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 */ 51980677318SStefano 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); 57780677318SStefano 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; 59980677318SStefano Zampini Mat auxmat; 60006656605SStefano Zampini IS is_aux; 60180677318SStefano Zampini PetscScalar *array,*array2; 60206656605SStefano Zampini 603f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 60480677318SStefano 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); 60980677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61088ebb749SStefano Zampini 61180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 61280677318SStefano 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 } 62580677318SStefano 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); 63080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 63206656605SStefano Zampini } else { 63380677318SStefano Zampini PetscScalar *marr; 63480677318SStefano Zampini 63580677318SStefano 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); 63880677318SStefano 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 } 64380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini } 64580677318SStefano Zampini if (!pcbddc->switch_static) { 64680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 64780677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 64880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 64980677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65080677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 65280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65680677318SStefano Zampini } 65780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65880677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66080677318SStefano Zampini } else { 66180677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 66280677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 66325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66480677318SStefano Zampini } 66580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 66680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 66780677318SStefano 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); 67080677318SStefano Zampini if (isCHOL) { 67180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 67280677318SStefano Zampini } else { 67325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 67480677318SStefano Zampini } 67580677318SStefano 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); 68080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68180677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 68280677318SStefano 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); 70180677318SStefano 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); 7063a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7073a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 708*b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 709*b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7103a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7113a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 71206656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7133a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7143a50541eSStefano Zampini } else { 715*b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7179577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7189577ea80SStefano Zampini } 71980677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 72080677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7213a50541eSStefano Zampini } 72225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 72388ebb749SStefano Zampini } 72488ebb749SStefano Zampini 72588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 726f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 72706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 72806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 72906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73006656605SStefano Zampini } 731f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 73206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 73306656605SStefano Zampini PetscScalar *marray; 73406656605SStefano Zampini 73506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 73606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 737f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 738f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 739f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 740f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 741f4ddd8eeSStefano Zampini } 742f4ddd8eeSStefano Zampini } 74306656605SStefano Zampini 744f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 74506656605SStefano Zampini PetscScalar *marray; 74688ebb749SStefano Zampini 74706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7488eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75088ebb749SStefano Zampini } 751b9d89cd5SStefano Zampini if (!pcbddc->issym) { 75206656605SStefano Zampini n *= 2; 753f4ddd8eeSStefano Zampini } 75406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 75506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 75606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 75906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76088ebb749SStefano Zampini } 76106656605SStefano Zampini if (!pcbddc->issym) { 76206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7631b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 76506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 76688ebb749SStefano Zampini } 76788ebb749SStefano Zampini } else { 768c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 769c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7701b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 771c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 772c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 773c0553b1fSStefano Zampini } 77488ebb749SStefano Zampini } 77506656605SStefano Zampini } 77606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 77706656605SStefano Zampini /* vertices */ 77806656605SStefano Zampini if (n_vertices) { 77916f15bc4SStefano Zampini 78016f15bc4SStefano Zampini if (n_R) { 78106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 78206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 78316f15bc4SStefano Zampini PetscScalar *x,*y; 78406656605SStefano Zampini 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 } 80180677318SStefano 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; 80580677318SStefano Zampini 806*b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 80780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 80880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 80980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 81080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 81380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 81480677318SStefano Zampini } 81580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 81780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 82206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 82306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82406656605SStefano Zampini } 82506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 82680677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 83006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 83106656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 83306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83406656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 83516f15bc4SStefano Zampini } else { 83616f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83716f15bc4SStefano Zampini } 83821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 83906656605SStefano Zampini /* coarse basis functions */ 84006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 84116f15bc4SStefano Zampini PetscScalar *y; 84216f15bc4SStefano Zampini 84306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 84806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 84906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 85106656605SStefano Zampini 85206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 85506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85906656605SStefano Zampini } 86006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 86106656605SStefano Zampini } 86206656605SStefano Zampini } 86306656605SStefano Zampini 86406656605SStefano Zampini if (n_constraints) { 86506656605SStefano Zampini Mat B; 86606656605SStefano Zampini 86706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 86980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87106656605SStefano Zampini if (n_vertices) { 87280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 87380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 87480677318SStefano Zampini } else { 87580677318SStefano Zampini Mat S_VCt; 87680677318SStefano Zampini 87780677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 87880677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 87980677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 88080677318SStefano Zampini } 88106656605SStefano Zampini } 88206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88306656605SStefano Zampini /* coarse basis functions */ 88406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 88506656605SStefano Zampini PetscScalar *y; 88606656605SStefano Zampini 88706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 89006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 89406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 89506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 89706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 90006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90106656605SStefano Zampini } 90206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90306656605SStefano Zampini } 90406656605SStefano Zampini } 90580677318SStefano Zampini if (n_constraints) { 90680677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 90780677318SStefano Zampini } 90806656605SStefano Zampini 90906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 91006656605SStefano Zampini if (!pcbddc->issym) { 91106656605SStefano Zampini 91206656605SStefano Zampini if (n_constraints) { 91316f15bc4SStefano Zampini Mat S_CCT,B_C; 91406656605SStefano Zampini 91580677318SStefano Zampini /* this is a lazy thing */ 91680677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 91906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 92016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 92106656605SStefano Zampini if (n_vertices) { 92216f15bc4SStefano Zampini Mat B_V,S_VCT; 92306656605SStefano Zampini 92406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 92606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 92706656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 92816f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 92906656605SStefano Zampini } 93006656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 93180677318SStefano Zampini } else { /* if there are no constraints, reset work */ 93280677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 93306656605SStefano Zampini } 93416f15bc4SStefano Zampini if (n_vertices && n_R) { 93506656605SStefano Zampini Mat A_VRT; 93680677318SStefano Zampini PetscScalar *marray; 93706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 93806656605SStefano Zampini 93980677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 94080677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 94380677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 94480677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 94516f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 94606656605SStefano Zampini } 94706656605SStefano Zampini 94806656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 94906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95506656605SStefano Zampini } 95606656605SStefano Zampini } else { 95706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96306656605SStefano Zampini } 96406656605SStefano Zampini } 96506656605SStefano Zampini /* coarse basis functions */ 96606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96706656605SStefano Zampini PetscScalar *y; 96806656605SStefano Zampini 96906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 97006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97406656605SStefano Zampini if (i<n_vertices) { 97506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 97606656605SStefano Zampini } 97706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 97906656605SStefano Zampini 98006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 98106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98706656605SStefano Zampini } 98806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98906656605SStefano Zampini } 99006656605SStefano Zampini } 99188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 99688ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 99788ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99888ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 99925084f0cSStefano Zampini if (pcbddc->dbg_flag) { 100088ebb749SStefano Zampini Mat coarse_sub_mat; 100125084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 100288ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 100388ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 100488ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10058bec7fa6SStefano Zampini Mat C_B,CPHI; 10068bec7fa6SStefano Zampini IS is_dummy; 10078bec7fa6SStefano Zampini Vec mones; 100888ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 100988ebb749SStefano Zampini PetscReal real_value; 101088ebb749SStefano Zampini 101188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 101288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 101388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 101488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 101588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 101688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1017c0553b1fSStefano Zampini if (unsymmetric_check) { 101888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 101988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 102088ebb749SStefano Zampini } 102188ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 102288ebb749SStefano Zampini 102325084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 102406656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 102525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1026c0553b1fSStefano Zampini if (unsymmetric_check) { 102788ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103088ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 103988ebb749SStefano Zampini } else { 104088ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104888ebb749SStefano Zampini } 104988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 105381d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10548bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10550fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 105606656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10578bec7fa6SStefano Zampini 10588bec7fa6SStefano Zampini /* check constraints */ 10598bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10608bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10618bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10628bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10638bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10648bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10658bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10668bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10678bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10688bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10698bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10718bec7fa6SStefano Zampini 107225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 107388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 107888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 107988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 108088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 108188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 108288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1083c0553b1fSStefano Zampini if (unsymmetric_check) { 108488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 108588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 108688ebb749SStefano Zampini } 108788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 108888ebb749SStefano Zampini } 108988ebb749SStefano Zampini /* free memory */ 109006656605SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 109188ebb749SStefano Zampini if (n_vertices) { 109288ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 109388ebb749SStefano Zampini } 109488ebb749SStefano Zampini if (n_constraints) { 109588ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 109688ebb749SStefano Zampini } 10978629588bSStefano Zampini /* get back data */ 10988629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 109988ebb749SStefano Zampini PetscFunctionReturn(0); 110088ebb749SStefano Zampini } 110188ebb749SStefano Zampini 110288ebb749SStefano Zampini #undef __FUNCT__ 1103d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1104d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1105d65f70fdSStefano Zampini { 1106d65f70fdSStefano Zampini Mat *work_mat; 1107d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1108d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1109d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1110d65f70fdSStefano Zampini PetscErrorCode ierr; 1111d65f70fdSStefano Zampini 1112d65f70fdSStefano Zampini PetscFunctionBegin; 1113d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1114d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1115d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1116d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1117d65f70fdSStefano Zampini 1118d65f70fdSStefano Zampini if (!rsorted) { 1119d65f70fdSStefano Zampini const PetscInt *idxs; 1120d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1121d65f70fdSStefano Zampini 1122d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1123d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1124d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1125d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1126d65f70fdSStefano Zampini } 1127d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1128d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1129d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1130d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1131d65f70fdSStefano Zampini } 1132d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1134d65f70fdSStefano Zampini } else { 1135d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1136d65f70fdSStefano Zampini isrow_s = isrow; 1137d65f70fdSStefano Zampini } 1138d65f70fdSStefano Zampini 1139d65f70fdSStefano Zampini if (!csorted) { 1140d65f70fdSStefano Zampini if (isrow == iscol) { 1141d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1142d65f70fdSStefano Zampini iscol_s = isrow_s; 1143d65f70fdSStefano Zampini } else { 1144d65f70fdSStefano Zampini const PetscInt *idxs; 1145d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1146d65f70fdSStefano Zampini 1147d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1149d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1150d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1151d65f70fdSStefano Zampini } 1152d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1155d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1156d65f70fdSStefano Zampini } 1157d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini } 1160d65f70fdSStefano Zampini } else { 1161d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1162d65f70fdSStefano Zampini iscol_s = iscol; 1163d65f70fdSStefano Zampini } 1164d65f70fdSStefano Zampini 1165d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini 1167d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1168d65f70fdSStefano Zampini Mat new_mat; 1169d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1170d65f70fdSStefano Zampini 1171d65f70fdSStefano Zampini if (!rsorted) { 1172d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1173d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1174d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1175d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1176d65f70fdSStefano Zampini } 1177d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini } else { 1180d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1181d65f70fdSStefano Zampini } 1182d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini 1184d65f70fdSStefano Zampini if (!csorted) { 1185d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1186d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1187d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1188d65f70fdSStefano Zampini } else { 1189d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1190d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1192d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini } 1197d65f70fdSStefano Zampini } else { 1198d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini } 1200d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1201d65f70fdSStefano Zampini 1202d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini work_mat[0] = new_mat; 1205d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1206d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini } 1208d65f70fdSStefano Zampini 1209d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1210d65f70fdSStefano Zampini *B = work_mat[0]; 1211d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini PetscFunctionReturn(0); 1215d65f70fdSStefano Zampini } 1216d65f70fdSStefano Zampini 1217d65f70fdSStefano Zampini #undef __FUNCT__ 12185e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12195e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1220aa0d41d4SStefano Zampini { 1221aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12225e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1223d65f70fdSStefano Zampini Mat new_mat; 12245e8657edSStefano Zampini IS is_local,is_global; 1225d65f70fdSStefano Zampini PetscInt local_size; 1226d65f70fdSStefano Zampini PetscBool isseqaij; 1227aa0d41d4SStefano Zampini PetscErrorCode ierr; 1228aa0d41d4SStefano Zampini 1229aa0d41d4SStefano Zampini PetscFunctionBegin; 1230aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12315e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12325e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1233906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1234906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1235d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1236906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1237906d46d4SStefano Zampini 1238906d46d4SStefano Zampini /* check */ 1239906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1240906d46d4SStefano Zampini Vec x,x_change; 1241906d46d4SStefano Zampini PetscReal error; 1242906d46d4SStefano Zampini 12435e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1244906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12455e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12465e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12475e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1248d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12505e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1252906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1253906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1254906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1255906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1256906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1257906d46d4SStefano Zampini } 1258906d46d4SStefano Zampini 125922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12609b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 126122d5777bSStefano Zampini if (isseqaij) { 1262d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1263aa0d41d4SStefano Zampini } else { 1264aa0d41d4SStefano Zampini Mat work_mat; 1265aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1266d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1267aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1268aa0d41d4SStefano Zampini } 1269d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1270e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1271e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1272e496cd5dSStefano Zampini #endif 127345a1bb75SStefano Zampini /* 127445a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1275d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 127645a1bb75SStefano Zampini */ 1277d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1278aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1279aa0d41d4SStefano Zampini } 1280aa0d41d4SStefano Zampini 1281aa0d41d4SStefano Zampini #undef __FUNCT__ 1282a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12838ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1284a64d13efSStefano Zampini { 1285a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1286a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12878ce42a96SStefano Zampini IS is_aux1,is_aux2; 1288e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 12893a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12903a50541eSStefano Zampini PetscInt vbs,bs; 12914641a718SStefano Zampini PetscBT bitmask; 1292a64d13efSStefano Zampini PetscErrorCode ierr; 1293a64d13efSStefano Zampini 1294a64d13efSStefano Zampini PetscFunctionBegin; 1295b23d619eSStefano Zampini /* 1296b23d619eSStefano Zampini No need to setup local scatters if 1297b23d619eSStefano Zampini - primal space is unchanged 1298b23d619eSStefano Zampini AND 1299b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1300b23d619eSStefano Zampini AND 1301b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1302b23d619eSStefano Zampini */ 1303b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1304f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1305f4ddd8eeSStefano Zampini } 1306f4ddd8eeSStefano Zampini /* destroy old objects */ 1307f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1308f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1309f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1310a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1311b371cd4fSStefano Zampini n_B = pcis->n_B; 1312b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1313b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13144641a718SStefano Zampini /* create auxiliary bitmask */ 13154641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13164641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13170e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13184641a718SStefano Zampini } 13193a50541eSStefano Zampini 1320a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 1321854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1322a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13234641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 1324a64d13efSStefano Zampini idx_R_local[n_R] = i; 1325a64d13efSStefano Zampini n_R++; 1326a64d13efSStefano Zampini } 1327a64d13efSStefano Zampini } 13283a50541eSStefano Zampini 13293a50541eSStefano Zampini /* Block code */ 13303a50541eSStefano Zampini vbs = 1; 13313a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13323a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13333a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13343a50541eSStefano Zampini PetscInt *vary; 13353a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1336785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13373a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13380e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1339964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13403a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13413a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13423a50541eSStefano Zampini break; 13433a50541eSStefano Zampini } 13443a50541eSStefano Zampini } 13453a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13463a50541eSStefano Zampini vbs = bs; 13473a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13483a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13493a50541eSStefano Zampini } 13503a50541eSStefano Zampini } 13513a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13523a50541eSStefano Zampini } 13533a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 13543a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1355a64d13efSStefano Zampini 1356a64d13efSStefano Zampini /* print some info if requested */ 1357a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1358a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1359a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13600fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1361a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1362a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13633a50541eSStefano 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); 1364a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1365a64d13efSStefano Zampini } 1366a64d13efSStefano Zampini 1367a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13683a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1369854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1370854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1371a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 13724641a718SStefano Zampini for (i=0; i<n_D; i++) { 13734641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 13744641a718SStefano Zampini } 1375a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1376a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13774641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 13784641a718SStefano Zampini aux_array1[j++] = i; 1379a64d13efSStefano Zampini } 1380a64d13efSStefano Zampini } 1381a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1382a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1383a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 13844641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 13854641a718SStefano Zampini aux_array2[j++] = i; 1386a64d13efSStefano Zampini } 1387a64d13efSStefano Zampini } 1388a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1389a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1390a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1391a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1392a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1393a64d13efSStefano Zampini 13948eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1395785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1396a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13974641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 13984641a718SStefano Zampini aux_array1[j++] = i; 1399a64d13efSStefano Zampini } 1400a64d13efSStefano Zampini } 1401a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1402a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1403a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1404a64d13efSStefano Zampini } 14054641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14063a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1407a64d13efSStefano Zampini PetscFunctionReturn(0); 1408a64d13efSStefano Zampini } 1409a64d13efSStefano Zampini 1410304d26faSStefano Zampini 1411304d26faSStefano Zampini #undef __FUNCT__ 1412304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1413684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1414304d26faSStefano Zampini { 1415304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1416304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1417304d26faSStefano Zampini PC pc_temp; 1418304d26faSStefano Zampini Mat A_RR; 1419f4ddd8eeSStefano Zampini MatReuse reuse; 1420304d26faSStefano Zampini PetscScalar m_one = -1.0; 1421304d26faSStefano Zampini PetscReal value; 1422af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14239577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1424304d26faSStefano Zampini PetscErrorCode ierr; 1425e604994aSStefano Zampini /* prefixes stuff */ 1426312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1427e604994aSStefano Zampini size_t len; 1428304d26faSStefano Zampini 1429304d26faSStefano Zampini PetscFunctionBegin; 1430304d26faSStefano Zampini 1431e604994aSStefano Zampini /* compute prefixes */ 1432e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1433e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1434e604994aSStefano Zampini if (!pcbddc->current_level) { 1435e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1436e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1437e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1438e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1439e604994aSStefano Zampini } else { 1440e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1441312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1442e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1443e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1444312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1445312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 144634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 144734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1448e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1449e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1450e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1451e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1452e604994aSStefano Zampini } 1453e604994aSStefano Zampini 1454304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1455684f6988SStefano Zampini if (dirichlet) { 1456d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1457964fefecSStefano Zampini if (pcbddc->issym) { 1458964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1459964fefecSStefano Zampini } 1460ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1461964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1462304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1463304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1464304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1465304d26faSStefano Zampini /* default */ 1466304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1467e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 14689577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1469304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 14709577ea80SStefano Zampini if (issbaij) { 14719577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 14729577ea80SStefano Zampini } else { 1473304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 14749577ea80SStefano Zampini } 1475304d26faSStefano Zampini /* Allow user's customization */ 1476304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1477304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1478304d26faSStefano Zampini } 1479d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1480d5574798SStefano Zampini if (sub_schurs->interior_solver) { 1481d5574798SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,sub_schurs->interior_solver);CHKERRQ(ierr); 1482d5574798SStefano Zampini } 1483304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1484304d26faSStefano Zampini if (!n_D) { 1485304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1486304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1487304d26faSStefano Zampini } 1488304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1489304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1490304d26faSStefano Zampini /* set ksp_D into pcis data */ 1491304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1492304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1493304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1494684f6988SStefano Zampini } 1495304d26faSStefano Zampini 1496304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1497684f6988SStefano Zampini A_RR = 0; 1498684f6988SStefano Zampini if (neumann) { 1499f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15008ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1501f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1502f4ddd8eeSStefano Zampini PetscInt nn_R; 150381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1504f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1505f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1506f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1507f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1508f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1509f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1510f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1511727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1512f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1513f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1514f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1515f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1516f4ddd8eeSStefano Zampini } 1517f4ddd8eeSStefano Zampini } 1518f4ddd8eeSStefano Zampini /* last check */ 1519d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1520f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1521f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1522f4ddd8eeSStefano Zampini } 1523f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1524f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1525f4ddd8eeSStefano Zampini } 1526f4ddd8eeSStefano Zampini /* extract A_RR */ 1527af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1528af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 15293a50541eSStefano Zampini if (ibs != mbs) { 1530af732b37SStefano Zampini Mat newmat; 1531af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1532f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1533af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1534af732b37SStefano Zampini } else { 1535f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1536af732b37SStefano Zampini } 1537964fefecSStefano Zampini if (pcbddc->issym) { 1538964fefecSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1539964fefecSStefano Zampini } 1540f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1541304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1542304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1543304d26faSStefano Zampini /* default */ 1544304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1545e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1546304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 15479577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 15489577ea80SStefano Zampini if (issbaij) { 15499577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15509577ea80SStefano Zampini } else { 1551304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15529577ea80SStefano Zampini } 1553304d26faSStefano Zampini /* Allow user's customization */ 1554304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1555304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1556304d26faSStefano Zampini } 1557d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1558304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1559304d26faSStefano Zampini if (!n_R) { 1560304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1561304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1562304d26faSStefano Zampini } 1563304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1564304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1565684f6988SStefano Zampini } 1566304d26faSStefano Zampini 1567304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 15680fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1569684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1570684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1571684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1572684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1573684f6988SStefano Zampini } 1574684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 15750fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 15760fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 15770fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 15780fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 15790fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1580304d26faSStefano Zampini /* need to be adapted? */ 1581b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1582b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1583b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1584304d26faSStefano Zampini /* print info */ 1585304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1586e604994aSStefano 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); 1587304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1588304d26faSStefano Zampini } 1589b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 15908ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1591304d26faSStefano Zampini } 1592684f6988SStefano Zampini } 1593684f6988SStefano Zampini if (neumann) { /* Neumann */ 15940fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 15950fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 15960fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 15970fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 15980fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1599304d26faSStefano Zampini /* need to be adapted? */ 1600b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1601b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1602304d26faSStefano Zampini /* print info */ 1603304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1604e604994aSStefano 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); 1605304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1606304d26faSStefano Zampini } 1607b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16088ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1609304d26faSStefano Zampini } 16100fccc4e9SStefano Zampini } 1611684f6988SStefano Zampini } 1612304d26faSStefano Zampini /* free Neumann problem's matrix */ 1613304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1614304d26faSStefano Zampini PetscFunctionReturn(0); 1615304d26faSStefano Zampini } 1616304d26faSStefano Zampini 1617304d26faSStefano Zampini #undef __FUNCT__ 1618ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 161980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1620674ae819SStefano Zampini { 1621674ae819SStefano Zampini PetscErrorCode ierr; 1622674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1623674ae819SStefano Zampini 1624674ae819SStefano Zampini PetscFunctionBegin; 162580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 162680677318SStefano Zampini if (!pcbddc->switch_static) { 162780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 162880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 162980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 163020c7b377SStefano Zampini } 163180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163320c7b377SStefano Zampini } else { 163480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 163880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 163980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 164080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 164180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 164280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 164320c7b377SStefano Zampini } 1644674ae819SStefano Zampini } 164580677318SStefano Zampini if (applytranspose) { 164680677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 164780677318SStefano Zampini } else { 164880677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 164980677318SStefano Zampini } 165080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 165180677318SStefano Zampini if (!pcbddc->switch_static) { 165280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 165580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 165680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 165780677318SStefano Zampini } 165880677318SStefano Zampini } else { 165980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 166480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 166580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 166680677318SStefano Zampini } 166780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167180677318SStefano Zampini } 1672674ae819SStefano Zampini PetscFunctionReturn(0); 1673674ae819SStefano Zampini } 1674674ae819SStefano Zampini 1675dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1676674ae819SStefano Zampini #undef __FUNCT__ 1677674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1678dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1679674ae819SStefano Zampini { 1680674ae819SStefano Zampini PetscErrorCode ierr; 1681674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1682674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1683674ae819SStefano Zampini const PetscScalar zero = 0.0; 1684674ae819SStefano Zampini 1685674ae819SStefano Zampini PetscFunctionBegin; 1686dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1687dc359a40SStefano Zampini if (applytranspose) { 1688674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 16898eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1690dc359a40SStefano Zampini } else { 1691674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1692674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 169315aaf578SStefano Zampini } 169412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 169512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 169612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 169712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 169812edc857SStefano Zampini 16999f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 170012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 170112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1702964fefecSStefano Zampini Vec rhs,sol; 1703964fefecSStefano Zampini 1704964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1705964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 170612edc857SStefano Zampini if (applytranspose) { 1707964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 170812edc857SStefano Zampini } else { 1709964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 171012edc857SStefano Zampini } 171112edc857SStefano Zampini } 1712674ae819SStefano Zampini 1713674ae819SStefano Zampini /* Local solution on R nodes */ 171480677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 171580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 17169f00e9b4SStefano Zampini } 1717674ae819SStefano Zampini 17189f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 17199f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1721674ae819SStefano Zampini 1722674ae819SStefano Zampini /* Sum contributions from two levels */ 1723dc359a40SStefano Zampini if (applytranspose) { 1724dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1725dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1726dc359a40SStefano Zampini } else { 1727674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 17288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1729dc359a40SStefano Zampini } 1730674ae819SStefano Zampini PetscFunctionReturn(0); 1731674ae819SStefano Zampini } 1732674ae819SStefano Zampini 1733674ae819SStefano Zampini #undef __FUNCT__ 1734674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 173512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1736674ae819SStefano Zampini { 1737674ae819SStefano Zampini PetscErrorCode ierr; 1738674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 173958da7f69SStefano Zampini PetscScalar *array; 174012edc857SStefano Zampini Vec from,to; 1741674ae819SStefano Zampini 1742674ae819SStefano Zampini PetscFunctionBegin; 174312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 174412edc857SStefano Zampini from = pcbddc->coarse_vec; 174512edc857SStefano Zampini to = pcbddc->vec1_P; 174612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 174712edc857SStefano Zampini Vec tvec; 174858da7f69SStefano Zampini 174958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 175058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 175112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 175258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 175358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 175458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 175512edc857SStefano Zampini } 175612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 175712edc857SStefano Zampini from = pcbddc->vec1_P; 175812edc857SStefano Zampini to = pcbddc->coarse_vec; 175912edc857SStefano Zampini } 176012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1761674ae819SStefano Zampini PetscFunctionReturn(0); 1762674ae819SStefano Zampini } 1763674ae819SStefano Zampini 1764674ae819SStefano Zampini #undef __FUNCT__ 1765674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 176612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1767674ae819SStefano Zampini { 1768674ae819SStefano Zampini PetscErrorCode ierr; 1769674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 177058da7f69SStefano Zampini PetscScalar *array; 177112edc857SStefano Zampini Vec from,to; 1772674ae819SStefano Zampini 1773674ae819SStefano Zampini PetscFunctionBegin; 177412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 177512edc857SStefano Zampini from = pcbddc->coarse_vec; 177612edc857SStefano Zampini to = pcbddc->vec1_P; 177712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 177812edc857SStefano Zampini from = pcbddc->vec1_P; 177912edc857SStefano Zampini to = pcbddc->coarse_vec; 178012edc857SStefano Zampini } 178112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 178212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 178312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 178412edc857SStefano Zampini Vec tvec; 178558da7f69SStefano Zampini 178612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 178758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 178858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 178958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 179058da7f69SStefano Zampini } 179158da7f69SStefano Zampini } else { 179258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 179358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 179412edc857SStefano Zampini } 179512edc857SStefano Zampini } 1796674ae819SStefano Zampini PetscFunctionReturn(0); 1797674ae819SStefano Zampini } 1798674ae819SStefano Zampini 1799984c4197SStefano Zampini /* uncomment for testing purposes */ 1800984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1801674ae819SStefano Zampini #undef __FUNCT__ 1802674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1803674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1804674ae819SStefano Zampini { 1805674ae819SStefano Zampini PetscErrorCode ierr; 1806674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1807674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1808674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1809984c4197SStefano Zampini /* one and zero */ 1810984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1811984c4197SStefano Zampini /* space to store constraints and their local indices */ 18129162d606SStefano Zampini PetscScalar *constraints_data; 18139162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 18149162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 18159162d606SStefano Zampini PetscInt *constraints_n; 1816984c4197SStefano Zampini /* iterators */ 1817*b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1818984c4197SStefano Zampini /* BLAS integers */ 1819e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1820e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1821c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1822727cdba6SStefano Zampini /* reuse */ 18230e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 18240e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1825984c4197SStefano Zampini /* change of basis */ 1826*b3d85658SStefano Zampini PetscBool qr_needed; 18279162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1828984c4197SStefano Zampini /* auxiliary stuff */ 182964efe560SStefano Zampini PetscInt *nnz,*is_indices; 18308a0068c3SStefano Zampini PetscInt ncc; 1831984c4197SStefano Zampini /* some quantities */ 183245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1833a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1834984c4197SStefano Zampini 1835674ae819SStefano Zampini PetscFunctionBegin; 18368e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 18378e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 18388e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1839088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1840088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 18410e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 18420e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 18430e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 18440e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 18450e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1846088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1847cf5a6209SStefano Zampini 1848cf5a6209SStefano Zampini /* print some info */ 1849cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1850cf5a6209SStefano Zampini IS vertices; 1851cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1852cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1853cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1854cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1855cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1856cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1857cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1858fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1859fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1860cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1861cf5a6209SStefano Zampini } 1862cf5a6209SStefano Zampini 1863cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 18649162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1865cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1866cf5a6209SStefano Zampini const Vec *nearnullvecs; 1867cf5a6209SStefano Zampini Vec *localnearnullsp; 1868cf5a6209SStefano Zampini PetscScalar *array; 1869cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1870cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1871cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1872*b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1873cf5a6209SStefano Zampini PetscScalar *work; 1874cf5a6209SStefano Zampini PetscReal *singular_vals; 1875cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1876cf5a6209SStefano Zampini PetscReal *rwork; 1877cf5a6209SStefano Zampini #endif 1878cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1879cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1880cf5a6209SStefano Zampini #else 1881964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1882964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1883cf5a6209SStefano Zampini #endif 1884cf5a6209SStefano Zampini 1885674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1886d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1887d06fc5fdSStefano Zampini /* free unneeded index sets */ 1888d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1889d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1890674ae819SStefano Zampini } 1891d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1892d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1893d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1894d06fc5fdSStefano Zampini } 1895d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1896d06fc5fdSStefano Zampini n_ISForEdges = 0; 1897d06fc5fdSStefano Zampini } 1898d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1899d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1900d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1901d06fc5fdSStefano Zampini } 1902d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1903d06fc5fdSStefano Zampini n_ISForFaces = 0; 1904d06fc5fdSStefano Zampini } 190570022509SStefano Zampini 190670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 190770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 190870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 190970022509SStefano Zampini if (pcbddc->NullSpace) { 191070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 191170022509SStefano Zampini 191270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 1913b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1914d06fc5fdSStefano Zampini if (!ISForEdges) { 1915d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1916d06fc5fdSStefano Zampini } 1917b8ffe317SStefano Zampini } 1918d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 1919d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 1920d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1921d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1922d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 192398a51de6SStefano Zampini } 192470022509SStefano Zampini #endif 192508122e43SStefano Zampini 1926674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1927674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1928674ae819SStefano Zampini if (nearnullsp) { 1929674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1930f4ddd8eeSStefano Zampini /* remove any stored info */ 1931f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1932f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1933f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1934f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1935f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1936473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1937f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1938f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1939f4ddd8eeSStefano Zampini } 1940984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1941984c4197SStefano Zampini nnsp_size = 0; 1942674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1943674ae819SStefano Zampini } 1944984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1945984c4197SStefano Zampini max_constraints = nnsp_size; 1946984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1947984c4197SStefano Zampini 1948674ae819SStefano Zampini /* 1949674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 19509162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 19519162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 19529162d606SStefano Zampini There can be multiple constraints per connected component 1953674ae819SStefano Zampini */ 1954674ae819SStefano Zampini n_vertices = 0; 1955674ae819SStefano Zampini if (ISForVertices) { 1956674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1957674ae819SStefano Zampini } 19589162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 19599162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 19609162d606SStefano Zampini 19619162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 19629162d606SStefano Zampini total_counts *= max_constraints; 1963674ae819SStefano Zampini total_counts += n_vertices; 19644641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 19659162d606SStefano Zampini 1966674ae819SStefano Zampini total_counts = 0; 1967674ae819SStefano Zampini max_size_of_constraint = 0; 1968674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 19699162d606SStefano Zampini IS used_is; 1970674ae819SStefano Zampini if (i<n_ISForEdges) { 19719162d606SStefano Zampini used_is = ISForEdges[i]; 1972674ae819SStefano Zampini } else { 19739162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 1974674ae819SStefano Zampini } 19759162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 1976674ae819SStefano Zampini total_counts += j; 1977674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1978674ae819SStefano Zampini } 19799162d606SStefano 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); 19809162d606SStefano Zampini 1981984c4197SStefano Zampini /* get local part of global near null space vectors */ 1982785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1983984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1984984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1985984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1986984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1987984c4197SStefano Zampini } 1988674ae819SStefano Zampini 1989242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1990242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1991a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1992242a89d7SStefano Zampini 1993984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1994a773dcb8SStefano Zampini if (!skip_lapack) { 1995674ae819SStefano Zampini PetscScalar temp_work; 1996911cabfeSStefano Zampini 1997674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1998984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1999785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2000785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2001785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2002674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2003785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2004674ae819SStefano Zampini #endif 2005674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2006c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2007c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2008674ae819SStefano Zampini lwork = -1; 2009674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2010674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2011c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2012674ae819SStefano Zampini #else 2013c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2014674ae819SStefano Zampini #endif 2015674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2016984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2017674ae819SStefano Zampini #else /* on missing GESVD */ 2018674ae819SStefano Zampini /* SVD */ 2019674ae819SStefano Zampini PetscInt max_n,min_n; 2020674ae819SStefano Zampini max_n = max_size_of_constraint; 2021984c4197SStefano Zampini min_n = max_constraints; 2022984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2023674ae819SStefano Zampini min_n = max_size_of_constraint; 2024984c4197SStefano Zampini max_n = max_constraints; 2025674ae819SStefano Zampini } 2026785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2027674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2028785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2029674ae819SStefano Zampini #endif 2030674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2031674ae819SStefano Zampini lwork = -1; 2032e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2033e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2034b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2035674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2036674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20379162d606SStefano 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)); 2038674ae819SStefano Zampini #else 20399162d606SStefano 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)); 2040674ae819SStefano Zampini #endif 2041674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2042984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2043984c4197SStefano Zampini #endif /* on missing GESVD */ 2044674ae819SStefano Zampini /* Allocate optimal workspace */ 2045674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2046854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2047674ae819SStefano Zampini } 2048674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2049674ae819SStefano Zampini total_counts = 0; 20509162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 20519162d606SStefano Zampini constraints_data_ptr[0] = 0; 2052674ae819SStefano Zampini /* vertices */ 20539162d606SStefano Zampini if (n_vertices) { 2054674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 20559162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 20569162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2057674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 20589162d606SStefano Zampini constraints_n[total_counts] = 1; 20599162d606SStefano Zampini constraints_data[total_counts] = 1.0; 20609162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 20619162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2062674ae819SStefano Zampini total_counts++; 2063674ae819SStefano Zampini } 2064674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2065984c4197SStefano Zampini PetscBool used_vertex; 2066674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2067674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2068674ae819SStefano Zampini k = 0; 2069674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2070984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2071984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 20729162d606SStefano Zampini constraints_n[total_counts] = 1; 20739162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 20749162d606SStefano Zampini constraints_data[total_counts] = 1.0; 20759162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 20769162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2077674ae819SStefano Zampini total_counts++; 2078674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2079674ae819SStefano Zampini } 2080984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2081674ae819SStefano Zampini k++; 2082674ae819SStefano Zampini } 2083674ae819SStefano Zampini } 2084674ae819SStefano Zampini } 2085674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2086674ae819SStefano Zampini n_vertices = total_counts; 2087674ae819SStefano Zampini } 2088984c4197SStefano Zampini 2089674ae819SStefano Zampini /* edges and faces */ 20909162d606SStefano Zampini total_counts_cc = total_counts; 2091911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 20929162d606SStefano Zampini IS used_is; 20939162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 20949162d606SStefano Zampini 2095911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 20969162d606SStefano Zampini used_is = ISForEdges[ncc]; 2097984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2098674ae819SStefano Zampini } else { 20999162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2100984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2101674ae819SStefano Zampini } 2102674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 21039162d606SStefano Zampini 21049162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 21059162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2106984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2107984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2108674ae819SStefano Zampini if (nnsp_has_cnst) { 21095b08dc53SStefano Zampini PetscScalar quad_value; 21109162d606SStefano Zampini 21119162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21129162d606SStefano Zampini idxs_copied = PETSC_TRUE; 21139162d606SStefano Zampini 2114a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2115674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2116a773dcb8SStefano Zampini } else { 2117a773dcb8SStefano Zampini quad_value = 1.0; 2118a773dcb8SStefano Zampini } 2119674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21209162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2121674ae819SStefano Zampini } 21229162d606SStefano Zampini temp_constraints++; 2123674ae819SStefano Zampini total_counts++; 2124674ae819SStefano Zampini } 2125674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2126984c4197SStefano Zampini PetscReal real_value; 21279162d606SStefano Zampini PetscScalar *ptr_to_data; 21289162d606SStefano Zampini 2129984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 21309162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2131674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 21329162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2133674ae819SStefano Zampini } 2134984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2135984c4197SStefano Zampini /* check if array is null on the connected component */ 2136e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21379162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 21385b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2139674ae819SStefano Zampini temp_constraints++; 2140674ae819SStefano Zampini total_counts++; 21419162d606SStefano Zampini if (!idxs_copied) { 21429162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 21439162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2144674ae819SStefano Zampini } 2145674ae819SStefano Zampini } 21469162d606SStefano Zampini } 21479162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 214845a1bb75SStefano Zampini valid_constraints = temp_constraints; 2149eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2150a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 21519162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 21529162d606SStefano Zampini 21539162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2154a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21559162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2156a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 21579162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2158a773dcb8SStefano Zampini } else { /* perform SVD */ 2159984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 21609162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2161674ae819SStefano Zampini 2162674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2163984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2164984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2165984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2166984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2167984c4197SStefano Zampini from that computed using LAPACKgesvd 2168984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2169984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2170984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2171674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2172e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2173984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2174674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2175674ae819SStefano Zampini for (k=0;k<j+1;k++) { 21769162d606SStefano 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)); 2177674ae819SStefano Zampini } 2178674ae819SStefano Zampini } 2179e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2180e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2181e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2182674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2183c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2184674ae819SStefano Zampini #else 2185c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2186674ae819SStefano Zampini #endif 2187674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2188984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2189984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2190674ae819SStefano Zampini j = 0; 2191984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2192674ae819SStefano Zampini total_counts = total_counts-j; 219345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2194e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2195c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2196c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2197c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2198c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2199c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2200c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2201674ae819SStefano Zampini if (j<temp_constraints) { 2202984c4197SStefano Zampini PetscInt ii; 2203984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2204674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 22059162d606SStefano 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)); 2206674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2207984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2208674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 22099162d606SStefano 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]; 2210674ae819SStefano Zampini } 2211674ae819SStefano Zampini } 2212674ae819SStefano Zampini } 2213674ae819SStefano Zampini #else /* on missing GESVD */ 2214e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2215e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2216b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2217674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2218674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 22199162d606SStefano 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)); 2220674ae819SStefano Zampini #else 22219162d606SStefano 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)); 2222674ae819SStefano Zampini #endif 2223984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2224674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2225984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2226e310c8b4SStefano Zampini k = temp_constraints; 2227e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2228674ae819SStefano Zampini j = 0; 2229e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 223045a1bb75SStefano Zampini valid_constraints = k-j; 2231911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2232984c4197SStefano Zampini #endif /* on missing GESVD */ 2233674ae819SStefano Zampini } 2234a773dcb8SStefano Zampini } 22359162d606SStefano Zampini /* update pointers information */ 22369162d606SStefano Zampini if (valid_constraints) { 22379162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 22389162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 22399162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 22409162d606SStefano Zampini /* set change_of_basis flag */ 224145a1bb75SStefano Zampini if (boolforchange) { 2242*b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 22439162d606SStefano Zampini } 2244*b3d85658SStefano Zampini total_counts_cc++; 224545a1bb75SStefano Zampini } 2246674ae819SStefano Zampini } 2247984c4197SStefano Zampini /* free workspace */ 22488f1c130eSStefano Zampini if (!skip_lapack) { 2249984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2250984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2251984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2252984c4197SStefano Zampini #endif 2253984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2254984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2255984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2256984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2257984c4197SStefano Zampini #endif 2258984c4197SStefano Zampini } 2259984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2260984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2261984c4197SStefano Zampini } 2262984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2263cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2264cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2265cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2266cf5a6209SStefano Zampini } 2267cf5a6209SStefano Zampini if (n_ISForFaces) { 2268cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2269cf5a6209SStefano Zampini } 2270cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2271cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2272cf5a6209SStefano Zampini } 2273cf5a6209SStefano Zampini if (n_ISForEdges) { 2274cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2275cf5a6209SStefano Zampini } 2276cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 227708122e43SStefano Zampini } else { 227808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 227908122e43SStefano Zampini 228008122e43SStefano Zampini total_counts = 0; 228108122e43SStefano Zampini n_vertices = 0; 22828bec7fa6SStefano Zampini if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { 228308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&n_vertices);CHKERRQ(ierr); 228408122e43SStefano Zampini } 228508122e43SStefano Zampini max_constraints = 0; 22869162d606SStefano Zampini total_counts_cc = 0; 228708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 228808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 22899162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 229008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 229108122e43SStefano Zampini } 22929162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 22939162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 22949162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 22959162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 229674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 22979162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 22989162d606SStefano Zampini total_counts_cc = 0; 22999162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 23009162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 23019162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 230208122e43SStefano Zampini } 230308122e43SStefano Zampini } 23049162d606SStefano Zampini #if 0 23059162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 23069162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23079162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 23089162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 23099162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 23109162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 23119162d606SStefano Zampini } 23129162d606SStefano Zampini printf("\n"); 23139162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 23149162d606SStefano Zampini } 23151b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 23168bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 23171b968477SStefano Zampini } 23181b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 23198bec7fa6SStefano 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]); 23201b968477SStefano Zampini } 232108122e43SStefano Zampini #endif 232208122e43SStefano Zampini 23238bec7fa6SStefano Zampini max_size_of_constraint = 0; 23249162d606SStefano 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]); 23259162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 232608122e43SStefano Zampini /* Change of basis */ 2327*b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 232808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 232908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 233008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2331*b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 233208122e43SStefano Zampini } 233308122e43SStefano Zampini } 233408122e43SStefano Zampini } 233508122e43SStefano Zampini } 2336b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 23370e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 233808122e43SStefano Zampini 23399162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 23409162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 23419162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 23429162d606SStefano 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); 234308122e43SStefano Zampini } 2344984c4197SStefano Zampini 2345674ae819SStefano Zampini /* Create constraint matrix */ 2346674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 234716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2348984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2349984c4197SStefano Zampini 2350984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 235174d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 235274d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 235374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2354984c4197SStefano Zampini total_primal_vertices=0; 2355*b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 23569162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 23579162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2358984c4197SStefano Zampini if (size_of_constraint == 1) { 23599162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2360*b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 236164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 23629162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 23639162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 23649162d606SStefano Zampini } 2365*b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 236674d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 236774d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 236874d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 236974d5cdf7SStefano Zampini } 2370*b3d85658SStefano Zampini } else { 2371*b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2372984c4197SStefano Zampini } 2373984c4197SStefano Zampini } 2374b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2375b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 237670022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 237770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2378*b3d85658SStefano Zampini 2379*b3d85658SStefano 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); 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; 2390*b3d85658SStefano Zampini cum = total_primal_vertices; 23919162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 23924641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2393*b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2394*b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2395*b3d85658SStefano Zampini cum++; 23969162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 239774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 239874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 239974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 240074d5cdf7SStefano Zampini } 24019162d606SStefano Zampini j += constraints_n[i]; 2402674ae819SStefano Zampini } 2403674ae819SStefano Zampini } 2404674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2405674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2406088faed8SStefano Zampini 2407674ae819SStefano Zampini /* set values in constraint matrix */ 2408984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 24090e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2410674ae819SStefano Zampini } 2411984c4197SStefano Zampini total_counts = total_primal_vertices; 24129162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24134641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 24149162d606SStefano Zampini PetscInt *cols; 24159162d606SStefano Zampini 24169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 24179162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 24189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24199162d606SStefano Zampini PetscInt row = total_counts+k; 24209162d606SStefano Zampini PetscScalar *vals; 24219162d606SStefano Zampini 24229162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 24239162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 24249162d606SStefano Zampini } 24259162d606SStefano Zampini total_counts += constraints_n[i]; 2426674ae819SStefano Zampini } 2427674ae819SStefano Zampini } 2428674ae819SStefano Zampini /* assembling */ 2429674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2430674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2431088faed8SStefano Zampini 2432984c4197SStefano Zampini /* 243345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2434984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2435984c4197SStefano Zampini */ 2436674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2437674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2438026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2439984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2440984c4197SStefano Zampini /* working stuff for GEQRF */ 244181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2442984c4197SStefano Zampini PetscBLASInt lqr_work; 2443984c4197SStefano Zampini /* working stuff for UNGQR */ 2444984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2445984c4197SStefano Zampini PetscBLASInt lgqr_work; 2446984c4197SStefano Zampini /* working stuff for TRTRS */ 2447984c4197SStefano Zampini PetscScalar *trs_rhs; 24483f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2449984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2450984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2451984c4197SStefano Zampini PetscScalar *start_vals; 2452984c4197SStefano Zampini /* working stuff for values insertion */ 24534641a718SStefano Zampini PetscBT is_primal; 245464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2455906d46d4SStefano Zampini /* matrix sizes */ 2456906d46d4SStefano Zampini PetscInt global_size,local_size; 2457906d46d4SStefano Zampini /* temporary change of basis */ 2458906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2459cf5a6209SStefano Zampini /* extra space for debugging */ 2460cf5a6209SStefano Zampini PetscScalar *dbg_work; 2461984c4197SStefano Zampini 2462906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2463906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 246416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2465bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2466906d46d4SStefano Zampini /* nonzeros for local mat */ 2467bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2468bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 24699162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2470a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 24719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2472a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 24739162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2474a717540cSStefano Zampini } else { 24759162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 24769162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2477a717540cSStefano Zampini } 2478a717540cSStefano Zampini } 2479a717540cSStefano Zampini } 2480906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2481bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2482a717540cSStefano Zampini /* Set initial identity in the matrix */ 2483bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2484906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2485a717540cSStefano Zampini } 2486a717540cSStefano Zampini 2487a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2488a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2489a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2490a717540cSStefano Zampini } 2491a717540cSStefano Zampini 2492a717540cSStefano Zampini 2493a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2494a717540cSStefano Zampini /* 2495a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2496a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2497a717540cSStefano Zampini 2498a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2499a717540cSStefano Zampini 2500a6b551f4SStefano 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) 2501a6b551f4SStefano Zampini 2502a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2503a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2504a717540cSStefano Zampini | ... | 2505a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2506a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2507a717540cSStefano Zampini 2508a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2509a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2510a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2511a6b551f4SStefano Zampini 2512a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2513a717540cSStefano Zampini */ 2514a717540cSStefano Zampini if (qr_needed) { 2515984c4197SStefano Zampini /* space to store Q */ 2516854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2517984c4197SStefano Zampini /* first we issue queries for optimal work */ 25183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25193f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 25203f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2521984c4197SStefano Zampini lqr_work = -1; 25223f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2523984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2524984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2525785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2526984c4197SStefano Zampini lgqr_work = -1; 25273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 25283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 25293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 25303f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25313f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 25323f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2533984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2534984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2535785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2536984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2537785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2538984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2539785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2540a717540cSStefano Zampini /* allocating workspace for check */ 2541a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2542cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2543a717540cSStefano Zampini } 2544a717540cSStefano Zampini } 2545984c4197SStefano Zampini /* array to store whether a node is primal or not */ 25464641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2547473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 25480e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 254939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 255039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 25514641a718SStefano Zampini } 255239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 255339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 255439e2fb2aSStefano Zampini } 255539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2556984c4197SStefano Zampini 2557a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 25589162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 25599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 25604641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2561984c4197SStefano Zampini /* get constraint info */ 25629162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2563984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2564984c4197SStefano Zampini 2565984c4197SStefano Zampini if (pcbddc->dbg_flag) { 25669162d606SStefano 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); 2567674ae819SStefano Zampini } 2568984c4197SStefano Zampini 2569fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2570a717540cSStefano Zampini 2571a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2572a717540cSStefano Zampini if (pcbddc->dbg_flag) { 25739162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2574a717540cSStefano Zampini } 2575984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 25769162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2577984c4197SStefano Zampini 2578984c4197SStefano Zampini /* compute QR decomposition of constraints */ 25793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25803f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2582674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25833f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2584984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2585674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2586984c4197SStefano Zampini 2587984c4197SStefano Zampini /* explictly compute R^-T */ 2588984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2589984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 25903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 25923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2594984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25953f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2596984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2597984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2598984c4197SStefano Zampini 2599a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 26003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2604984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2606984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2607984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2608984c4197SStefano Zampini 2609984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2610984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2611984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 26123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 26153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 26173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2618984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26199162d606SStefano 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)); 2620984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 26219162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2622984c4197SStefano Zampini 2623984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 26249162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2625984c4197SStefano Zampini /* insert cols for primal dofs */ 2626984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2627984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 26289162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2629906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2630984c4197SStefano Zampini } 2631984c4197SStefano Zampini /* insert cols for dual dofs */ 2632984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 26339162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2634984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 26359162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2636906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2637984c4197SStefano Zampini j++; 2638674ae819SStefano Zampini } 2639674ae819SStefano Zampini } 2640984c4197SStefano Zampini 2641984c4197SStefano Zampini /* check change of basis */ 2642984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2643984c4197SStefano Zampini PetscInt ii,jj; 2644984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2645c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2646c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2647c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2648c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2649c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2650c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2651984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2652cf5a6209SStefano 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)); 2653984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2654984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2655984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2656cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2657cf5a6209SStefano 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; 2658674ae819SStefano Zampini } 2659674ae819SStefano Zampini } 2660984c4197SStefano Zampini if (!valid_qr) { 266122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2662984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2663984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2664cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2665cf5a6209SStefano 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])); 2666674ae819SStefano Zampini } 2667cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2668cf5a6209SStefano 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])); 2669984c4197SStefano Zampini } 2670984c4197SStefano Zampini } 2671984c4197SStefano Zampini } 2672674ae819SStefano Zampini } else { 267322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2674674ae819SStefano Zampini } 2675674ae819SStefano Zampini } 2676a717540cSStefano Zampini } else { /* simple transformation block */ 2677a717540cSStefano Zampini PetscInt row,col; 2678a6b551f4SStefano Zampini PetscScalar val,norm; 2679a6b551f4SStefano Zampini 2680a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26819162d606SStefano 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)); 2682a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 26839162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 26849162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2685bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 26869162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2687906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 26889162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2689a717540cSStefano Zampini } else { 2690a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 26919162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2692a717540cSStefano Zampini if (row != col) { 26939162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2694a717540cSStefano Zampini } else { 26959162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2696a717540cSStefano Zampini } 2697906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2698a717540cSStefano Zampini } 2699a717540cSStefano Zampini } 2700a717540cSStefano Zampini } 270198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 270222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2703a717540cSStefano Zampini } 2704674ae819SStefano Zampini } 2705984c4197SStefano Zampini } else { 2706984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27079162d606SStefano 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); 2708674ae819SStefano Zampini } 2709674ae819SStefano Zampini } 2710674ae819SStefano Zampini } 2711a717540cSStefano Zampini 2712a717540cSStefano Zampini /* free workspace */ 2713a717540cSStefano Zampini if (qr_needed) { 2714984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2715cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2716984c4197SStefano Zampini } 2717984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2718984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2719984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2720984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2721984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2722674ae819SStefano Zampini } 2723a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2724906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2725906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2726906d46d4SStefano Zampini 2727906d46d4SStefano Zampini /* assembling of global change of variable */ 2728bbb9e6c6SStefano Zampini { 2729bbb9e6c6SStefano Zampini Mat tmat; 273016f15bc4SStefano Zampini PetscInt bs; 273116f15bc4SStefano Zampini 2732906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2733906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2734bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2735bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2736bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2737bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 273816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 273916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2740906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2741bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2742bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2743bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2744bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2745bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2746bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2747bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2748bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2749bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2750906d46d4SStefano Zampini } 2751906d46d4SStefano Zampini /* check */ 2752906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2753906d46d4SStefano Zampini PetscReal error; 2754906d46d4SStefano Zampini Vec x,x_change; 2755906d46d4SStefano Zampini 2756906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2757906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2758906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2759906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2760bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2761bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2762bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2763bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2764bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2765906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2766906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2767906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2768906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2769bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2770906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2771906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2772906d46d4SStefano Zampini } 2773b96c3477SStefano Zampini 2774b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2775b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2776b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2777b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2778ac632422SStefano Zampini Mat S_new,tmat; 2779bbb9e6c6SStefano Zampini IS is_all_N; 2780bbb9e6c6SStefano Zampini 2781bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 2782d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr); 2783bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2784ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2785b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2786ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2787ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2788ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2789ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2790ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2791b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2792ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2793ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2794ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2795ac632422SStefano Zampini } 2796b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2797b96c3477SStefano Zampini } 2798b96c3477SStefano Zampini } 2799906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2800906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2801b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2802b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2803b9b85e73SStefano Zampini } 2804906d46d4SStefano Zampini 2805906d46d4SStefano Zampini /* set up change of basis context */ 2806906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2807906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2808906d46d4SStefano Zampini 2809906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2810906d46d4SStefano Zampini PetscInt global_size,local_size; 2811906d46d4SStefano Zampini 2812906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2813906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2814906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2815906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2816906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2817906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2818906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2819906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2820906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2821906d46d4SStefano Zampini } else { 2822906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2823906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2824906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2825906d46d4SStefano Zampini } 2826906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2827906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2828906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2829906d46d4SStefano Zampini } else { 2830906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2831906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2832906d46d4SStefano Zampini } 2833906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2834906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2835906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2836906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2837b9b85e73SStefano Zampini } 2838a717540cSStefano Zampini 2839727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2840727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2841727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 28420e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 28430e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 28440e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 28450e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2846c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2847727cdba6SStefano Zampini } 28480e6343abSStefano Zampini } 28490e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2850727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2851727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2852727cdba6SStefano Zampini 2853a717540cSStefano Zampini /* flush dbg viewer */ 2854b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2855b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2856b8ffe317SStefano Zampini } 2857a717540cSStefano Zampini 2858e310c8b4SStefano Zampini /* free workspace */ 2859a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 28604641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 286108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 28629162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 28639162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 286408122e43SStefano Zampini } else { 28659162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 28669162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 28679162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 286808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 286908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 28709162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 28719162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 287208122e43SStefano Zampini } 2873674ae819SStefano Zampini PetscFunctionReturn(0); 2874674ae819SStefano Zampini } 2875674ae819SStefano Zampini 2876674ae819SStefano Zampini #undef __FUNCT__ 2877674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2878674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2879674ae819SStefano Zampini { 2880674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2881674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2882674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 28837fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2884674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2885674ae819SStefano Zampini 2886674ae819SStefano Zampini PetscFunctionBegin; 28878e61c736SStefano Zampini /* Reset previously computed graph */ 28888e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2889674ae819SStefano Zampini /* Init local Graph struct */ 28907fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 28917fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 2892674ae819SStefano Zampini 2893575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2894575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2895575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2896575ad6abSStefano Zampini } 28979577ea80SStefano Zampini 2898674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 28994d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 29004d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 29014d379d7bSStefano Zampini PetscInt nvtxs; 2902e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 2903674ae819SStefano Zampini 29044d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 29052fffb893SStefano Zampini 29062fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29072fffb893SStefano Zampini if (flg_row) { 29084d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2909b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 29102fffb893SStefano Zampini } 29112fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29129b28b941SStefano 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 */ 29134d379d7bSStefano Zampini IS is_dummy; 29144d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 29154d379d7bSStefano Zampini PetscInt j,sum; 29164d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 29174d379d7bSStefano Zampini const PetscInt *idxs; 29184d379d7bSStefano Zampini PCBDDCGraph graph; 29194d379d7bSStefano Zampini PetscBT is_on_boundary; 29204d379d7bSStefano Zampini 29214d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 29224d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 29234d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 29244d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 29257fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 29264d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 2927e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2928e496cd5dSStefano Zampini if (flg_row) { 29294d379d7bSStefano Zampini graph->xadj = xadj; 29304d379d7bSStefano Zampini graph->adjncy = adjncy; 2931e496cd5dSStefano Zampini } 29324d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 29334d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2934e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 29354d379d7bSStefano Zampini 29364d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 29379b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 29384d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29394d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 29404d379d7bSStefano Zampini } 29414d379d7bSStefano Zampini } 29424d379d7bSStefano Zampini 2943e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 29444d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29454d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 29464d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 29474d379d7bSStefano Zampini } 29484d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29494d379d7bSStefano Zampini 2950e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 29514d379d7bSStefano Zampini sum = 0; 29524d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29534d379d7bSStefano Zampini PetscInt sizecc = 0; 29544d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29554d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29564d379d7bSStefano Zampini sizecc++; 29574d379d7bSStefano Zampini } 29584d379d7bSStefano Zampini } 29594d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29604d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29614d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 29624d379d7bSStefano Zampini } 29634d379d7bSStefano Zampini } 29644d379d7bSStefano Zampini sum += sizecc*sizecc; 29654d379d7bSStefano Zampini } 29664d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 29674d379d7bSStefano Zampini sum = 0; 2968e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 29694d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 29704d379d7bSStefano Zampini cxadj[i] = sum; 29714d379d7bSStefano Zampini sum += temp; 29724d379d7bSStefano Zampini } 2973e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 29744d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29754d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29764d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29774d379d7bSStefano Zampini PetscInt k,sizecc = 0; 29784d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 29794d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 29804d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 29814d379d7bSStefano Zampini sizecc++; 29824d379d7bSStefano Zampini } 29834d379d7bSStefano Zampini } 29844d379d7bSStefano Zampini } 29854d379d7bSStefano Zampini } 29864d379d7bSStefano Zampini } 29879b28b941SStefano Zampini if (sum) { 2988e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 29894d379d7bSStefano Zampini } else { 29904d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 29914d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 29924d379d7bSStefano Zampini } 29934d379d7bSStefano Zampini graph->xadj = 0; 29944d379d7bSStefano Zampini graph->adjncy = 0; 29954d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 29964d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 29974d379d7bSStefano Zampini } 2998674ae819SStefano Zampini } 29999b28b941SStefano Zampini if (pcbddc->dbg_flag) { 30009b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30019b28b941SStefano Zampini } 3002674ae819SStefano Zampini 300363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3004674ae819SStefano Zampini vertex_size = 1; 300563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 300663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 300795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 300863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3009a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 301063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3011674ae819SStefano Zampini } 301263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 301363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 301463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3015674ae819SStefano Zampini } 301663602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3017674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 301863602bcaSStefano Zampini } else { 301963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 302063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3021854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 302263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 302363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 302463602bcaSStefano Zampini } 302563602bcaSStefano Zampini } 3026674ae819SStefano Zampini } 3027674ae819SStefano Zampini 3028674ae819SStefano Zampini /* Setup of Graph */ 3029785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3030a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3031785d1243SStefano Zampini } 3032785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3033a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3034785d1243SStefano Zampini } 3035302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3036674ae819SStefano Zampini 3037674ae819SStefano Zampini /* Graph's connected components analysis */ 3038674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3039674ae819SStefano Zampini 3040674ae819SStefano Zampini /* print some info to stdout */ 3041674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3042302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3043674ae819SStefano Zampini } 3044fb180af4SStefano Zampini 3045fb180af4SStefano Zampini /* mark topography has done */ 3046fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3047674ae819SStefano Zampini PetscFunctionReturn(0); 3048674ae819SStefano Zampini } 3049674ae819SStefano Zampini 3050674ae819SStefano Zampini #undef __FUNCT__ 3051674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3052674ae819SStefano 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[]) 3053674ae819SStefano Zampini { 3054674ae819SStefano Zampini Vec local_vec,global_vec; 3055674ae819SStefano Zampini IS seqis,paris; 3056674ae819SStefano Zampini VecScatter scatter_ctx; 3057674ae819SStefano Zampini PetscScalar *array; 3058674ae819SStefano Zampini PetscInt *temp_global_dofs; 3059674ae819SStefano Zampini PetscScalar globalsum; 3060674ae819SStefano Zampini PetscInt i,j,s; 306117ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 306217ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 306317ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3064674ae819SStefano Zampini PetscBool first_found; 3065674ae819SStefano Zampini PetscErrorCode ierr; 3066674ae819SStefano Zampini 3067674ae819SStefano Zampini PetscFunctionBegin; 3068674ae819SStefano Zampini /* mpi buffers */ 3069b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3070b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3071674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 307217ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3073674ae819SStefano Zampini /* get maximum size of subset */ 3074785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3075674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3076674ae819SStefano Zampini max_local = 0; 3077b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3078674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3079674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3080674ae819SStefano Zampini } 3081674ae819SStefano Zampini } 3082b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3083674ae819SStefano Zampini max_global++; 3084674ae819SStefano Zampini max_local = 0; 3085b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3086674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3087674ae819SStefano Zampini max_local = local_dofs[i]; 3088674ae819SStefano Zampini } 3089674ae819SStefano Zampini } 3090674ae819SStefano Zampini max_local++; 3091674ae819SStefano Zampini /* allocate workspace */ 3092674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3093674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3094674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3095674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3096674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3097674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3098674ae819SStefano Zampini /* create scatter */ 3099674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3100674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3101674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3102674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3103674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3104674ae819SStefano Zampini /* init array */ 3105674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3106674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3107674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3108674ae819SStefano Zampini if (local_dofs_mult) { 3109674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3110674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3111674ae819SStefano Zampini } 3112674ae819SStefano Zampini } else { 3113674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3114674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3115674ae819SStefano Zampini } 3116674ae819SStefano Zampini } 3117674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3118674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3119674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3120674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3121674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 31225b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3123674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3124674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3125674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3126674ae819SStefano Zampini nlocals = 0; 3127674ae819SStefano Zampini first_index = -1; 3128674ae819SStefano Zampini first_found = PETSC_FALSE; 3129674ae819SStefano Zampini for (i=0;i<s;i++) { 3130b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3131674ae819SStefano Zampini first_found = PETSC_TRUE; 3132674ae819SStefano Zampini first_index = i; 3133674ae819SStefano Zampini } 31345b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3135674ae819SStefano Zampini } 3136674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3137674ae819SStefano Zampini if (!rank_prec_comm) { 3138674ae819SStefano Zampini dof_displs[0]=0; 3139674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3140674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3141674ae819SStefano Zampini } 3142674ae819SStefano Zampini } 3143674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3144674ae819SStefano Zampini if (first_found) { 3145674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3146674ae819SStefano Zampini old_index = first_index; 3147674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3148b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3149674ae819SStefano Zampini array[i] += array[old_index]; 3150674ae819SStefano Zampini old_index = i; 3151674ae819SStefano Zampini } 3152674ae819SStefano Zampini } 3153674ae819SStefano Zampini } 3154674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3155674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3156674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3157674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3158674ae819SStefano Zampini /* get global ordering of local dofs */ 31592b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3160674ae819SStefano Zampini if (local_dofs_mult) { 3161674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31625b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3163674ae819SStefano Zampini } 3164674ae819SStefano Zampini } else { 3165674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31665b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3167674ae819SStefano Zampini } 3168674ae819SStefano Zampini } 31692b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3170674ae819SStefano Zampini /* free workspace */ 3171674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3172674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3173674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 317417ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3175674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3176674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3177674ae819SStefano Zampini PetscFunctionReturn(0); 3178674ae819SStefano Zampini } 31799a7d3425SStefano Zampini 31809a7d3425SStefano Zampini #undef __FUNCT__ 31819a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 31829a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 31839a7d3425SStefano Zampini { 31849a7d3425SStefano Zampini PetscInt i,j; 31859a7d3425SStefano Zampini PetscScalar *alphas; 31869a7d3425SStefano Zampini PetscErrorCode ierr; 31879a7d3425SStefano Zampini 31889a7d3425SStefano Zampini PetscFunctionBegin; 31899a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3190785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 31919a7d3425SStefano Zampini for (i=0;i<n;i++) { 31929a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 31939a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 31949a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 31959a7d3425SStefano Zampini } 31969a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 31979a7d3425SStefano Zampini PetscFunctionReturn(0); 31989a7d3425SStefano Zampini } 31999a7d3425SStefano Zampini 3200e7931f94SStefano Zampini #undef __FUNCT__ 320170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 320228143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3203e7931f94SStefano Zampini { 3204e7931f94SStefano Zampini Mat subdomain_adj; 3205e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3206e7931f94SStefano Zampini MatPartitioning partitioner; 3207e7931f94SStefano Zampini Mat_IS *matis; 3208e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3209e7931f94SStefano Zampini PetscInt prank; 3210e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3211e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3212e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 32133837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3214e7931f94SStefano Zampini PetscErrorCode ierr; 32152b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3216e7931f94SStefano Zampini PetscSubcomm subcomm; 3217a57a6d2fSStefano Zampini 3218e7931f94SStefano Zampini PetscFunctionBegin; 32192b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 32202b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 32212b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3222e7931f94SStefano Zampini 3223e7931f94SStefano Zampini /* Get info on mapping */ 3224e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3225e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3226e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3227e7931f94SStefano Zampini 3228e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3229785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3230e7931f94SStefano Zampini xadj[0] = 0; 3231e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3232785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3233785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3234e7931f94SStefano Zampini 32352b510759SStefano Zampini if (threshold) { 3236d023bfaeSStefano Zampini PetscInt xadj_count = 0; 32372b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3238d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3239d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3240d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3241d023bfaeSStefano Zampini xadj_count++; 3242e7931f94SStefano Zampini } 3243e7931f94SStefano Zampini } 3244d023bfaeSStefano Zampini xadj[1] = xadj_count; 3245c8587f34SStefano Zampini } else { 3246e7931f94SStefano Zampini if (xadj[1]) { 3247e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3248e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3249c8587f34SStefano Zampini } 3250e7931f94SStefano Zampini } 3251e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3252e7931f94SStefano Zampini if (use_square) { 3253e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3254e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3255e7931f94SStefano Zampini } 3256e7931f94SStefano Zampini } 3257e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3258e7931f94SStefano Zampini 32593837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3260e7931f94SStefano Zampini 3261e7931f94SStefano Zampini /* 3262e7931f94SStefano Zampini Restrict work on active processes only. 3263e7931f94SStefano Zampini */ 3264e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3265e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3266e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 32672b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3268d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3269e7931f94SStefano Zampini if (color) { 3270e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3271e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3272e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3273c8587f34SStefano Zampini } else { 327428143c3dSStefano Zampini PetscInt coarsening_ratio; 3275306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3276785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3277e7931f94SStefano Zampini prank = rank; 3278306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 32798002ef2cSStefano Zampini /* 3280e7931f94SStefano Zampini for (i=0;i<size;i++) { 3281e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3282c8587f34SStefano Zampini } 32838002ef2cSStefano Zampini */ 3284e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3285e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3286c8587f34SStefano Zampini } 3287e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3288306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 328922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3290e7931f94SStefano Zampini 3291e7931f94SStefano Zampini /* Partition */ 3292306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3293e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3294e7931f94SStefano Zampini if (use_vwgt) { 32953837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3296e7931f94SStefano Zampini v_wgt[0] = local_size; 3297e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3298c8587f34SStefano Zampini } 329928143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 330028143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 330128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3302e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3303e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 330422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3305e7931f94SStefano Zampini 3306e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 330728143c3dSStefano Zampini if (contiguous) { 330828143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 330928143c3dSStefano Zampini } else { 331084ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 331128143c3dSStefano Zampini } 3312e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3313e7931f94SStefano Zampini /* clean up */ 3314e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3315e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3316e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3317e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3318e7931f94SStefano Zampini } 3319e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3320e7931f94SStefano Zampini 3321e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3322e7931f94SStefano Zampini i = 1; 3323e7931f94SStefano Zampini if (color) i=0; 3324e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3325e7931f94SStefano Zampini 3326e7931f94SStefano Zampini /* get back IS */ 3327e7931f94SStefano Zampini *is_sends = ranks_send_to; 3328e7931f94SStefano Zampini PetscFunctionReturn(0); 3329e7931f94SStefano Zampini } 3330e7931f94SStefano Zampini 3331e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3332e7931f94SStefano Zampini 3333e7931f94SStefano Zampini #undef __FUNCT__ 3334e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 333528143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3336e7931f94SStefano Zampini { 333770cf5478SStefano Zampini Mat local_mat; 3338e7931f94SStefano Zampini Mat_IS *matis; 3339e7931f94SStefano Zampini IS is_sends_internal; 33409d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 334128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 33429d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3343e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3344e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3345e7931f94SStefano Zampini const PetscInt* is_indices; 3346e7931f94SStefano Zampini MatType new_local_type; 3347e7931f94SStefano Zampini /* buffers */ 3348e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 334928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 33509d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3351e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3352e7931f94SStefano Zampini /* MPI */ 335328143c3dSStefano Zampini MPI_Comm comm,comm_n; 335428143c3dSStefano Zampini PetscSubcomm subcomm; 3355e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 335628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 335728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 335828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 335928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 336028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3361e7931f94SStefano Zampini PetscErrorCode ierr; 3362e7931f94SStefano Zampini 3363e7931f94SStefano Zampini PetscFunctionBegin; 336428143c3dSStefano Zampini /* TODO: add missing checks */ 336528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 336628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 336728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 336828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3369e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 337028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3371e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3372e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3373e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3374e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3375e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 337628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 337770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 337870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 337928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 338070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 338170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 338270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 338370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 338470cf5478SStefano Zampini } 3385e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3386e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3387e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3388e7931f94SStefano Zampini if (!is_sends) { 338928143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 339028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 339128143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3392c8587f34SStefano Zampini } else { 3393e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3394e7931f94SStefano Zampini is_sends_internal = is_sends; 3395c8587f34SStefano Zampini } 3396e7931f94SStefano Zampini 3397e7931f94SStefano Zampini /* get pointer of MATIS data */ 3398e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3399e7931f94SStefano Zampini 3400e7931f94SStefano Zampini /* get comm */ 3401a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3402e7931f94SStefano Zampini 3403e7931f94SStefano Zampini /* compute number of sends */ 3404e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3405e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3406e7931f94SStefano Zampini 3407e7931f94SStefano Zampini /* compute number of receives */ 3408e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3409785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3410e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3411e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3412e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3413e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3414e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3415e7931f94SStefano Zampini 341628143c3dSStefano Zampini /* restrict comm if requested */ 341728143c3dSStefano Zampini subcomm = 0; 341828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 341928143c3dSStefano Zampini if (restrict_comm) { 3420779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3421779c1cceSStefano Zampini 342228143c3dSStefano Zampini color = 0; 3423779c1cceSStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */ 342428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 342528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 342628143c3dSStefano Zampini /* check if reuse has been requested */ 342728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 342828143c3dSStefano Zampini if (*mat_n) { 342928143c3dSStefano Zampini PetscMPIInt subcommsize2; 343028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 343128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 343228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 343328143c3dSStefano Zampini } else { 343428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 343528143c3dSStefano Zampini } 343628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3437779c1cceSStefano Zampini PetscMPIInt rank; 3438779c1cceSStefano Zampini 3439779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 344028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 344128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 344228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3443306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 344428143c3dSStefano Zampini } 344528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 344628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 344728143c3dSStefano Zampini } else { 344828143c3dSStefano Zampini comm_n = comm; 344928143c3dSStefano Zampini } 345028143c3dSStefano Zampini 3451e7931f94SStefano Zampini /* prepare send/receive buffers */ 3452785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3453e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3454785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3455e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 345628143c3dSStefano Zampini if (nis) { 3457854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 345828143c3dSStefano Zampini } 3459e7931f94SStefano Zampini 346028143c3dSStefano Zampini /* Get data from local matrices */ 3461e7931f94SStefano Zampini if (!isdense) { 3462a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3463e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3464e7931f94SStefano Zampini /* 3465e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3466e7931f94SStefano Zampini send_buffer_idxs should contain: 3467e7931f94SStefano Zampini - MatType_PRIVATE type 3468e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3469e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3470e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3471e7931f94SStefano Zampini */ 3472e7931f94SStefano Zampini } else { 3473e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3474e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3475854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3476e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3477e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3478e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3479e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3480e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3481e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3482e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3483e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3484e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3485c8587f34SStefano Zampini } 3486c8587f34SStefano Zampini } 3487e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 348828143c3dSStefano Zampini /* additional is (if any) */ 348928143c3dSStefano Zampini if (nis) { 349028143c3dSStefano Zampini PetscMPIInt psum; 349128143c3dSStefano Zampini PetscInt j; 349228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 349328143c3dSStefano Zampini PetscInt plen; 349428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 349528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 349628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 349728143c3dSStefano Zampini } 3498854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 349928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 350028143c3dSStefano Zampini PetscInt plen; 350128143c3dSStefano Zampini const PetscInt *is_array_idxs; 350228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 350328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 350428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 350528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 350628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 350728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 350828143c3dSStefano Zampini } 350928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 351028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 351128143c3dSStefano Zampini } 351228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 351328143c3dSStefano Zampini } 351428143c3dSStefano Zampini 3515e7931f94SStefano Zampini buf_size_idxs = 0; 3516e7931f94SStefano Zampini buf_size_vals = 0; 351728143c3dSStefano Zampini buf_size_idxs_is = 0; 3518e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3519e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3520e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 352128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3522e7931f94SStefano Zampini } 3523785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3524785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 352595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3526e7931f94SStefano Zampini 3527e7931f94SStefano Zampini /* get new tags for clean communications */ 3528e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3529e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 353028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3531e7931f94SStefano Zampini 3532e7931f94SStefano Zampini /* allocate for requests */ 3533785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3534785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 353595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3536785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3537785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 353895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3539e7931f94SStefano Zampini 3540e7931f94SStefano Zampini /* communications */ 3541e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3542e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 354328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3544e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3545e7931f94SStefano Zampini source_dest = onodes[i]; 3546e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3547e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3548e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3549e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 355028143c3dSStefano Zampini if (nis) { 355128143c3dSStefano 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); 355228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 355328143c3dSStefano Zampini } 3554e7931f94SStefano Zampini } 3555e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3556e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3557e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3558e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 355928143c3dSStefano Zampini if (nis) { 356028143c3dSStefano 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); 356128143c3dSStefano Zampini } 3562e7931f94SStefano Zampini } 3563e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3564e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3565e7931f94SStefano Zampini 3566e7931f94SStefano Zampini /* assemble new l2g map */ 3567e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3568e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35699d30be91SStefano Zampini new_local_rows = 0; 3570e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35719d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3572e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3573e7931f94SStefano Zampini } 35749d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3575e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35769d30be91SStefano Zampini new_local_rows = 0; 3577e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35789d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 35799d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3580e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3581e7931f94SStefano Zampini } 35829d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 35839d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3584e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3585e7931f94SStefano Zampini 3586e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3587e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3588e7931f94SStefano 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) */ 3589e7931f94SStefano Zampini if (n_recvs) { 359028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3591e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3592e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3593e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3594e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3595e7931f94SStefano Zampini break; 3596e7931f94SStefano Zampini } 3597e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3598e7931f94SStefano Zampini } 3599e7931f94SStefano Zampini switch (new_local_type_private) { 360028143c3dSStefano Zampini case MATDENSE_PRIVATE: 360128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3602e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3603e7931f94SStefano Zampini bs = 1; 360428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 360528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 360628143c3dSStefano Zampini bs = 1; 360728143c3dSStefano Zampini } 3608e7931f94SStefano Zampini break; 3609e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3610e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3611e7931f94SStefano Zampini bs = 1; 3612e7931f94SStefano Zampini break; 3613e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3614e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3615e7931f94SStefano Zampini break; 3616e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3617e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3618e7931f94SStefano Zampini break; 3619e7931f94SStefano Zampini default: 36209d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3621e7931f94SStefano Zampini break; 3622e7931f94SStefano Zampini } 362328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 362428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 362528143c3dSStefano Zampini bs = 1; 3626e7931f94SStefano Zampini } 3627e7931f94SStefano Zampini 362870cf5478SStefano Zampini /* create MATIS object if needed */ 362970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3630e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 363128143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 363270cf5478SStefano Zampini } else { 363370cf5478SStefano Zampini /* it also destroys the local matrices */ 363470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 363570cf5478SStefano Zampini } 363670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3637e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 36389d30be91SStefano Zampini 36399d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 36409d30be91SStefano Zampini 36419d30be91SStefano Zampini /* Global to local map of received indices */ 36429d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 36439d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 36449d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 36459d30be91SStefano Zampini 36469d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 36479d30be91SStefano Zampini buf_size_idxs = 0; 36489d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 36499d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 36509d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 36519d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 36529d30be91SStefano Zampini } 36539d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 36549d30be91SStefano Zampini 36559d30be91SStefano Zampini /* set preallocation */ 36569d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 36579d30be91SStefano Zampini if (!newisdense) { 36589d30be91SStefano Zampini PetscInt *new_local_nnz=0; 36599d30be91SStefano Zampini 36609d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 36619d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 36629d30be91SStefano Zampini if (n_recvs) { 36639d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 36649d30be91SStefano Zampini } 36659d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 36669d30be91SStefano Zampini PetscInt j; 36679d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 36689d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 36699d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 36709d30be91SStefano Zampini } 36719d30be91SStefano Zampini } else { 36729d30be91SStefano Zampini /* TODO */ 36739d30be91SStefano Zampini } 36749d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 36759d30be91SStefano Zampini } 36769d30be91SStefano Zampini if (new_local_nnz) { 36779d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 36789d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 36799d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 36809d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36819d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 36829d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36839d30be91SStefano Zampini } else { 36849d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36859d30be91SStefano Zampini } 36869d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 36879d30be91SStefano Zampini } else { 36889d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36899d30be91SStefano Zampini } 3690e7931f94SStefano Zampini 3691e7931f94SStefano Zampini /* set values */ 3692e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 36939d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3694e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3695e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3696e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 36979d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3698e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3699e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3700e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 370128143c3dSStefano Zampini } else { 370228143c3dSStefano Zampini /* TODO */ 3703e7931f94SStefano Zampini } 3704e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3705e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3706e7931f94SStefano Zampini } 3707e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3708e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 370970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 371070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 37119d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 37129d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3713e7931f94SStefano Zampini 3714dfd14d43SStefano Zampini #if 0 371528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3716e7931f94SStefano Zampini Vec lvec,rvec; 3717e7931f94SStefano Zampini PetscReal infty_error; 3718e7931f94SStefano Zampini 37192a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3720e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3721e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3722e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 372370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3724e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3725e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3726e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3727e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3728e7931f94SStefano Zampini } 372928143c3dSStefano Zampini #endif 3730e7931f94SStefano Zampini 373128143c3dSStefano Zampini /* assemble new additional is (if any) */ 373228143c3dSStefano Zampini if (nis) { 373328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 373428143c3dSStefano Zampini 373528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3736854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 373728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 373828143c3dSStefano Zampini psum = 0; 373928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 374028143c3dSStefano Zampini for (j=0;j<nis;j++) { 374128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 374228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 374328143c3dSStefano Zampini psum += plen; 374428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 374528143c3dSStefano Zampini } 374628143c3dSStefano Zampini } 3747854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3748854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 374928143c3dSStefano Zampini for (i=1;i<nis;i++) { 375028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 375128143c3dSStefano Zampini } 375228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 375328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 375428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 375528143c3dSStefano Zampini for (j=0;j<nis;j++) { 375628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 375728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 375828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 375928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 376028143c3dSStefano Zampini } 376128143c3dSStefano Zampini } 376228143c3dSStefano Zampini for (i=0;i<nis;i++) { 376328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 376428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 376528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 376628143c3dSStefano Zampini } 376728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 376828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 376928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 377028143c3dSStefano Zampini } 3771e7931f94SStefano Zampini /* free workspace */ 377228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3773e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3774e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3775e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3776e7931f94SStefano Zampini if (isdense) { 3777e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3778e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3779e7931f94SStefano Zampini } else { 3780e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3781e7931f94SStefano Zampini } 378228143c3dSStefano Zampini if (nis) { 378328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 378428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 378528143c3dSStefano Zampini } 3786e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3787e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 378828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3789e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3790e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 379128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3792e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3793e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3794e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3795e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3796e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 379728143c3dSStefano Zampini if (nis) { 379828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 379928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 380028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 380128143c3dSStefano Zampini } 380228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 380328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 380428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 380528143c3dSStefano Zampini for (i=0;i<nis;i++) { 380628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 380728143c3dSStefano Zampini } 380828143c3dSStefano Zampini } 3809e7931f94SStefano Zampini PetscFunctionReturn(0); 3810e7931f94SStefano Zampini } 3811a57a6d2fSStefano Zampini 381212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3813af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 381412edc857SStefano Zampini 3815c8587f34SStefano Zampini #undef __FUNCT__ 3816c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3817c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3818c8587f34SStefano Zampini { 3819c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3820c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 382120a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 38229881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 382320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 38246e683305SStefano Zampini IS coarse_is,*isarray; 38256e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 38266e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3827f9eb5b7dSStefano Zampini PC pc_temp; 3828c8587f34SStefano Zampini PCType coarse_pc_type; 3829c8587f34SStefano Zampini KSPType coarse_ksp_type; 3830f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 38314f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 38326e683305SStefano Zampini Mat t_coarse_mat_is; 38336e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 38346e683305SStefano Zampini PetscMPIInt all_procs; 383574e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 383668457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 383722bc73bbSStefano Zampini PetscScalar *array; 38389881197aSStefano Zampini PetscErrorCode ierr; 3839fdc09c96SStefano Zampini 3840c8587f34SStefano Zampini PetscFunctionBegin; 3841c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 384268457ee5SStefano 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 */ 3843fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 38445a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3845fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3846f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3847f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3848f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3849fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3850fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3851727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3852fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3853fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3854fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3855f4ddd8eeSStefano Zampini } 3856fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3857fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3858f4ddd8eeSStefano Zampini } 385970cf5478SStefano Zampini /* reset any subassembling information */ 386070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 38616e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 38626e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3863fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3864f4ddd8eeSStefano Zampini } 3865c8587f34SStefano Zampini 38666e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 38672b510759SStefano Zampini im_active = !!(pcis->n); 38682b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 38696e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 38706e683305SStefano Zampini void_procs = all_procs-active_procs; 38716e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 387274e2c79eSStefano Zampini redist = PETSC_FALSE; 387322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 38746e683305SStefano Zampini csin_ml = PETSC_TRUE; 38756e683305SStefano Zampini ncoarse_ml = void_procs; 3876779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3877779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 387818a45a71SStefano Zampini csin_ds = PETSC_TRUE; 387918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 388018a45a71SStefano Zampini redist = PETSC_TRUE; 388118a45a71SStefano Zampini } else { 38826e683305SStefano Zampini csin_ds = PETSC_TRUE; 3883779c1cceSStefano Zampini ncoarse_ds = active_procs; 3884779c1cceSStefano Zampini redist = PETSC_TRUE; 388518a45a71SStefano Zampini } 38866e683305SStefano Zampini } else { 38876e683305SStefano Zampini csin_ml = PETSC_FALSE; 38886e683305SStefano Zampini ncoarse_ml = all_procs; 38896e683305SStefano Zampini if (void_procs) { 38906e683305SStefano Zampini csin_ds = PETSC_TRUE; 38916e683305SStefano Zampini ncoarse_ds = void_procs; 38926e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 38936e683305SStefano Zampini } else { 3894779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 389574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 389674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 389774e2c79eSStefano Zampini redist = PETSC_TRUE; 389874e2c79eSStefano Zampini } else { 38996e683305SStefano Zampini csin_ds = PETSC_FALSE; 39006e683305SStefano Zampini ncoarse_ds = all_procs; 39016e683305SStefano Zampini } 39026e683305SStefano Zampini } 390374e2c79eSStefano Zampini } 39046e683305SStefano Zampini 39056e683305SStefano Zampini /* 39066e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 39076e683305SStefano Zampini - we have not exceeded the number of levels requested 39086e683305SStefano Zampini - we can actually subassemble the active processes 39096e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 39106e683305SStefano Zampini */ 39116e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 39126e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 39136e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 39146e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 39156e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3916f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 39172b510759SStefano Zampini } else { 3918f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3919c8587f34SStefano Zampini } 3920c8587f34SStefano Zampini } 39216e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 39226e683305SStefano Zampini if (multilevel_allowed) { 39236e683305SStefano Zampini ncoarse = ncoarse_ml; 39246e683305SStefano Zampini csin = csin_ml; 392558da7f69SStefano Zampini redist = PETSC_FALSE; 39266e683305SStefano Zampini } else { 39276e683305SStefano Zampini ncoarse = ncoarse_ds; 39286e683305SStefano Zampini csin = csin_ds; 39296e683305SStefano Zampini } 3930e7931f94SStefano Zampini 3931abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3932abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3933abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3934abbbba34SStefano Zampini 3935abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 393622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 393722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 393822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 393922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3940b9b85e73SStefano Zampini #if 0 3941b9b85e73SStefano Zampini { 3942b9b85e73SStefano Zampini PetscViewer viewer; 3943b9b85e73SStefano Zampini char filename[256]; 3944b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 3945b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3946b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3947b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 3948b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3949b9b85e73SStefano Zampini } 3950b9b85e73SStefano Zampini #endif 39516e683305SStefano 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); 39526e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 39536e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 39546e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3955abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 3956abbbba34SStefano Zampini 39576e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 39586e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 39596e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 39606e683305SStefano Zampini const PetscInt *idxs; 39616e683305SStefano Zampini ISLocalToGlobalMapping tmap; 39626e683305SStefano Zampini 39636e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 39640be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 39656e683305SStefano Zampini /* allocate space for temporary storage */ 3966854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 3967854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 39686e683305SStefano Zampini /* allocate for IS array */ 39696e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 39706e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 39716e683305SStefano Zampini nis = nisdofs + nisneu; 3972854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 39736e683305SStefano Zampini /* dofs splitting */ 39746e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 39756e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 39766e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 39776e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39786e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39796e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39806e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39816e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 39826e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 39836e683305SStefano Zampini } 39846e683305SStefano Zampini /* neumann boundaries */ 39856e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 39866e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 39876e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 39886e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39896e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39906e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39916e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39926e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 39936e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 39946e683305SStefano Zampini } 39956e683305SStefano Zampini /* free memory */ 39966e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 39976e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 39986e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 39996e683305SStefano Zampini } else { 40006e683305SStefano Zampini nis = 0; 40016e683305SStefano Zampini nisdofs = 0; 40026e683305SStefano Zampini nisneu = 0; 40036e683305SStefano Zampini isarray = NULL; 40046e683305SStefano Zampini } 40056e683305SStefano Zampini /* destroy no longer needed map */ 40066e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 40076e683305SStefano Zampini 40086e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 40096e683305SStefano Zampini coarse_mat_is = NULL; 40106e683305SStefano Zampini if (csin) { 40116e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 401274e2c79eSStefano Zampini if (redist) { 401374e2c79eSStefano Zampini PetscMPIInt rank; 4014779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 401574e2c79eSStefano Zampini 401674e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 401758da7f69SStefano Zampini spc = active_procs/ncoarse; 401858da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4019779c1cceSStefano Zampini if (im_active) { 4020779c1cceSStefano Zampini destsize = 1; 402174e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 402274e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 402374e2c79eSStefano Zampini } else { 402474e2c79eSStefano Zampini dest[0] = rank/(spc+1); 402574e2c79eSStefano Zampini } 402674e2c79eSStefano Zampini } else { 4027779c1cceSStefano Zampini destsize = 0; 40286e683305SStefano Zampini } 4029779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4030779c1cceSStefano Zampini } else if (csin_type_simple) { 40316e683305SStefano Zampini PetscMPIInt rank; 40326e683305SStefano Zampini PetscInt issize,isidx; 4033779c1cceSStefano Zampini 40346e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 40356e683305SStefano Zampini if (im_active) { 40366e683305SStefano Zampini issize = 1; 40376e683305SStefano Zampini isidx = (PetscInt)rank; 40386e683305SStefano Zampini } else { 40396e683305SStefano Zampini issize = 0; 40406e683305SStefano Zampini isidx = -1; 40416e683305SStefano Zampini } 40426e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4043779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 40446e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40456e683305SStefano Zampini } 4046779c1cceSStefano Zampini 4047779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4048779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4049779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4050779c1cceSStefano Zampini PetscInt *coarse_candidates; 4051779c1cceSStefano Zampini const PetscInt* tisindices; 4052779c1cceSStefano Zampini 4053779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4054779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4055779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4056779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4057779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4058779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4059779c1cceSStefano Zampini } 4060779c1cceSStefano Zampini } 4061779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4062779c1cceSStefano Zampini 4063779c1cceSStefano Zampini 40646e683305SStefano Zampini if (pcbddc->dbg_flag) { 40656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 40676e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4069779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 40706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 40716e683305SStefano Zampini } 40726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 40736e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40746e683305SStefano Zampini } 40756e683305SStefano Zampini /* shift the pattern on coarse candidates */ 40766e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 40776e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4078854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 40796e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 40806e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 40816e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 40826e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 40836e683305SStefano Zampini } 40846e683305SStefano Zampini if (pcbddc->dbg_flag) { 40856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 40876e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40886e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40896e683305SStefano Zampini } 4090779c1cceSStefano Zampini } 40916e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 40926e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 40936e683305SStefano Zampini } else { 40946e683305SStefano Zampini if (pcbddc->dbg_flag) { 40956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 40976e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40986e683305SStefano Zampini } 40996e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 41006e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 41016e683305SStefano Zampini } 41026e683305SStefano Zampini 41036e683305SStefano Zampini /* create local to global scatters for coarse problem */ 410468457ee5SStefano Zampini if (compute_vecs) { 41056e683305SStefano Zampini PetscInt lrows; 41066e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41076e683305SStefano Zampini if (coarse_mat_is) { 41086e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 41096e683305SStefano Zampini } else { 41106e683305SStefano Zampini lrows = 0; 41116e683305SStefano Zampini } 41126e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 41156e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41166e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41176e683305SStefano Zampini } 41186e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 41196e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4120c8587f34SStefano Zampini 4121f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4122f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4123f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4124f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4125f9eb5b7dSStefano Zampini } else { 4126f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4127f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4128c8587f34SStefano Zampini } 4129c8587f34SStefano Zampini 41306e683305SStefano Zampini /* print some info if requested */ 41316e683305SStefano Zampini if (pcbddc->dbg_flag) { 41326e683305SStefano Zampini if (!multilevel_allowed) { 41336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41346e683305SStefano Zampini if (multilevel_requested) { 41356e683305SStefano 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); 41366e683305SStefano Zampini } else if (pcbddc->max_levels) { 41376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 41386e683305SStefano Zampini } 41396e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41406e683305SStefano Zampini } 41416e683305SStefano Zampini } 41426e683305SStefano Zampini 4143f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 41446e683305SStefano Zampini if (coarse_mat_is) { 41456e683305SStefano Zampini MatReuse coarse_mat_reuse; 41466a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 41476e683305SStefano Zampini if (pcbddc->dbg_flag) { 41486e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 41496e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 41506e683305SStefano Zampini } 4151f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4152312be037SStefano Zampini char prefix[256],str_level[16]; 4153e604994aSStefano Zampini size_t len; 41546e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4155c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4156f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 41575f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4158c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 41596e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4160c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4161c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 41623e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4163e604994aSStefano Zampini /* prefix */ 4164e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4165e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4166e604994aSStefano Zampini if (!pcbddc->current_level) { 4167e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4168e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4169c8587f34SStefano Zampini } else { 4170e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4171312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4172312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 417334d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4174312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4175e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4176e604994aSStefano Zampini } 4177e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 41783e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41793e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 41803e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 41813e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4182f9eb5b7dSStefano Zampini /* allow user customization */ 4183f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 41843e3c6dadSStefano Zampini } 41853e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41863e3c6dadSStefano Zampini if (nisdofs) { 41873e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 41883e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 41893e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 41903e3c6dadSStefano Zampini } 41913e3c6dadSStefano Zampini } 41923e3c6dadSStefano Zampini if (nisneu) { 41933e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 41943e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4195312be037SStefano Zampini } 4196f9eb5b7dSStefano Zampini 4197f9eb5b7dSStefano Zampini /* get some info after set from options */ 4198f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4199f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4200f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 42014f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 42026e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4203f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4204f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4205f9eb5b7dSStefano Zampini } 42064f3a063dSStefano Zampini if (isredundant) { 42074f3a063dSStefano Zampini KSP inner_ksp; 42084f3a063dSStefano Zampini PC inner_pc; 42094f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 42104f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 42114f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 42124f3a063dSStefano Zampini } 4213f9eb5b7dSStefano Zampini 4214f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4215fa7f1dd8SStefano Zampini if (coarse_reuse) { 421681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4217fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 42186e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4219fa7f1dd8SStefano Zampini } else { 42206e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4221fa7f1dd8SStefano Zampini } 4222c8587f34SStefano Zampini if (isbddc || isnn) { 422322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 422470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 422528143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 422622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 42276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 42296e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 42306e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 423122b6e8a2SStefano Zampini } 423270cf5478SStefano Zampini } 42336e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 423470cf5478SStefano Zampini } else { 423522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 423622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 423722bc73bbSStefano Zampini } 423822bc73bbSStefano Zampini } else { 42392e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4240c8587f34SStefano Zampini } 4241c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4242c8587f34SStefano Zampini 4243c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4244b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 42455a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4246c8587f34SStefano Zampini 42476e683305SStefano Zampini /* set operators */ 42485f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 42496e683305SStefano Zampini if (pcbddc->dbg_flag) { 42506e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42516e683305SStefano Zampini } 42526e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 42536e683305SStefano Zampini coarse_mat = 0; 42546e683305SStefano Zampini } 42556e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4256b9b85e73SStefano Zampini #if 0 4257b9b85e73SStefano Zampini { 4258b9b85e73SStefano Zampini PetscViewer viewer; 4259b9b85e73SStefano Zampini char filename[256]; 4260b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4261b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4262b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4263b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4264b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4265b9b85e73SStefano Zampini } 4266b9b85e73SStefano Zampini #endif 4267c8587f34SStefano Zampini 4268c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4269c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4270c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 427198a51de6SStefano Zampini } 427298a51de6SStefano Zampini 427398a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 427498a51de6SStefano Zampini Vec crhs,csol; 427598a51de6SStefano Zampini PetscBool ispreonly; 427698a51de6SStefano Zampini if (CoarseNullSpace) { 4277c8587f34SStefano Zampini if (isbddc) { 4278c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4279c8587f34SStefano Zampini } else { 4280c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4281c8587f34SStefano Zampini } 4282c8587f34SStefano Zampini } 4283f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4284f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4285f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4286f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 42876e683305SStefano Zampini /* hack */ 4288f347579bSStefano Zampini if (!csol) { 42892a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4290f9eb5b7dSStefano Zampini } 4291f347579bSStefano Zampini if (!crhs) { 42922a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4293f347579bSStefano Zampini } 4294cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4295cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 42966e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4297c8587f34SStefano Zampini KSP check_ksp; 42982b510759SStefano Zampini KSPType check_ksp_type; 4299c8587f34SStefano Zampini PC check_pc; 43006e683305SStefano Zampini Vec check_vec,coarse_vec; 43016a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 43022b510759SStefano Zampini PetscInt its; 43036e683305SStefano Zampini PetscBool compute_eigs; 43046e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 43056e683305SStefano Zampini PetscInt neigs; 43068e185a42SStefano Zampini const char *prefix; 4307c8587f34SStefano Zampini 43082b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 43096e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 431023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4311f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 43122b510759SStefano Zampini if (ispreonly) { 43132b510759SStefano Zampini check_ksp_type = KSPPREONLY; 43146e683305SStefano Zampini compute_eigs = PETSC_FALSE; 43152b510759SStefano Zampini } else { 4316cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 43176e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4318c8587f34SStefano Zampini } 4319c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 43206e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 43216e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 43226e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4323a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4324a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4325a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4326a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4327c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4328c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4329c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4330c8587f34SStefano Zampini /* create random vec */ 43316e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 43326e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4333c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4334c8587f34SStefano Zampini if (CoarseNullSpace) { 4335c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4336c8587f34SStefano Zampini } 43376e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4338c8587f34SStefano Zampini /* solve coarse problem */ 43396e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4340c8587f34SStefano Zampini if (CoarseNullSpace) { 43416e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4342c8587f34SStefano Zampini } 4343cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 43446e683305SStefano Zampini if (compute_eigs) { 4345854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4346854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 43476e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 43486e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 43496e683305SStefano Zampini lambda_min = eigs_r[0]; 43506e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 43516e683305SStefano Zampini if (lambda_max>lambda_min) { 4352cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4353cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4354cbcc2c2aSStefano Zampini } 4355c8587f34SStefano Zampini } 4356c8587f34SStefano Zampini } 4357cbcc2c2aSStefano Zampini 4358c8587f34SStefano Zampini /* check coarse problem residual error */ 43596e683305SStefano Zampini if (pcbddc->dbg_flag) { 43606e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 43616e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43626e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4363c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 43646e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 43656e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4366c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4367779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 43686e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 43696e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 43706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 43716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 43726e683305SStefano Zampini if (compute_eigs) { 43736e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4374deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4375c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 43766e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 43776e683305SStefano 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); 43786e683305SStefano Zampini for (i=0;i<neigs;i++) { 43796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4380c8587f34SStefano Zampini } 43816e683305SStefano Zampini } 43826e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 43836e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43846e683305SStefano Zampini } 4385c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 43866e683305SStefano Zampini if (compute_eigs) { 43876e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 43886e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4389c8587f34SStefano Zampini } 43906e683305SStefano Zampini } 43916e683305SStefano Zampini } 4392cbcc2c2aSStefano Zampini /* print additional info */ 4393cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 43946e683305SStefano Zampini /* waits until all processes reaches this point */ 43956e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4396cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4397cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4398cbcc2c2aSStefano Zampini } 4399cbcc2c2aSStefano Zampini 44002b510759SStefano Zampini /* free memory */ 4401c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4402fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4403c8587f34SStefano Zampini PetscFunctionReturn(0); 4404c8587f34SStefano Zampini } 4405674ae819SStefano Zampini 4406f34684f1SStefano Zampini #undef __FUNCT__ 4407f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4408f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4409f34684f1SStefano Zampini { 4410f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4411f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4412f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 44130e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 441473be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 44150e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4416f34684f1SStefano Zampini PetscErrorCode ierr; 4417f34684f1SStefano Zampini 4418f34684f1SStefano Zampini PetscFunctionBegin; 4419f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 44200e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 44210e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4422727cdba6SStefano Zampini } 44230e6343abSStefano 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); 44240e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 44250e6343abSStefano Zampini local_size = 0; 44260e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 44270e6343abSStefano Zampini PetscInt j; 44280e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 44290e6343abSStefano Zampini } 44300e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 44310e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 44320e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 44330e6343abSStefano Zampini } 4434f34684f1SStefano Zampini 4435f34684f1SStefano Zampini /* check numbering */ 4436f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4437f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4438b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4439f34684f1SStefano Zampini 4440f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4441f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4442f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 44430fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4444f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4445f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4446727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4447f34684f1SStefano Zampini } 4448f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4449f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4450f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4451f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4452f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4453f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4454f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4455f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4456f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4457f34684f1SStefano Zampini if (array[i] == 1.0) { 4458b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4459f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4460f34684f1SStefano Zampini } 4461f34684f1SStefano Zampini } 4462b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4463f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4464f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4465f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4466f34684f1SStefano Zampini } 4467f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4468f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4469f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4470f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4471f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4472f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4473b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4474ca8b9ea9SStefano Zampini PetscInt *gidxs; 4475ca8b9ea9SStefano Zampini 4476ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4477ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4478f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4479f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4480f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4481f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 44824bc2dc4bSStefano 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); 4483f34684f1SStefano Zampini } 4484f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4485ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4486f34684f1SStefano Zampini } 4487f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4488302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4489f34684f1SStefano Zampini } 44908bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4491f34684f1SStefano Zampini /* get back data */ 4492f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4493f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4494674ae819SStefano Zampini PetscFunctionReturn(0); 4495674ae819SStefano Zampini } 4496674ae819SStefano Zampini 4497e456f2a8SStefano Zampini #undef __FUNCT__ 4498e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4499a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4500e456f2a8SStefano Zampini { 4501e456f2a8SStefano Zampini IS localis_t; 4502a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4503e456f2a8SStefano Zampini PetscScalar *vals; 4504e456f2a8SStefano Zampini PetscErrorCode ierr; 4505e456f2a8SStefano Zampini 4506e456f2a8SStefano Zampini PetscFunctionBegin; 4507a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4508e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4509854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4510e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4511e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4512a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4513a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 45141035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4515a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 45161035eff8SStefano Zampini } 4517a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4518e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4519e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4520a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4521a7dc3881SStefano Zampini /* now compute set in local ordering */ 4522a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4523a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4524a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4525a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4526a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4527ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4528e456f2a8SStefano Zampini lsize++; 4529e456f2a8SStefano Zampini } 4530e456f2a8SStefano Zampini } 4531854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4532a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4533ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4534e456f2a8SStefano Zampini idxs[lsize++] = i; 4535e456f2a8SStefano Zampini } 4536e456f2a8SStefano Zampini } 4537a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4538a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4539e456f2a8SStefano Zampini *localis = localis_t; 4540e456f2a8SStefano Zampini PetscFunctionReturn(0); 4541e456f2a8SStefano Zampini } 4542906d46d4SStefano Zampini 4543906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4544906d46d4SStefano Zampini #undef __FUNCT__ 4545906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4546906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4547906d46d4SStefano Zampini { 4548906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4549906d46d4SStefano Zampini PetscErrorCode ierr; 4550906d46d4SStefano Zampini 4551906d46d4SStefano Zampini PetscFunctionBegin; 4552906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4553906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4554906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4555906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4556906d46d4SStefano Zampini PetscFunctionReturn(0); 4557906d46d4SStefano Zampini } 4558906d46d4SStefano Zampini 4559906d46d4SStefano Zampini #undef __FUNCT__ 4560906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4561906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4562906d46d4SStefano Zampini { 4563906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4564906d46d4SStefano Zampini PetscErrorCode ierr; 4565906d46d4SStefano Zampini 4566906d46d4SStefano Zampini PetscFunctionBegin; 4567906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4568906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4569906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4570906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4571906d46d4SStefano Zampini PetscFunctionReturn(0); 4572906d46d4SStefano Zampini } 4573b96c3477SStefano Zampini 4574b96c3477SStefano Zampini #undef __FUNCT__ 4575b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 457608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4577b96c3477SStefano Zampini { 4578a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4579b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4580b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4581a64f4aa4SStefano Zampini Mat S_j; 4582b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4583b96c3477SStefano Zampini PetscBool free_used_adj; 4584b96c3477SStefano Zampini PetscErrorCode ierr; 4585b96c3477SStefano Zampini 4586b96c3477SStefano Zampini PetscFunctionBegin; 4587b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4588b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 458908122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4590b96c3477SStefano Zampini used_xadj = NULL; 4591b96c3477SStefano Zampini used_adjncy = NULL; 4592b96c3477SStefano Zampini } else { 459308122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 459408122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 459508122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 459608122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4597b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4598b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4599b96c3477SStefano Zampini } else { 46002fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4601b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4602b96c3477SStefano Zampini PetscInt nvtxs; 4603b96c3477SStefano Zampini 46042fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 46052fffb893SStefano Zampini if (flg_row) { 4606b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4607b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4608b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4609b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 46102fffb893SStefano Zampini } else { 46112fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 46122fffb893SStefano Zampini used_xadj = NULL; 46132fffb893SStefano Zampini used_adjncy = NULL; 46142fffb893SStefano Zampini } 46152fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4616b96c3477SStefano Zampini } 4617b96c3477SStefano Zampini } 4618d5574798SStefano Zampini 4619d5574798SStefano Zampini /* setup sub_schurs data */ 4620a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4621a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4622a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4623a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4624a64f4aa4SStefano 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); 4625a64f4aa4SStefano Zampini } else { 4626a64f4aa4SStefano 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); 4627a64f4aa4SStefano Zampini } 4628a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4629b96c3477SStefano Zampini 4630b96c3477SStefano Zampini /* free adjacency */ 4631b96c3477SStefano Zampini if (free_used_adj) { 4632b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4633b96c3477SStefano Zampini } 4634b96c3477SStefano Zampini PetscFunctionReturn(0); 4635b96c3477SStefano Zampini } 4636b96c3477SStefano Zampini 4637b96c3477SStefano Zampini #undef __FUNCT__ 4638b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 463908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4640b96c3477SStefano Zampini { 4641b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4642b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4643b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4644b96c3477SStefano Zampini PCBDDCGraph graph; 4645b96c3477SStefano Zampini PetscErrorCode ierr; 4646b96c3477SStefano Zampini 4647b96c3477SStefano Zampini PetscFunctionBegin; 4648b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 464908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4650b96c3477SStefano Zampini IS verticesIS; 4651b96c3477SStefano Zampini 4652b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4653b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 46547fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4655b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4656b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4657b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4658b96c3477SStefano Zampini /* 4659b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4660b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4661b96c3477SStefano Zampini } 4662b96c3477SStefano Zampini */ 4663b96c3477SStefano Zampini } else { 4664b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4665b96c3477SStefano Zampini } 4666b96c3477SStefano Zampini 4667b96c3477SStefano Zampini /* sub_schurs init */ 4668a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4669a64f4aa4SStefano Zampini 4670b96c3477SStefano Zampini /* free graph struct */ 467108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4672b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4673b96c3477SStefano Zampini } 4674b96c3477SStefano Zampini PetscFunctionReturn(0); 4675b96c3477SStefano Zampini } 4676