fa6SStefano Zampini if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { 13408122e43SStefano Zampini const PetscInt *idxs; 13508122e43SStefano Zampini 13608122e43SStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_com,&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; 14508122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_com,&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; 16208122e43SStefano Zampini 163f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 164f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1659ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 16608122e43SStefano Zampini PetscInt j; 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 16808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 16908122e43SStefano Zampini } 17008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 17108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 17208122e43SStefano Zampini } 17308122e43SStefano Zampini } else { 17408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17608122e43SStefano Zampini } 1778bec7fa6SStefano Zampini } else { 178f6f667cfSStefano Zampini S = Sarray + cumarray; 179f6f667cfSStefano Zampini St = Starray + cumarray; 1808bec7fa6SStefano Zampini } 18108122e43SStefano Zampini 182f6f667cfSStefano Zampini /* Threshold: this is an heuristic for edges */ 183f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 184f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 185f6f667cfSStefano Zampini 1869ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 18708122e43SStefano Zampini PetscBLASInt B_itype = 1; 188f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1894c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1909552c7c7SStefano Zampini PetscInt nmin_s; 19108122e43SStefano Zampini 192f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 193fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1948bec7fa6SStefano 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]]); 195fd14bc51SStefano Zampini } 19608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 19708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 198f6f667cfSStefano 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)); 19908122e43SStefano Zampini #else 200f6f667cfSStefano 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)); 20108122e43SStefano Zampini #endif 20208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 20308122e43SStefano Zampini if (B_ierr) { 20408122e43SStefano Zampini if (B_ierr < 0 ) { 20508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 20608122e43SStefano Zampini } else if (B_ierr <= B_N) { 20708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 20808122e43SStefano Zampini } else { 2099552c7c7SStefano 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); 21008122e43SStefano Zampini } 21108122e43SStefano Zampini } 21208122e43SStefano Zampini 21308122e43SStefano Zampini if (B_neigs > nmax) { 214fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 215fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 216fd14bc51SStefano Zampini } 217f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 21808122e43SStefano Zampini B_neigs = nmax; 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 2219552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2229552c7c7SStefano Zampini if (B_neigs < nmin_s) { 22308122e43SStefano Zampini PetscBLASInt B_neigs2; 22408122e43SStefano Zampini 225f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 226f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 227fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 228fd14bc51SStefano 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); 229fd14bc51SStefano Zampini } 2309ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23108122e43SStefano Zampini PetscInt j; 23208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23408122e43SStefano Zampini } 23508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23708122e43SStefano Zampini } 23808122e43SStefano Zampini } else { 23908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24108122e43SStefano Zampini } 24208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 244f6f667cfSStefano 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)); 24508122e43SStefano Zampini #else 246f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 24708122e43SStefano Zampini #endif 24808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 24908122e43SStefano Zampini B_neigs += B_neigs2; 25008122e43SStefano Zampini } 25108122e43SStefano Zampini if (B_ierr) { 25208122e43SStefano Zampini if (B_ierr < 0 ) { 25308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 25408122e43SStefano Zampini } else if (B_ierr <= B_N) { 25508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25608122e43SStefano Zampini } else { 2579552c7c7SStefano 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); 25808122e43SStefano Zampini } 25908122e43SStefano Zampini } 260fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 261fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 26208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 26308122e43SStefano Zampini if (eigs[j] == 0.0) { 264fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 26508122e43SStefano Zampini } else { 266f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 267fd14bc51SStefano Zampini } 26808122e43SStefano Zampini } 26908122e43SStefano Zampini } 27008122e43SStefano Zampini } else { 27108122e43SStefano Zampini /* TODO */ 27208122e43SStefano Zampini } 27308122e43SStefano Zampini 2748bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2758bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2769162d606SStefano Zampini if (B_neigs) { 2779162d606SStefano 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); 278fd14bc51SStefano Zampini 279fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2809552c7c7SStefano Zampini PetscInt ii; 2819552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 282fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2839552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2849162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2859552c7c7SStefano Zampini } 2869552c7c7SStefano Zampini } 287fd14bc51SStefano Zampini } 28808122e43SStefano Zampini #if 0 2899162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29008122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29108122e43SStefano Zampini PetscScalar norm; 29208122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 2939162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 2949162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 29508122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 29608122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 29708122e43SStefano Zampini } else { 29808122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 29908122e43SStefano Zampini } 3009162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 301b1b3d7a2SStefano Zampini } 302b1b3d7a2SStefano Zampini #endif 3039162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3049162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3059162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3069162d606SStefano Zampini cum++; 30708122e43SStefano Zampini } 30808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 30908122e43SStefano Zampini } 31008122e43SStefano Zampini /* shift for next computation */ 31108122e43SStefano Zampini cumarray += subset_size*subset_size; 31208122e43SStefano Zampini } 313fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 314fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 315fd14bc51SStefano Zampini } 31608122e43SStefano Zampini 31708122e43SStefano Zampini if (mss) { 31808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 31908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 320f6f667cfSStefano Zampini /* destroy matrices (junk) */ 321f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 322f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 32308122e43SStefano Zampini } 324f6f667cfSStefano Zampini if (allocated_S_St) { 325f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 326f6f667cfSStefano Zampini } 327f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 32808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33008122e43SStefano Zampini #endif 33108122e43SStefano Zampini if (pcbddc->dbg_flag) { 3321b968477SStefano Zampini PetscInt maxneigs_r; 33308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3349b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33508122e43SStefano Zampini } 33608122e43SStefano Zampini PetscFunctionReturn(0); 33708122e43SStefano Zampini } 338b1b3d7a2SStefano Zampini 339674ae819SStefano Zampini #undef __FUNCT__ 340c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 341c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 342c8587f34SStefano Zampini { 343c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3448629588bSStefano Zampini PetscScalar *coarse_submat_vals; 345c8587f34SStefano Zampini PetscErrorCode ierr; 346c8587f34SStefano Zampini 347c8587f34SStefano Zampini PetscFunctionBegin; 348f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3495e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 350c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 351c8587f34SStefano Zampini 352684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3530fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 354684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 355c8587f34SStefano Zampini 356c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 357b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 358c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 359c8587f34SStefano Zampini } 360c8587f34SStefano Zampini 3618629588bSStefano Zampini /* 3628629588bSStefano Zampini Setup local correction and local part of coarse basis. 3638629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3648629588bSStefano Zampini */ 36547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3668629588bSStefano Zampini 3678629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3688629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3698629588bSStefano Zampini 3708629588bSStefano Zampini /* free */ 3718629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 372c8587f34SStefano Zampini PetscFunctionReturn(0); 373c8587f34SStefano Zampini } 374c8587f34SStefano Zampini 375c8587f34SStefano Zampini #undef __FUNCT__ 376674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 377674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 378674ae819SStefano Zampini { 379674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 380674ae819SStefano Zampini PetscErrorCode ierr; 381674ae819SStefano Zampini 382674ae819SStefano Zampini PetscFunctionBegin; 383674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 384674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 385674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 386674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 387785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 388674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 389f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 390f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 391785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 39363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 394674ae819SStefano Zampini PetscFunctionReturn(0); 395674ae819SStefano Zampini } 396674ae819SStefano Zampini 397674ae819SStefano Zampini #undef __FUNCT__ 398674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 399674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 400674ae819SStefano Zampini { 401674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 402674ae819SStefano Zampini PetscErrorCode ierr; 403674ae819SStefano Zampini 404674ae819SStefano Zampini PetscFunctionBegin; 405b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 406674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 407674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 408674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 409b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 410674ae819SStefano Zampini PetscFunctionReturn(0); 411674ae819SStefano Zampini } 412674ae819SStefano Zampini 413674ae819SStefano Zampini #undef __FUNCT__ 414674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 415674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 416674ae819SStefano Zampini { 417674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 41806656605SStefano Zampini PetscScalar *array; 419674ae819SStefano Zampini PetscErrorCode ierr; 420674ae819SStefano Zampini 421674ae819SStefano Zampini PetscFunctionBegin; 422674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42658da7f69SStefano Zampini } 427674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 428674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 42915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 431674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 432674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 433674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4378ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 441f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 442f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 443f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 444f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 445727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4460e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 44870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4496e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 450674ae819SStefano Zampini PetscFunctionReturn(0); 451674ae819SStefano Zampini } 452674ae819SStefano Zampini 453674ae819SStefano Zampini #undef __FUNCT__ 454f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 455f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4566bfb1811SStefano Zampini { 4576bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4586bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4596bfb1811SStefano Zampini VecType impVecType; 460e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4616bfb1811SStefano Zampini PetscErrorCode ierr; 4626bfb1811SStefano Zampini 4636bfb1811SStefano Zampini PetscFunctionBegin; 464f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 465f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 466f4ddd8eeSStefano Zampini } 467e7b262bdSStefano Zampini /* get sizes */ 468b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 469b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4706bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 471e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 472e7b262bdSStefano Zampini /* R nodes */ 473e7b262bdSStefano Zampini old_size = -1; 474e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 475e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 476e7b262bdSStefano Zampini } 477e7b262bdSStefano Zampini if (n_R != old_size) { 478e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 479e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4806bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4816bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 484e7b262bdSStefano Zampini } 485e7b262bdSStefano Zampini /* local primal dofs */ 486e7b262bdSStefano Zampini old_size = -1; 487e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 493e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4946bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 495e7b262bdSStefano Zampini } 496e7b262bdSStefano Zampini /* local explicit constraints */ 497e7b262bdSStefano Zampini old_size = -1; 498e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 499e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 500e7b262bdSStefano Zampini } 501e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 502e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50683b7ccabSStefano Zampini } 5076bfb1811SStefano Zampini PetscFunctionReturn(0); 5086bfb1811SStefano Zampini } 5096bfb1811SStefano Zampini 5106bfb1811SStefano Zampini #undef __FUNCT__ 51147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51388ebb749SStefano Zampini { 51425084f0cSStefano Zampini PetscErrorCode ierr; 51525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 51825084f0cSStefano Zampini /* submatrices of local problem */ 519*80677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52006656605SStefano Zampini /* submatrices of local coarse problem */ 52106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52225084f0cSStefano Zampini /* working matrices */ 52306656605SStefano Zampini Mat C_CR; 52425084f0cSStefano Zampini /* additional working stuff */ 52506656605SStefano Zampini PC pc_R; 52606656605SStefano Zampini Mat F; 52706656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 52806656605SStefano Zampini 52925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53006656605SStefano Zampini PetscScalar *work; 53106656605SStefano Zampini PetscInt *idx_V_B; 53206656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 534b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53545a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53688ebb749SStefano Zampini MatType impMatType; 53725084f0cSStefano Zampini /* some shortcuts to scalars */ 53806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 53988ebb749SStefano Zampini 54088ebb749SStefano Zampini PetscFunctionBegin; 541b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54288ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54388ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 544b371cd4fSStefano Zampini n_B = pcis->n_B; 545b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54688ebb749SStefano Zampini n_R = pcis->n - n_vertices; 54788ebb749SStefano Zampini 54888ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 54988ebb749SStefano Zampini impMatType = MATSEQDENSE; 55088ebb749SStefano Zampini 55188ebb749SStefano Zampini /* vertices in boundary numbering */ 552785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5530e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 55488ebb749SStefano Zampini if (i != n_vertices) { 55522d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 55688ebb749SStefano Zampini } 55788ebb749SStefano Zampini 55806656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 55906656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56806656605SStefano Zampini 56906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57006656605SStefano Zampini /* allocate workspace */ 57106656605SStefano Zampini n = 0; 57206656605SStefano Zampini if (n_constraints) { 57306656605SStefano Zampini n += n_R*n_constraints; 57406656605SStefano Zampini } 57506656605SStefano Zampini if (n_vertices) { 57606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 577*80677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57806656605SStefano Zampini } 57906656605SStefano Zampini if (!pcbddc->issym) { 58006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58206656605SStefano Zampini } 58306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58406656605SStefano Zampini 58506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59006656605SStefano Zampini if (isLU || isILU || isCHOL) { 59106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 59206656605SStefano Zampini } else { 59306656605SStefano Zampini F = NULL; 59406656605SStefano Zampini } 59506656605SStefano Zampini 59688ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 59788ebb749SStefano Zampini if (n_constraints) { 59806656605SStefano Zampini Mat M1,M2,M3; 599*80677318SStefano Zampini Mat auxmat; 60006656605SStefano Zampini IS is_aux; 601*80677318SStefano Zampini PetscScalar *array,*array2; 60206656605SStefano Zampini 603f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 604*80677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 60588ebb749SStefano Zampini 60625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 60725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6088ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 609*80677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61088ebb749SStefano Zampini 611*80677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 612*80677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 61306656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 61488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 61506656605SStefano Zampini const PetscScalar *row_cmat_values; 61606656605SStefano Zampini const PetscInt *row_cmat_indices; 61706656605SStefano Zampini PetscInt size_of_constraint,j; 61888ebb749SStefano Zampini 61906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62106656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 62206656605SStefano Zampini } 62306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62406656605SStefano Zampini } 625*80677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 62606656605SStefano Zampini if (F) { 62706656605SStefano Zampini Mat B; 62806656605SStefano Zampini 62906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 630*80677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 63206656605SStefano Zampini } else { 633*80677318SStefano Zampini PetscScalar *marr; 634*80677318SStefano Zampini 635*80677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 63606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 63706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 638*80677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 64206656605SStefano Zampini } 643*80677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini } 645*80677318SStefano Zampini if (!pcbddc->switch_static) { 646*80677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 647*80677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 648*80677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 649*80677318SStefano Zampini for (i=0;i<n_constraints;i++) { 650*80677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 651*80677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 652*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 653*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 654*80677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 655*80677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 656*80677318SStefano Zampini } 657*80677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 658*80677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 659*80677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 660*80677318SStefano Zampini } else { 661*80677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 662*80677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 66325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 664*80677318SStefano Zampini } 665*80677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 666*80677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 667*80677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 66806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 66906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 670*80677318SStefano Zampini if (isCHOL) { 671*80677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 672*80677318SStefano Zampini } else { 67325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 674*80677318SStefano Zampini } 675*80677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 67725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 67825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 67925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 680*80677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 681*80677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 682*80677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 68306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 68406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 68588ebb749SStefano Zampini } 68688ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 68788ebb749SStefano Zampini if (n_vertices) { 68806656605SStefano Zampini Mat newmat; 68906656605SStefano Zampini IS is_aux; 6903a50541eSStefano Zampini PetscInt ibs,mbs; 6913a50541eSStefano Zampini PetscBool issbaij; 6923a50541eSStefano Zampini 6933a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 6949577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 6959577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 6963a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 6979577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 6989577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 6999577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7009577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 701*80677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7029577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7039577ea80SStefano Zampini } else { 7043a50541eSStefano Zampini /* this is safe */ 7059577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 706*80677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7073a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7083a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 7093a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7103a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 71106656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7123a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 71303af80d8SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7143a50541eSStefano Zampini } else { 7159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7179577ea80SStefano Zampini } 718*80677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 719*80677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7203a50541eSStefano Zampini } 72125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 72288ebb749SStefano Zampini } 72388ebb749SStefano Zampini 72488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 725f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 72606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 72706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 72806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 72906656605SStefano Zampini } 730f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 73106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 73206656605SStefano Zampini PetscScalar *marray; 73306656605SStefano Zampini 73406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 73506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 736f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 737f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 738f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 739f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 740f4ddd8eeSStefano Zampini } 741f4ddd8eeSStefano Zampini } 74206656605SStefano Zampini 743f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 74406656605SStefano Zampini PetscScalar *marray; 74588ebb749SStefano Zampini 74606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7478eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74988ebb749SStefano Zampini } 750b9d89cd5SStefano Zampini if (!pcbddc->issym) { 75106656605SStefano Zampini n *= 2; 752f4ddd8eeSStefano Zampini } 75306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 75406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 75506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 75806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75988ebb749SStefano Zampini } 76006656605SStefano Zampini if (!pcbddc->issym) { 76106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7621b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 76406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 76588ebb749SStefano Zampini } 76688ebb749SStefano Zampini } else { 767c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 768c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7691b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 770c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 771c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 772c0553b1fSStefano Zampini } 77388ebb749SStefano Zampini } 77406656605SStefano Zampini } 77506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 77606656605SStefano Zampini /* vertices */ 77706656605SStefano Zampini if (n_vertices) { 77816f15bc4SStefano Zampini 77916f15bc4SStefano Zampini if (n_R) { 78006656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 78106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 78216f15bc4SStefano Zampini PetscScalar *x,*y; 78306656605SStefano Zampini 784*80677318SStefano Zampini ierr = PetscMemzero(work,n_R*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 78506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 78621eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 78706656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 78806656605SStefano Zampini if (F) { 78906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 79006656605SStefano Zampini } else { 79106656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 79206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 79306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 79406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 79506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 79606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 79706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 79806656605SStefano Zampini } 79906656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 80006656605SStefano Zampini } 801*80677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 80206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 80306656605SStefano Zampini if (n_constraints) { 80406656605SStefano Zampini Mat B; 805*80677318SStefano Zampini 806*80677318SStefano Zampini for (i=0;i<n_vertices;i++) { 807*80677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 808*80677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 809*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 810*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 811*80677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 812*80677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 813*80677318SStefano Zampini } 814*80677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 815*80677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 816*80677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 818*80677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 81906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 82106656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 82206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82306656605SStefano Zampini } 82406656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 825*80677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 82606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 82906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 83006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83306656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 83416f15bc4SStefano Zampini } else { 83516f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83616f15bc4SStefano Zampini } 83721eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 83806656605SStefano Zampini /* coarse basis functions */ 83906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 84016f15bc4SStefano Zampini PetscScalar *y; 84116f15bc4SStefano Zampini 84206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 84806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 85006656605SStefano Zampini 85106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85806656605SStefano Zampini } 85906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 86006656605SStefano Zampini } 86106656605SStefano Zampini } 86206656605SStefano Zampini 86306656605SStefano Zampini if (n_constraints) { 86406656605SStefano Zampini Mat B; 86506656605SStefano Zampini 86606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 868*80677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87006656605SStefano Zampini if (n_vertices) { 871*80677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 872*80677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 873*80677318SStefano Zampini } else { 874*80677318SStefano Zampini Mat S_VCt; 875*80677318SStefano Zampini 876*80677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 877*80677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 878*80677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 879*80677318SStefano Zampini } 88006656605SStefano Zampini } 88106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88206656605SStefano Zampini /* coarse basis functions */ 88306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 88406656605SStefano Zampini PetscScalar *y; 88506656605SStefano Zampini 88606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 88906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 89306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 89406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 89606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 89906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90006656605SStefano Zampini } 90106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90206656605SStefano Zampini } 90306656605SStefano Zampini } 904*80677318SStefano Zampini if (n_constraints) { 905*80677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 906*80677318SStefano Zampini } 90706656605SStefano Zampini 90806656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 90906656605SStefano Zampini if (!pcbddc->issym) { 91006656605SStefano Zampini 91106656605SStefano Zampini if (n_constraints) { 91216f15bc4SStefano Zampini Mat S_CCT,B_C; 91306656605SStefano Zampini 914*80677318SStefano Zampini /* this is a lazy thing */ 915*80677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 91916f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 92006656605SStefano Zampini if (n_vertices) { 92116f15bc4SStefano Zampini Mat B_V,S_VCT; 92206656605SStefano Zampini 92306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 92606656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 92716f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 92806656605SStefano Zampini } 92906656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 930*80677318SStefano Zampini } else { /* if there are no constraints, reset work */ 931*80677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 93206656605SStefano Zampini } 93316f15bc4SStefano Zampini if (n_vertices && n_R) { 93406656605SStefano Zampini Mat A_VRT; 935*80677318SStefano Zampini PetscScalar *marray; 93606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 93706656605SStefano Zampini 938*80677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 939*80677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 940*80677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 942*80677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 943*80677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 94416f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 94506656605SStefano Zampini } 94606656605SStefano Zampini 94706656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 94806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 94906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95406656605SStefano Zampini } 95506656605SStefano Zampini } else { 95606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96206656605SStefano Zampini } 96306656605SStefano Zampini } 96406656605SStefano Zampini /* coarse basis functions */ 96506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96606656605SStefano Zampini PetscScalar *y; 96706656605SStefano Zampini 96806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 96906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97306656605SStefano Zampini if (i<n_vertices) { 97406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 97506656605SStefano Zampini } 97606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 97806656605SStefano Zampini 97906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 98006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 98506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98606656605SStefano Zampini } 98706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98806656605SStefano Zampini } 98906656605SStefano Zampini } 99088ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 99588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 99688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99825084f0cSStefano Zampini if (pcbddc->dbg_flag) { 99988ebb749SStefano Zampini Mat coarse_sub_mat; 100025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 100188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 100288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 100388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10048bec7fa6SStefano Zampini Mat C_B,CPHI; 10058bec7fa6SStefano Zampini IS is_dummy; 10068bec7fa6SStefano Zampini Vec mones; 100788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 100888ebb749SStefano Zampini PetscReal real_value; 100988ebb749SStefano Zampini 101088ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 101188ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 101288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 101388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 101488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 101588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1016c0553b1fSStefano Zampini if (unsymmetric_check) { 101788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 101888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 101988ebb749SStefano Zampini } 102088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 102188ebb749SStefano Zampini 102225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 102306656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 102425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1025c0553b1fSStefano Zampini if (unsymmetric_check) { 102688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 102788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103888ebb749SStefano Zampini } else { 103988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104788ebb749SStefano Zampini } 104888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 105281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10538bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10540fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 105506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10568bec7fa6SStefano Zampini 10578bec7fa6SStefano Zampini /* check constraints */ 10588bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10598bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10608bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10618bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10628bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10638bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10648bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10658bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10668bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10678bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10688bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10698bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini 107125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 107388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 107888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 107988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 108088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 108188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1082c0553b1fSStefano Zampini if (unsymmetric_check) { 108388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 108488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 108588ebb749SStefano Zampini } 108688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 108788ebb749SStefano Zampini } 108888ebb749SStefano Zampini /* free memory */ 108906656605SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 109088ebb749SStefano Zampini if (n_vertices) { 109188ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 109288ebb749SStefano Zampini } 109388ebb749SStefano Zampini if (n_constraints) { 109488ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 109588ebb749SStefano Zampini } 10968629588bSStefano Zampini /* get back data */ 10978629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 109888ebb749SStefano Zampini PetscFunctionReturn(0); 109988ebb749SStefano Zampini } 110088ebb749SStefano Zampini 110188ebb749SStefano Zampini #undef __FUNCT__ 1102d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1103d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1104d65f70fdSStefano Zampini { 1105d65f70fdSStefano Zampini Mat *work_mat; 1106d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1107d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1108d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1109d65f70fdSStefano Zampini PetscErrorCode ierr; 1110d65f70fdSStefano Zampini 1111d65f70fdSStefano Zampini PetscFunctionBegin; 1112d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1113d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1114d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1115d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1116d65f70fdSStefano Zampini 1117d65f70fdSStefano Zampini if (!rsorted) { 1118d65f70fdSStefano Zampini const PetscInt *idxs; 1119d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1120d65f70fdSStefano Zampini 1121d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1122d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1123d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1124d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1125d65f70fdSStefano Zampini } 1126d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1127d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1128d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1129d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1130d65f70fdSStefano Zampini } 1131d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini } else { 1134d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini isrow_s = isrow; 1136d65f70fdSStefano Zampini } 1137d65f70fdSStefano Zampini 1138d65f70fdSStefano Zampini if (!csorted) { 1139d65f70fdSStefano Zampini if (isrow == iscol) { 1140d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini iscol_s = isrow_s; 1142d65f70fdSStefano Zampini } else { 1143d65f70fdSStefano Zampini const PetscInt *idxs; 1144d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1145d65f70fdSStefano Zampini 1146d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1149d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1150d65f70fdSStefano Zampini } 1151d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1154d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1155d65f70fdSStefano Zampini } 1156d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1157d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini } 1159d65f70fdSStefano Zampini } else { 1160d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1161d65f70fdSStefano Zampini iscol_s = iscol; 1162d65f70fdSStefano Zampini } 1163d65f70fdSStefano Zampini 1164d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini 1166d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1167d65f70fdSStefano Zampini Mat new_mat; 1168d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1169d65f70fdSStefano Zampini 1170d65f70fdSStefano Zampini if (!rsorted) { 1171d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1172d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1173d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1174d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1175d65f70fdSStefano Zampini } 1176d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1177d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini } else { 1179d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1182d65f70fdSStefano Zampini 1183d65f70fdSStefano Zampini if (!csorted) { 1184d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1185d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1187d65f70fdSStefano Zampini } else { 1188d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1189d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1190d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1191d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1192d65f70fdSStefano Zampini } 1193d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1194d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini } 1196d65f70fdSStefano Zampini } else { 1197d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini } 1199d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini 1201d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1202d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini work_mat[0] = new_mat; 1204d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1205d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1206d65f70fdSStefano Zampini } 1207d65f70fdSStefano Zampini 1208d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini *B = work_mat[0]; 1210d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini PetscFunctionReturn(0); 1214d65f70fdSStefano Zampini } 1215d65f70fdSStefano Zampini 1216d65f70fdSStefano Zampini #undef __FUNCT__ 12175e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12185e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1219aa0d41d4SStefano Zampini { 1220aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12215e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1222d65f70fdSStefano Zampini Mat new_mat; 12235e8657edSStefano Zampini IS is_local,is_global; 1224d65f70fdSStefano Zampini PetscInt local_size; 1225d65f70fdSStefano Zampini PetscBool isseqaij; 1226aa0d41d4SStefano Zampini PetscErrorCode ierr; 1227aa0d41d4SStefano Zampini 1228aa0d41d4SStefano Zampini PetscFunctionBegin; 1229aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12305e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12315e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1232906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1233906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1234d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1235906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1236906d46d4SStefano Zampini 1237906d46d4SStefano Zampini /* check */ 1238906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1239906d46d4SStefano Zampini Vec x,x_change; 1240906d46d4SStefano Zampini PetscReal error; 1241906d46d4SStefano Zampini 12425e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1243906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12445e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12455e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12465e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1247d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12485e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1250906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1252906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1253906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1254906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1255906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1256906d46d4SStefano Zampini } 1257906d46d4SStefano Zampini 125822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12599b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 126022d5777bSStefano Zampini if (isseqaij) { 1261d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1262aa0d41d4SStefano Zampini } else { 1263aa0d41d4SStefano Zampini Mat work_mat; 1264aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1265d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1266aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1267aa0d41d4SStefano Zampini } 1268d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1269e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1270e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1271e496cd5dSStefano Zampini #endif 127245a1bb75SStefano Zampini /* 127345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1274d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 127545a1bb75SStefano Zampini */ 1276d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1277aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1278aa0d41d4SStefano Zampini } 1279aa0d41d4SStefano Zampini 1280aa0d41d4SStefano Zampini #undef __FUNCT__ 1281a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12828ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1283a64d13efSStefano Zampini { 1284a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1285a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12868ce42a96SStefano Zampini IS is_aux1,is_aux2; 1287e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 12883a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12893a50541eSStefano Zampini PetscInt vbs,bs; 12904641a718SStefano Zampini PetscBT bitmask; 1291a64d13efSStefano Zampini PetscErrorCode ierr; 1292a64d13efSStefano Zampini 1293a64d13efSStefano Zampini PetscFunctionBegin; 1294b23d619eSStefano Zampini /* 1295b23d619eSStefano Zampini No need to setup local scatters if 1296b23d619eSStefano Zampini - primal space is unchanged 1297b23d619eSStefano Zampini AND 1298b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1299b23d619eSStefano Zampini AND 1300b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1301b23d619eSStefano Zampini */ 1302b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1303f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1304f4ddd8eeSStefano Zampini } 1305f4ddd8eeSStefano Zampini /* destroy old objects */ 1306f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1307f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1308f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1309a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1310b371cd4fSStefano Zampini n_B = pcis->n_B; 1311b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1312b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13134641a718SStefano Zampini /* create auxiliary bitmask */ 13144641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13154641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13160e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13174641a718SStefano Zampini } 13183a50541eSStefano Zampini 1319a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 1320854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1321a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13224641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 1323a64d13efSStefano Zampini idx_R_local[n_R] = i; 1324a64d13efSStefano Zampini n_R++; 1325a64d13efSStefano Zampini } 1326a64d13efSStefano Zampini } 13273a50541eSStefano Zampini 13283a50541eSStefano Zampini /* Block code */ 13293a50541eSStefano Zampini vbs = 1; 13303a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13313a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13323a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13333a50541eSStefano Zampini PetscInt *vary; 13343a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1335785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13363a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13370e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1338964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13393a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13403a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13413a50541eSStefano Zampini break; 13423a50541eSStefano Zampini } 13433a50541eSStefano Zampini } 13443a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13453a50541eSStefano Zampini vbs = bs; 13463a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13473a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13483a50541eSStefano Zampini } 13493a50541eSStefano Zampini } 13503a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13513a50541eSStefano Zampini } 13523a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 13533a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1354a64d13efSStefano Zampini 1355a64d13efSStefano Zampini /* print some info if requested */ 1356a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1357a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1358a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13590fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1360a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1361a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13623a50541eSStefano 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); 1363a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1364a64d13efSStefano Zampini } 1365a64d13efSStefano Zampini 1366a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13673a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1368854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1369854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1370a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 13714641a718SStefano Zampini for (i=0; i<n_D; i++) { 13724641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 13734641a718SStefano Zampini } 1374a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1375a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13764641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 13774641a718SStefano Zampini aux_array1[j++] = i; 1378a64d13efSStefano Zampini } 1379a64d13efSStefano Zampini } 1380a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1381a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1382a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 13834641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 13844641a718SStefano Zampini aux_array2[j++] = i; 1385a64d13efSStefano Zampini } 1386a64d13efSStefano Zampini } 1387a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1388a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1389a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1390a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1391a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1392a64d13efSStefano Zampini 13938eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1394785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1395a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13964641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 13974641a718SStefano Zampini aux_array1[j++] = i; 1398a64d13efSStefano Zampini } 1399a64d13efSStefano Zampini } 1400a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1401a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1402a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1403a64d13efSStefano Zampini } 14044641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14053a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1406a64d13efSStefano Zampini PetscFunctionReturn(0); 1407a64d13efSStefano Zampini } 1408a64d13efSStefano Zampini 1409304d26faSStefano Zampini 1410304d26faSStefano Zampini #undef __FUNCT__ 1411304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1412684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1413304d26faSStefano Zampini { 1414304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1415304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1416304d26faSStefano Zampini PC pc_temp; 1417304d26faSStefano Zampini Mat A_RR; 1418f4ddd8eeSStefano Zampini MatReuse reuse; 1419304d26faSStefano Zampini PetscScalar m_one = -1.0; 1420304d26faSStefano Zampini PetscReal value; 1421af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14229577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1423304d26faSStefano Zampini PetscErrorCode ierr; 1424e604994aSStefano Zampini /* prefixes stuff */ 1425312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1426e604994aSStefano Zampini size_t len; 1427304d26faSStefano Zampini 1428304d26faSStefano Zampini PetscFunctionBegin; 1429304d26faSStefano Zampini 1430e604994aSStefano Zampini /* compute prefixes */ 1431e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1432e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1433e604994aSStefano Zampini if (!pcbddc->current_level) { 1434e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1435e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1436e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1437e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1438e604994aSStefano Zampini } else { 1439e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1440312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1441e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1442e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1443312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1444312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 144534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 144634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1447e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1448e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1449e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1450e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1451e604994aSStefano Zampini } 1452e604994aSStefano Zampini 1453304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1454684f6988SStefano Zampini if (dirichlet) { 1455d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1456964fefecSStefano Zampini if (pcbddc->issym) { 1457964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1458964fefecSStefano Zampini } 1459ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1460964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1461304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1462304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1463304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1464304d26faSStefano Zampini /* default */ 1465304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1466e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 14679577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1468304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 14699577ea80SStefano Zampini if (issbaij) { 14709577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 14719577ea80SStefano Zampini } else { 1472304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 14739577ea80SStefano Zampini } 1474304d26faSStefano Zampini /* Allow user's customization */ 1475304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1476304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1477304d26faSStefano Zampini } 1478d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1479d5574798SStefano Zampini if (sub_schurs->interior_solver) { 1480d5574798SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,sub_schurs->interior_solver);CHKERRQ(ierr); 1481d5574798SStefano Zampini } 1482304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1483304d26faSStefano Zampini if (!n_D) { 1484304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1485304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1486304d26faSStefano Zampini } 1487304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1488304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1489304d26faSStefano Zampini /* set ksp_D into pcis data */ 1490304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1491304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1492304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1493684f6988SStefano Zampini } 1494304d26faSStefano Zampini 1495304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1496684f6988SStefano Zampini A_RR = 0; 1497684f6988SStefano Zampini if (neumann) { 1498f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 14998ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1500f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1501f4ddd8eeSStefano Zampini PetscInt nn_R; 150281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1503f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1504f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1505f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1506f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1507f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1508f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1509f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1510727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1511f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1512f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1513f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1514f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1515f4ddd8eeSStefano Zampini } 1516f4ddd8eeSStefano Zampini } 1517f4ddd8eeSStefano Zampini /* last check */ 1518d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1519f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1520f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1521f4ddd8eeSStefano Zampini } 1522f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1523f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1524f4ddd8eeSStefano Zampini } 1525f4ddd8eeSStefano Zampini /* extract A_RR */ 1526af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1527af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 15283a50541eSStefano Zampini if (ibs != mbs) { 1529af732b37SStefano Zampini Mat newmat; 1530af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1531f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1532af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1533af732b37SStefano Zampini } else { 1534f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1535af732b37SStefano Zampini } 1536964fefecSStefano Zampini if (pcbddc->issym) { 1537964fefecSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1538964fefecSStefano Zampini } 1539f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1540304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1541304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1542304d26faSStefano Zampini /* default */ 1543304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1544e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1545304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 15469577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 15479577ea80SStefano Zampini if (issbaij) { 15489577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15499577ea80SStefano Zampini } else { 1550304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15519577ea80SStefano Zampini } 1552304d26faSStefano Zampini /* Allow user's customization */ 1553304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1554304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1555304d26faSStefano Zampini } 1556d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1557304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1558304d26faSStefano Zampini if (!n_R) { 1559304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1560304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1561304d26faSStefano Zampini } 1562304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1563304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1564684f6988SStefano Zampini } 1565304d26faSStefano Zampini 1566304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 15670fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1568684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1569684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1570684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1571684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1572684f6988SStefano Zampini } 1573684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 15740fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 15750fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 15760fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 15770fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 15780fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1579304d26faSStefano Zampini /* need to be adapted? */ 1580b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1581b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1582b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1583304d26faSStefano Zampini /* print info */ 1584304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1585e604994aSStefano 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); 1586304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1587304d26faSStefano Zampini } 1588b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 15898ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1590304d26faSStefano Zampini } 1591684f6988SStefano Zampini } 1592684f6988SStefano Zampini if (neumann) { /* Neumann */ 15930fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 15940fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 15950fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 15960fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 15970fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1598304d26faSStefano Zampini /* need to be adapted? */ 1599b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1600b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1601304d26faSStefano Zampini /* print info */ 1602304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1603e604994aSStefano 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); 1604304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1605304d26faSStefano Zampini } 1606b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16078ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1608304d26faSStefano Zampini } 16090fccc4e9SStefano Zampini } 1610684f6988SStefano Zampini } 1611304d26faSStefano Zampini /* free Neumann problem's matrix */ 1612304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1613304d26faSStefano Zampini PetscFunctionReturn(0); 1614304d26faSStefano Zampini } 1615304d26faSStefano Zampini 1616304d26faSStefano Zampini #undef __FUNCT__ 1617ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 1618*80677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1619674ae819SStefano Zampini { 1620674ae819SStefano Zampini PetscErrorCode ierr; 1621674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1622674ae819SStefano Zampini 1623674ae819SStefano Zampini PetscFunctionBegin; 1624*80677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1625*80677318SStefano Zampini if (!pcbddc->switch_static) { 1626*80677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 1627*80677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 1628*80677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 162920c7b377SStefano Zampini } 1630*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1631*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163220c7b377SStefano Zampini } else { 1633*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1634*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1635*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1636*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1637*80677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 1638*80677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 1639*80677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 1640*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1641*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 164220c7b377SStefano Zampini } 1643674ae819SStefano Zampini } 1644*80677318SStefano Zampini if (applytranspose) { 1645*80677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 1646*80677318SStefano Zampini } else { 1647*80677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 1648*80677318SStefano Zampini } 1649*80677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 1650*80677318SStefano Zampini if (!pcbddc->switch_static) { 1651*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1652*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1653*80677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 1654*80677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 1655*80677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 1656*80677318SStefano Zampini } 1657*80677318SStefano Zampini } else { 1658*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1659*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1660*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1661*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1662*80677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 1663*80677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 1664*80677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 1665*80677318SStefano Zampini } 1666*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1667*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1668*80677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1669*80677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1670*80677318SStefano Zampini } 1671674ae819SStefano Zampini PetscFunctionReturn(0); 1672674ae819SStefano Zampini } 1673674ae819SStefano Zampini 1674dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1675674ae819SStefano Zampini #undef __FUNCT__ 1676674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1677dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1678674ae819SStefano Zampini { 1679674ae819SStefano Zampini PetscErrorCode ierr; 1680674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1681674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1682674ae819SStefano Zampini const PetscScalar zero = 0.0; 1683674ae819SStefano Zampini 1684674ae819SStefano Zampini PetscFunctionBegin; 1685dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1686dc359a40SStefano Zampini if (applytranspose) { 1687674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 16888eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1689dc359a40SStefano Zampini } else { 1690674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1691674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 169215aaf578SStefano Zampini } 169312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 169412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 169512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 169612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 169712edc857SStefano Zampini 16989f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 169912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 170012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1701964fefecSStefano Zampini Vec rhs,sol; 1702964fefecSStefano Zampini 1703964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1704964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 170512edc857SStefano Zampini if (applytranspose) { 1706964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 170712edc857SStefano Zampini } else { 1708964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 170912edc857SStefano Zampini } 171012edc857SStefano Zampini } 1711674ae819SStefano Zampini 1712674ae819SStefano Zampini /* Local solution on R nodes */ 1713*80677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 1714*80677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 17159f00e9b4SStefano Zampini } 1716674ae819SStefano Zampini 17179f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 17189f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1720674ae819SStefano Zampini 1721674ae819SStefano Zampini /* Sum contributions from two levels */ 1722dc359a40SStefano Zampini if (applytranspose) { 1723dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1724dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1725dc359a40SStefano Zampini } else { 1726674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 17278eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1728dc359a40SStefano Zampini } 1729674ae819SStefano Zampini PetscFunctionReturn(0); 1730674ae819SStefano Zampini } 1731674ae819SStefano Zampini 1732674ae819SStefano Zampini #undef __FUNCT__ 1733674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 173412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1735674ae819SStefano Zampini { 1736674ae819SStefano Zampini PetscErrorCode ierr; 1737674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 173858da7f69SStefano Zampini PetscScalar *array; 173912edc857SStefano Zampini Vec from,to; 1740674ae819SStefano Zampini 1741674ae819SStefano Zampini PetscFunctionBegin; 174212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 174312edc857SStefano Zampini from = pcbddc->coarse_vec; 174412edc857SStefano Zampini to = pcbddc->vec1_P; 174512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 174612edc857SStefano Zampini Vec tvec; 174758da7f69SStefano Zampini 174858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 174958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 175012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 175158da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 175258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 175358da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 175412edc857SStefano Zampini } 175512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 175612edc857SStefano Zampini from = pcbddc->vec1_P; 175712edc857SStefano Zampini to = pcbddc->coarse_vec; 175812edc857SStefano Zampini } 175912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1760674ae819SStefano Zampini PetscFunctionReturn(0); 1761674ae819SStefano Zampini } 1762674ae819SStefano Zampini 1763674ae819SStefano Zampini #undef __FUNCT__ 1764674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 176512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1766674ae819SStefano Zampini { 1767674ae819SStefano Zampini PetscErrorCode ierr; 1768674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 176958da7f69SStefano Zampini PetscScalar *array; 177012edc857SStefano Zampini Vec from,to; 1771674ae819SStefano Zampini 1772674ae819SStefano Zampini PetscFunctionBegin; 177312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 177412edc857SStefano Zampini from = pcbddc->coarse_vec; 177512edc857SStefano Zampini to = pcbddc->vec1_P; 177612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 177712edc857SStefano Zampini from = pcbddc->vec1_P; 177812edc857SStefano Zampini to = pcbddc->coarse_vec; 177912edc857SStefano Zampini } 178012edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 178112edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 178212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 178312edc857SStefano Zampini Vec tvec; 178458da7f69SStefano Zampini 178512edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 178658da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 178758da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 178858da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 178958da7f69SStefano Zampini } 179058da7f69SStefano Zampini } else { 179158da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 179258da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 179312edc857SStefano Zampini } 179412edc857SStefano Zampini } 1795674ae819SStefano Zampini PetscFunctionReturn(0); 1796674ae819SStefano Zampini } 1797674ae819SStefano Zampini 1798984c4197SStefano Zampini /* uncomment for testing purposes */ 1799984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1800674ae819SStefano Zampini #undef __FUNCT__ 1801674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1802674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1803674ae819SStefano Zampini { 1804674ae819SStefano Zampini PetscErrorCode ierr; 1805674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1806674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1807674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1808984c4197SStefano Zampini /* one and zero */ 1809984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1810984c4197SStefano Zampini /* space to store constraints and their local indices */ 18119162d606SStefano Zampini PetscScalar *constraints_data; 18129162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 18139162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 18149162d606SStefano Zampini PetscInt *constraints_n; 1815984c4197SStefano Zampini /* iterators */ 18169162d606SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc; 1817984c4197SStefano Zampini /* BLAS integers */ 1818e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1819e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1820c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1821727cdba6SStefano Zampini /* reuse */ 18220e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 18230e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1824984c4197SStefano Zampini /* change of basis */ 1825a717540cSStefano Zampini PetscBool boolforchange,qr_needed; 18269162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1827984c4197SStefano Zampini /* auxiliary stuff */ 182864efe560SStefano Zampini PetscInt *nnz,*is_indices; 18298a0068c3SStefano Zampini PetscInt ncc; 1830984c4197SStefano Zampini /* some quantities */ 183145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1832a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1833984c4197SStefano Zampini 1834674ae819SStefano Zampini PetscFunctionBegin; 18358e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 18368e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 18378e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1838088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1839088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 18400e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 18410e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 18420e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 18430e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 18440e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1845088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1846cf5a6209SStefano Zampini 1847cf5a6209SStefano Zampini /* print some info */ 1848cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1849cf5a6209SStefano Zampini IS vertices; 1850cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1851cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1852cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1853cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1854cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1855cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1856cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1857fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1858fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1859cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1860cf5a6209SStefano Zampini } 1861cf5a6209SStefano Zampini 1862cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 18639162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1864cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1865cf5a6209SStefano Zampini const Vec *nearnullvecs; 1866cf5a6209SStefano Zampini Vec *localnearnullsp; 1867cf5a6209SStefano Zampini PetscScalar *array; 1868cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1869cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1870cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1871cf5a6209SStefano Zampini PetscBool skip_lapack; 1872cf5a6209SStefano Zampini PetscScalar *work; 1873cf5a6209SStefano Zampini PetscReal *singular_vals; 1874cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1875cf5a6209SStefano Zampini PetscReal *rwork; 1876cf5a6209SStefano Zampini #endif 1877cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1878cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1879cf5a6209SStefano Zampini #else 1880964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1881964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1882cf5a6209SStefano Zampini #endif 1883cf5a6209SStefano Zampini 1884674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1885d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1886d06fc5fdSStefano Zampini /* free unneeded index sets */ 1887d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1888d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1889674ae819SStefano Zampini } 1890d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1891d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1892d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1893d06fc5fdSStefano Zampini } 1894d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1895d06fc5fdSStefano Zampini n_ISForEdges = 0; 1896d06fc5fdSStefano Zampini } 1897d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1898d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1899d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1900d06fc5fdSStefano Zampini } 1901d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1902d06fc5fdSStefano Zampini n_ISForFaces = 0; 1903d06fc5fdSStefano Zampini } 190470022509SStefano Zampini 190570022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 190670022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 190770022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 190870022509SStefano Zampini if (pcbddc->NullSpace) { 190970022509SStefano Zampini PetscBool tbool[2],gbool[2]; 191070022509SStefano Zampini 191170022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 1912b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1913d06fc5fdSStefano Zampini if (!ISForEdges) { 1914d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1915d06fc5fdSStefano Zampini } 1916b8ffe317SStefano Zampini } 1917d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 1918d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 1919d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1920d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1921d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 192298a51de6SStefano Zampini } 192370022509SStefano Zampini #endif 192408122e43SStefano Zampini 1925674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1926674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1927674ae819SStefano Zampini if (nearnullsp) { 1928674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1929f4ddd8eeSStefano Zampini /* remove any stored info */ 1930f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1931f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1932f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1933f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1934f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1935473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1936f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1937f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1938f4ddd8eeSStefano Zampini } 1939984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1940984c4197SStefano Zampini nnsp_size = 0; 1941674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1942674ae819SStefano Zampini } 1943984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1944984c4197SStefano Zampini max_constraints = nnsp_size; 1945984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1946984c4197SStefano Zampini 1947674ae819SStefano Zampini /* 1948674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 19499162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 19509162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 19519162d606SStefano Zampini There can be multiple constraints per connected component 1952674ae819SStefano Zampini */ 1953674ae819SStefano Zampini n_vertices = 0; 1954674ae819SStefano Zampini if (ISForVertices) { 1955674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1956674ae819SStefano Zampini } 19579162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 19589162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 19599162d606SStefano Zampini 19609162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 19619162d606SStefano Zampini total_counts *= max_constraints; 1962674ae819SStefano Zampini total_counts += n_vertices; 19634641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 19649162d606SStefano Zampini 1965674ae819SStefano Zampini total_counts = 0; 1966674ae819SStefano Zampini max_size_of_constraint = 0; 1967674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 19689162d606SStefano Zampini IS used_is; 1969674ae819SStefano Zampini if (i<n_ISForEdges) { 19709162d606SStefano Zampini used_is = ISForEdges[i]; 1971674ae819SStefano Zampini } else { 19729162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 1973674ae819SStefano Zampini } 19749162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 1975674ae819SStefano Zampini total_counts += j; 1976674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1977674ae819SStefano Zampini } 19789162d606SStefano 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); 19799162d606SStefano Zampini 1980984c4197SStefano Zampini /* get local part of global near null space vectors */ 1981785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1982984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1983984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1984984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1985984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1986984c4197SStefano Zampini } 1987674ae819SStefano Zampini 1988242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1989242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1990a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1991242a89d7SStefano Zampini 1992984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1993a773dcb8SStefano Zampini if (!skip_lapack) { 1994674ae819SStefano Zampini PetscScalar temp_work; 1995911cabfeSStefano Zampini 1996674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1997984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1998785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 1999785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2000785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2001674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2002785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2003674ae819SStefano Zampini #endif 2004674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2005c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2006c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2007674ae819SStefano Zampini lwork = -1; 2008674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2009674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2010c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2011674ae819SStefano Zampini #else 2012c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2013674ae819SStefano Zampini #endif 2014674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2015984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2016674ae819SStefano Zampini #else /* on missing GESVD */ 2017674ae819SStefano Zampini /* SVD */ 2018674ae819SStefano Zampini PetscInt max_n,min_n; 2019674ae819SStefano Zampini max_n = max_size_of_constraint; 2020984c4197SStefano Zampini min_n = max_constraints; 2021984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2022674ae819SStefano Zampini min_n = max_size_of_constraint; 2023984c4197SStefano Zampini max_n = max_constraints; 2024674ae819SStefano Zampini } 2025785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2026674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2027785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2028674ae819SStefano Zampini #endif 2029674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2030674ae819SStefano Zampini lwork = -1; 2031e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2032e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2033b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2034674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2035674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20369162d606SStefano 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)); 2037674ae819SStefano Zampini #else 20389162d606SStefano 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)); 2039674ae819SStefano Zampini #endif 2040674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2041984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2042984c4197SStefano Zampini #endif /* on missing GESVD */ 2043674ae819SStefano Zampini /* Allocate optimal workspace */ 2044674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2045854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2046674ae819SStefano Zampini } 2047674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2048674ae819SStefano Zampini total_counts = 0; 20499162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 20509162d606SStefano Zampini constraints_data_ptr[0] = 0; 2051674ae819SStefano Zampini /* vertices */ 20529162d606SStefano Zampini if (n_vertices) { 2053674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 20549162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 20559162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2056674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 20579162d606SStefano Zampini constraints_n[total_counts] = 1; 20589162d606SStefano Zampini constraints_data[total_counts] = 1.0; 20599162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 20609162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2061674ae819SStefano Zampini total_counts++; 2062674ae819SStefano Zampini } 2063674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2064984c4197SStefano Zampini PetscBool used_vertex; 2065674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2066674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2067674ae819SStefano Zampini k = 0; 2068674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2069984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2070984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 20719162d606SStefano Zampini constraints_n[total_counts] = 1; 20729162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 20739162d606SStefano Zampini constraints_data[total_counts] = 1.0; 20749162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 20759162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2076674ae819SStefano Zampini total_counts++; 2077674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2078674ae819SStefano Zampini } 2079984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2080674ae819SStefano Zampini k++; 2081674ae819SStefano Zampini } 2082674ae819SStefano Zampini } 2083674ae819SStefano Zampini } 2084674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2085674ae819SStefano Zampini n_vertices = total_counts; 2086674ae819SStefano Zampini } 2087984c4197SStefano Zampini 2088674ae819SStefano Zampini /* edges and faces */ 20899162d606SStefano Zampini total_counts_cc = total_counts; 2090911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 20919162d606SStefano Zampini IS used_is; 20929162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 20939162d606SStefano Zampini 2094911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 20959162d606SStefano Zampini used_is = ISForEdges[ncc]; 2096984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2097674ae819SStefano Zampini } else { 20989162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2099984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2100674ae819SStefano Zampini } 2101674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 21029162d606SStefano Zampini 21039162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 21049162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2105984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2106984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2107674ae819SStefano Zampini if (nnsp_has_cnst) { 21085b08dc53SStefano Zampini PetscScalar quad_value; 21099162d606SStefano Zampini 21109162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21119162d606SStefano Zampini idxs_copied = PETSC_TRUE; 21129162d606SStefano Zampini 2113a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2114674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2115a773dcb8SStefano Zampini } else { 2116a773dcb8SStefano Zampini quad_value = 1.0; 2117a773dcb8SStefano Zampini } 2118674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21199162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2120674ae819SStefano Zampini } 21219162d606SStefano Zampini temp_constraints++; 2122674ae819SStefano Zampini total_counts++; 2123674ae819SStefano Zampini } 2124674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2125984c4197SStefano Zampini PetscReal real_value; 21269162d606SStefano Zampini PetscScalar *ptr_to_data; 21279162d606SStefano Zampini 2128984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 21299162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2130674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21319162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2132674ae819SStefano Zampini } 2133984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2134984c4197SStefano Zampini /* check if array is null on the connected component */ 2135e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21369162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 21375b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2138674ae819SStefano Zampini temp_constraints++; 2139674ae819SStefano Zampini total_counts++; 21409162d606SStefano Zampini if (!idxs_copied) { 21419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21429162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2143674ae819SStefano Zampini } 2144674ae819SStefano Zampini } 21459162d606SStefano Zampini } 21469162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 214745a1bb75SStefano Zampini valid_constraints = temp_constraints; 2148eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2149a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 21509162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 21519162d606SStefano Zampini 21529162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2153a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21549162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2155a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 21569162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2157a773dcb8SStefano Zampini } else { /* perform SVD */ 2158984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 21599162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2160674ae819SStefano Zampini 2161674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2162984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2163984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2164984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2165984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2166984c4197SStefano Zampini from that computed using LAPACKgesvd 2167984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2168984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2169984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2170674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2171e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2172984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2173674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2174674ae819SStefano Zampini for (k=0;k<j+1;k++) { 21759162d606SStefano 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)); 2176674ae819SStefano Zampini } 2177674ae819SStefano Zampini } 2178e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2179e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2180e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2181674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2182c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2183674ae819SStefano Zampini #else 2184c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2185674ae819SStefano Zampini #endif 2186674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2187984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2188984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2189674ae819SStefano Zampini j = 0; 2190984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2191674ae819SStefano Zampini total_counts = total_counts-j; 219245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2193e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2194c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2195c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2196c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2197c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2198c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2199c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2200674ae819SStefano Zampini if (j<temp_constraints) { 2201984c4197SStefano Zampini PetscInt ii; 2202984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2203674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 22049162d606SStefano 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)); 2205674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2206984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2207674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 22089162d606SStefano 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]; 2209674ae819SStefano Zampini } 2210674ae819SStefano Zampini } 2211674ae819SStefano Zampini } 2212674ae819SStefano Zampini #else /* on missing GESVD */ 2213e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2214e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2215b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2216674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2217674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 22189162d606SStefano 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)); 2219674ae819SStefano Zampini #else 22209162d606SStefano 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)); 2221674ae819SStefano Zampini #endif 2222984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2223674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2224984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2225e310c8b4SStefano Zampini k = temp_constraints; 2226e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2227674ae819SStefano Zampini j = 0; 2228e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 222945a1bb75SStefano Zampini valid_constraints = k-j; 2230911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2231984c4197SStefano Zampini #endif /* on missing GESVD */ 2232674ae819SStefano Zampini } 2233a773dcb8SStefano Zampini } 22349162d606SStefano Zampini /* update pointers information */ 22359162d606SStefano Zampini if (valid_constraints) { 22369162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 22379162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 22389162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 22399162d606SStefano Zampini total_counts_cc++; 22409162d606SStefano Zampini /* set change_of_basis flag */ 224145a1bb75SStefano Zampini if (boolforchange) { 22429162d606SStefano Zampini for (j=valid_constraints;j>0;j--) { 22439162d606SStefano Zampini PetscBTSet(change_basis,total_counts-j); 22449162d606SStefano Zampini } 224545a1bb75SStefano Zampini } 224645a1bb75SStefano Zampini } 2247674ae819SStefano Zampini } 2248984c4197SStefano Zampini /* free workspace */ 22498f1c130eSStefano Zampini if (!skip_lapack) { 2250984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2251984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2252984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2253984c4197SStefano Zampini #endif 2254984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2255984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2256984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2257984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2258984c4197SStefano Zampini #endif 2259984c4197SStefano Zampini } 2260984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2261984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2262984c4197SStefano Zampini } 2263984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2264cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2265cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2266cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2267cf5a6209SStefano Zampini } 2268cf5a6209SStefano Zampini if (n_ISForFaces) { 2269cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2270cf5a6209SStefano Zampini } 2271cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2272cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2273cf5a6209SStefano Zampini } 2274cf5a6209SStefano Zampini if (n_ISForEdges) { 2275cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2276cf5a6209SStefano Zampini } 2277cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 227808122e43SStefano Zampini } else { 227908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 228008122e43SStefano Zampini 228108122e43SStefano Zampini total_counts = 0; 228208122e43SStefano Zampini n_vertices = 0; 22838bec7fa6SStefano Zampini if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { 228408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&n_vertices);CHKERRQ(ierr); 228508122e43SStefano Zampini } 228608122e43SStefano Zampini max_constraints = 0; 22879162d606SStefano Zampini total_counts_cc = 0; 228808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 228908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 22909162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 229108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 229208122e43SStefano Zampini } 22939162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 22949162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 22959162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 22969162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 229774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 22989162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 22999162d606SStefano Zampini total_counts_cc = 0; 23009162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 23019162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 23029162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 230308122e43SStefano Zampini } 230408122e43SStefano Zampini } 23059162d606SStefano Zampini #if 0 23069162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 23079162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23089162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 23099162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 23109162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 23119162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 23129162d606SStefano Zampini } 23139162d606SStefano Zampini printf("\n"); 23149162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 23159162d606SStefano Zampini } 23161b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 23178bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 23181b968477SStefano Zampini } 23191b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 23208bec7fa6SStefano 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]); 23211b968477SStefano Zampini } 232208122e43SStefano Zampini #endif 232308122e43SStefano Zampini 23248bec7fa6SStefano Zampini max_size_of_constraint = 0; 23259162d606SStefano 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]); 23269162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 232708122e43SStefano Zampini /* Change of basis */ 232808122e43SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 232908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 23309162d606SStefano Zampini PetscInt cum = n_vertices; 233108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 233208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2333bbb9e6c6SStefano Zampini for (j=0;j<pcbddc->adaptive_constraints_n[i+n_vertices];j++) { 233408122e43SStefano Zampini ierr = PetscBTSet(change_basis,cum+j);CHKERRQ(ierr); 233508122e43SStefano Zampini } 233608122e43SStefano Zampini } 2337bbb9e6c6SStefano Zampini cum += pcbddc->adaptive_constraints_n[i+n_vertices]; 233808122e43SStefano Zampini } 233908122e43SStefano Zampini } 234008122e43SStefano Zampini } 2341b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 23420e6343abSStefano Zampini pcbddc->local_primal_size_cc = total_counts_cc; 23430e6343abSStefano 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); 23440e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 234508122e43SStefano Zampini 23469162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 23479162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 23489162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 23499162d606SStefano 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); 235008122e43SStefano Zampini } 2351984c4197SStefano Zampini 2352674ae819SStefano Zampini /* Create constraint matrix */ 2353674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 235416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2355984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2356984c4197SStefano Zampini 2357984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 235874d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 235974d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 236074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2361984c4197SStefano Zampini total_primal_vertices=0; 23629162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23639162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2364984c4197SStefano Zampini if (size_of_constraint == 1) { 23659162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 236664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 23679162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 23689162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 23699162d606SStefano Zampini } 237074d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 237174d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 237274d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 237374d5cdf7SStefano Zampini } 2374984c4197SStefano Zampini } 2375984c4197SStefano Zampini } 2376b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2377b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 237870022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 237970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 23800e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 23810e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 238270022509SStefano Zampini 2383984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 238474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2385785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2386984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 238774d5cdf7SStefano Zampini 2388984c4197SStefano Zampini j = total_primal_vertices; 238974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 23909162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 23914641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 23920e6343abSStefano Zampini pcbddc->local_primal_ref_node[i] = constraints_idxs[constraints_idxs_ptr[i]]; 23930e6343abSStefano Zampini pcbddc->local_primal_ref_mult[i] = constraints_n[i]; 23949162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 239574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 239674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 239774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 239874d5cdf7SStefano Zampini } 23999162d606SStefano Zampini j += constraints_n[i]; 2400674ae819SStefano Zampini } 2401674ae819SStefano Zampini } 2402674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2403674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2404088faed8SStefano Zampini 2405674ae819SStefano Zampini /* set values in constraint matrix */ 2406984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 24070e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2408674ae819SStefano Zampini } 2409984c4197SStefano Zampini total_counts = total_primal_vertices; 24109162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24114641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 24129162d606SStefano Zampini PetscInt *cols; 24139162d606SStefano Zampini 24149162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 24159162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 24169162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24179162d606SStefano Zampini PetscInt row = total_counts+k; 24189162d606SStefano Zampini PetscScalar *vals; 24199162d606SStefano Zampini 24209162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 24219162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 24229162d606SStefano Zampini } 24239162d606SStefano Zampini total_counts += constraints_n[i]; 2424674ae819SStefano Zampini } 2425674ae819SStefano Zampini } 2426674ae819SStefano Zampini /* assembling */ 2427674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2428674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2429088faed8SStefano Zampini 2430984c4197SStefano Zampini /* 243145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2432984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2433984c4197SStefano Zampini */ 2434674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2435674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2436026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2437984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2438984c4197SStefano Zampini /* working stuff for GEQRF */ 243981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2440984c4197SStefano Zampini PetscBLASInt lqr_work; 2441984c4197SStefano Zampini /* working stuff for UNGQR */ 2442984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2443984c4197SStefano Zampini PetscBLASInt lgqr_work; 2444984c4197SStefano Zampini /* working stuff for TRTRS */ 2445984c4197SStefano Zampini PetscScalar *trs_rhs; 24463f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2447984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2448984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2449984c4197SStefano Zampini PetscScalar *start_vals; 2450984c4197SStefano Zampini /* working stuff for values insertion */ 24514641a718SStefano Zampini PetscBT is_primal; 245264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2453906d46d4SStefano Zampini /* matrix sizes */ 2454906d46d4SStefano Zampini PetscInt global_size,local_size; 2455906d46d4SStefano Zampini /* temporary change of basis */ 2456906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2457cf5a6209SStefano Zampini /* extra space for debugging */ 2458cf5a6209SStefano Zampini PetscScalar *dbg_work; 2459984c4197SStefano Zampini 2460906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2461906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 246216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2463bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2464906d46d4SStefano Zampini /* nonzeros for local mat */ 2465bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2466bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 24679162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2468a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 24699162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2470a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 24719162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2472a717540cSStefano Zampini } else { 24739162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 24749162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2475a717540cSStefano Zampini } 2476a717540cSStefano Zampini } 2477a717540cSStefano Zampini } 2478906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2479bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2480a717540cSStefano Zampini /* Set initial identity in the matrix */ 2481bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2482906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2483a717540cSStefano Zampini } 2484a717540cSStefano Zampini 2485a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2486a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2487a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2488a717540cSStefano Zampini } 2489a717540cSStefano Zampini 2490a717540cSStefano Zampini 2491a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2492a717540cSStefano Zampini /* 2493a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2494a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2495a717540cSStefano Zampini 2496a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2497a717540cSStefano Zampini 2498a6b551f4SStefano 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) 2499a6b551f4SStefano Zampini 2500a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2501a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2502a717540cSStefano Zampini | ... | 2503a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2504a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2505a717540cSStefano Zampini 2506a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2507a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2508a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2509a6b551f4SStefano Zampini 2510a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2511a717540cSStefano Zampini */ 2512a717540cSStefano Zampini if (qr_needed) { 2513984c4197SStefano Zampini /* space to store Q */ 2514854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2515984c4197SStefano Zampini /* first we issue queries for optimal work */ 25163f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25173f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 25183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2519984c4197SStefano Zampini lqr_work = -1; 25203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2521984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2522984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2523785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2524984c4197SStefano Zampini lgqr_work = -1; 25253f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25263f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 25273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 25283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25293f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 25303f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2531984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2532984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2533785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2534984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2535785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2536984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2537785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2538a717540cSStefano Zampini /* allocating workspace for check */ 2539a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2540cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2541a717540cSStefano Zampini } 2542a717540cSStefano Zampini } 2543984c4197SStefano Zampini /* array to store whether a node is primal or not */ 25444641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2545473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 25460e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 254739e2fb2aSStefano Zampini if (i != total_primal_vertices) { 254839e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 25494641a718SStefano Zampini } 255039e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 255139e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 255239e2fb2aSStefano Zampini } 255339e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2554984c4197SStefano Zampini 2555a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 25569162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 25579162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 25584641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2559984c4197SStefano Zampini /* get constraint info */ 25609162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2561984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2562984c4197SStefano Zampini 2563984c4197SStefano Zampini if (pcbddc->dbg_flag) { 25649162d606SStefano 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); 2565674ae819SStefano Zampini } 2566984c4197SStefano Zampini 2567fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2568a717540cSStefano Zampini 2569a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2570a717540cSStefano Zampini if (pcbddc->dbg_flag) { 25719162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2572a717540cSStefano Zampini } 2573984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 25749162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2575984c4197SStefano Zampini 2576984c4197SStefano Zampini /* compute QR decomposition of constraints */ 25773f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25783f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2580674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25813f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2582984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2583674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2584984c4197SStefano Zampini 2585984c4197SStefano Zampini /* explictly compute R^-T */ 2586984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2587984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 25883f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 25903f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2592984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25933f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2594984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2595984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2596984c4197SStefano Zampini 2597a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 25983f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26003f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2602984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26033f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2604984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2605984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2606984c4197SStefano Zampini 2607984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2608984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2609984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 26103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26113f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 26153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2616984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26179162d606SStefano 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)); 2618984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 26199162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2620984c4197SStefano Zampini 2621984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 26229162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2623984c4197SStefano Zampini /* insert cols for primal dofs */ 2624984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2625984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 26269162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2627906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2628984c4197SStefano Zampini } 2629984c4197SStefano Zampini /* insert cols for dual dofs */ 2630984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 26319162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2632984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 26339162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2634906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2635984c4197SStefano Zampini j++; 2636674ae819SStefano Zampini } 2637674ae819SStefano Zampini } 2638984c4197SStefano Zampini 2639984c4197SStefano Zampini /* check change of basis */ 2640984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2641984c4197SStefano Zampini PetscInt ii,jj; 2642984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2643c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2644c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2645c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2646c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2647c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2648c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2649984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2650cf5a6209SStefano 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)); 2651984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2652984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2653984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2654cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2655cf5a6209SStefano 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; 2656674ae819SStefano Zampini } 2657674ae819SStefano Zampini } 2658984c4197SStefano Zampini if (!valid_qr) { 265922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2660984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2661984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2662cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2663cf5a6209SStefano 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])); 2664674ae819SStefano Zampini } 2665cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2666cf5a6209SStefano 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])); 2667984c4197SStefano Zampini } 2668984c4197SStefano Zampini } 2669984c4197SStefano Zampini } 2670674ae819SStefano Zampini } else { 267122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2672674ae819SStefano Zampini } 2673674ae819SStefano Zampini } 2674a717540cSStefano Zampini } else { /* simple transformation block */ 2675a717540cSStefano Zampini PetscInt row,col; 2676a6b551f4SStefano Zampini PetscScalar val,norm; 2677a6b551f4SStefano Zampini 2678a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26799162d606SStefano 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)); 2680a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 26819162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 26829162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2683bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 26849162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2685906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 26869162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2687a717540cSStefano Zampini } else { 2688a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 26899162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2690a717540cSStefano Zampini if (row != col) { 26919162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2692a717540cSStefano Zampini } else { 26939162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2694a717540cSStefano Zampini } 2695906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2696a717540cSStefano Zampini } 2697a717540cSStefano Zampini } 2698a717540cSStefano Zampini } 269998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 270022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2701a717540cSStefano Zampini } 2702674ae819SStefano Zampini } 2703984c4197SStefano Zampini } else { 2704984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27059162d606SStefano 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); 2706674ae819SStefano Zampini } 2707674ae819SStefano Zampini } 2708674ae819SStefano Zampini } 2709a717540cSStefano Zampini 2710a717540cSStefano Zampini /* free workspace */ 2711a717540cSStefano Zampini if (qr_needed) { 2712984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2713cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2714984c4197SStefano Zampini } 2715984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2716984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2717984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2718984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2719984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2720674ae819SStefano Zampini } 2721a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2722906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2723906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2724906d46d4SStefano Zampini 2725906d46d4SStefano Zampini /* assembling of global change of variable */ 2726bbb9e6c6SStefano Zampini { 2727bbb9e6c6SStefano Zampini Mat tmat; 272816f15bc4SStefano Zampini PetscInt bs; 272916f15bc4SStefano Zampini 2730906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2731906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2732bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2733bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2734bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2735bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 273616f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 273716f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2738906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2739bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2740bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2741bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2742bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2743bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2744bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2745bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2746bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2747bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2748906d46d4SStefano Zampini } 2749906d46d4SStefano Zampini /* check */ 2750906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2751906d46d4SStefano Zampini PetscReal error; 2752906d46d4SStefano Zampini Vec x,x_change; 2753906d46d4SStefano Zampini 2754906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2755906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2756906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2757906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2758bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2759bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2760bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2761bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2762bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2763906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2764906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2765906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2766906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2767bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2768906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2769906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2770906d46d4SStefano Zampini } 2771b96c3477SStefano Zampini 2772b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2773b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2774b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2775b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2776ac632422SStefano Zampini Mat S_new,tmat; 2777bbb9e6c6SStefano Zampini IS is_all_N; 2778bbb9e6c6SStefano Zampini 2779bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 2780d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr); 2781bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2782ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2783b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2784ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2785ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2786ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2787ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2788ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2789b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2790ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2791ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2792ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2793ac632422SStefano Zampini } 2794b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2795b96c3477SStefano Zampini } 2796b96c3477SStefano Zampini } 2797906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2798906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2799b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2800b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2801b9b85e73SStefano Zampini } 2802906d46d4SStefano Zampini 2803906d46d4SStefano Zampini /* set up change of basis context */ 2804906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2805906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2806906d46d4SStefano Zampini 2807906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2808906d46d4SStefano Zampini PetscInt global_size,local_size; 2809906d46d4SStefano Zampini 2810906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2811906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2812906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2813906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2814906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2815906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2816906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2817906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2818906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2819906d46d4SStefano Zampini } else { 2820906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2821906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2822906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2823906d46d4SStefano Zampini } 2824906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2825906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2826906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2827906d46d4SStefano Zampini } else { 2828906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2829906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2830906d46d4SStefano Zampini } 2831906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2832906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2833906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2834906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2835b9b85e73SStefano Zampini } 2836a717540cSStefano Zampini 2837727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2838727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2839727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 28400e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 28410e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 28420e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 28430e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2844c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2845727cdba6SStefano Zampini } 28460e6343abSStefano Zampini } 28470e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2848727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2849727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2850727cdba6SStefano Zampini 2851a717540cSStefano Zampini /* flush dbg viewer */ 2852b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2853b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2854b8ffe317SStefano Zampini } 2855a717540cSStefano Zampini 2856e310c8b4SStefano Zampini /* free workspace */ 2857a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 28584641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 285908122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 28609162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 28619162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 286208122e43SStefano Zampini } else { 28639162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 28649162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 28659162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 286608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 286708122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 28689162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 28699162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 287008122e43SStefano Zampini } 2871674ae819SStefano Zampini PetscFunctionReturn(0); 2872674ae819SStefano Zampini } 2873674ae819SStefano Zampini 2874674ae819SStefano Zampini #undef __FUNCT__ 2875674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2876674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2877674ae819SStefano Zampini { 2878674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2879674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2880674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 28817fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2882674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2883674ae819SStefano Zampini 2884674ae819SStefano Zampini PetscFunctionBegin; 28858e61c736SStefano Zampini /* Reset previously computed graph */ 28868e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2887674ae819SStefano Zampini /* Init local Graph struct */ 28887fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 28897fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 2890674ae819SStefano Zampini 2891575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2892575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2893575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2894575ad6abSStefano Zampini } 28959577ea80SStefano Zampini 2896674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 28974d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 28984d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 28994d379d7bSStefano Zampini PetscInt nvtxs; 2900e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 2901674ae819SStefano Zampini 29024d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 29032fffb893SStefano Zampini 29042fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29052fffb893SStefano Zampini if (flg_row) { 29064d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2907b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 29082fffb893SStefano Zampini } 29092fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29109b28b941SStefano 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 */ 29114d379d7bSStefano Zampini IS is_dummy; 29124d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 29134d379d7bSStefano Zampini PetscInt j,sum; 29144d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 29154d379d7bSStefano Zampini const PetscInt *idxs; 29164d379d7bSStefano Zampini PCBDDCGraph graph; 29174d379d7bSStefano Zampini PetscBT is_on_boundary; 29184d379d7bSStefano Zampini 29194d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 29204d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 29214d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 29224d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 29237fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 29244d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 2925e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2926e496cd5dSStefano Zampini if (flg_row) { 29274d379d7bSStefano Zampini graph->xadj = xadj; 29284d379d7bSStefano Zampini graph->adjncy = adjncy; 2929e496cd5dSStefano Zampini } 29304d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 29314d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2932e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29334d379d7bSStefano Zampini 29344d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 29359b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 29364d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29374d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 29384d379d7bSStefano Zampini } 29394d379d7bSStefano Zampini } 29404d379d7bSStefano Zampini 2941e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 29424d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29434d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 29444d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 29454d379d7bSStefano Zampini } 29464d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29474d379d7bSStefano Zampini 2948e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 29494d379d7bSStefano Zampini sum = 0; 29504d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29514d379d7bSStefano Zampini PetscInt sizecc = 0; 29524d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29534d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29544d379d7bSStefano Zampini sizecc++; 29554d379d7bSStefano Zampini } 29564d379d7bSStefano Zampini } 29574d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29584d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29594d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 29604d379d7bSStefano Zampini } 29614d379d7bSStefano Zampini } 29624d379d7bSStefano Zampini sum += sizecc*sizecc; 29634d379d7bSStefano Zampini } 29644d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 29654d379d7bSStefano Zampini sum = 0; 2966e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 29674d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 29684d379d7bSStefano Zampini cxadj[i] = sum; 29694d379d7bSStefano Zampini sum += temp; 29704d379d7bSStefano Zampini } 2971e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 29724d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29734d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29744d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29754d379d7bSStefano Zampini PetscInt k,sizecc = 0; 29764d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 29774d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 29784d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 29794d379d7bSStefano Zampini sizecc++; 29804d379d7bSStefano Zampini } 29814d379d7bSStefano Zampini } 29824d379d7bSStefano Zampini } 29834d379d7bSStefano Zampini } 29844d379d7bSStefano Zampini } 29859b28b941SStefano Zampini if (sum) { 2986e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 29874d379d7bSStefano Zampini } else { 29884d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 29894d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 29904d379d7bSStefano Zampini } 29914d379d7bSStefano Zampini graph->xadj = 0; 29924d379d7bSStefano Zampini graph->adjncy = 0; 29934d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 29944d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 29954d379d7bSStefano Zampini } 2996674ae819SStefano Zampini } 29979b28b941SStefano Zampini if (pcbddc->dbg_flag) { 29989b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29999b28b941SStefano Zampini } 3000674ae819SStefano Zampini 300163602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3002674ae819SStefano Zampini vertex_size = 1; 300363602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 300463602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 300595ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 300663602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3007a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 300863602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3009674ae819SStefano Zampini } 301063602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 301163602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 301263602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3013674ae819SStefano Zampini } 301463602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3015674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 301663602bcaSStefano Zampini } else { 301763602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 301863602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3019854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 302063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 302163602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 302263602bcaSStefano Zampini } 302363602bcaSStefano Zampini } 3024674ae819SStefano Zampini } 3025674ae819SStefano Zampini 3026674ae819SStefano Zampini /* Setup of Graph */ 3027785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3028a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3029785d1243SStefano Zampini } 3030785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3031a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3032785d1243SStefano Zampini } 3033302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3034674ae819SStefano Zampini 3035674ae819SStefano Zampini /* Graph's connected components analysis */ 3036674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3037674ae819SStefano Zampini 3038674ae819SStefano Zampini /* print some info to stdout */ 3039674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3040302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3041674ae819SStefano Zampini } 3042fb180af4SStefano Zampini 3043fb180af4SStefano Zampini /* mark topography has done */ 3044fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3045674ae819SStefano Zampini PetscFunctionReturn(0); 3046674ae819SStefano Zampini } 3047674ae819SStefano Zampini 3048674ae819SStefano Zampini #undef __FUNCT__ 3049674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3050674ae819SStefano 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[]) 3051674ae819SStefano Zampini { 3052674ae819SStefano Zampini Vec local_vec,global_vec; 3053674ae819SStefano Zampini IS seqis,paris; 3054674ae819SStefano Zampini VecScatter scatter_ctx; 3055674ae819SStefano Zampini PetscScalar *array; 3056674ae819SStefano Zampini PetscInt *temp_global_dofs; 3057674ae819SStefano Zampini PetscScalar globalsum; 3058674ae819SStefano Zampini PetscInt i,j,s; 305917ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 306017ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 306117ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3062674ae819SStefano Zampini PetscBool first_found; 3063674ae819SStefano Zampini PetscErrorCode ierr; 3064674ae819SStefano Zampini 3065674ae819SStefano Zampini PetscFunctionBegin; 3066674ae819SStefano Zampini /* mpi buffers */ 3067b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3068b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3069674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 307017ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3071674ae819SStefano Zampini /* get maximum size of subset */ 3072785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3073674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3074674ae819SStefano Zampini max_local = 0; 3075b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3076674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3077674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3078674ae819SStefano Zampini } 3079674ae819SStefano Zampini } 3080b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3081674ae819SStefano Zampini max_global++; 3082674ae819SStefano Zampini max_local = 0; 3083b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3084674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3085674ae819SStefano Zampini max_local = local_dofs[i]; 3086674ae819SStefano Zampini } 3087674ae819SStefano Zampini } 3088674ae819SStefano Zampini max_local++; 3089674ae819SStefano Zampini /* allocate workspace */ 3090674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3091674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3092674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3093674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3094674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3095674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3096674ae819SStefano Zampini /* create scatter */ 3097674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3098674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3099674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3100674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3101674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3102674ae819SStefano Zampini /* init array */ 3103674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3104674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3105674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3106674ae819SStefano Zampini if (local_dofs_mult) { 3107674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3108674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3109674ae819SStefano Zampini } 3110674ae819SStefano Zampini } else { 3111674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3112674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3113674ae819SStefano Zampini } 3114674ae819SStefano Zampini } 3115674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3116674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3117674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3118674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3119674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 31205b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3121674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3122674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3123674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3124674ae819SStefano Zampini nlocals = 0; 3125674ae819SStefano Zampini first_index = -1; 3126674ae819SStefano Zampini first_found = PETSC_FALSE; 3127674ae819SStefano Zampini for (i=0;i<s;i++) { 3128b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3129674ae819SStefano Zampini first_found = PETSC_TRUE; 3130674ae819SStefano Zampini first_index = i; 3131674ae819SStefano Zampini } 31325b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3133674ae819SStefano Zampini } 3134674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3135674ae819SStefano Zampini if (!rank_prec_comm) { 3136674ae819SStefano Zampini dof_displs[0]=0; 3137674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3138674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3139674ae819SStefano Zampini } 3140674ae819SStefano Zampini } 3141674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3142674ae819SStefano Zampini if (first_found) { 3143674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3144674ae819SStefano Zampini old_index = first_index; 3145674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3146b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3147674ae819SStefano Zampini array[i] += array[old_index]; 3148674ae819SStefano Zampini old_index = i; 3149674ae819SStefano Zampini } 3150674ae819SStefano Zampini } 3151674ae819SStefano Zampini } 3152674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3153674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3154674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3155674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3156674ae819SStefano Zampini /* get global ordering of local dofs */ 31572b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3158674ae819SStefano Zampini if (local_dofs_mult) { 3159674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31605b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3161674ae819SStefano Zampini } 3162674ae819SStefano Zampini } else { 3163674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31645b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3165674ae819SStefano Zampini } 3166674ae819SStefano Zampini } 31672b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3168674ae819SStefano Zampini /* free workspace */ 3169674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3170674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3171674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 317217ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3173674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3174674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3175674ae819SStefano Zampini PetscFunctionReturn(0); 3176674ae819SStefano Zampini } 31779a7d3425SStefano Zampini 31789a7d3425SStefano Zampini #undef __FUNCT__ 31799a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 31809a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 31819a7d3425SStefano Zampini { 31829a7d3425SStefano Zampini PetscInt i,j; 31839a7d3425SStefano Zampini PetscScalar *alphas; 31849a7d3425SStefano Zampini PetscErrorCode ierr; 31859a7d3425SStefano Zampini 31869a7d3425SStefano Zampini PetscFunctionBegin; 31879a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3188785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 31899a7d3425SStefano Zampini for (i=0;i<n;i++) { 31909a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 31919a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 31929a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 31939a7d3425SStefano Zampini } 31949a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 31959a7d3425SStefano Zampini PetscFunctionReturn(0); 31969a7d3425SStefano Zampini } 31979a7d3425SStefano Zampini 3198e7931f94SStefano Zampini #undef __FUNCT__ 319970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 320028143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3201e7931f94SStefano Zampini { 3202e7931f94SStefano Zampini Mat subdomain_adj; 3203e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3204e7931f94SStefano Zampini MatPartitioning partitioner; 3205e7931f94SStefano Zampini Mat_IS *matis; 3206e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3207e7931f94SStefano Zampini PetscInt prank; 3208e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3209e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3210e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 32113837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3212e7931f94SStefano Zampini PetscErrorCode ierr; 32132b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3214e7931f94SStefano Zampini PetscSubcomm subcomm; 3215a57a6d2fSStefano Zampini 3216e7931f94SStefano Zampini PetscFunctionBegin; 32172b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 32182b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 32192b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3220e7931f94SStefano Zampini 3221e7931f94SStefano Zampini /* Get info on mapping */ 3222e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3223e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3224e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3225e7931f94SStefano Zampini 3226e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3227785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3228e7931f94SStefano Zampini xadj[0] = 0; 3229e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3230785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3231785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3232e7931f94SStefano Zampini 32332b510759SStefano Zampini if (threshold) { 3234d023bfaeSStefano Zampini PetscInt xadj_count = 0; 32352b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3236d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3237d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3238d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3239d023bfaeSStefano Zampini xadj_count++; 3240e7931f94SStefano Zampini } 3241e7931f94SStefano Zampini } 3242d023bfaeSStefano Zampini xadj[1] = xadj_count; 3243c8587f34SStefano Zampini } else { 3244e7931f94SStefano Zampini if (xadj[1]) { 3245e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3246e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3247c8587f34SStefano Zampini } 3248e7931f94SStefano Zampini } 3249e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3250e7931f94SStefano Zampini if (use_square) { 3251e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3252e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3253e7931f94SStefano Zampini } 3254e7931f94SStefano Zampini } 3255e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3256e7931f94SStefano Zampini 32573837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3258e7931f94SStefano Zampini 3259e7931f94SStefano Zampini /* 3260e7931f94SStefano Zampini Restrict work on active processes only. 3261e7931f94SStefano Zampini */ 3262e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3263e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3264e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 32652b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3266d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3267e7931f94SStefano Zampini if (color) { 3268e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3269e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3270e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3271c8587f34SStefano Zampini } else { 327228143c3dSStefano Zampini PetscInt coarsening_ratio; 3273306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3274785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3275e7931f94SStefano Zampini prank = rank; 3276306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 32778002ef2cSStefano Zampini /* 3278e7931f94SStefano Zampini for (i=0;i<size;i++) { 3279e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3280c8587f34SStefano Zampini } 32818002ef2cSStefano Zampini */ 3282e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3283e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3284c8587f34SStefano Zampini } 3285e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3286306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 328722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3288e7931f94SStefano Zampini 3289e7931f94SStefano Zampini /* Partition */ 3290306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3291e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3292e7931f94SStefano Zampini if (use_vwgt) { 32933837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3294e7931f94SStefano Zampini v_wgt[0] = local_size; 3295e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3296c8587f34SStefano Zampini } 329728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 329828143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 329928143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3300e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3301e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 330222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3303e7931f94SStefano Zampini 3304e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 330528143c3dSStefano Zampini if (contiguous) { 330628143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 330728143c3dSStefano Zampini } else { 330884ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 330928143c3dSStefano Zampini } 3310e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3311e7931f94SStefano Zampini /* clean up */ 3312e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3313e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3314e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3315e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3316e7931f94SStefano Zampini } 3317e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3318e7931f94SStefano Zampini 3319e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3320e7931f94SStefano Zampini i = 1; 3321e7931f94SStefano Zampini if (color) i=0; 3322e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3323e7931f94SStefano Zampini 3324e7931f94SStefano Zampini /* get back IS */ 3325e7931f94SStefano Zampini *is_sends = ranks_send_to; 3326e7931f94SStefano Zampini PetscFunctionReturn(0); 3327e7931f94SStefano Zampini } 3328e7931f94SStefano Zampini 3329e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3330e7931f94SStefano Zampini 3331e7931f94SStefano Zampini #undef __FUNCT__ 3332e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 333328143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3334e7931f94SStefano Zampini { 333570cf5478SStefano Zampini Mat local_mat; 3336e7931f94SStefano Zampini Mat_IS *matis; 3337e7931f94SStefano Zampini IS is_sends_internal; 33389d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 333928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 33409d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3341e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3342e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3343e7931f94SStefano Zampini const PetscInt* is_indices; 3344e7931f94SStefano Zampini MatType new_local_type; 3345e7931f94SStefano Zampini /* buffers */ 3346e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 334728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 33489d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3349e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3350e7931f94SStefano Zampini /* MPI */ 335128143c3dSStefano Zampini MPI_Comm comm,comm_n; 335228143c3dSStefano Zampini PetscSubcomm subcomm; 3353e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 335428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 335528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 335628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 335728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 335828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3359e7931f94SStefano Zampini PetscErrorCode ierr; 3360e7931f94SStefano Zampini 3361e7931f94SStefano Zampini PetscFunctionBegin; 336228143c3dSStefano Zampini /* TODO: add missing checks */ 336328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 336428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 336528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 336628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3367e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 336828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3369e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3370e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3371e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3372e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3373e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 337428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 337570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 337670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 337728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 337870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 337970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 338070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 338170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 338270cf5478SStefano Zampini } 3383e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3384e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3385e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3386e7931f94SStefano Zampini if (!is_sends) { 338728143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 338828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 338928143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3390c8587f34SStefano Zampini } else { 3391e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3392e7931f94SStefano Zampini is_sends_internal = is_sends; 3393c8587f34SStefano Zampini } 3394e7931f94SStefano Zampini 3395e7931f94SStefano Zampini /* get pointer of MATIS data */ 3396e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3397e7931f94SStefano Zampini 3398e7931f94SStefano Zampini /* get comm */ 3399a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3400e7931f94SStefano Zampini 3401e7931f94SStefano Zampini /* compute number of sends */ 3402e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3403e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3404e7931f94SStefano Zampini 3405e7931f94SStefano Zampini /* compute number of receives */ 3406e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3407785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3408e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3409e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3410e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3411e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3412e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3413e7931f94SStefano Zampini 341428143c3dSStefano Zampini /* restrict comm if requested */ 341528143c3dSStefano Zampini subcomm = 0; 341628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 341728143c3dSStefano Zampini if (restrict_comm) { 3418779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3419779c1cceSStefano Zampini 342028143c3dSStefano Zampini color = 0; 3421779c1cceSStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */ 342228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 342328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 342428143c3dSStefano Zampini /* check if reuse has been requested */ 342528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 342628143c3dSStefano Zampini if (*mat_n) { 342728143c3dSStefano Zampini PetscMPIInt subcommsize2; 342828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 342928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 343028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 343128143c3dSStefano Zampini } else { 343228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 343328143c3dSStefano Zampini } 343428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3435779c1cceSStefano Zampini PetscMPIInt rank; 3436779c1cceSStefano Zampini 3437779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 343828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 343928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 344028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3441306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 344228143c3dSStefano Zampini } 344328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 344428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 344528143c3dSStefano Zampini } else { 344628143c3dSStefano Zampini comm_n = comm; 344728143c3dSStefano Zampini } 344828143c3dSStefano Zampini 3449e7931f94SStefano Zampini /* prepare send/receive buffers */ 3450785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3451e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3452785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3453e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 345428143c3dSStefano Zampini if (nis) { 3455854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 345628143c3dSStefano Zampini } 3457e7931f94SStefano Zampini 345828143c3dSStefano Zampini /* Get data from local matrices */ 3459e7931f94SStefano Zampini if (!isdense) { 3460a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3461e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3462e7931f94SStefano Zampini /* 3463e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3464e7931f94SStefano Zampini send_buffer_idxs should contain: 3465e7931f94SStefano Zampini - MatType_PRIVATE type 3466e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3467e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3468e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3469e7931f94SStefano Zampini */ 3470e7931f94SStefano Zampini } else { 3471e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3472e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3473854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3474e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3475e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3476e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3477e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3478e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3479e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3480e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3481e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3482e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3483c8587f34SStefano Zampini } 3484c8587f34SStefano Zampini } 3485e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 348628143c3dSStefano Zampini /* additional is (if any) */ 348728143c3dSStefano Zampini if (nis) { 348828143c3dSStefano Zampini PetscMPIInt psum; 348928143c3dSStefano Zampini PetscInt j; 349028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 349128143c3dSStefano Zampini PetscInt plen; 349228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 349328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 349428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 349528143c3dSStefano Zampini } 3496854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 349728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 349828143c3dSStefano Zampini PetscInt plen; 349928143c3dSStefano Zampini const PetscInt *is_array_idxs; 350028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 350128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 350228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 350328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 350428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 350528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 350628143c3dSStefano Zampini } 350728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 350828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 350928143c3dSStefano Zampini } 351028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 351128143c3dSStefano Zampini } 351228143c3dSStefano Zampini 3513e7931f94SStefano Zampini buf_size_idxs = 0; 3514e7931f94SStefano Zampini buf_size_vals = 0; 351528143c3dSStefano Zampini buf_size_idxs_is = 0; 3516e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3517e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3518e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 351928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3520e7931f94SStefano Zampini } 3521785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3522785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 352395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3524e7931f94SStefano Zampini 3525e7931f94SStefano Zampini /* get new tags for clean communications */ 3526e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3527e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 352828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3529e7931f94SStefano Zampini 3530e7931f94SStefano Zampini /* allocate for requests */ 3531785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3532785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 353395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3534785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3535785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 353695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3537e7931f94SStefano Zampini 3538e7931f94SStefano Zampini /* communications */ 3539e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3540e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 354128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3542e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3543e7931f94SStefano Zampini source_dest = onodes[i]; 3544e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3545e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3546e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3547e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 354828143c3dSStefano Zampini if (nis) { 354928143c3dSStefano 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); 355028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 355128143c3dSStefano Zampini } 3552e7931f94SStefano Zampini } 3553e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3554e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3555e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3556e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 355728143c3dSStefano Zampini if (nis) { 355828143c3dSStefano 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); 355928143c3dSStefano Zampini } 3560e7931f94SStefano Zampini } 3561e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3562e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3563e7931f94SStefano Zampini 3564e7931f94SStefano Zampini /* assemble new l2g map */ 3565e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3566e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35679d30be91SStefano Zampini new_local_rows = 0; 3568e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35699d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3570e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3571e7931f94SStefano Zampini } 35729d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3573e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35749d30be91SStefano Zampini new_local_rows = 0; 3575e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35769d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 35779d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3578e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3579e7931f94SStefano Zampini } 35809d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 35819d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3582e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3583e7931f94SStefano Zampini 3584e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3585e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3586e7931f94SStefano 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) */ 3587e7931f94SStefano Zampini if (n_recvs) { 358828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3589e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3590e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3591e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3592e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3593e7931f94SStefano Zampini break; 3594e7931f94SStefano Zampini } 3595e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3596e7931f94SStefano Zampini } 3597e7931f94SStefano Zampini switch (new_local_type_private) { 359828143c3dSStefano Zampini case MATDENSE_PRIVATE: 359928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3600e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3601e7931f94SStefano Zampini bs = 1; 360228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 360328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 360428143c3dSStefano Zampini bs = 1; 360528143c3dSStefano Zampini } 3606e7931f94SStefano Zampini break; 3607e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3608e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3609e7931f94SStefano Zampini bs = 1; 3610e7931f94SStefano Zampini break; 3611e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3612e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3613e7931f94SStefano Zampini break; 3614e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3615e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3616e7931f94SStefano Zampini break; 3617e7931f94SStefano Zampini default: 36189d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3619e7931f94SStefano Zampini break; 3620e7931f94SStefano Zampini } 362128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 362228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 362328143c3dSStefano Zampini bs = 1; 3624e7931f94SStefano Zampini } 3625e7931f94SStefano Zampini 362670cf5478SStefano Zampini /* create MATIS object if needed */ 362770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3628e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 362928143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 363070cf5478SStefano Zampini } else { 363170cf5478SStefano Zampini /* it also destroys the local matrices */ 363270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 363370cf5478SStefano Zampini } 363470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3635e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 36369d30be91SStefano Zampini 36379d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 36389d30be91SStefano Zampini 36399d30be91SStefano Zampini /* Global to local map of received indices */ 36409d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 36419d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 36429d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 36439d30be91SStefano Zampini 36449d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 36459d30be91SStefano Zampini buf_size_idxs = 0; 36469d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 36479d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 36489d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 36499d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 36509d30be91SStefano Zampini } 36519d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 36529d30be91SStefano Zampini 36539d30be91SStefano Zampini /* set preallocation */ 36549d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 36559d30be91SStefano Zampini if (!newisdense) { 36569d30be91SStefano Zampini PetscInt *new_local_nnz=0; 36579d30be91SStefano Zampini 36589d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 36599d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 36609d30be91SStefano Zampini if (n_recvs) { 36619d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 36629d30be91SStefano Zampini } 36639d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 36649d30be91SStefano Zampini PetscInt j; 36659d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 36669d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 36679d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 36689d30be91SStefano Zampini } 36699d30be91SStefano Zampini } else { 36709d30be91SStefano Zampini /* TODO */ 36719d30be91SStefano Zampini } 36729d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 36739d30be91SStefano Zampini } 36749d30be91SStefano Zampini if (new_local_nnz) { 36759d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 36769d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 36779d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 36789d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36799d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 36809d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36819d30be91SStefano Zampini } else { 36829d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36839d30be91SStefano Zampini } 36849d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 36859d30be91SStefano Zampini } else { 36869d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36879d30be91SStefano Zampini } 3688e7931f94SStefano Zampini 3689e7931f94SStefano Zampini /* set values */ 3690e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 36919d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3692e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3693e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3694e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 36959d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3696e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3697e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3698e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 369928143c3dSStefano Zampini } else { 370028143c3dSStefano Zampini /* TODO */ 3701e7931f94SStefano Zampini } 3702e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3703e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3704e7931f94SStefano Zampini } 3705e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3706e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 370770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 370870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 37099d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 37109d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3711e7931f94SStefano Zampini 3712dfd14d43SStefano Zampini #if 0 371328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3714e7931f94SStefano Zampini Vec lvec,rvec; 3715e7931f94SStefano Zampini PetscReal infty_error; 3716e7931f94SStefano Zampini 37172a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3718e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3719e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3720e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 372170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3722e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3723e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3724e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3725e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3726e7931f94SStefano Zampini } 372728143c3dSStefano Zampini #endif 3728e7931f94SStefano Zampini 372928143c3dSStefano Zampini /* assemble new additional is (if any) */ 373028143c3dSStefano Zampini if (nis) { 373128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 373228143c3dSStefano Zampini 373328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3734854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 373528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 373628143c3dSStefano Zampini psum = 0; 373728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 373828143c3dSStefano Zampini for (j=0;j<nis;j++) { 373928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 374028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 374128143c3dSStefano Zampini psum += plen; 374228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 374328143c3dSStefano Zampini } 374428143c3dSStefano Zampini } 3745854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3746854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 374728143c3dSStefano Zampini for (i=1;i<nis;i++) { 374828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 374928143c3dSStefano Zampini } 375028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 375128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 375228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 375328143c3dSStefano Zampini for (j=0;j<nis;j++) { 375428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 375528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 375628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 375728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 375828143c3dSStefano Zampini } 375928143c3dSStefano Zampini } 376028143c3dSStefano Zampini for (i=0;i<nis;i++) { 376128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 376228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 376328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 376428143c3dSStefano Zampini } 376528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 376628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 376728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 376828143c3dSStefano Zampini } 3769e7931f94SStefano Zampini /* free workspace */ 377028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3771e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3772e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3773e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3774e7931f94SStefano Zampini if (isdense) { 3775e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3776e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3777e7931f94SStefano Zampini } else { 3778e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3779e7931f94SStefano Zampini } 378028143c3dSStefano Zampini if (nis) { 378128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 378228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 378328143c3dSStefano Zampini } 3784e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3785e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 378628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3787e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3788e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 378928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3790e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3791e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3792e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3793e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3794e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 379528143c3dSStefano Zampini if (nis) { 379628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 379728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 379828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 379928143c3dSStefano Zampini } 380028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 380128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 380228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 380328143c3dSStefano Zampini for (i=0;i<nis;i++) { 380428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 380528143c3dSStefano Zampini } 380628143c3dSStefano Zampini } 3807e7931f94SStefano Zampini PetscFunctionReturn(0); 3808e7931f94SStefano Zampini } 3809a57a6d2fSStefano Zampini 381012edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3811af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 381212edc857SStefano Zampini 3813c8587f34SStefano Zampini #undef __FUNCT__ 3814c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3815c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3816c8587f34SStefano Zampini { 3817c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3818c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 381920a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 38209881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 382120a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 38226e683305SStefano Zampini IS coarse_is,*isarray; 38236e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 38246e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3825f9eb5b7dSStefano Zampini PC pc_temp; 3826c8587f34SStefano Zampini PCType coarse_pc_type; 3827c8587f34SStefano Zampini KSPType coarse_ksp_type; 3828f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 38294f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 38306e683305SStefano Zampini Mat t_coarse_mat_is; 38316e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 38326e683305SStefano Zampini PetscMPIInt all_procs; 383374e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 383468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 383522bc73bbSStefano Zampini PetscScalar *array; 38369881197aSStefano Zampini PetscErrorCode ierr; 3837fdc09c96SStefano Zampini 3838c8587f34SStefano Zampini PetscFunctionBegin; 3839c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 384068457ee5SStefano 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 */ 3841fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 38425a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3843fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3844f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3845f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3846f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3847fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3848fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3849727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3850fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3851fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3852fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3853f4ddd8eeSStefano Zampini } 3854fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3855fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3856f4ddd8eeSStefano Zampini } 385770cf5478SStefano Zampini /* reset any subassembling information */ 385870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 38596e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 38606e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3861fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3862f4ddd8eeSStefano Zampini } 3863c8587f34SStefano Zampini 38646e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 38652b510759SStefano Zampini im_active = !!(pcis->n); 38662b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 38676e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 38686e683305SStefano Zampini void_procs = all_procs-active_procs; 38696e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 387074e2c79eSStefano Zampini redist = PETSC_FALSE; 387122bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 38726e683305SStefano Zampini csin_ml = PETSC_TRUE; 38736e683305SStefano Zampini ncoarse_ml = void_procs; 3874779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3875779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 387618a45a71SStefano Zampini csin_ds = PETSC_TRUE; 387718a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 387818a45a71SStefano Zampini redist = PETSC_TRUE; 387918a45a71SStefano Zampini } else { 38806e683305SStefano Zampini csin_ds = PETSC_TRUE; 3881779c1cceSStefano Zampini ncoarse_ds = active_procs; 3882779c1cceSStefano Zampini redist = PETSC_TRUE; 388318a45a71SStefano Zampini } 38846e683305SStefano Zampini } else { 38856e683305SStefano Zampini csin_ml = PETSC_FALSE; 38866e683305SStefano Zampini ncoarse_ml = all_procs; 38876e683305SStefano Zampini if (void_procs) { 38886e683305SStefano Zampini csin_ds = PETSC_TRUE; 38896e683305SStefano Zampini ncoarse_ds = void_procs; 38906e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 38916e683305SStefano Zampini } else { 3892779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 389374e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 389474e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 389574e2c79eSStefano Zampini redist = PETSC_TRUE; 389674e2c79eSStefano Zampini } else { 38976e683305SStefano Zampini csin_ds = PETSC_FALSE; 38986e683305SStefano Zampini ncoarse_ds = all_procs; 38996e683305SStefano Zampini } 39006e683305SStefano Zampini } 390174e2c79eSStefano Zampini } 39026e683305SStefano Zampini 39036e683305SStefano Zampini /* 39046e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 39056e683305SStefano Zampini - we have not exceeded the number of levels requested 39066e683305SStefano Zampini - we can actually subassemble the active processes 39076e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 39086e683305SStefano Zampini */ 39096e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 39106e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 39116e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 39126e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 39136e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3914f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 39152b510759SStefano Zampini } else { 3916f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3917c8587f34SStefano Zampini } 3918c8587f34SStefano Zampini } 39196e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 39206e683305SStefano Zampini if (multilevel_allowed) { 39216e683305SStefano Zampini ncoarse = ncoarse_ml; 39226e683305SStefano Zampini csin = csin_ml; 392358da7f69SStefano Zampini redist = PETSC_FALSE; 39246e683305SStefano Zampini } else { 39256e683305SStefano Zampini ncoarse = ncoarse_ds; 39266e683305SStefano Zampini csin = csin_ds; 39276e683305SStefano Zampini } 3928e7931f94SStefano Zampini 3929abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3930abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3931abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3932abbbba34SStefano Zampini 3933abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 393422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 393522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 393622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 393722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3938b9b85e73SStefano Zampini #if 0 3939b9b85e73SStefano Zampini { 3940b9b85e73SStefano Zampini PetscViewer viewer; 3941b9b85e73SStefano Zampini char filename[256]; 3942b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 3943b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3944b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3945b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 3946b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3947b9b85e73SStefano Zampini } 3948b9b85e73SStefano Zampini #endif 39496e683305SStefano 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); 39506e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 39516e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 39526e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3953abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 3954abbbba34SStefano Zampini 39556e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 39566e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 39576e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 39586e683305SStefano Zampini const PetscInt *idxs; 39596e683305SStefano Zampini ISLocalToGlobalMapping tmap; 39606e683305SStefano Zampini 39616e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 39620be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 39636e683305SStefano Zampini /* allocate space for temporary storage */ 3964854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 3965854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 39666e683305SStefano Zampini /* allocate for IS array */ 39676e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 39686e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 39696e683305SStefano Zampini nis = nisdofs + nisneu; 3970854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 39716e683305SStefano Zampini /* dofs splitting */ 39726e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 39736e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 39746e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 39756e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39766e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39776e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39786e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39796e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 39806e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 39816e683305SStefano Zampini } 39826e683305SStefano Zampini /* neumann boundaries */ 39836e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 39846e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 39856e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 39866e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39876e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39886e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39896e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39906e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 39916e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 39926e683305SStefano Zampini } 39936e683305SStefano Zampini /* free memory */ 39946e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 39956e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 39966e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 39976e683305SStefano Zampini } else { 39986e683305SStefano Zampini nis = 0; 39996e683305SStefano Zampini nisdofs = 0; 40006e683305SStefano Zampini nisneu = 0; 40016e683305SStefano Zampini isarray = NULL; 40026e683305SStefano Zampini } 40036e683305SStefano Zampini /* destroy no longer needed map */ 40046e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 40056e683305SStefano Zampini 40066e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 40076e683305SStefano Zampini coarse_mat_is = NULL; 40086e683305SStefano Zampini if (csin) { 40096e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 401074e2c79eSStefano Zampini if (redist) { 401174e2c79eSStefano Zampini PetscMPIInt rank; 4012779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 401374e2c79eSStefano Zampini 401474e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 401558da7f69SStefano Zampini spc = active_procs/ncoarse; 401658da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4017779c1cceSStefano Zampini if (im_active) { 4018779c1cceSStefano Zampini destsize = 1; 401974e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 402074e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 402174e2c79eSStefano Zampini } else { 402274e2c79eSStefano Zampini dest[0] = rank/(spc+1); 402374e2c79eSStefano Zampini } 402474e2c79eSStefano Zampini } else { 4025779c1cceSStefano Zampini destsize = 0; 40266e683305SStefano Zampini } 4027779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4028779c1cceSStefano Zampini } else if (csin_type_simple) { 40296e683305SStefano Zampini PetscMPIInt rank; 40306e683305SStefano Zampini PetscInt issize,isidx; 4031779c1cceSStefano Zampini 40326e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 40336e683305SStefano Zampini if (im_active) { 40346e683305SStefano Zampini issize = 1; 40356e683305SStefano Zampini isidx = (PetscInt)rank; 40366e683305SStefano Zampini } else { 40376e683305SStefano Zampini issize = 0; 40386e683305SStefano Zampini isidx = -1; 40396e683305SStefano Zampini } 40406e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4041779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 40426e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40436e683305SStefano Zampini } 4044779c1cceSStefano Zampini 4045779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4046779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4047779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4048779c1cceSStefano Zampini PetscInt *coarse_candidates; 4049779c1cceSStefano Zampini const PetscInt* tisindices; 4050779c1cceSStefano Zampini 4051779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4052779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4053779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4054779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4055779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4056779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4057779c1cceSStefano Zampini } 4058779c1cceSStefano Zampini } 4059779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4060779c1cceSStefano Zampini 4061779c1cceSStefano Zampini 40626e683305SStefano Zampini if (pcbddc->dbg_flag) { 40636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 40656e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4067779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 40686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 40696e683305SStefano Zampini } 40706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 40716e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40726e683305SStefano Zampini } 40736e683305SStefano Zampini /* shift the pattern on coarse candidates */ 40746e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 40756e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4076854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 40776e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 40786e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 40796e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 40806e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 40816e683305SStefano Zampini } 40826e683305SStefano Zampini if (pcbddc->dbg_flag) { 40836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 40856e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40866e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40876e683305SStefano Zampini } 4088779c1cceSStefano Zampini } 40896e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 40906e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 40916e683305SStefano Zampini } else { 40926e683305SStefano Zampini if (pcbddc->dbg_flag) { 40936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 40956e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40966e683305SStefano Zampini } 40976e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 40986e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 40996e683305SStefano Zampini } 41006e683305SStefano Zampini 41016e683305SStefano Zampini /* create local to global scatters for coarse problem */ 410268457ee5SStefano Zampini if (compute_vecs) { 41036e683305SStefano Zampini PetscInt lrows; 41046e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41056e683305SStefano Zampini if (coarse_mat_is) { 41066e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 41076e683305SStefano Zampini } else { 41086e683305SStefano Zampini lrows = 0; 41096e683305SStefano Zampini } 41106e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 41116e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 41126e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41156e683305SStefano Zampini } 41166e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 41176e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4118c8587f34SStefano Zampini 4119f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4120f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4121f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4122f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4123f9eb5b7dSStefano Zampini } else { 4124f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4125f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4126c8587f34SStefano Zampini } 4127c8587f34SStefano Zampini 41286e683305SStefano Zampini /* print some info if requested */ 41296e683305SStefano Zampini if (pcbddc->dbg_flag) { 41306e683305SStefano Zampini if (!multilevel_allowed) { 41316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41326e683305SStefano Zampini if (multilevel_requested) { 41336e683305SStefano 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); 41346e683305SStefano Zampini } else if (pcbddc->max_levels) { 41356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 41366e683305SStefano Zampini } 41376e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41386e683305SStefano Zampini } 41396e683305SStefano Zampini } 41406e683305SStefano Zampini 4141f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 41426e683305SStefano Zampini if (coarse_mat_is) { 41436e683305SStefano Zampini MatReuse coarse_mat_reuse; 41446a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 41456e683305SStefano Zampini if (pcbddc->dbg_flag) { 41466e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 41476e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 41486e683305SStefano Zampini } 4149f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4150312be037SStefano Zampini char prefix[256],str_level[16]; 4151e604994aSStefano Zampini size_t len; 41526e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4153c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4154f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 41555f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4156c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 41576e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4158c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4159c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 41603e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4161e604994aSStefano Zampini /* prefix */ 4162e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4163e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4164e604994aSStefano Zampini if (!pcbddc->current_level) { 4165e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4166e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4167c8587f34SStefano Zampini } else { 4168e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4169312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4170312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 417134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4172312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4173e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4174e604994aSStefano Zampini } 4175e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 41763e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41773e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 41783e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 41793e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4180f9eb5b7dSStefano Zampini /* allow user customization */ 4181f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 41823e3c6dadSStefano Zampini } 41833e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41843e3c6dadSStefano Zampini if (nisdofs) { 41853e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 41863e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 41873e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 41883e3c6dadSStefano Zampini } 41893e3c6dadSStefano Zampini } 41903e3c6dadSStefano Zampini if (nisneu) { 41913e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 41923e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4193312be037SStefano Zampini } 4194f9eb5b7dSStefano Zampini 4195f9eb5b7dSStefano Zampini /* get some info after set from options */ 4196f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4197f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4198f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 41994f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 42006e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4201f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4202f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4203f9eb5b7dSStefano Zampini } 42044f3a063dSStefano Zampini if (isredundant) { 42054f3a063dSStefano Zampini KSP inner_ksp; 42064f3a063dSStefano Zampini PC inner_pc; 42074f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 42084f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 42094f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 42104f3a063dSStefano Zampini } 4211f9eb5b7dSStefano Zampini 4212f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4213fa7f1dd8SStefano Zampini if (coarse_reuse) { 421481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4215fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 42166e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4217fa7f1dd8SStefano Zampini } else { 42186e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4219fa7f1dd8SStefano Zampini } 4220c8587f34SStefano Zampini if (isbddc || isnn) { 422122bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 422270cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 422328143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 422422b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 42256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 42276e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 42286e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 422922b6e8a2SStefano Zampini } 423070cf5478SStefano Zampini } 42316e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 423270cf5478SStefano Zampini } else { 423322bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 423422bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 423522bc73bbSStefano Zampini } 423622bc73bbSStefano Zampini } else { 42372e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4238c8587f34SStefano Zampini } 4239c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4240c8587f34SStefano Zampini 4241c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4242b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 42435a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4244c8587f34SStefano Zampini 42456e683305SStefano Zampini /* set operators */ 42465f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 42476e683305SStefano Zampini if (pcbddc->dbg_flag) { 42486e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42496e683305SStefano Zampini } 42506e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 42516e683305SStefano Zampini coarse_mat = 0; 42526e683305SStefano Zampini } 42536e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4254b9b85e73SStefano Zampini #if 0 4255b9b85e73SStefano Zampini { 4256b9b85e73SStefano Zampini PetscViewer viewer; 4257b9b85e73SStefano Zampini char filename[256]; 4258b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4259b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4260b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4261b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4262b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4263b9b85e73SStefano Zampini } 4264b9b85e73SStefano Zampini #endif 4265c8587f34SStefano Zampini 4266c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4267c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4268c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 426998a51de6SStefano Zampini } 427098a51de6SStefano Zampini 427198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 427298a51de6SStefano Zampini Vec crhs,csol; 427398a51de6SStefano Zampini PetscBool ispreonly; 427498a51de6SStefano Zampini if (CoarseNullSpace) { 4275c8587f34SStefano Zampini if (isbddc) { 4276c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4277c8587f34SStefano Zampini } else { 4278c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4279c8587f34SStefano Zampini } 4280c8587f34SStefano Zampini } 4281f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4282f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4283f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4284f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 42856e683305SStefano Zampini /* hack */ 4286f347579bSStefano Zampini if (!csol) { 42872a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4288f9eb5b7dSStefano Zampini } 4289f347579bSStefano Zampini if (!crhs) { 42902a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4291f347579bSStefano Zampini } 4292cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4293cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 42946e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4295c8587f34SStefano Zampini KSP check_ksp; 42962b510759SStefano Zampini KSPType check_ksp_type; 4297c8587f34SStefano Zampini PC check_pc; 42986e683305SStefano Zampini Vec check_vec,coarse_vec; 42996a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 43002b510759SStefano Zampini PetscInt its; 43016e683305SStefano Zampini PetscBool compute_eigs; 43026e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 43036e683305SStefano Zampini PetscInt neigs; 43048e185a42SStefano Zampini const char *prefix; 4305c8587f34SStefano Zampini 43062b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 43076e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 430823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4309f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 43102b510759SStefano Zampini if (ispreonly) { 43112b510759SStefano Zampini check_ksp_type = KSPPREONLY; 43126e683305SStefano Zampini compute_eigs = PETSC_FALSE; 43132b510759SStefano Zampini } else { 4314cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 43156e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4316c8587f34SStefano Zampini } 4317c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 43186e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 43196e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 43206e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4321a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4322a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4323a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4324a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4325c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4326c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4327c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4328c8587f34SStefano Zampini /* create random vec */ 43296e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 43306e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4331c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4332c8587f34SStefano Zampini if (CoarseNullSpace) { 4333c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4334c8587f34SStefano Zampini } 43356e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4336c8587f34SStefano Zampini /* solve coarse problem */ 43376e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4338c8587f34SStefano Zampini if (CoarseNullSpace) { 43396e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4340c8587f34SStefano Zampini } 4341cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 43426e683305SStefano Zampini if (compute_eigs) { 4343854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4344854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 43456e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 43466e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 43476e683305SStefano Zampini lambda_min = eigs_r[0]; 43486e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 43496e683305SStefano Zampini if (lambda_max>lambda_min) { 4350cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4351cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4352cbcc2c2aSStefano Zampini } 4353c8587f34SStefano Zampini } 4354c8587f34SStefano Zampini } 4355cbcc2c2aSStefano Zampini 4356c8587f34SStefano Zampini /* check coarse problem residual error */ 43576e683305SStefano Zampini if (pcbddc->dbg_flag) { 43586e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 43596e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43606e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4361c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 43626e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 43636e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4364c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4365779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 43666e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 43676e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 43686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 43696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 43706e683305SStefano Zampini if (compute_eigs) { 43716e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4372deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4373c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 43746e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 43756e683305SStefano 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); 43766e683305SStefano Zampini for (i=0;i<neigs;i++) { 43776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4378c8587f34SStefano Zampini } 43796e683305SStefano Zampini } 43806e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 43816e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43826e683305SStefano Zampini } 4383c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 43846e683305SStefano Zampini if (compute_eigs) { 43856e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 43866e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4387c8587f34SStefano Zampini } 43886e683305SStefano Zampini } 43896e683305SStefano Zampini } 4390cbcc2c2aSStefano Zampini /* print additional info */ 4391cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 43926e683305SStefano Zampini /* waits until all processes reaches this point */ 43936e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4394cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4395cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4396cbcc2c2aSStefano Zampini } 4397cbcc2c2aSStefano Zampini 43982b510759SStefano Zampini /* free memory */ 4399c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4400fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4401c8587f34SStefano Zampini PetscFunctionReturn(0); 4402c8587f34SStefano Zampini } 4403674ae819SStefano Zampini 4404f34684f1SStefano Zampini #undef __FUNCT__ 4405f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4406f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4407f34684f1SStefano Zampini { 4408f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4409f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4410f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 44110e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 441273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 44130e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4414f34684f1SStefano Zampini PetscErrorCode ierr; 4415f34684f1SStefano Zampini 4416f34684f1SStefano Zampini PetscFunctionBegin; 4417f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 44180e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 44190e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4420727cdba6SStefano Zampini } 44210e6343abSStefano 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); 44220e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 44230e6343abSStefano Zampini local_size = 0; 44240e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 44250e6343abSStefano Zampini PetscInt j; 44260e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 44270e6343abSStefano Zampini } 44280e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 44290e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 44300e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 44310e6343abSStefano Zampini } 4432f34684f1SStefano Zampini 4433f34684f1SStefano Zampini /* check numbering */ 4434f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4435f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4436b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4437f34684f1SStefano Zampini 4438f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4439f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4440f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 44410fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4442f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4443f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4444727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4445f34684f1SStefano Zampini } 4446f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4447f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4448f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4449f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4450f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4451f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4452f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4453f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4454f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4455f34684f1SStefano Zampini if (array[i] == 1.0) { 4456b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4457f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4458f34684f1SStefano Zampini } 4459f34684f1SStefano Zampini } 4460b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4461f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4462f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4463f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4464f34684f1SStefano Zampini } 4465f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4466f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4467f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4468f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4469f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4470f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4471b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4472ca8b9ea9SStefano Zampini PetscInt *gidxs; 4473ca8b9ea9SStefano Zampini 4474ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4475ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4476f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4477f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4478f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4479f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 44804bc2dc4bSStefano 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); 4481f34684f1SStefano Zampini } 4482f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4483ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4484f34684f1SStefano Zampini } 4485f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4486302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4487f34684f1SStefano Zampini } 44888bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4489f34684f1SStefano Zampini /* get back data */ 4490f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4491f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4492674ae819SStefano Zampini PetscFunctionReturn(0); 4493674ae819SStefano Zampini } 4494674ae819SStefano Zampini 4495e456f2a8SStefano Zampini #undef __FUNCT__ 4496e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4497a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4498e456f2a8SStefano Zampini { 4499e456f2a8SStefano Zampini IS localis_t; 4500a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4501e456f2a8SStefano Zampini PetscScalar *vals; 4502e456f2a8SStefano Zampini PetscErrorCode ierr; 4503e456f2a8SStefano Zampini 4504e456f2a8SStefano Zampini PetscFunctionBegin; 4505a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4506e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4507854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4508e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4509e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4510a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4511a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 45121035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4513a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 45141035eff8SStefano Zampini } 4515a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4516e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4517e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4518a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4519a7dc3881SStefano Zampini /* now compute set in local ordering */ 4520a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4521a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4522a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4523a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4524a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4525ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4526e456f2a8SStefano Zampini lsize++; 4527e456f2a8SStefano Zampini } 4528e456f2a8SStefano Zampini } 4529854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4530a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4531ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4532e456f2a8SStefano Zampini idxs[lsize++] = i; 4533e456f2a8SStefano Zampini } 4534e456f2a8SStefano Zampini } 4535a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4536a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4537e456f2a8SStefano Zampini *localis = localis_t; 4538e456f2a8SStefano Zampini PetscFunctionReturn(0); 4539e456f2a8SStefano Zampini } 4540906d46d4SStefano Zampini 4541906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4542906d46d4SStefano Zampini #undef __FUNCT__ 4543906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4544906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4545906d46d4SStefano Zampini { 4546906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4547906d46d4SStefano Zampini PetscErrorCode ierr; 4548906d46d4SStefano Zampini 4549906d46d4SStefano Zampini PetscFunctionBegin; 4550906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4551906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4552906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4553906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4554906d46d4SStefano Zampini PetscFunctionReturn(0); 4555906d46d4SStefano Zampini } 4556906d46d4SStefano Zampini 4557906d46d4SStefano Zampini #undef __FUNCT__ 4558906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4559906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4560906d46d4SStefano Zampini { 4561906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4562906d46d4SStefano Zampini PetscErrorCode ierr; 4563906d46d4SStefano Zampini 4564906d46d4SStefano Zampini PetscFunctionBegin; 4565906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4566906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4567906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4568906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4569906d46d4SStefano Zampini PetscFunctionReturn(0); 4570906d46d4SStefano Zampini } 4571b96c3477SStefano Zampini 4572b96c3477SStefano Zampini #undef __FUNCT__ 4573b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 457408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4575b96c3477SStefano Zampini { 4576a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4577b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4578b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4579a64f4aa4SStefano Zampini Mat S_j; 4580b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4581b96c3477SStefano Zampini PetscBool free_used_adj; 4582b96c3477SStefano Zampini PetscErrorCode ierr; 4583b96c3477SStefano Zampini 4584b96c3477SStefano Zampini PetscFunctionBegin; 4585b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4586b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 458708122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4588b96c3477SStefano Zampini used_xadj = NULL; 4589b96c3477SStefano Zampini used_adjncy = NULL; 4590b96c3477SStefano Zampini } else { 459108122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 459208122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 459308122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 459408122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4595b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4596b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4597b96c3477SStefano Zampini } else { 45982fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4599b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4600b96c3477SStefano Zampini PetscInt nvtxs; 4601b96c3477SStefano Zampini 46022fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 46032fffb893SStefano Zampini if (flg_row) { 4604b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4605b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4606b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4607b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 46082fffb893SStefano Zampini } else { 46092fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 46102fffb893SStefano Zampini used_xadj = NULL; 46112fffb893SStefano Zampini used_adjncy = NULL; 46122fffb893SStefano Zampini } 46132fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4614b96c3477SStefano Zampini } 4615b96c3477SStefano Zampini } 4616d5574798SStefano Zampini 4617d5574798SStefano Zampini /* setup sub_schurs data */ 4618a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4619a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4620a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4621a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4622a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4623a64f4aa4SStefano Zampini } else { 4624a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4625a64f4aa4SStefano Zampini } 4626a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4627b96c3477SStefano Zampini 4628b96c3477SStefano Zampini /* free adjacency */ 4629b96c3477SStefano Zampini if (free_used_adj) { 4630b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4631b96c3477SStefano Zampini } 4632b96c3477SStefano Zampini PetscFunctionReturn(0); 4633b96c3477SStefano Zampini } 4634b96c3477SStefano Zampini 4635b96c3477SStefano Zampini #undef __FUNCT__ 4636b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 463708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4638b96c3477SStefano Zampini { 4639b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4640b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4641b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4642b96c3477SStefano Zampini PCBDDCGraph graph; 4643b96c3477SStefano Zampini PetscErrorCode ierr; 4644b96c3477SStefano Zampini 4645b96c3477SStefano Zampini PetscFunctionBegin; 4646b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 464708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4648b96c3477SStefano Zampini IS verticesIS; 4649b96c3477SStefano Zampini 4650b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4651b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 46527fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4653b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4654b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4655b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4656b96c3477SStefano Zampini /* 4657b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4658b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4659b96c3477SStefano Zampini } 4660b96c3477SStefano Zampini */ 4661b96c3477SStefano Zampini } else { 4662b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4663b96c3477SStefano Zampini } 4664b96c3477SStefano Zampini 4665b96c3477SStefano Zampini /* sub_schurs init */ 4666a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4667a64f4aa4SStefano Zampini 4668b96c3477SStefano Zampini /* free graph struct */ 466908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4670b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4671b96c3477SStefano Zampini } 4672b96c3477SStefano Zampini PetscFunctionReturn(0); 4673b96c3477SStefano Zampini } 4674