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); 75*9162d606SStefano Zampini cum += subset_size; 76*9162d606SStefano 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 122*9162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 123*9162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 124*9162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 126*9162d606SStefano 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; 131*9162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 132*9162d606SStefano 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; 141*9162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 142*9162d606SStefano 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; 276*9162d606SStefano Zampini if (B_neigs) { 277*9162d606SStefano 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++) { 284*9162d606SStefano 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 0 289*9162d606SStefano 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); 293*9162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 294*9162d606SStefano 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 } 300*9162d606SStefano 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 303*9162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 304*9162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 305*9162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 306*9162d606SStefano 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); 43306656605SStefano Zampini if (pcbddc->local_auxmat2) { 43406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 43506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43606656605SStefano Zampini } 437674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43806656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4418ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 442674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 443674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 444674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 445f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 446f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 448f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 449727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 450f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 45170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4526e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 453674ae819SStefano Zampini PetscFunctionReturn(0); 454674ae819SStefano Zampini } 455674ae819SStefano Zampini 456674ae819SStefano Zampini #undef __FUNCT__ 457f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 458f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4596bfb1811SStefano Zampini { 4606bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4616bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4626bfb1811SStefano Zampini VecType impVecType; 463e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4646bfb1811SStefano Zampini PetscErrorCode ierr; 4656bfb1811SStefano Zampini 4666bfb1811SStefano Zampini PetscFunctionBegin; 467f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 468f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 469f4ddd8eeSStefano Zampini } 470e7b262bdSStefano Zampini /* get sizes */ 471b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 472b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4736bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 474e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 475e7b262bdSStefano Zampini /* R nodes */ 476e7b262bdSStefano Zampini old_size = -1; 477e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 478e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 479e7b262bdSStefano Zampini } 480e7b262bdSStefano Zampini if (n_R != old_size) { 481e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 482e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4846bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4856bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4866bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 487e7b262bdSStefano Zampini } 488e7b262bdSStefano Zampini /* local primal dofs */ 489e7b262bdSStefano Zampini old_size = -1; 490e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 491e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 492e7b262bdSStefano Zampini } 493e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 494e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 496e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4976bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 498e7b262bdSStefano Zampini } 499e7b262bdSStefano Zampini /* local explicit constraints */ 500e7b262bdSStefano Zampini old_size = -1; 501e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 502e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 503e7b262bdSStefano Zampini } 504e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 505e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50983b7ccabSStefano Zampini } 5106bfb1811SStefano Zampini PetscFunctionReturn(0); 5116bfb1811SStefano Zampini } 5126bfb1811SStefano Zampini 5136bfb1811SStefano Zampini #undef __FUNCT__ 51447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51688ebb749SStefano Zampini { 51725084f0cSStefano Zampini PetscErrorCode ierr; 51825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 52088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 52125084f0cSStefano Zampini /* submatrices of local problem */ 52288ebb749SStefano Zampini Mat A_RV,A_VR,A_VV; 52306656605SStefano Zampini /* submatrices of local coarse problem */ 52406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52525084f0cSStefano Zampini /* working matrices */ 52606656605SStefano Zampini Mat C_CR; 52725084f0cSStefano Zampini /* additional working stuff */ 52806656605SStefano Zampini PC pc_R; 52906656605SStefano Zampini Mat F; 53006656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 53106656605SStefano Zampini 53225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53306656605SStefano Zampini PetscScalar *work; 53406656605SStefano Zampini PetscInt *idx_V_B; 53506656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 537b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53845a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53988ebb749SStefano Zampini MatType impMatType; 54025084f0cSStefano Zampini /* some shortcuts to scalars */ 54106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54288ebb749SStefano Zampini 54388ebb749SStefano Zampini PetscFunctionBegin; 544b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54588ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54688ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 547b371cd4fSStefano Zampini n_B = pcis->n_B; 548b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54988ebb749SStefano Zampini n_R = pcis->n - n_vertices; 55088ebb749SStefano Zampini 55188ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55288ebb749SStefano Zampini impMatType = MATSEQDENSE; 55388ebb749SStefano Zampini 55488ebb749SStefano Zampini /* vertices in boundary numbering */ 555785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5565e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->primal_indices_local_idxs,&i,idx_V_B);CHKERRQ(ierr); 55788ebb749SStefano Zampini if (i != n_vertices) { 55822d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 55988ebb749SStefano Zampini } 56088ebb749SStefano Zampini 56106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56206656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 57006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 57106656605SStefano Zampini 57206656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57306656605SStefano Zampini /* allocate workspace */ 57406656605SStefano Zampini n = 0; 57506656605SStefano Zampini if (n_constraints) { 57606656605SStefano Zampini n += n_R*n_constraints; 57706656605SStefano Zampini } 57806656605SStefano Zampini if (n_vertices) { 57906656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 58006656605SStefano Zampini } 58106656605SStefano Zampini if (!pcbddc->issym) { 58206656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58306656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58406656605SStefano Zampini } 58506656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58606656605SStefano Zampini 58706656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58806656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 59006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59206656605SStefano Zampini if (isLU || isILU || isCHOL) { 59306656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 59406656605SStefano Zampini } else { 59506656605SStefano Zampini F = NULL; 59606656605SStefano Zampini } 59706656605SStefano Zampini 59888ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 59988ebb749SStefano Zampini if (n_constraints) { 60006656605SStefano Zampini Mat M1,M2,M3; 60106656605SStefano Zampini IS is_aux; 602f4ddd8eeSStefano Zampini /* see if we can save some allocations */ 603f4ddd8eeSStefano Zampini if (pcbddc->local_auxmat2) { 604f4ddd8eeSStefano Zampini PetscInt on_R,on_constraints; 605f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->local_auxmat2,&on_R,&on_constraints);CHKERRQ(ierr); 606f4ddd8eeSStefano Zampini if (on_R != n_R || on_constraints != n_constraints) { 60706656605SStefano Zampini PetscScalar *marray; 60806656605SStefano Zampini 60906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&marray);CHKERRQ(ierr); 61006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 611f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 612f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 613f4ddd8eeSStefano Zampini } 614f4ddd8eeSStefano Zampini } 61545a1bb75SStefano Zampini /* auxiliary matrices */ 616f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat2) { 61706656605SStefano Zampini PetscScalar *marray; 61806656605SStefano Zampini 61906656605SStefano Zampini ierr = PetscMalloc1(2*n_R*n_constraints,&marray);CHKERRQ(ierr); 62006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,marray,&pcbddc->local_auxmat2);CHKERRQ(ierr); 62106656605SStefano Zampini marray += n_R*n_constraints; 62206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_R,marray,&pcbddc->local_auxmat1);CHKERRQ(ierr); 623f4ddd8eeSStefano Zampini } 62488ebb749SStefano Zampini 62525084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62625084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6278ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 62825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 62988ebb749SStefano Zampini 63088ebb749SStefano Zampini /* Assemble local_auxmat2 = - A_{RR}^{-1} C^T_{CR} needed by BDDC application */ 63106656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 63288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63306656605SStefano Zampini const PetscScalar *row_cmat_values; 63406656605SStefano Zampini const PetscInt *row_cmat_indices; 63506656605SStefano Zampini PetscInt size_of_constraint,j; 63688ebb749SStefano Zampini 63706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63906656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64006656605SStefano Zampini } 64106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64206656605SStefano Zampini } 64306656605SStefano Zampini if (F) { 64406656605SStefano Zampini Mat B; 64506656605SStefano Zampini 64606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 64706656605SStefano Zampini ierr = MatMatSolve(F,B,pcbddc->local_auxmat2);CHKERRQ(ierr); 64806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64906656605SStefano Zampini } else { 65006656605SStefano Zampini PetscScalar *xarray; 65106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr); 65206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 65306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,xarray+i*n_R);CHKERRQ(ierr); 65506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 65606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65806656605SStefano Zampini } 65906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr); 66006656605SStefano Zampini } 66106656605SStefano Zampini 66206656605SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 66306656605SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 66425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 66606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 66725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 66806656605SStefano Zampini ierr = VecSet(pcbddc->vec1_C,m_one);CHKERRQ(ierr); 66906656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 67025084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 67125084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 67225084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 67306656605SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CR} needed by BDDC application in KSP and in preproc */ 674f4ddd8eeSStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_REUSE_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 67506656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 67606656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 67788ebb749SStefano Zampini } 67888ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 67988ebb749SStefano Zampini if (n_vertices) { 68006656605SStefano Zampini Mat newmat; 68106656605SStefano Zampini IS is_aux; 6823a50541eSStefano Zampini PetscInt ibs,mbs; 6833a50541eSStefano Zampini PetscBool issbaij; 6843a50541eSStefano Zampini 6853a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 6869577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 6879577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 6883a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 6899577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 6909577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 6919577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 6929577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 6939577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 6949577ea80SStefano Zampini } else { 6953a50541eSStefano Zampini /* this is safe */ 6969577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 6973a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 6983a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 6993a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7003a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 70106656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7023a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 70303af80d8SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7043a50541eSStefano Zampini } else { 7059577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7069577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7079577ea80SStefano Zampini } 7083a50541eSStefano Zampini } 70925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 71088ebb749SStefano Zampini } 71188ebb749SStefano Zampini 71288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 713f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 71406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 71506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 71606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 71706656605SStefano Zampini } 718f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 71906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 72006656605SStefano Zampini PetscScalar *marray; 72106656605SStefano Zampini 72206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 72306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 724f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 725f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 726f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 727f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 728f4ddd8eeSStefano Zampini } 729f4ddd8eeSStefano Zampini } 73006656605SStefano Zampini 731f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 73206656605SStefano Zampini PetscScalar *marray; 73388ebb749SStefano Zampini 73406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7358eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 73788ebb749SStefano Zampini } 738b9d89cd5SStefano Zampini if (!pcbddc->issym) { 73906656605SStefano Zampini n *= 2; 740f4ddd8eeSStefano Zampini } 74106656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 74206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 74306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 74406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 74606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74788ebb749SStefano Zampini } 74806656605SStefano Zampini if (!pcbddc->issym) { 74906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7501b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 75388ebb749SStefano Zampini } 75488ebb749SStefano Zampini } else { 755c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 756c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7571b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 758c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 759c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 760c0553b1fSStefano Zampini } 76188ebb749SStefano Zampini } 76206656605SStefano Zampini } 76306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 76406656605SStefano Zampini /* vertices */ 76506656605SStefano Zampini if (n_vertices) { 76616f15bc4SStefano Zampini 76716f15bc4SStefano Zampini if (n_R) { 76806656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 76906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 77016f15bc4SStefano Zampini PetscScalar *x,*y; 77106656605SStefano Zampini 77206656605SStefano Zampini ierr = PetscMemzero(work,2*n_R*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 77306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 77406656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 77506656605SStefano Zampini if (F) { 77606656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 77706656605SStefano Zampini } else { 77806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 77906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 78006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 78106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 78206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 78306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 78406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 78506656605SStefano Zampini } 78606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 78706656605SStefano Zampini } 78806656605SStefano Zampini ierr = MatScale(A_RRmA_RV,m_one);CHKERRQ(ierr); 78906656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 79006656605SStefano Zampini if (n_constraints) { 79106656605SStefano Zampini Mat B; 79206656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,A_RRmA_RV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 79306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 79406656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat2,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 79506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 79606656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 79706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 79806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 79906656605SStefano Zampini } 80006656605SStefano Zampini ierr = MatConvert(A_VR,impMatType,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 80106656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 80206656605SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 80306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 80606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 80706656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 80806656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 80906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 81006656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 81116f15bc4SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 81216f15bc4SStefano Zampini } else { 81316f15bc4SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 81416f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 81516f15bc4SStefano Zampini } 81606656605SStefano Zampini /* coarse basis functions */ 81706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 81816f15bc4SStefano Zampini PetscScalar *y; 81916f15bc4SStefano Zampini 82006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 82106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 82206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 82306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 82606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82806656605SStefano Zampini 82906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 83006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 83606656605SStefano Zampini } 83706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 83806656605SStefano Zampini } 83916f15bc4SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 84106656605SStefano Zampini } 84206656605SStefano Zampini 84306656605SStefano Zampini if (n_constraints) { 84406656605SStefano Zampini Mat B; 84506656605SStefano Zampini 84606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 84706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 84806656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat2,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 85006656605SStefano Zampini if (n_vertices) { 85106656605SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_VC);CHKERRQ(ierr); 85206656605SStefano Zampini } 85306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 85406656605SStefano Zampini /* coarse basis functions */ 85506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 85606656605SStefano Zampini PetscScalar *y; 85706656605SStefano Zampini 85806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 86106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86506656605SStefano Zampini 86606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 86906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87306656605SStefano Zampini } 87406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87506656605SStefano Zampini } 87606656605SStefano Zampini } 87706656605SStefano Zampini 87806656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 87906656605SStefano Zampini if (!pcbddc->issym) { 88006656605SStefano Zampini Mat B,X; 88106656605SStefano Zampini 88206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work,&B);CHKERRQ(ierr); 88306656605SStefano Zampini 88406656605SStefano Zampini if (n_constraints) { 88516f15bc4SStefano Zampini Mat S_CCT,B_C; 88606656605SStefano Zampini 88706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 89016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 89106656605SStefano Zampini if (n_vertices) { 89216f15bc4SStefano Zampini Mat B_V,S_VCT; 89306656605SStefano Zampini 89406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 89506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 89816f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 89906656605SStefano Zampini } 90006656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 90106656605SStefano Zampini } 90216f15bc4SStefano Zampini if (n_vertices && n_R) { 90306656605SStefano Zampini Mat A_VRT; 90406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 90506656605SStefano Zampini 90606656605SStefano Zampini if (!n_constraints) { /* if there are no constraints, reset work */ 90706656605SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 90806656605SStefano Zampini } 90906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+pcbddc->local_primal_size*n_R,&A_VRT);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 91106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 91206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,work+pcbddc->local_primal_size*n_R,&B_one,work,&B_one)); 91316f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 91406656605SStefano Zampini } 91506656605SStefano Zampini 91606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work+pcbddc->local_primal_size*n_R,&X);CHKERRQ(ierr); 91706656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 91806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 91906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 92006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 92106656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 92206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 92406656605SStefano Zampini } 92506656605SStefano Zampini } else { 92606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 92706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 92806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 92906656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 93006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 93106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 93206656605SStefano Zampini } 93306656605SStefano Zampini } 93406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 93506656605SStefano Zampini /* coarse basis functions */ 93606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 93706656605SStefano Zampini PetscScalar *y; 93806656605SStefano Zampini 93906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 94006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94406656605SStefano Zampini if (i<n_vertices) { 94506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 94606656605SStefano Zampini } 94706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 94806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 94906656605SStefano Zampini 95006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 95106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 95306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 95706656605SStefano Zampini } 95806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95906656605SStefano Zampini } 96006656605SStefano Zampini ierr = MatDestroy(&X);CHKERRQ(ierr); 96106656605SStefano Zampini } 96206656605SStefano Zampini 96388ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 96406656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 96506656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 96606656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 96706656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 96888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 96988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 97088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 97125084f0cSStefano Zampini if (pcbddc->dbg_flag) { 97288ebb749SStefano Zampini Mat coarse_sub_mat; 97325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 97488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 97588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 97688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 9778bec7fa6SStefano Zampini Mat C_B,CPHI; 9788bec7fa6SStefano Zampini IS is_dummy; 9798bec7fa6SStefano Zampini Vec mones; 98088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 98188ebb749SStefano Zampini PetscReal real_value; 98288ebb749SStefano Zampini 98388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 98488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 98588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 98688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 98788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 98888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 989c0553b1fSStefano Zampini if (unsymmetric_check) { 99088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 99188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 99288ebb749SStefano Zampini } 99388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 99488ebb749SStefano Zampini 99525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 99606656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 99725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 998c0553b1fSStefano Zampini if (unsymmetric_check) { 99988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 100188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 100288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 100488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 100588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 100788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 100888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 101088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 101188ebb749SStefano Zampini } else { 101288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 101388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 101488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 101588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 101688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 101788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 101888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 101988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 102088ebb749SStefano Zampini } 102188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 102288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 102388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 102488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 102581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10268bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10270fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 102806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10298bec7fa6SStefano Zampini 10308bec7fa6SStefano Zampini /* check constraints */ 10318bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10328bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10338bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10348bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10358bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10368bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10378bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10388bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10398bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10408bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10418bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10428bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10438bec7fa6SStefano Zampini 104425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1055c0553b1fSStefano Zampini if (unsymmetric_check) { 105688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 105888ebb749SStefano Zampini } 105988ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 106088ebb749SStefano Zampini } 106106656605SStefano Zampini 106288ebb749SStefano Zampini /* free memory */ 106306656605SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 106488ebb749SStefano Zampini if (n_vertices) { 106588ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 106688ebb749SStefano Zampini } 106788ebb749SStefano Zampini if (n_constraints) { 106888ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 106988ebb749SStefano Zampini } 10708629588bSStefano Zampini /* get back data */ 10718629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 107288ebb749SStefano Zampini PetscFunctionReturn(0); 107388ebb749SStefano Zampini } 107488ebb749SStefano Zampini 107588ebb749SStefano Zampini #undef __FUNCT__ 1076d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1077d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1078d65f70fdSStefano Zampini { 1079d65f70fdSStefano Zampini Mat *work_mat; 1080d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1081d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1082d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1083d65f70fdSStefano Zampini PetscErrorCode ierr; 1084d65f70fdSStefano Zampini 1085d65f70fdSStefano Zampini PetscFunctionBegin; 1086d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1087d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1088d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1089d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1090d65f70fdSStefano Zampini 1091d65f70fdSStefano Zampini if (!rsorted) { 1092d65f70fdSStefano Zampini const PetscInt *idxs; 1093d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1094d65f70fdSStefano Zampini 1095d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1096d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1097d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1098d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1099d65f70fdSStefano Zampini } 1100d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1101d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1102d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1103d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1104d65f70fdSStefano Zampini } 1105d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1106d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1107d65f70fdSStefano Zampini } else { 1108d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1109d65f70fdSStefano Zampini isrow_s = isrow; 1110d65f70fdSStefano Zampini } 1111d65f70fdSStefano Zampini 1112d65f70fdSStefano Zampini if (!csorted) { 1113d65f70fdSStefano Zampini if (isrow == iscol) { 1114d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1115d65f70fdSStefano Zampini iscol_s = isrow_s; 1116d65f70fdSStefano Zampini } else { 1117d65f70fdSStefano Zampini const PetscInt *idxs; 1118d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1119d65f70fdSStefano Zampini 1120d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1121d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1122d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1123d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1124d65f70fdSStefano Zampini } 1125d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1126d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1127d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1128d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1129d65f70fdSStefano Zampini } 1130d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini } 1133d65f70fdSStefano Zampini } else { 1134d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini iscol_s = iscol; 1136d65f70fdSStefano Zampini } 1137d65f70fdSStefano Zampini 1138d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini 1140d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1141d65f70fdSStefano Zampini Mat new_mat; 1142d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1143d65f70fdSStefano Zampini 1144d65f70fdSStefano Zampini if (!rsorted) { 1145d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1146d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1148d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1149d65f70fdSStefano Zampini } 1150d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini } else { 1153d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini } 1155d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1156d65f70fdSStefano Zampini 1157d65f70fdSStefano Zampini if (!csorted) { 1158d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1159d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1161d65f70fdSStefano Zampini } else { 1162d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1163d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1164d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1165d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1166d65f70fdSStefano Zampini } 1167d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1168d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1169d65f70fdSStefano Zampini } 1170d65f70fdSStefano Zampini } else { 1171d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1172d65f70fdSStefano Zampini } 1173d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1174d65f70fdSStefano Zampini 1175d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1177d65f70fdSStefano Zampini work_mat[0] = new_mat; 1178d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini 1182d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini *B = work_mat[0]; 1184d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1185d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1187d65f70fdSStefano Zampini PetscFunctionReturn(0); 1188d65f70fdSStefano Zampini } 1189d65f70fdSStefano Zampini 1190d65f70fdSStefano Zampini #undef __FUNCT__ 11915e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 11925e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1193aa0d41d4SStefano Zampini { 1194aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 11955e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1196d65f70fdSStefano Zampini Mat new_mat; 11975e8657edSStefano Zampini IS is_local,is_global; 1198d65f70fdSStefano Zampini PetscInt local_size; 1199d65f70fdSStefano Zampini PetscBool isseqaij; 1200aa0d41d4SStefano Zampini PetscErrorCode ierr; 1201aa0d41d4SStefano Zampini 1202aa0d41d4SStefano Zampini PetscFunctionBegin; 1203aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12045e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12055e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1206906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1207906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1208d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1209906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1210906d46d4SStefano Zampini 1211906d46d4SStefano Zampini /* check */ 1212906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1213906d46d4SStefano Zampini Vec x,x_change; 1214906d46d4SStefano Zampini PetscReal error; 1215906d46d4SStefano Zampini 12165e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1217906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12185e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12195e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12205e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12225e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12235e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1224906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1225906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1226906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1227906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1228906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1229906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1230906d46d4SStefano Zampini } 1231906d46d4SStefano Zampini 123222d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12339b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 123422d5777bSStefano Zampini if (isseqaij) { 1235d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1236aa0d41d4SStefano Zampini } else { 1237aa0d41d4SStefano Zampini Mat work_mat; 1238aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1239d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1240aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1241aa0d41d4SStefano Zampini } 1242d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1243e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1244e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1245e496cd5dSStefano Zampini #endif 124645a1bb75SStefano Zampini /* 124745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1248d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 124945a1bb75SStefano Zampini */ 1250d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1251aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1252aa0d41d4SStefano Zampini } 1253aa0d41d4SStefano Zampini 1254aa0d41d4SStefano Zampini #undef __FUNCT__ 1255a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12568ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1257a64d13efSStefano Zampini { 1258a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1259a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12608ce42a96SStefano Zampini IS is_aux1,is_aux2; 1261e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 12623a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12633a50541eSStefano Zampini PetscInt vbs,bs; 12644641a718SStefano Zampini PetscBT bitmask; 1265a64d13efSStefano Zampini PetscErrorCode ierr; 1266a64d13efSStefano Zampini 1267a64d13efSStefano Zampini PetscFunctionBegin; 1268b23d619eSStefano Zampini /* 1269b23d619eSStefano Zampini No need to setup local scatters if 1270b23d619eSStefano Zampini - primal space is unchanged 1271b23d619eSStefano Zampini AND 1272b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1273b23d619eSStefano Zampini AND 1274b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1275b23d619eSStefano Zampini */ 1276b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1277f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1278f4ddd8eeSStefano Zampini } 1279f4ddd8eeSStefano Zampini /* destroy old objects */ 1280f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1281f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1282f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1283a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1284b371cd4fSStefano Zampini n_B = pcis->n_B; 1285b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1286b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 12874641a718SStefano Zampini /* create auxiliary bitmask */ 12884641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 12894641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 1290e9189074SStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->primal_indices_local_idxs[i]);CHKERRQ(ierr); 12914641a718SStefano Zampini } 12923a50541eSStefano Zampini 1293a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 1294854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1295a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 12964641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 1297a64d13efSStefano Zampini idx_R_local[n_R] = i; 1298a64d13efSStefano Zampini n_R++; 1299a64d13efSStefano Zampini } 1300a64d13efSStefano Zampini } 13013a50541eSStefano Zampini 13023a50541eSStefano Zampini /* Block code */ 13033a50541eSStefano Zampini vbs = 1; 13043a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13053a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13063a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13073a50541eSStefano Zampini PetscInt *vary; 13083a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1309785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13103a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1311e9189074SStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->primal_indices_local_idxs[i]/bs]++; 1312964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13133a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13143a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13153a50541eSStefano Zampini break; 13163a50541eSStefano Zampini } 13173a50541eSStefano Zampini } 13183a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13193a50541eSStefano Zampini vbs = bs; 13203a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13213a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13223a50541eSStefano Zampini } 13233a50541eSStefano Zampini } 13243a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13253a50541eSStefano Zampini } 13263a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 13273a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1328a64d13efSStefano Zampini 1329a64d13efSStefano Zampini /* print some info if requested */ 1330a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1331a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1332a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13330fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1334a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1335a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13363a50541eSStefano 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); 1337a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1338a64d13efSStefano Zampini } 1339a64d13efSStefano Zampini 1340a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13413a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1342854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1343854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1344a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 13454641a718SStefano Zampini for (i=0; i<n_D; i++) { 13464641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 13474641a718SStefano Zampini } 1348a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1349a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13504641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 13514641a718SStefano Zampini aux_array1[j++] = i; 1352a64d13efSStefano Zampini } 1353a64d13efSStefano Zampini } 1354a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1355a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1356a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 13574641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 13584641a718SStefano Zampini aux_array2[j++] = i; 1359a64d13efSStefano Zampini } 1360a64d13efSStefano Zampini } 1361a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1362a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1363a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1364a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1365a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1366a64d13efSStefano Zampini 13678eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1368785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1369a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13704641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 13714641a718SStefano Zampini aux_array1[j++] = i; 1372a64d13efSStefano Zampini } 1373a64d13efSStefano Zampini } 1374a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1375a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1376a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1377a64d13efSStefano Zampini } 13784641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 13793a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1380a64d13efSStefano Zampini PetscFunctionReturn(0); 1381a64d13efSStefano Zampini } 1382a64d13efSStefano Zampini 1383304d26faSStefano Zampini 1384304d26faSStefano Zampini #undef __FUNCT__ 1385304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1386684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1387304d26faSStefano Zampini { 1388304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1389304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1390304d26faSStefano Zampini PC pc_temp; 1391304d26faSStefano Zampini Mat A_RR; 1392f4ddd8eeSStefano Zampini MatReuse reuse; 1393304d26faSStefano Zampini PetscScalar m_one = -1.0; 1394304d26faSStefano Zampini PetscReal value; 1395af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 13969577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1397304d26faSStefano Zampini PetscErrorCode ierr; 1398e604994aSStefano Zampini /* prefixes stuff */ 1399312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1400e604994aSStefano Zampini size_t len; 1401304d26faSStefano Zampini 1402304d26faSStefano Zampini PetscFunctionBegin; 1403304d26faSStefano Zampini 1404e604994aSStefano Zampini /* compute prefixes */ 1405e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1406e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1407e604994aSStefano Zampini if (!pcbddc->current_level) { 1408e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1409e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1410e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1411e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1412e604994aSStefano Zampini } else { 1413e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1414312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1415e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1416e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1417312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1418312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 141934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 142034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1421e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1422e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1423e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1424e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1425e604994aSStefano Zampini } 1426e604994aSStefano Zampini 1427304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1428684f6988SStefano Zampini if (dirichlet) { 1429d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1430964fefecSStefano Zampini if (pcbddc->issym) { 1431964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1432964fefecSStefano Zampini } 1433ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1434964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1435304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1436304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1437304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1438304d26faSStefano Zampini /* default */ 1439304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1440e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 14419577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1442304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 14439577ea80SStefano Zampini if (issbaij) { 14449577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 14459577ea80SStefano Zampini } else { 1446304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 14479577ea80SStefano Zampini } 1448304d26faSStefano Zampini /* Allow user's customization */ 1449304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1450304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1451304d26faSStefano Zampini } 1452d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1453d5574798SStefano Zampini if (sub_schurs->interior_solver) { 1454d5574798SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,sub_schurs->interior_solver);CHKERRQ(ierr); 1455d5574798SStefano Zampini } 1456304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1457304d26faSStefano Zampini if (!n_D) { 1458304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1459304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1460304d26faSStefano Zampini } 1461304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1462304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1463304d26faSStefano Zampini /* set ksp_D into pcis data */ 1464304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1465304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1466304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1467684f6988SStefano Zampini } 1468304d26faSStefano Zampini 1469304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1470684f6988SStefano Zampini A_RR = 0; 1471684f6988SStefano Zampini if (neumann) { 1472f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 14738ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1474f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1475f4ddd8eeSStefano Zampini PetscInt nn_R; 147681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1477f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1478f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1479f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1480f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1481f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1482f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1483f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1484727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1485f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1486f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1487f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1488f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1489f4ddd8eeSStefano Zampini } 1490f4ddd8eeSStefano Zampini } 1491f4ddd8eeSStefano Zampini /* last check */ 1492d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1493f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1494f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1495f4ddd8eeSStefano Zampini } 1496f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1497f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1498f4ddd8eeSStefano Zampini } 1499f4ddd8eeSStefano Zampini /* extract A_RR */ 1500af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1501af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 15023a50541eSStefano Zampini if (ibs != mbs) { 1503af732b37SStefano Zampini Mat newmat; 1504af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1505f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1506af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1507af732b37SStefano Zampini } else { 1508f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1509af732b37SStefano Zampini } 1510964fefecSStefano Zampini if (pcbddc->issym) { 1511964fefecSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1512964fefecSStefano Zampini } 1513f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1514304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1515304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1516304d26faSStefano Zampini /* default */ 1517304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1518e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1519304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 15209577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 15219577ea80SStefano Zampini if (issbaij) { 15229577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15239577ea80SStefano Zampini } else { 1524304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15259577ea80SStefano Zampini } 1526304d26faSStefano Zampini /* Allow user's customization */ 1527304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1528304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1529304d26faSStefano Zampini } 1530d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1531304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1532304d26faSStefano Zampini if (!n_R) { 1533304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1534304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1535304d26faSStefano Zampini } 1536304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1537304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1538684f6988SStefano Zampini } 1539304d26faSStefano Zampini 1540304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 15410fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1542684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1543684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1544684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1545684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1546684f6988SStefano Zampini } 1547684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 15480fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 15490fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 15500fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 15510fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 15520fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1553304d26faSStefano Zampini /* need to be adapted? */ 1554b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1555b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1556b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1557304d26faSStefano Zampini /* print info */ 1558304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1559e604994aSStefano 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); 1560304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1561304d26faSStefano Zampini } 1562b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 15638ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1564304d26faSStefano Zampini } 1565684f6988SStefano Zampini } 1566684f6988SStefano Zampini if (neumann) { /* Neumann */ 15670fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 15680fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 15690fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 15700fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 15710fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1572304d26faSStefano Zampini /* need to be adapted? */ 1573b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1574b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1575304d26faSStefano Zampini /* print info */ 1576304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1577e604994aSStefano 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); 1578304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1579304d26faSStefano Zampini } 1580b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 15818ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1582304d26faSStefano Zampini } 15830fccc4e9SStefano Zampini } 1584684f6988SStefano Zampini } 1585304d26faSStefano Zampini /* free Neumann problem's matrix */ 1586304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1587304d26faSStefano Zampini PetscFunctionReturn(0); 1588304d26faSStefano Zampini } 1589304d26faSStefano Zampini 1590304d26faSStefano Zampini #undef __FUNCT__ 1591ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 159220c7b377SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec rhs, Vec sol, Vec work, PetscBool applytranspose) 1593674ae819SStefano Zampini { 1594674ae819SStefano Zampini PetscErrorCode ierr; 1595674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1596674ae819SStefano Zampini 1597674ae819SStefano Zampini PetscFunctionBegin; 159820c7b377SStefano Zampini if (applytranspose) { 1599674ae819SStefano Zampini if (pcbddc->local_auxmat1) { 160020c7b377SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,rhs,work);CHKERRQ(ierr); 160120c7b377SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,work,rhs,rhs);CHKERRQ(ierr); 160220c7b377SStefano Zampini } 160320c7b377SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 160420c7b377SStefano Zampini } else { 160520c7b377SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 160620c7b377SStefano Zampini if (pcbddc->local_auxmat1) { 160720c7b377SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,sol,work);CHKERRQ(ierr); 160820c7b377SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,work,sol,sol);CHKERRQ(ierr); 160920c7b377SStefano Zampini } 1610674ae819SStefano Zampini } 1611674ae819SStefano Zampini PetscFunctionReturn(0); 1612674ae819SStefano Zampini } 1613674ae819SStefano Zampini 1614dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1615674ae819SStefano Zampini #undef __FUNCT__ 1616674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1617dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1618674ae819SStefano Zampini { 1619674ae819SStefano Zampini PetscErrorCode ierr; 1620674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1621674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1622674ae819SStefano Zampini const PetscScalar zero = 0.0; 1623674ae819SStefano Zampini 1624674ae819SStefano Zampini PetscFunctionBegin; 1625dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1626dc359a40SStefano Zampini if (applytranspose) { 1627674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 16288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1629dc359a40SStefano Zampini } else { 1630674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1631674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 163215aaf578SStefano Zampini } 163312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 163412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 163512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 163612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 163712edc857SStefano Zampini 16389f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 163912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 164012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1641964fefecSStefano Zampini Vec rhs,sol; 1642964fefecSStefano Zampini 1643964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1644964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 164512edc857SStefano Zampini if (applytranspose) { 1646964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 164712edc857SStefano Zampini } else { 1648964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 164912edc857SStefano Zampini } 165012edc857SStefano Zampini } 1651674ae819SStefano Zampini 1652674ae819SStefano Zampini /* Local solution on R nodes */ 16539f00e9b4SStefano Zampini if (pcis->n) { 1654674ae819SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 1655674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1656674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 16578eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1658674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1659674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1660674ae819SStefano Zampini } 166120c7b377SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcbddc->vec1_R,pcbddc->vec2_R,pcbddc->vec1_C,applytranspose);CHKERRQ(ierr); 1662674ae819SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 1663674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1664674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 16658eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1666674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1667674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1668674ae819SStefano Zampini } 16699f00e9b4SStefano Zampini } 1670674ae819SStefano Zampini 16719f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 16729f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 167312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1674674ae819SStefano Zampini 1675674ae819SStefano Zampini /* Sum contributions from two levels */ 1676dc359a40SStefano Zampini if (applytranspose) { 1677dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1678dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1679dc359a40SStefano Zampini } else { 1680674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 16818eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1682dc359a40SStefano Zampini } 1683674ae819SStefano Zampini PetscFunctionReturn(0); 1684674ae819SStefano Zampini } 1685674ae819SStefano Zampini 1686674ae819SStefano Zampini #undef __FUNCT__ 1687674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 168812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1689674ae819SStefano Zampini { 1690674ae819SStefano Zampini PetscErrorCode ierr; 1691674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 169258da7f69SStefano Zampini PetscScalar *array; 169312edc857SStefano Zampini Vec from,to; 1694674ae819SStefano Zampini 1695674ae819SStefano Zampini PetscFunctionBegin; 169612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 169712edc857SStefano Zampini from = pcbddc->coarse_vec; 169812edc857SStefano Zampini to = pcbddc->vec1_P; 169912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 170012edc857SStefano Zampini Vec tvec; 170158da7f69SStefano Zampini 170258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 170358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 170412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 170558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 170658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 170758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 170812edc857SStefano Zampini } 170912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 171012edc857SStefano Zampini from = pcbddc->vec1_P; 171112edc857SStefano Zampini to = pcbddc->coarse_vec; 171212edc857SStefano Zampini } 171312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1714674ae819SStefano Zampini PetscFunctionReturn(0); 1715674ae819SStefano Zampini } 1716674ae819SStefano Zampini 1717674ae819SStefano Zampini #undef __FUNCT__ 1718674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 171912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1720674ae819SStefano Zampini { 1721674ae819SStefano Zampini PetscErrorCode ierr; 1722674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 172358da7f69SStefano Zampini PetscScalar *array; 172412edc857SStefano Zampini Vec from,to; 1725674ae819SStefano Zampini 1726674ae819SStefano Zampini PetscFunctionBegin; 172712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 172812edc857SStefano Zampini from = pcbddc->coarse_vec; 172912edc857SStefano Zampini to = pcbddc->vec1_P; 173012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 173112edc857SStefano Zampini from = pcbddc->vec1_P; 173212edc857SStefano Zampini to = pcbddc->coarse_vec; 173312edc857SStefano Zampini } 173412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 173512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 173612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 173712edc857SStefano Zampini Vec tvec; 173858da7f69SStefano Zampini 173912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 174058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 174158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 174258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 174358da7f69SStefano Zampini } 174458da7f69SStefano Zampini } else { 174558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 174658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 174712edc857SStefano Zampini } 174812edc857SStefano Zampini } 1749674ae819SStefano Zampini PetscFunctionReturn(0); 1750674ae819SStefano Zampini } 1751674ae819SStefano Zampini 1752984c4197SStefano Zampini /* uncomment for testing purposes */ 1753984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1754674ae819SStefano Zampini #undef __FUNCT__ 1755674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1756674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1757674ae819SStefano Zampini { 1758674ae819SStefano Zampini PetscErrorCode ierr; 1759674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1760674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1761674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1762984c4197SStefano Zampini /* one and zero */ 1763984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1764984c4197SStefano Zampini /* space to store constraints and their local indices */ 1765*9162d606SStefano Zampini PetscScalar *constraints_data; 1766*9162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 1767*9162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 1768*9162d606SStefano Zampini PetscInt *constraints_n; 1769984c4197SStefano Zampini /* iterators */ 1770*9162d606SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc; 1771984c4197SStefano Zampini /* BLAS integers */ 1772e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1773e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1774c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1775727cdba6SStefano Zampini /* reuse */ 1776727cdba6SStefano Zampini PetscInt olocal_primal_size; 1777727cdba6SStefano Zampini PetscInt *oprimal_indices_local_idxs; 1778984c4197SStefano Zampini /* change of basis */ 1779a717540cSStefano Zampini PetscBool boolforchange,qr_needed; 1780*9162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1781984c4197SStefano Zampini /* auxiliary stuff */ 178264efe560SStefano Zampini PetscInt *nnz,*is_indices; 17838a0068c3SStefano Zampini PetscInt ncc; 1784984c4197SStefano Zampini /* some quantities */ 178545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1786a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1787984c4197SStefano Zampini 1788674ae819SStefano Zampini PetscFunctionBegin; 17898e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 17908e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 17918e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1792088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1793088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 1794088faed8SStefano Zampini ierr = PetscMalloc1(olocal_primal_size,&oprimal_indices_local_idxs);CHKERRQ(ierr); 1795088faed8SStefano Zampini ierr = PetscMemcpy(oprimal_indices_local_idxs,pcbddc->primal_indices_local_idxs,olocal_primal_size*sizeof(PetscInt));CHKERRQ(ierr); 1796088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1797cf5a6209SStefano Zampini 1798cf5a6209SStefano Zampini /* print some info */ 1799cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1800cf5a6209SStefano Zampini IS vertices; 1801cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1802cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1803cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1804cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1805cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1806cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1807cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1808fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1809fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1810cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1811cf5a6209SStefano Zampini } 1812cf5a6209SStefano Zampini 1813cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 1814*9162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1815cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1816cf5a6209SStefano Zampini const Vec *nearnullvecs; 1817cf5a6209SStefano Zampini Vec *localnearnullsp; 1818cf5a6209SStefano Zampini PetscScalar *array; 1819cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1820cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1821cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1822cf5a6209SStefano Zampini PetscBool skip_lapack; 1823cf5a6209SStefano Zampini PetscScalar *work; 1824cf5a6209SStefano Zampini PetscReal *singular_vals; 1825cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1826cf5a6209SStefano Zampini PetscReal *rwork; 1827cf5a6209SStefano Zampini #endif 1828cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1829cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1830cf5a6209SStefano Zampini #else 1831964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1832964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1833cf5a6209SStefano Zampini #endif 1834cf5a6209SStefano Zampini 1835674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1836d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1837d06fc5fdSStefano Zampini /* free unneeded index sets */ 1838d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1839d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1840674ae819SStefano Zampini } 1841d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1842d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1843d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1844d06fc5fdSStefano Zampini } 1845d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1846d06fc5fdSStefano Zampini n_ISForEdges = 0; 1847d06fc5fdSStefano Zampini } 1848d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1849d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1850d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1851d06fc5fdSStefano Zampini } 1852d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1853d06fc5fdSStefano Zampini n_ISForFaces = 0; 1854d06fc5fdSStefano Zampini } 185570022509SStefano Zampini 185670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 185770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 185870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 185970022509SStefano Zampini if (pcbddc->NullSpace) { 186070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 186170022509SStefano Zampini 186270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 1863b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1864d06fc5fdSStefano Zampini if (!ISForEdges) { 1865d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1866d06fc5fdSStefano Zampini } 1867b8ffe317SStefano Zampini } 1868d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 1869d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 1870d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1871d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1872d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 187398a51de6SStefano Zampini } 187470022509SStefano Zampini #endif 187508122e43SStefano Zampini 1876674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1877674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1878674ae819SStefano Zampini if (nearnullsp) { 1879674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1880f4ddd8eeSStefano Zampini /* remove any stored info */ 1881f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1882f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1883f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1884f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1885f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1886473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1887f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1888f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1889f4ddd8eeSStefano Zampini } 1890984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1891984c4197SStefano Zampini nnsp_size = 0; 1892674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1893674ae819SStefano Zampini } 1894984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1895984c4197SStefano Zampini max_constraints = nnsp_size; 1896984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1897984c4197SStefano Zampini 1898674ae819SStefano Zampini /* 1899674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 1900*9162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 1901*9162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 1902*9162d606SStefano Zampini There can be multiple constraints per connected component 1903674ae819SStefano Zampini */ 1904674ae819SStefano Zampini n_vertices = 0; 1905674ae819SStefano Zampini if (ISForVertices) { 1906674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1907674ae819SStefano Zampini } 1908*9162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 1909*9162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 1910*9162d606SStefano Zampini 1911*9162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 1912*9162d606SStefano Zampini total_counts *= max_constraints; 1913674ae819SStefano Zampini total_counts += n_vertices; 19144641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 1915*9162d606SStefano Zampini 1916674ae819SStefano Zampini total_counts = 0; 1917674ae819SStefano Zampini max_size_of_constraint = 0; 1918674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 1919*9162d606SStefano Zampini IS used_is; 1920674ae819SStefano Zampini if (i<n_ISForEdges) { 1921*9162d606SStefano Zampini used_is = ISForEdges[i]; 1922674ae819SStefano Zampini } else { 1923*9162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 1924674ae819SStefano Zampini } 1925*9162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 1926674ae819SStefano Zampini total_counts += j; 1927674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1928674ae819SStefano Zampini } 1929*9162d606SStefano 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); 1930*9162d606SStefano Zampini 1931984c4197SStefano Zampini /* get local part of global near null space vectors */ 1932785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1933984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1934984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1935984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1936984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1937984c4197SStefano Zampini } 1938674ae819SStefano Zampini 1939242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1940242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1941a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1942242a89d7SStefano Zampini 1943984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1944a773dcb8SStefano Zampini if (!skip_lapack) { 1945674ae819SStefano Zampini PetscScalar temp_work; 1946911cabfeSStefano Zampini 1947674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1948984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1949785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 1950785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 1951785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 1952674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1953785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 1954674ae819SStefano Zampini #endif 1955674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1956c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 1957c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 1958674ae819SStefano Zampini lwork = -1; 1959674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1960674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1961c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 1962674ae819SStefano Zampini #else 1963c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 1964674ae819SStefano Zampini #endif 1965674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1966984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 1967674ae819SStefano Zampini #else /* on missing GESVD */ 1968674ae819SStefano Zampini /* SVD */ 1969674ae819SStefano Zampini PetscInt max_n,min_n; 1970674ae819SStefano Zampini max_n = max_size_of_constraint; 1971984c4197SStefano Zampini min_n = max_constraints; 1972984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 1973674ae819SStefano Zampini min_n = max_size_of_constraint; 1974984c4197SStefano Zampini max_n = max_constraints; 1975674ae819SStefano Zampini } 1976785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 1977674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1978785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 1979674ae819SStefano Zampini #endif 1980674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1981674ae819SStefano Zampini lwork = -1; 1982e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 1983e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 1984b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 1985674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1986674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1987*9162d606SStefano 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)); 1988674ae819SStefano Zampini #else 1989*9162d606SStefano 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)); 1990674ae819SStefano Zampini #endif 1991674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1992984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 1993984c4197SStefano Zampini #endif /* on missing GESVD */ 1994674ae819SStefano Zampini /* Allocate optimal workspace */ 1995674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 1996854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 1997674ae819SStefano Zampini } 1998674ae819SStefano Zampini /* Now we can loop on constraining sets */ 1999674ae819SStefano Zampini total_counts = 0; 2000*9162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 2001*9162d606SStefano Zampini constraints_data_ptr[0] = 0; 2002674ae819SStefano Zampini /* vertices */ 2003*9162d606SStefano Zampini if (n_vertices) { 2004674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2005*9162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 2006*9162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2007674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2008*9162d606SStefano Zampini constraints_n[total_counts] = 1; 2009*9162d606SStefano Zampini constraints_data[total_counts] = 1.0; 2010*9162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 2011*9162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2012674ae819SStefano Zampini total_counts++; 2013674ae819SStefano Zampini } 2014674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2015984c4197SStefano Zampini PetscBool used_vertex; 2016674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2017674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2018674ae819SStefano Zampini k = 0; 2019674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2020984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2021984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 2022*9162d606SStefano Zampini constraints_n[total_counts] = 1; 2023*9162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 2024*9162d606SStefano Zampini constraints_data[total_counts] = 1.0; 2025*9162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 2026*9162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2027674ae819SStefano Zampini total_counts++; 2028674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2029674ae819SStefano Zampini } 2030984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2031674ae819SStefano Zampini k++; 2032674ae819SStefano Zampini } 2033674ae819SStefano Zampini } 2034674ae819SStefano Zampini } 2035674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2036674ae819SStefano Zampini n_vertices = total_counts; 2037674ae819SStefano Zampini } 2038984c4197SStefano Zampini 2039674ae819SStefano Zampini /* edges and faces */ 2040*9162d606SStefano Zampini total_counts_cc = total_counts; 2041911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 2042*9162d606SStefano Zampini IS used_is; 2043*9162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 2044*9162d606SStefano Zampini 2045911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 2046*9162d606SStefano Zampini used_is = ISForEdges[ncc]; 2047984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2048674ae819SStefano Zampini } else { 2049*9162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2050984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2051674ae819SStefano Zampini } 2052674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 2053*9162d606SStefano Zampini 2054*9162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 2055*9162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2056984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2057984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2058674ae819SStefano Zampini if (nnsp_has_cnst) { 20595b08dc53SStefano Zampini PetscScalar quad_value; 2060*9162d606SStefano Zampini 2061*9162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2062*9162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2063*9162d606SStefano Zampini 2064a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2065674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2066a773dcb8SStefano Zampini } else { 2067a773dcb8SStefano Zampini quad_value = 1.0; 2068a773dcb8SStefano Zampini } 2069674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 2070*9162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2071674ae819SStefano Zampini } 2072*9162d606SStefano Zampini temp_constraints++; 2073674ae819SStefano Zampini total_counts++; 2074674ae819SStefano Zampini } 2075674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2076984c4197SStefano Zampini PetscReal real_value; 2077*9162d606SStefano Zampini PetscScalar *ptr_to_data; 2078*9162d606SStefano Zampini 2079984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2080*9162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2081674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 2082*9162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2083674ae819SStefano Zampini } 2084984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2085984c4197SStefano Zampini /* check if array is null on the connected component */ 2086e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2087*9162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 20885b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2089674ae819SStefano Zampini temp_constraints++; 2090674ae819SStefano Zampini total_counts++; 2091*9162d606SStefano Zampini if (!idxs_copied) { 2092*9162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2093*9162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2094674ae819SStefano Zampini } 2095674ae819SStefano Zampini } 2096*9162d606SStefano Zampini } 2097*9162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 209845a1bb75SStefano Zampini valid_constraints = temp_constraints; 2099eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2100a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 2101*9162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 2102*9162d606SStefano Zampini 2103*9162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2104a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2105*9162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2106a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 2107*9162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2108a773dcb8SStefano Zampini } else { /* perform SVD */ 2109984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 2110*9162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2111674ae819SStefano Zampini 2112674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2113984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2114984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2115984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2116984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2117984c4197SStefano Zampini from that computed using LAPACKgesvd 2118984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2119984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2120984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2121674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2122e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2123984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2124674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2125674ae819SStefano Zampini for (k=0;k<j+1;k++) { 2126*9162d606SStefano 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)); 2127674ae819SStefano Zampini } 2128674ae819SStefano Zampini } 2129e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2130e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2131e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2132674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2133c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2134674ae819SStefano Zampini #else 2135c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2136674ae819SStefano Zampini #endif 2137674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2138984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2139984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2140674ae819SStefano Zampini j = 0; 2141984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2142674ae819SStefano Zampini total_counts = total_counts-j; 214345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2144e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2145c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2146c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2147c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2148c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2149c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2150c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2151674ae819SStefano Zampini if (j<temp_constraints) { 2152984c4197SStefano Zampini PetscInt ii; 2153984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2154674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2155*9162d606SStefano 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)); 2156674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2157984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2158674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 2159*9162d606SStefano 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]; 2160674ae819SStefano Zampini } 2161674ae819SStefano Zampini } 2162674ae819SStefano Zampini } 2163674ae819SStefano Zampini #else /* on missing GESVD */ 2164e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2165e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2166b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2167674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2168674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2169*9162d606SStefano 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)); 2170674ae819SStefano Zampini #else 2171*9162d606SStefano 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)); 2172674ae819SStefano Zampini #endif 2173984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2174674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2175984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2176e310c8b4SStefano Zampini k = temp_constraints; 2177e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2178674ae819SStefano Zampini j = 0; 2179e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 218045a1bb75SStefano Zampini valid_constraints = k-j; 2181911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2182984c4197SStefano Zampini #endif /* on missing GESVD */ 2183674ae819SStefano Zampini } 2184a773dcb8SStefano Zampini } 2185*9162d606SStefano Zampini /* update pointers information */ 2186*9162d606SStefano Zampini if (valid_constraints) { 2187*9162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 2188*9162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 2189*9162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 2190*9162d606SStefano Zampini total_counts_cc++; 2191*9162d606SStefano Zampini /* set change_of_basis flag */ 219245a1bb75SStefano Zampini if (boolforchange) { 2193*9162d606SStefano Zampini for (j=valid_constraints;j>0;j--) { 2194*9162d606SStefano Zampini PetscBTSet(change_basis,total_counts-j); 2195*9162d606SStefano Zampini } 219645a1bb75SStefano Zampini } 219745a1bb75SStefano Zampini } 2198674ae819SStefano Zampini } 2199984c4197SStefano Zampini /* free workspace */ 22008f1c130eSStefano Zampini if (!skip_lapack) { 2201984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2202984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2203984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2204984c4197SStefano Zampini #endif 2205984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2206984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2207984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2208984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2209984c4197SStefano Zampini #endif 2210984c4197SStefano Zampini } 2211984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2212984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2213984c4197SStefano Zampini } 2214984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2215cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2216cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2217cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2218cf5a6209SStefano Zampini } 2219cf5a6209SStefano Zampini if (n_ISForFaces) { 2220cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2221cf5a6209SStefano Zampini } 2222cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2223cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2224cf5a6209SStefano Zampini } 2225cf5a6209SStefano Zampini if (n_ISForEdges) { 2226cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2227cf5a6209SStefano Zampini } 2228cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 222908122e43SStefano Zampini } else { 223008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 223108122e43SStefano Zampini 223208122e43SStefano Zampini total_counts = 0; 223308122e43SStefano Zampini n_vertices = 0; 22348bec7fa6SStefano Zampini if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { 223508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&n_vertices);CHKERRQ(ierr); 223608122e43SStefano Zampini } 223708122e43SStefano Zampini max_constraints = 0; 2238*9162d606SStefano Zampini total_counts_cc = 0; 223908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 224008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 2241*9162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 224208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 224308122e43SStefano Zampini } 2244*9162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 2245*9162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 2246*9162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 2247*9162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 2248*9162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 2249*9162d606SStefano Zampini total_counts_cc = 0; 2250*9162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 2251*9162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 2252*9162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 225308122e43SStefano Zampini } 225408122e43SStefano Zampini } 2255*9162d606SStefano Zampini #if 0 2256*9162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 2257*9162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 2258*9162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 2259*9162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 2260*9162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 2261*9162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 2262*9162d606SStefano Zampini } 2263*9162d606SStefano Zampini printf("\n"); 2264*9162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 2265*9162d606SStefano Zampini } 22661b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 22678bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 22681b968477SStefano Zampini } 22691b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 22708bec7fa6SStefano 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]); 22711b968477SStefano Zampini } 227208122e43SStefano Zampini #endif 227308122e43SStefano Zampini 22748bec7fa6SStefano Zampini max_size_of_constraint = 0; 2275*9162d606SStefano 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]); 2276*9162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 227708122e43SStefano Zampini /* Change of basis */ 227808122e43SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 227908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 2280*9162d606SStefano Zampini PetscInt cum = n_vertices; 228108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 228208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2283bbb9e6c6SStefano Zampini for (j=0;j<pcbddc->adaptive_constraints_n[i+n_vertices];j++) { 228408122e43SStefano Zampini ierr = PetscBTSet(change_basis,cum+j);CHKERRQ(ierr); 228508122e43SStefano Zampini } 228608122e43SStefano Zampini } 2287bbb9e6c6SStefano Zampini cum += pcbddc->adaptive_constraints_n[i+n_vertices]; 228808122e43SStefano Zampini } 228908122e43SStefano Zampini } 229008122e43SStefano Zampini } 2291b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 229208122e43SStefano Zampini 2293*9162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 2294*9162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 2295*9162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 2296*9162d606SStefano 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); 229708122e43SStefano Zampini } 2298984c4197SStefano Zampini 2299674ae819SStefano Zampini /* Create constraint matrix */ 2300674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 230116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2302984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2303984c4197SStefano Zampini /* array to compute a local numbering of constraints : vertices first then constraints */ 230464efe560SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2305984c4197SStefano Zampini 2306984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2307984c4197SStefano Zampini total_primal_vertices=0; 2308*9162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 2309*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2310984c4197SStefano Zampini if (size_of_constraint == 1) { 2311*9162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 231264efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 2313*9162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 2314*9162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2315*9162d606SStefano Zampini } 2316984c4197SStefano Zampini } 2317984c4197SStefano Zampini } 2318b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2319b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 232070022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 232170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 232270022509SStefano Zampini 2323a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2324a717540cSStefano Zampini qr_needed = PETSC_FALSE; 2325*9162d606SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2326*9162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2327a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 2328*9162d606SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2329fa434743SStefano Zampini PetscBTSet(qr_needed_idx,i); 2330fa434743SStefano Zampini qr_needed = PETSC_TRUE; 2331fa434743SStefano Zampini } 2332a717540cSStefano Zampini } 2333a717540cSStefano Zampini } 233470022509SStefano Zampini 233570022509SStefano Zampini /* get reference dof for local constraints */ 2336088faed8SStefano Zampini total_counts = total_primal_vertices; 2337*9162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2338*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2339*9162d606SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2340*9162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 2341*9162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2342*9162d606SStefano Zampini } 2343088faed8SStefano Zampini } 2344088faed8SStefano Zampini } 234545a1bb75SStefano Zampini 2346984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 2347785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2348984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i]=1; 2349984c4197SStefano Zampini j = total_primal_vertices; 2350*9162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 23514641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2352*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2353*9162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) nnz[j+k] = size_of_constraint; 2354*9162d606SStefano Zampini j += constraints_n[i]; 2355674ae819SStefano Zampini } 2356674ae819SStefano Zampini } 2357674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2358674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2359088faed8SStefano Zampini 2360674ae819SStefano Zampini /* set values in constraint matrix */ 2361984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 236264efe560SStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2363674ae819SStefano Zampini } 2364984c4197SStefano Zampini total_counts = total_primal_vertices; 2365*9162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 23664641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2367*9162d606SStefano Zampini PetscInt *cols; 2368*9162d606SStefano Zampini 2369*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2370*9162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 2371*9162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 2372*9162d606SStefano Zampini PetscInt row = total_counts+k; 2373*9162d606SStefano Zampini PetscScalar *vals; 2374*9162d606SStefano Zampini 2375*9162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 2376*9162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2377*9162d606SStefano Zampini } 2378*9162d606SStefano Zampini total_counts += constraints_n[i]; 2379674ae819SStefano Zampini } 2380674ae819SStefano Zampini } 2381674ae819SStefano Zampini /* assembling */ 2382674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2383674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2384088faed8SStefano Zampini 2385984c4197SStefano Zampini /* 238645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2387984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2388984c4197SStefano Zampini */ 2389674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2390674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2391026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2392984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2393984c4197SStefano Zampini /* working stuff for GEQRF */ 239481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2395984c4197SStefano Zampini PetscBLASInt lqr_work; 2396984c4197SStefano Zampini /* working stuff for UNGQR */ 2397984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2398984c4197SStefano Zampini PetscBLASInt lgqr_work; 2399984c4197SStefano Zampini /* working stuff for TRTRS */ 2400984c4197SStefano Zampini PetscScalar *trs_rhs; 24013f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2402984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2403984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2404984c4197SStefano Zampini PetscScalar *start_vals; 2405984c4197SStefano Zampini /* working stuff for values insertion */ 24064641a718SStefano Zampini PetscBT is_primal; 240764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2408906d46d4SStefano Zampini /* matrix sizes */ 2409906d46d4SStefano Zampini PetscInt global_size,local_size; 2410906d46d4SStefano Zampini /* temporary change of basis */ 2411906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2412cf5a6209SStefano Zampini /* extra space for debugging */ 2413cf5a6209SStefano Zampini PetscScalar *dbg_work; 2414984c4197SStefano Zampini 2415906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2416906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 241716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2418bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2419906d46d4SStefano Zampini /* nonzeros for local mat */ 2420bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2421bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 2422*9162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2423a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 2424*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2425a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 2426*9162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2427a717540cSStefano Zampini } else { 2428*9162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 2429*9162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2430a717540cSStefano Zampini } 2431a717540cSStefano Zampini } 2432a717540cSStefano Zampini } 2433906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2434bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2435a717540cSStefano Zampini /* Set initial identity in the matrix */ 2436bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2437906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2438a717540cSStefano Zampini } 2439a717540cSStefano Zampini 2440a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2441a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2442a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2443a717540cSStefano Zampini } 2444a717540cSStefano Zampini 2445a717540cSStefano Zampini 2446a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2447a717540cSStefano Zampini /* 2448a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2449a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2450a717540cSStefano Zampini 2451a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2452a717540cSStefano Zampini 2453a6b551f4SStefano 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) 2454a6b551f4SStefano Zampini 2455a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2456a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2457a717540cSStefano Zampini | ... | 2458a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2459a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2460a717540cSStefano Zampini 2461a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2462a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2463a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2464a6b551f4SStefano Zampini 2465a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2466a717540cSStefano Zampini */ 2467a717540cSStefano Zampini if (qr_needed) { 2468984c4197SStefano Zampini /* space to store Q */ 2469854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2470984c4197SStefano Zampini /* first we issue queries for optimal work */ 24713f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 24723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 24733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2474984c4197SStefano Zampini lqr_work = -1; 24753f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2476984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2477984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2478785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2479984c4197SStefano Zampini lgqr_work = -1; 24803f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 24813f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 24823f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 24833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 24843f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 24853f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2486984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2487984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2488785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2489984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2490785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2491984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2492785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2493a717540cSStefano Zampini /* allocating workspace for check */ 2494a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2495cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2496a717540cSStefano Zampini } 2497a717540cSStefano Zampini } 2498984c4197SStefano Zampini /* array to store whether a node is primal or not */ 24994641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2500473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 250164efe560SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->primal_indices_local_idxs,&i,aux_primal_numbering_B);CHKERRQ(ierr); 250239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 250339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 25044641a718SStefano Zampini } 250539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 250639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 250739e2fb2aSStefano Zampini } 250839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2509984c4197SStefano Zampini 2510a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 2511*9162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 2512*9162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 25134641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2514984c4197SStefano Zampini /* get constraint info */ 2515*9162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2516984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2517984c4197SStefano Zampini 2518984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2519*9162d606SStefano 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); 2520674ae819SStefano Zampini } 2521984c4197SStefano Zampini 2522fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2523a717540cSStefano Zampini 2524a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2525a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2526*9162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2527a717540cSStefano Zampini } 2528984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 2529*9162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2530984c4197SStefano Zampini 2531984c4197SStefano Zampini /* compute QR decomposition of constraints */ 25323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25333f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2535674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25363f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2537984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2538674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2539984c4197SStefano Zampini 2540984c4197SStefano Zampini /* explictly compute R^-T */ 2541984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2542984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 25433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 25453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2547984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2549984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2550984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2551984c4197SStefano Zampini 2552a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 25533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25543f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25553f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 25563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2557984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25583f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2559984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2560984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2561984c4197SStefano Zampini 2562984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2563984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2564984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 25653f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25663f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25673f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 25683f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 25703f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2571984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2572*9162d606SStefano 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)); 2573984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2574*9162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2575984c4197SStefano Zampini 2576984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 2577*9162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2578984c4197SStefano Zampini /* insert cols for primal dofs */ 2579984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2580984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 2581*9162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2582906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2583984c4197SStefano Zampini } 2584984c4197SStefano Zampini /* insert cols for dual dofs */ 2585984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 2586*9162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2587984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 2588*9162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2589906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2590984c4197SStefano Zampini j++; 2591674ae819SStefano Zampini } 2592674ae819SStefano Zampini } 2593984c4197SStefano Zampini 2594984c4197SStefano Zampini /* check change of basis */ 2595984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2596984c4197SStefano Zampini PetscInt ii,jj; 2597984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2598c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2599c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2600c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2601c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2602c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2603c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2604984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2605cf5a6209SStefano 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)); 2606984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2607984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2608984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2609cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2610cf5a6209SStefano 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; 2611674ae819SStefano Zampini } 2612674ae819SStefano Zampini } 2613984c4197SStefano Zampini if (!valid_qr) { 261422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2615984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2616984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2617cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2618cf5a6209SStefano 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])); 2619674ae819SStefano Zampini } 2620cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2621cf5a6209SStefano 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])); 2622984c4197SStefano Zampini } 2623984c4197SStefano Zampini } 2624984c4197SStefano Zampini } 2625674ae819SStefano Zampini } else { 262622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2627674ae819SStefano Zampini } 2628674ae819SStefano Zampini } 2629a717540cSStefano Zampini } else { /* simple transformation block */ 2630a717540cSStefano Zampini PetscInt row,col; 2631a6b551f4SStefano Zampini PetscScalar val,norm; 2632a6b551f4SStefano Zampini 2633a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2634*9162d606SStefano 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)); 2635a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2636*9162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 2637*9162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2638bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 2639*9162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2640906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 2641*9162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2642a717540cSStefano Zampini } else { 2643a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 2644*9162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2645a717540cSStefano Zampini if (row != col) { 2646*9162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2647a717540cSStefano Zampini } else { 2648*9162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2649a717540cSStefano Zampini } 2650906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2651a717540cSStefano Zampini } 2652a717540cSStefano Zampini } 2653a717540cSStefano Zampini } 265498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 265522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2656a717540cSStefano Zampini } 2657674ae819SStefano Zampini } 2658984c4197SStefano Zampini } else { 2659984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2660*9162d606SStefano 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); 2661674ae819SStefano Zampini } 2662674ae819SStefano Zampini } 2663674ae819SStefano Zampini } 2664a717540cSStefano Zampini 2665a717540cSStefano Zampini /* free workspace */ 2666a717540cSStefano Zampini if (qr_needed) { 2667984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2668cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2669984c4197SStefano Zampini } 2670984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2671984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2672984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2673984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2674984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2675674ae819SStefano Zampini } 2676a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2677906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2678906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2679906d46d4SStefano Zampini 2680906d46d4SStefano Zampini /* assembling of global change of variable */ 2681bbb9e6c6SStefano Zampini { 2682bbb9e6c6SStefano Zampini Mat tmat; 268316f15bc4SStefano Zampini PetscInt bs; 268416f15bc4SStefano Zampini 2685906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2686906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2687bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2688bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2689bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2690bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 269116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 269216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2693906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2694bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2695bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2696bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2697bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2698bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2699bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2700bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2701bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2702bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2703906d46d4SStefano Zampini } 2704906d46d4SStefano Zampini /* check */ 2705906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2706906d46d4SStefano Zampini PetscReal error; 2707906d46d4SStefano Zampini Vec x,x_change; 2708906d46d4SStefano Zampini 2709906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2710906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2711906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2712906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2713bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2714bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2715bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2716bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2717bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2718906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2719906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2720906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2721906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2722bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2723906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2724906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2725906d46d4SStefano Zampini } 2726b96c3477SStefano Zampini 2727b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2728b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2729b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2730b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2731ac632422SStefano Zampini Mat S_new,tmat; 2732bbb9e6c6SStefano Zampini IS is_all_N; 2733bbb9e6c6SStefano Zampini 2734bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 2735d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr); 2736bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2737ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2738b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2739ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2740ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2741ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2742ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2743ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2744b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2745ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2746ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2747ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2748ac632422SStefano Zampini } 2749b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2750b96c3477SStefano Zampini } 2751b96c3477SStefano Zampini } 2752906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2753906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2754b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2755b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2756b9b85e73SStefano Zampini } 2757906d46d4SStefano Zampini 2758906d46d4SStefano Zampini /* set up change of basis context */ 2759906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2760906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2761906d46d4SStefano Zampini 2762906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2763906d46d4SStefano Zampini PetscInt global_size,local_size; 2764906d46d4SStefano Zampini 2765906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2766906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2767906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2768906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2769906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2770906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2771906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2772906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2773906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2774906d46d4SStefano Zampini } else { 2775906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2776906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2777906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2778906d46d4SStefano Zampini } 2779906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2780906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2781906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2782906d46d4SStefano Zampini } else { 2783906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2784906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2785906d46d4SStefano Zampini } 2786906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2787906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2788906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2789906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2790b9b85e73SStefano Zampini } 2791a717540cSStefano Zampini 2792727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2793727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2794727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2795727cdba6SStefano Zampini ierr = PetscMemcmp(pcbddc->primal_indices_local_idxs,oprimal_indices_local_idxs,olocal_primal_size,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2796c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2797727cdba6SStefano Zampini } 2798088faed8SStefano Zampini ierr = PetscFree(oprimal_indices_local_idxs);CHKERRQ(ierr); 2799727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2800727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2801727cdba6SStefano Zampini 2802a717540cSStefano Zampini /* flush dbg viewer */ 2803b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2804b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2805b8ffe317SStefano Zampini } 2806a717540cSStefano Zampini 2807e310c8b4SStefano Zampini /* free workspace */ 2808a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 28094641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 281008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 2811*9162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 2812*9162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 281308122e43SStefano Zampini } else { 2814*9162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 2815*9162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 2816*9162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 281708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 281808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 2819*9162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 2820*9162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 282108122e43SStefano Zampini } 2822674ae819SStefano Zampini PetscFunctionReturn(0); 2823674ae819SStefano Zampini } 2824674ae819SStefano Zampini 2825674ae819SStefano Zampini #undef __FUNCT__ 2826674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2827674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2828674ae819SStefano Zampini { 2829674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2830674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2831674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 28327fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2833674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2834674ae819SStefano Zampini 2835674ae819SStefano Zampini PetscFunctionBegin; 28368e61c736SStefano Zampini /* Reset previously computed graph */ 28378e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2838674ae819SStefano Zampini /* Init local Graph struct */ 28397fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 28407fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 2841674ae819SStefano Zampini 2842575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2843575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2844575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2845575ad6abSStefano Zampini } 28469577ea80SStefano Zampini 2847674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 28484d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 28494d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 28504d379d7bSStefano Zampini PetscInt nvtxs; 2851e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 2852674ae819SStefano Zampini 28534d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 28542fffb893SStefano Zampini 28552fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 28562fffb893SStefano Zampini if (flg_row) { 28574d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2858b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 28592fffb893SStefano Zampini } 28602fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 28619b28b941SStefano 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 */ 28624d379d7bSStefano Zampini IS is_dummy; 28634d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 28644d379d7bSStefano Zampini PetscInt j,sum; 28654d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 28664d379d7bSStefano Zampini const PetscInt *idxs; 28674d379d7bSStefano Zampini PCBDDCGraph graph; 28684d379d7bSStefano Zampini PetscBT is_on_boundary; 28694d379d7bSStefano Zampini 28704d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 28714d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 28724d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 28734d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 28747fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 28754d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 2876e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2877e496cd5dSStefano Zampini if (flg_row) { 28784d379d7bSStefano Zampini graph->xadj = xadj; 28794d379d7bSStefano Zampini graph->adjncy = adjncy; 2880e496cd5dSStefano Zampini } 28814d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 28824d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2883e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 28844d379d7bSStefano Zampini 28854d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 28869b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 28874d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 28884d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 28894d379d7bSStefano Zampini } 28904d379d7bSStefano Zampini } 28914d379d7bSStefano Zampini 2892e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 28934d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 28944d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 28954d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 28964d379d7bSStefano Zampini } 28974d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 28984d379d7bSStefano Zampini 2899e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 29004d379d7bSStefano Zampini sum = 0; 29014d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29024d379d7bSStefano Zampini PetscInt sizecc = 0; 29034d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29044d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29054d379d7bSStefano Zampini sizecc++; 29064d379d7bSStefano Zampini } 29074d379d7bSStefano Zampini } 29084d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29094d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29104d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 29114d379d7bSStefano Zampini } 29124d379d7bSStefano Zampini } 29134d379d7bSStefano Zampini sum += sizecc*sizecc; 29144d379d7bSStefano Zampini } 29154d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 29164d379d7bSStefano Zampini sum = 0; 2917e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 29184d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 29194d379d7bSStefano Zampini cxadj[i] = sum; 29204d379d7bSStefano Zampini sum += temp; 29214d379d7bSStefano Zampini } 2922e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 29234d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29244d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29254d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29264d379d7bSStefano Zampini PetscInt k,sizecc = 0; 29274d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 29284d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 29294d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 29304d379d7bSStefano Zampini sizecc++; 29314d379d7bSStefano Zampini } 29324d379d7bSStefano Zampini } 29334d379d7bSStefano Zampini } 29344d379d7bSStefano Zampini } 29354d379d7bSStefano Zampini } 29369b28b941SStefano Zampini if (sum) { 2937e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 29384d379d7bSStefano Zampini } else { 29394d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 29404d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 29414d379d7bSStefano Zampini } 29424d379d7bSStefano Zampini graph->xadj = 0; 29434d379d7bSStefano Zampini graph->adjncy = 0; 29444d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 29454d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 29464d379d7bSStefano Zampini } 2947674ae819SStefano Zampini } 29489b28b941SStefano Zampini if (pcbddc->dbg_flag) { 29499b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29509b28b941SStefano Zampini } 2951674ae819SStefano Zampini 295263602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 2953674ae819SStefano Zampini vertex_size = 1; 295463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 295563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 295695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 295763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 2958a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 295963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 2960674ae819SStefano Zampini } 296163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 296263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 296363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 2964674ae819SStefano Zampini } 296563602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 2966674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 296763602bcaSStefano Zampini } else { 296863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 296963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 2970854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 297163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 297263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 297363602bcaSStefano Zampini } 297463602bcaSStefano Zampini } 2975674ae819SStefano Zampini } 2976674ae819SStefano Zampini 2977674ae819SStefano Zampini /* Setup of Graph */ 2978785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 2979a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 2980785d1243SStefano Zampini } 2981785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 2982a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2983785d1243SStefano Zampini } 2984302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 2985674ae819SStefano Zampini 2986674ae819SStefano Zampini /* Graph's connected components analysis */ 2987674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 2988674ae819SStefano Zampini 2989674ae819SStefano Zampini /* print some info to stdout */ 2990674ae819SStefano Zampini if (pcbddc->dbg_flag) { 2991302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 2992674ae819SStefano Zampini } 2993fb180af4SStefano Zampini 2994fb180af4SStefano Zampini /* mark topography has done */ 2995fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 2996674ae819SStefano Zampini PetscFunctionReturn(0); 2997674ae819SStefano Zampini } 2998674ae819SStefano Zampini 2999674ae819SStefano Zampini #undef __FUNCT__ 3000674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3001674ae819SStefano 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[]) 3002674ae819SStefano Zampini { 3003674ae819SStefano Zampini Vec local_vec,global_vec; 3004674ae819SStefano Zampini IS seqis,paris; 3005674ae819SStefano Zampini VecScatter scatter_ctx; 3006674ae819SStefano Zampini PetscScalar *array; 3007674ae819SStefano Zampini PetscInt *temp_global_dofs; 3008674ae819SStefano Zampini PetscScalar globalsum; 3009674ae819SStefano Zampini PetscInt i,j,s; 301017ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 301117ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 301217ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3013674ae819SStefano Zampini PetscBool first_found; 3014674ae819SStefano Zampini PetscErrorCode ierr; 3015674ae819SStefano Zampini 3016674ae819SStefano Zampini PetscFunctionBegin; 3017674ae819SStefano Zampini /* mpi buffers */ 3018b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3019b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3020674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 302117ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3022674ae819SStefano Zampini /* get maximum size of subset */ 3023785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3024674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3025674ae819SStefano Zampini max_local = 0; 3026b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3027674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3028674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3029674ae819SStefano Zampini } 3030674ae819SStefano Zampini } 3031b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3032674ae819SStefano Zampini max_global++; 3033674ae819SStefano Zampini max_local = 0; 3034b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3035674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3036674ae819SStefano Zampini max_local = local_dofs[i]; 3037674ae819SStefano Zampini } 3038674ae819SStefano Zampini } 3039674ae819SStefano Zampini max_local++; 3040674ae819SStefano Zampini /* allocate workspace */ 3041674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3042674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3043674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3044674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3045674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3046674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3047674ae819SStefano Zampini /* create scatter */ 3048674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3049674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3050674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3051674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3052674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3053674ae819SStefano Zampini /* init array */ 3054674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3055674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3056674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3057674ae819SStefano Zampini if (local_dofs_mult) { 3058674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3059674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3060674ae819SStefano Zampini } 3061674ae819SStefano Zampini } else { 3062674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3063674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3064674ae819SStefano Zampini } 3065674ae819SStefano Zampini } 3066674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3067674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3068674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3069674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3070674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 30715b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3072674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3073674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3074674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3075674ae819SStefano Zampini nlocals = 0; 3076674ae819SStefano Zampini first_index = -1; 3077674ae819SStefano Zampini first_found = PETSC_FALSE; 3078674ae819SStefano Zampini for (i=0;i<s;i++) { 3079b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3080674ae819SStefano Zampini first_found = PETSC_TRUE; 3081674ae819SStefano Zampini first_index = i; 3082674ae819SStefano Zampini } 30835b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3084674ae819SStefano Zampini } 3085674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3086674ae819SStefano Zampini if (!rank_prec_comm) { 3087674ae819SStefano Zampini dof_displs[0]=0; 3088674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3089674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3090674ae819SStefano Zampini } 3091674ae819SStefano Zampini } 3092674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3093674ae819SStefano Zampini if (first_found) { 3094674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3095674ae819SStefano Zampini old_index = first_index; 3096674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3097b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3098674ae819SStefano Zampini array[i] += array[old_index]; 3099674ae819SStefano Zampini old_index = i; 3100674ae819SStefano Zampini } 3101674ae819SStefano Zampini } 3102674ae819SStefano Zampini } 3103674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3104674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3105674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3106674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3107674ae819SStefano Zampini /* get global ordering of local dofs */ 31082b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3109674ae819SStefano Zampini if (local_dofs_mult) { 3110674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31115b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3112674ae819SStefano Zampini } 3113674ae819SStefano Zampini } else { 3114674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 31155b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3116674ae819SStefano Zampini } 3117674ae819SStefano Zampini } 31182b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3119674ae819SStefano Zampini /* free workspace */ 3120674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3121674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3122674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 312317ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3124674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3125674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3126674ae819SStefano Zampini PetscFunctionReturn(0); 3127674ae819SStefano Zampini } 31289a7d3425SStefano Zampini 31299a7d3425SStefano Zampini #undef __FUNCT__ 31309a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 31319a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 31329a7d3425SStefano Zampini { 31339a7d3425SStefano Zampini PetscInt i,j; 31349a7d3425SStefano Zampini PetscScalar *alphas; 31359a7d3425SStefano Zampini PetscErrorCode ierr; 31369a7d3425SStefano Zampini 31379a7d3425SStefano Zampini PetscFunctionBegin; 31389a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3139785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 31409a7d3425SStefano Zampini for (i=0;i<n;i++) { 31419a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 31429a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 31439a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 31449a7d3425SStefano Zampini } 31459a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 31469a7d3425SStefano Zampini PetscFunctionReturn(0); 31479a7d3425SStefano Zampini } 31489a7d3425SStefano Zampini 3149e7931f94SStefano Zampini #undef __FUNCT__ 315070cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 315128143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3152e7931f94SStefano Zampini { 3153e7931f94SStefano Zampini Mat subdomain_adj; 3154e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3155e7931f94SStefano Zampini MatPartitioning partitioner; 3156e7931f94SStefano Zampini Mat_IS *matis; 3157e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3158e7931f94SStefano Zampini PetscInt prank; 3159e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3160e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3161e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 31623837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3163e7931f94SStefano Zampini PetscErrorCode ierr; 31642b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3165e7931f94SStefano Zampini PetscSubcomm subcomm; 3166a57a6d2fSStefano Zampini 3167e7931f94SStefano Zampini PetscFunctionBegin; 31682b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 31692b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 31702b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3171e7931f94SStefano Zampini 3172e7931f94SStefano Zampini /* Get info on mapping */ 3173e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3174e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3175e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3176e7931f94SStefano Zampini 3177e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3178785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3179e7931f94SStefano Zampini xadj[0] = 0; 3180e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3181785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3182785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3183e7931f94SStefano Zampini 31842b510759SStefano Zampini if (threshold) { 3185d023bfaeSStefano Zampini PetscInt xadj_count = 0; 31862b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3187d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3188d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3189d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3190d023bfaeSStefano Zampini xadj_count++; 3191e7931f94SStefano Zampini } 3192e7931f94SStefano Zampini } 3193d023bfaeSStefano Zampini xadj[1] = xadj_count; 3194c8587f34SStefano Zampini } else { 3195e7931f94SStefano Zampini if (xadj[1]) { 3196e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3197e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3198c8587f34SStefano Zampini } 3199e7931f94SStefano Zampini } 3200e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3201e7931f94SStefano Zampini if (use_square) { 3202e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3203e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3204e7931f94SStefano Zampini } 3205e7931f94SStefano Zampini } 3206e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3207e7931f94SStefano Zampini 32083837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3209e7931f94SStefano Zampini 3210e7931f94SStefano Zampini /* 3211e7931f94SStefano Zampini Restrict work on active processes only. 3212e7931f94SStefano Zampini */ 3213e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3214e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3215e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 32162b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3217d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3218e7931f94SStefano Zampini if (color) { 3219e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3220e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3221e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3222c8587f34SStefano Zampini } else { 322328143c3dSStefano Zampini PetscInt coarsening_ratio; 3224306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3225785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3226e7931f94SStefano Zampini prank = rank; 3227306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 32288002ef2cSStefano Zampini /* 3229e7931f94SStefano Zampini for (i=0;i<size;i++) { 3230e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3231c8587f34SStefano Zampini } 32328002ef2cSStefano Zampini */ 3233e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3234e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3235c8587f34SStefano Zampini } 3236e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3237306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 323822b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3239e7931f94SStefano Zampini 3240e7931f94SStefano Zampini /* Partition */ 3241306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3242e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3243e7931f94SStefano Zampini if (use_vwgt) { 32443837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3245e7931f94SStefano Zampini v_wgt[0] = local_size; 3246e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3247c8587f34SStefano Zampini } 324828143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 324928143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 325028143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3251e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3252e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 325322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3254e7931f94SStefano Zampini 3255e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 325628143c3dSStefano Zampini if (contiguous) { 325728143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 325828143c3dSStefano Zampini } else { 325984ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 326028143c3dSStefano Zampini } 3261e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3262e7931f94SStefano Zampini /* clean up */ 3263e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3264e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3265e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3266e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3267e7931f94SStefano Zampini } 3268e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3269e7931f94SStefano Zampini 3270e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3271e7931f94SStefano Zampini i = 1; 3272e7931f94SStefano Zampini if (color) i=0; 3273e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3274e7931f94SStefano Zampini 3275e7931f94SStefano Zampini /* get back IS */ 3276e7931f94SStefano Zampini *is_sends = ranks_send_to; 3277e7931f94SStefano Zampini PetscFunctionReturn(0); 3278e7931f94SStefano Zampini } 3279e7931f94SStefano Zampini 3280e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3281e7931f94SStefano Zampini 3282e7931f94SStefano Zampini #undef __FUNCT__ 3283e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 328428143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3285e7931f94SStefano Zampini { 328670cf5478SStefano Zampini Mat local_mat; 3287e7931f94SStefano Zampini Mat_IS *matis; 3288e7931f94SStefano Zampini IS is_sends_internal; 32899d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 329028143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 32919d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3292e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3293e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3294e7931f94SStefano Zampini const PetscInt* is_indices; 3295e7931f94SStefano Zampini MatType new_local_type; 3296e7931f94SStefano Zampini /* buffers */ 3297e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 329828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 32999d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3300e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3301e7931f94SStefano Zampini /* MPI */ 330228143c3dSStefano Zampini MPI_Comm comm,comm_n; 330328143c3dSStefano Zampini PetscSubcomm subcomm; 3304e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 330528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 330628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 330728143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 330828143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 330928143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3310e7931f94SStefano Zampini PetscErrorCode ierr; 3311e7931f94SStefano Zampini 3312e7931f94SStefano Zampini PetscFunctionBegin; 331328143c3dSStefano Zampini /* TODO: add missing checks */ 331428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 331528143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 331628143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 331728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3318e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 331928143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3320e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3321e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3322e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3323e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3324e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 332528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 332670cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 332770cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 332828143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 332970cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 333070cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 333170cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 333270cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 333370cf5478SStefano Zampini } 3334e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3335e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3336e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3337e7931f94SStefano Zampini if (!is_sends) { 333828143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 333928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 334028143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3341c8587f34SStefano Zampini } else { 3342e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3343e7931f94SStefano Zampini is_sends_internal = is_sends; 3344c8587f34SStefano Zampini } 3345e7931f94SStefano Zampini 3346e7931f94SStefano Zampini /* get pointer of MATIS data */ 3347e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3348e7931f94SStefano Zampini 3349e7931f94SStefano Zampini /* get comm */ 3350a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3351e7931f94SStefano Zampini 3352e7931f94SStefano Zampini /* compute number of sends */ 3353e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3354e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3355e7931f94SStefano Zampini 3356e7931f94SStefano Zampini /* compute number of receives */ 3357e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3358785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3359e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3360e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3361e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3362e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3363e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3364e7931f94SStefano Zampini 336528143c3dSStefano Zampini /* restrict comm if requested */ 336628143c3dSStefano Zampini subcomm = 0; 336728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 336828143c3dSStefano Zampini if (restrict_comm) { 3369779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3370779c1cceSStefano Zampini 337128143c3dSStefano Zampini color = 0; 3372779c1cceSStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */ 337328143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 337428143c3dSStefano Zampini subcommsize = commsize - subcommsize; 337528143c3dSStefano Zampini /* check if reuse has been requested */ 337628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 337728143c3dSStefano Zampini if (*mat_n) { 337828143c3dSStefano Zampini PetscMPIInt subcommsize2; 337928143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 338028143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 338128143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 338228143c3dSStefano Zampini } else { 338328143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 338428143c3dSStefano Zampini } 338528143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3386779c1cceSStefano Zampini PetscMPIInt rank; 3387779c1cceSStefano Zampini 3388779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 338928143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 339028143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 339128143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3392306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 339328143c3dSStefano Zampini } 339428143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 339528143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 339628143c3dSStefano Zampini } else { 339728143c3dSStefano Zampini comm_n = comm; 339828143c3dSStefano Zampini } 339928143c3dSStefano Zampini 3400e7931f94SStefano Zampini /* prepare send/receive buffers */ 3401785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3402e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3403785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3404e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 340528143c3dSStefano Zampini if (nis) { 3406854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 340728143c3dSStefano Zampini } 3408e7931f94SStefano Zampini 340928143c3dSStefano Zampini /* Get data from local matrices */ 3410e7931f94SStefano Zampini if (!isdense) { 3411a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3412e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3413e7931f94SStefano Zampini /* 3414e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3415e7931f94SStefano Zampini send_buffer_idxs should contain: 3416e7931f94SStefano Zampini - MatType_PRIVATE type 3417e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3418e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3419e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3420e7931f94SStefano Zampini */ 3421e7931f94SStefano Zampini } else { 3422e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3423e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3424854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3425e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3426e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3427e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3428e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3429e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3431e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3432e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3433e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3434c8587f34SStefano Zampini } 3435c8587f34SStefano Zampini } 3436e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 343728143c3dSStefano Zampini /* additional is (if any) */ 343828143c3dSStefano Zampini if (nis) { 343928143c3dSStefano Zampini PetscMPIInt psum; 344028143c3dSStefano Zampini PetscInt j; 344128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 344228143c3dSStefano Zampini PetscInt plen; 344328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 344428143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 344528143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 344628143c3dSStefano Zampini } 3447854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 344828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 344928143c3dSStefano Zampini PetscInt plen; 345028143c3dSStefano Zampini const PetscInt *is_array_idxs; 345128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 345228143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 345328143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 345428143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 345528143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 345628143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 345728143c3dSStefano Zampini } 345828143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 345928143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 346028143c3dSStefano Zampini } 346128143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 346228143c3dSStefano Zampini } 346328143c3dSStefano Zampini 3464e7931f94SStefano Zampini buf_size_idxs = 0; 3465e7931f94SStefano Zampini buf_size_vals = 0; 346628143c3dSStefano Zampini buf_size_idxs_is = 0; 3467e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3468e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3469e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 347028143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3471e7931f94SStefano Zampini } 3472785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3473785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 347495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3475e7931f94SStefano Zampini 3476e7931f94SStefano Zampini /* get new tags for clean communications */ 3477e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3478e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 347928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3480e7931f94SStefano Zampini 3481e7931f94SStefano Zampini /* allocate for requests */ 3482785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3483785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 348495ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3485785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3486785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 348795ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3488e7931f94SStefano Zampini 3489e7931f94SStefano Zampini /* communications */ 3490e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3491e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 349228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3493e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3494e7931f94SStefano Zampini source_dest = onodes[i]; 3495e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3496e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3497e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3498e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 349928143c3dSStefano Zampini if (nis) { 350028143c3dSStefano 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); 350128143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 350228143c3dSStefano Zampini } 3503e7931f94SStefano Zampini } 3504e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3505e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3506e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3507e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 350828143c3dSStefano Zampini if (nis) { 350928143c3dSStefano 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); 351028143c3dSStefano Zampini } 3511e7931f94SStefano Zampini } 3512e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3513e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3514e7931f94SStefano Zampini 3515e7931f94SStefano Zampini /* assemble new l2g map */ 3516e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3517e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35189d30be91SStefano Zampini new_local_rows = 0; 3519e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35209d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3521e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3522e7931f94SStefano Zampini } 35239d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3524e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 35259d30be91SStefano Zampini new_local_rows = 0; 3526e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 35279d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 35289d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3529e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3530e7931f94SStefano Zampini } 35319d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 35329d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3533e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3534e7931f94SStefano Zampini 3535e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3536e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3537e7931f94SStefano 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) */ 3538e7931f94SStefano Zampini if (n_recvs) { 353928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3540e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3541e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3542e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3543e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3544e7931f94SStefano Zampini break; 3545e7931f94SStefano Zampini } 3546e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3547e7931f94SStefano Zampini } 3548e7931f94SStefano Zampini switch (new_local_type_private) { 354928143c3dSStefano Zampini case MATDENSE_PRIVATE: 355028143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3551e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3552e7931f94SStefano Zampini bs = 1; 355328143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 355428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 355528143c3dSStefano Zampini bs = 1; 355628143c3dSStefano Zampini } 3557e7931f94SStefano Zampini break; 3558e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3559e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3560e7931f94SStefano Zampini bs = 1; 3561e7931f94SStefano Zampini break; 3562e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3563e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3564e7931f94SStefano Zampini break; 3565e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3566e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3567e7931f94SStefano Zampini break; 3568e7931f94SStefano Zampini default: 35699d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3570e7931f94SStefano Zampini break; 3571e7931f94SStefano Zampini } 357228143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 357328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 357428143c3dSStefano Zampini bs = 1; 3575e7931f94SStefano Zampini } 3576e7931f94SStefano Zampini 357770cf5478SStefano Zampini /* create MATIS object if needed */ 357870cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3579e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 358028143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 358170cf5478SStefano Zampini } else { 358270cf5478SStefano Zampini /* it also destroys the local matrices */ 358370cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 358470cf5478SStefano Zampini } 358570cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3586e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 35879d30be91SStefano Zampini 35889d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 35899d30be91SStefano Zampini 35909d30be91SStefano Zampini /* Global to local map of received indices */ 35919d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 35929d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 35939d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 35949d30be91SStefano Zampini 35959d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 35969d30be91SStefano Zampini buf_size_idxs = 0; 35979d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 35989d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 35999d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 36009d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 36019d30be91SStefano Zampini } 36029d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 36039d30be91SStefano Zampini 36049d30be91SStefano Zampini /* set preallocation */ 36059d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 36069d30be91SStefano Zampini if (!newisdense) { 36079d30be91SStefano Zampini PetscInt *new_local_nnz=0; 36089d30be91SStefano Zampini 36099d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 36109d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 36119d30be91SStefano Zampini if (n_recvs) { 36129d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 36139d30be91SStefano Zampini } 36149d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 36159d30be91SStefano Zampini PetscInt j; 36169d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 36179d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 36189d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 36199d30be91SStefano Zampini } 36209d30be91SStefano Zampini } else { 36219d30be91SStefano Zampini /* TODO */ 36229d30be91SStefano Zampini } 36239d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 36249d30be91SStefano Zampini } 36259d30be91SStefano Zampini if (new_local_nnz) { 36269d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 36279d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 36289d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 36299d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36309d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 36319d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 36329d30be91SStefano Zampini } else { 36339d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36349d30be91SStefano Zampini } 36359d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 36369d30be91SStefano Zampini } else { 36379d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 36389d30be91SStefano Zampini } 3639e7931f94SStefano Zampini 3640e7931f94SStefano Zampini /* set values */ 3641e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 36429d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3643e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3644e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3645e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 36469d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3647e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3648e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3649e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 365028143c3dSStefano Zampini } else { 365128143c3dSStefano Zampini /* TODO */ 3652e7931f94SStefano Zampini } 3653e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3654e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3655e7931f94SStefano Zampini } 3656e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3657e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 365870cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 365970cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 36609d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 36619d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3662e7931f94SStefano Zampini 3663dfd14d43SStefano Zampini #if 0 366428143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3665e7931f94SStefano Zampini Vec lvec,rvec; 3666e7931f94SStefano Zampini PetscReal infty_error; 3667e7931f94SStefano Zampini 36682a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3669e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3670e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3671e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 367270cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3673e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3674e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3675e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3676e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3677e7931f94SStefano Zampini } 367828143c3dSStefano Zampini #endif 3679e7931f94SStefano Zampini 368028143c3dSStefano Zampini /* assemble new additional is (if any) */ 368128143c3dSStefano Zampini if (nis) { 368228143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 368328143c3dSStefano Zampini 368428143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3685854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 368628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 368728143c3dSStefano Zampini psum = 0; 368828143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 368928143c3dSStefano Zampini for (j=0;j<nis;j++) { 369028143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 369128143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 369228143c3dSStefano Zampini psum += plen; 369328143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 369428143c3dSStefano Zampini } 369528143c3dSStefano Zampini } 3696854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3697854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 369828143c3dSStefano Zampini for (i=1;i<nis;i++) { 369928143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 370028143c3dSStefano Zampini } 370128143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 370228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 370328143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 370428143c3dSStefano Zampini for (j=0;j<nis;j++) { 370528143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 370628143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 370728143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 370828143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 370928143c3dSStefano Zampini } 371028143c3dSStefano Zampini } 371128143c3dSStefano Zampini for (i=0;i<nis;i++) { 371228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 371328143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 371428143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 371528143c3dSStefano Zampini } 371628143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 371728143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 371828143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 371928143c3dSStefano Zampini } 3720e7931f94SStefano Zampini /* free workspace */ 372128143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3722e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3723e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3724e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3725e7931f94SStefano Zampini if (isdense) { 3726e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3727e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3728e7931f94SStefano Zampini } else { 3729e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3730e7931f94SStefano Zampini } 373128143c3dSStefano Zampini if (nis) { 373228143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 373328143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 373428143c3dSStefano Zampini } 3735e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3736e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 373728143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3738e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3739e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 374028143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3741e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3742e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3743e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3744e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3745e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 374628143c3dSStefano Zampini if (nis) { 374728143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 374828143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 374928143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 375028143c3dSStefano Zampini } 375128143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 375228143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 375328143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 375428143c3dSStefano Zampini for (i=0;i<nis;i++) { 375528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 375628143c3dSStefano Zampini } 375728143c3dSStefano Zampini } 3758e7931f94SStefano Zampini PetscFunctionReturn(0); 3759e7931f94SStefano Zampini } 3760a57a6d2fSStefano Zampini 376112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3762af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 376312edc857SStefano Zampini 3764c8587f34SStefano Zampini #undef __FUNCT__ 3765c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3766c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3767c8587f34SStefano Zampini { 3768c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3769c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 377020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 37719881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 377220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 37736e683305SStefano Zampini IS coarse_is,*isarray; 37746e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 37756e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3776f9eb5b7dSStefano Zampini PC pc_temp; 3777c8587f34SStefano Zampini PCType coarse_pc_type; 3778c8587f34SStefano Zampini KSPType coarse_ksp_type; 3779f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 37804f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 37816e683305SStefano Zampini Mat t_coarse_mat_is; 37826e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 37836e683305SStefano Zampini PetscMPIInt all_procs; 378474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 378568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 378622bc73bbSStefano Zampini PetscScalar *array; 37879881197aSStefano Zampini PetscErrorCode ierr; 3788fdc09c96SStefano Zampini 3789c8587f34SStefano Zampini PetscFunctionBegin; 3790c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 379168457ee5SStefano 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 */ 3792fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 37935a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3794fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3795f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3796f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3797f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3798fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3799fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3800727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3801fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3802fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3803fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3804f4ddd8eeSStefano Zampini } 3805fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3806fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3807f4ddd8eeSStefano Zampini } 380870cf5478SStefano Zampini /* reset any subassembling information */ 380970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 38106e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 38116e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3812fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3813f4ddd8eeSStefano Zampini } 3814c8587f34SStefano Zampini 38156e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 38162b510759SStefano Zampini im_active = !!(pcis->n); 38172b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 38186e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 38196e683305SStefano Zampini void_procs = all_procs-active_procs; 38206e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 382174e2c79eSStefano Zampini redist = PETSC_FALSE; 382222bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 38236e683305SStefano Zampini csin_ml = PETSC_TRUE; 38246e683305SStefano Zampini ncoarse_ml = void_procs; 3825779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3826779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 382718a45a71SStefano Zampini csin_ds = PETSC_TRUE; 382818a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 382918a45a71SStefano Zampini redist = PETSC_TRUE; 383018a45a71SStefano Zampini } else { 38316e683305SStefano Zampini csin_ds = PETSC_TRUE; 3832779c1cceSStefano Zampini ncoarse_ds = active_procs; 3833779c1cceSStefano Zampini redist = PETSC_TRUE; 383418a45a71SStefano Zampini } 38356e683305SStefano Zampini } else { 38366e683305SStefano Zampini csin_ml = PETSC_FALSE; 38376e683305SStefano Zampini ncoarse_ml = all_procs; 38386e683305SStefano Zampini if (void_procs) { 38396e683305SStefano Zampini csin_ds = PETSC_TRUE; 38406e683305SStefano Zampini ncoarse_ds = void_procs; 38416e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 38426e683305SStefano Zampini } else { 3843779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 384474e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 384574e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 384674e2c79eSStefano Zampini redist = PETSC_TRUE; 384774e2c79eSStefano Zampini } else { 38486e683305SStefano Zampini csin_ds = PETSC_FALSE; 38496e683305SStefano Zampini ncoarse_ds = all_procs; 38506e683305SStefano Zampini } 38516e683305SStefano Zampini } 385274e2c79eSStefano Zampini } 38536e683305SStefano Zampini 38546e683305SStefano Zampini /* 38556e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 38566e683305SStefano Zampini - we have not exceeded the number of levels requested 38576e683305SStefano Zampini - we can actually subassemble the active processes 38586e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 38596e683305SStefano Zampini */ 38606e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 38616e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 38626e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 38636e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 38646e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3865f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 38662b510759SStefano Zampini } else { 3867f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3868c8587f34SStefano Zampini } 3869c8587f34SStefano Zampini } 38706e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 38716e683305SStefano Zampini if (multilevel_allowed) { 38726e683305SStefano Zampini ncoarse = ncoarse_ml; 38736e683305SStefano Zampini csin = csin_ml; 387458da7f69SStefano Zampini redist = PETSC_FALSE; 38756e683305SStefano Zampini } else { 38766e683305SStefano Zampini ncoarse = ncoarse_ds; 38776e683305SStefano Zampini csin = csin_ds; 38786e683305SStefano Zampini } 3879e7931f94SStefano Zampini 3880abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3881abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3882abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3883abbbba34SStefano Zampini 3884abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 388522bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 388622bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 388722bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 388822bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3889b9b85e73SStefano Zampini #if 0 3890b9b85e73SStefano Zampini { 3891b9b85e73SStefano Zampini PetscViewer viewer; 3892b9b85e73SStefano Zampini char filename[256]; 3893b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 3894b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3895b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3896b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 3897b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3898b9b85e73SStefano Zampini } 3899b9b85e73SStefano Zampini #endif 39006e683305SStefano 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); 39016e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 39026e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 39036e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3904abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 3905abbbba34SStefano Zampini 39066e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 39076e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 39086e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 39096e683305SStefano Zampini const PetscInt *idxs; 39106e683305SStefano Zampini ISLocalToGlobalMapping tmap; 39116e683305SStefano Zampini 39126e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 39130be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 39146e683305SStefano Zampini /* allocate space for temporary storage */ 3915854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 3916854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 39176e683305SStefano Zampini /* allocate for IS array */ 39186e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 39196e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 39206e683305SStefano Zampini nis = nisdofs + nisneu; 3921854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 39226e683305SStefano Zampini /* dofs splitting */ 39236e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 39246e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 39256e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 39266e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39276e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39286e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 39296e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39306e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 39316e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 39326e683305SStefano Zampini } 39336e683305SStefano Zampini /* neumann boundaries */ 39346e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 39356e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 39366e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 39376e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39386e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 39396e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 39406e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 39416e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 39426e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 39436e683305SStefano Zampini } 39446e683305SStefano Zampini /* free memory */ 39456e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 39466e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 39476e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 39486e683305SStefano Zampini } else { 39496e683305SStefano Zampini nis = 0; 39506e683305SStefano Zampini nisdofs = 0; 39516e683305SStefano Zampini nisneu = 0; 39526e683305SStefano Zampini isarray = NULL; 39536e683305SStefano Zampini } 39546e683305SStefano Zampini /* destroy no longer needed map */ 39556e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 39566e683305SStefano Zampini 39576e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 39586e683305SStefano Zampini coarse_mat_is = NULL; 39596e683305SStefano Zampini if (csin) { 39606e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 396174e2c79eSStefano Zampini if (redist) { 396274e2c79eSStefano Zampini PetscMPIInt rank; 3963779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 396474e2c79eSStefano Zampini 396574e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 396658da7f69SStefano Zampini spc = active_procs/ncoarse; 396758da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 3968779c1cceSStefano Zampini if (im_active) { 3969779c1cceSStefano Zampini destsize = 1; 397074e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 397174e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 397274e2c79eSStefano Zampini } else { 397374e2c79eSStefano Zampini dest[0] = rank/(spc+1); 397474e2c79eSStefano Zampini } 397574e2c79eSStefano Zampini } else { 3976779c1cceSStefano Zampini destsize = 0; 39776e683305SStefano Zampini } 3978779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 3979779c1cceSStefano Zampini } else if (csin_type_simple) { 39806e683305SStefano Zampini PetscMPIInt rank; 39816e683305SStefano Zampini PetscInt issize,isidx; 3982779c1cceSStefano Zampini 39836e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 39846e683305SStefano Zampini if (im_active) { 39856e683305SStefano Zampini issize = 1; 39866e683305SStefano Zampini isidx = (PetscInt)rank; 39876e683305SStefano Zampini } else { 39886e683305SStefano Zampini issize = 0; 39896e683305SStefano Zampini isidx = -1; 39906e683305SStefano Zampini } 39916e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 3992779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 39936e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39946e683305SStefano Zampini } 3995779c1cceSStefano Zampini 3996779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 3997779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 3998779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 3999779c1cceSStefano Zampini PetscInt *coarse_candidates; 4000779c1cceSStefano Zampini const PetscInt* tisindices; 4001779c1cceSStefano Zampini 4002779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4003779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4004779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4005779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4006779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4007779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4008779c1cceSStefano Zampini } 4009779c1cceSStefano Zampini } 4010779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4011779c1cceSStefano Zampini 4012779c1cceSStefano Zampini 40136e683305SStefano Zampini if (pcbddc->dbg_flag) { 40146e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 40166e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4018779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 40196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 40206e683305SStefano Zampini } 40216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 40226e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40236e683305SStefano Zampini } 40246e683305SStefano Zampini /* shift the pattern on coarse candidates */ 40256e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 40266e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4027854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 40286e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 40296e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 40306e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 40316e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 40326e683305SStefano Zampini } 40336e683305SStefano Zampini if (pcbddc->dbg_flag) { 40346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 40366e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 40376e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40386e683305SStefano Zampini } 4039779c1cceSStefano Zampini } 40406e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 40416e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 40426e683305SStefano Zampini } else { 40436e683305SStefano Zampini if (pcbddc->dbg_flag) { 40446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40456e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 40466e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40476e683305SStefano Zampini } 40486e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 40496e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 40506e683305SStefano Zampini } 40516e683305SStefano Zampini 40526e683305SStefano Zampini /* create local to global scatters for coarse problem */ 405368457ee5SStefano Zampini if (compute_vecs) { 40546e683305SStefano Zampini PetscInt lrows; 40556e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 40566e683305SStefano Zampini if (coarse_mat_is) { 40576e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 40586e683305SStefano Zampini } else { 40596e683305SStefano Zampini lrows = 0; 40606e683305SStefano Zampini } 40616e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 40626e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 40636e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 40646e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 40656e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 40666e683305SStefano Zampini } 40676e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 40686e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4069c8587f34SStefano Zampini 4070f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4071f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4072f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4073f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4074f9eb5b7dSStefano Zampini } else { 4075f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4076f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4077c8587f34SStefano Zampini } 4078c8587f34SStefano Zampini 40796e683305SStefano Zampini /* print some info if requested */ 40806e683305SStefano Zampini if (pcbddc->dbg_flag) { 40816e683305SStefano Zampini if (!multilevel_allowed) { 40826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40836e683305SStefano Zampini if (multilevel_requested) { 40846e683305SStefano 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); 40856e683305SStefano Zampini } else if (pcbddc->max_levels) { 40866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 40876e683305SStefano Zampini } 40886e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40896e683305SStefano Zampini } 40906e683305SStefano Zampini } 40916e683305SStefano Zampini 4092f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 40936e683305SStefano Zampini if (coarse_mat_is) { 40946e683305SStefano Zampini MatReuse coarse_mat_reuse; 40956a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 40966e683305SStefano Zampini if (pcbddc->dbg_flag) { 40976e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 40986e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 40996e683305SStefano Zampini } 4100f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4101312be037SStefano Zampini char prefix[256],str_level[16]; 4102e604994aSStefano Zampini size_t len; 41036e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4104c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4105f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 41065f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4107c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 41086e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4109c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4110c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 41113e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4112e604994aSStefano Zampini /* prefix */ 4113e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4114e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4115e604994aSStefano Zampini if (!pcbddc->current_level) { 4116e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4117e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4118c8587f34SStefano Zampini } else { 4119e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4120312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4121312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 412234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4123312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4124e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4125e604994aSStefano Zampini } 4126e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 41273e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41283e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 41293e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 41303e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4131f9eb5b7dSStefano Zampini /* allow user customization */ 4132f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 41333e3c6dadSStefano Zampini } 41343e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 41353e3c6dadSStefano Zampini if (nisdofs) { 41363e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 41373e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 41383e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 41393e3c6dadSStefano Zampini } 41403e3c6dadSStefano Zampini } 41413e3c6dadSStefano Zampini if (nisneu) { 41423e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 41433e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4144312be037SStefano Zampini } 4145f9eb5b7dSStefano Zampini 4146f9eb5b7dSStefano Zampini /* get some info after set from options */ 4147f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4148f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4149f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 41504f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 41516e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4152f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4153f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4154f9eb5b7dSStefano Zampini } 41554f3a063dSStefano Zampini if (isredundant) { 41564f3a063dSStefano Zampini KSP inner_ksp; 41574f3a063dSStefano Zampini PC inner_pc; 41584f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 41594f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 41604f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 41614f3a063dSStefano Zampini } 4162f9eb5b7dSStefano Zampini 4163f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4164fa7f1dd8SStefano Zampini if (coarse_reuse) { 416581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4166fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 41676e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4168fa7f1dd8SStefano Zampini } else { 41696e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4170fa7f1dd8SStefano Zampini } 4171c8587f34SStefano Zampini if (isbddc || isnn) { 417222bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 417370cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 417428143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 417522b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 41766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 41786e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 41796e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 418022b6e8a2SStefano Zampini } 418170cf5478SStefano Zampini } 41826e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 418370cf5478SStefano Zampini } else { 418422bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 418522bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 418622bc73bbSStefano Zampini } 418722bc73bbSStefano Zampini } else { 41882e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4189c8587f34SStefano Zampini } 4190c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4191c8587f34SStefano Zampini 4192c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4193b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 41945a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4195c8587f34SStefano Zampini 41966e683305SStefano Zampini /* set operators */ 41975f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 41986e683305SStefano Zampini if (pcbddc->dbg_flag) { 41996e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42006e683305SStefano Zampini } 42016e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 42026e683305SStefano Zampini coarse_mat = 0; 42036e683305SStefano Zampini } 42046e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4205b9b85e73SStefano Zampini #if 0 4206b9b85e73SStefano Zampini { 4207b9b85e73SStefano Zampini PetscViewer viewer; 4208b9b85e73SStefano Zampini char filename[256]; 4209b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4210b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4211b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4212b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4213b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4214b9b85e73SStefano Zampini } 4215b9b85e73SStefano Zampini #endif 4216c8587f34SStefano Zampini 4217c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4218c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4219c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 422098a51de6SStefano Zampini } 422198a51de6SStefano Zampini 422298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 422398a51de6SStefano Zampini Vec crhs,csol; 422498a51de6SStefano Zampini PetscBool ispreonly; 422598a51de6SStefano Zampini if (CoarseNullSpace) { 4226c8587f34SStefano Zampini if (isbddc) { 4227c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4228c8587f34SStefano Zampini } else { 4229c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4230c8587f34SStefano Zampini } 4231c8587f34SStefano Zampini } 4232f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4233f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4234f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4235f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 42366e683305SStefano Zampini /* hack */ 4237f347579bSStefano Zampini if (!csol) { 42382a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4239f9eb5b7dSStefano Zampini } 4240f347579bSStefano Zampini if (!crhs) { 42412a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4242f347579bSStefano Zampini } 4243cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4244cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 42456e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4246c8587f34SStefano Zampini KSP check_ksp; 42472b510759SStefano Zampini KSPType check_ksp_type; 4248c8587f34SStefano Zampini PC check_pc; 42496e683305SStefano Zampini Vec check_vec,coarse_vec; 42506a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 42512b510759SStefano Zampini PetscInt its; 42526e683305SStefano Zampini PetscBool compute_eigs; 42536e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 42546e683305SStefano Zampini PetscInt neigs; 42558e185a42SStefano Zampini const char *prefix; 4256c8587f34SStefano Zampini 42572b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 42586e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 425923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4260f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 42612b510759SStefano Zampini if (ispreonly) { 42622b510759SStefano Zampini check_ksp_type = KSPPREONLY; 42636e683305SStefano Zampini compute_eigs = PETSC_FALSE; 42642b510759SStefano Zampini } else { 4265cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 42666e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4267c8587f34SStefano Zampini } 4268c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 42696e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 42706e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 42716e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4272a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4273a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4274a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4275a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4276c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4277c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4278c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4279c8587f34SStefano Zampini /* create random vec */ 42806e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 42816e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4282c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4283c8587f34SStefano Zampini if (CoarseNullSpace) { 4284c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4285c8587f34SStefano Zampini } 42866e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4287c8587f34SStefano Zampini /* solve coarse problem */ 42886e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4289c8587f34SStefano Zampini if (CoarseNullSpace) { 42906e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4291c8587f34SStefano Zampini } 4292cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 42936e683305SStefano Zampini if (compute_eigs) { 4294854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4295854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 42966e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 42976e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 42986e683305SStefano Zampini lambda_min = eigs_r[0]; 42996e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 43006e683305SStefano Zampini if (lambda_max>lambda_min) { 4301cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4302cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4303cbcc2c2aSStefano Zampini } 4304c8587f34SStefano Zampini } 4305c8587f34SStefano Zampini } 4306cbcc2c2aSStefano Zampini 4307c8587f34SStefano Zampini /* check coarse problem residual error */ 43086e683305SStefano Zampini if (pcbddc->dbg_flag) { 43096e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 43106e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43116e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4312c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 43136e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 43146e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4315c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4316779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 43176e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 43186e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 43196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 43206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 43216e683305SStefano Zampini if (compute_eigs) { 43226e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4323deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4324c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 43256e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 43266e683305SStefano 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); 43276e683305SStefano Zampini for (i=0;i<neigs;i++) { 43286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4329c8587f34SStefano Zampini } 43306e683305SStefano Zampini } 43316e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 43326e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 43336e683305SStefano Zampini } 4334c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 43356e683305SStefano Zampini if (compute_eigs) { 43366e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 43376e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4338c8587f34SStefano Zampini } 43396e683305SStefano Zampini } 43406e683305SStefano Zampini } 4341cbcc2c2aSStefano Zampini /* print additional info */ 4342cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 43436e683305SStefano Zampini /* waits until all processes reaches this point */ 43446e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4345cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4346cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4347cbcc2c2aSStefano Zampini } 4348cbcc2c2aSStefano Zampini 43492b510759SStefano Zampini /* free memory */ 4350c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4351fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4352c8587f34SStefano Zampini PetscFunctionReturn(0); 4353c8587f34SStefano Zampini } 4354674ae819SStefano Zampini 4355f34684f1SStefano Zampini #undef __FUNCT__ 4356f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4357f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4358f34684f1SStefano Zampini { 4359f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4360f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4361f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 436273be2a3aSStefano Zampini PetscInt i,coarse_size=0; 436373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4364f34684f1SStefano Zampini PetscErrorCode ierr; 4365f34684f1SStefano Zampini 4366f34684f1SStefano Zampini PetscFunctionBegin; 4367f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 436889c96988SStefano Zampini if (!pcbddc->primal_indices_local_idxs && pcbddc->local_primal_size) { 436989c96988SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Local primal indices have not been created"); 4370727cdba6SStefano Zampini } 4371727cdba6SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,NULL,&coarse_size,&local_primal_indices);CHKERRQ(ierr); 4372f34684f1SStefano Zampini 4373f34684f1SStefano Zampini /* check numbering */ 4374f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4375f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4376b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4377f34684f1SStefano Zampini 4378f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4379f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4380f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 43810fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4382f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4383f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4384727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4385f34684f1SStefano Zampini } 4386f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4387f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4388f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4389f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4390f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4391f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4392f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4393f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4394f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4395f34684f1SStefano Zampini if (array[i] == 1.0) { 4396b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4397f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4398f34684f1SStefano Zampini } 4399f34684f1SStefano Zampini } 4400b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4401f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4402f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4403f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4404f34684f1SStefano Zampini } 4405f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4406f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4407f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4408f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4409f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4410f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4411b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4412ca8b9ea9SStefano Zampini PetscInt *gidxs; 4413ca8b9ea9SStefano Zampini 4414ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4415ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4416f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4417f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4418f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4419f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 44204bc2dc4bSStefano 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); 4421f34684f1SStefano Zampini } 4422f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4423ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4424f34684f1SStefano Zampini } 4425f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4426302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4427f34684f1SStefano Zampini } 44288bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4429f34684f1SStefano Zampini /* get back data */ 4430f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4431f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4432674ae819SStefano Zampini PetscFunctionReturn(0); 4433674ae819SStefano Zampini } 4434674ae819SStefano Zampini 4435e456f2a8SStefano Zampini #undef __FUNCT__ 4436e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4437a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4438e456f2a8SStefano Zampini { 4439e456f2a8SStefano Zampini IS localis_t; 4440a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4441e456f2a8SStefano Zampini PetscScalar *vals; 4442e456f2a8SStefano Zampini PetscErrorCode ierr; 4443e456f2a8SStefano Zampini 4444e456f2a8SStefano Zampini PetscFunctionBegin; 4445a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4446e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4447854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4448e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4449e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4450a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4451a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 44521035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4453a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 44541035eff8SStefano Zampini } 4455a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4456e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4457e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4458a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4459a7dc3881SStefano Zampini /* now compute set in local ordering */ 4460a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4461a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4462a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4463a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4464a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4465ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4466e456f2a8SStefano Zampini lsize++; 4467e456f2a8SStefano Zampini } 4468e456f2a8SStefano Zampini } 4469854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4470a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4471ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4472e456f2a8SStefano Zampini idxs[lsize++] = i; 4473e456f2a8SStefano Zampini } 4474e456f2a8SStefano Zampini } 4475a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4476a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4477e456f2a8SStefano Zampini *localis = localis_t; 4478e456f2a8SStefano Zampini PetscFunctionReturn(0); 4479e456f2a8SStefano Zampini } 4480906d46d4SStefano Zampini 4481906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4482906d46d4SStefano Zampini #undef __FUNCT__ 4483906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4484906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4485906d46d4SStefano Zampini { 4486906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4487906d46d4SStefano Zampini PetscErrorCode ierr; 4488906d46d4SStefano Zampini 4489906d46d4SStefano Zampini PetscFunctionBegin; 4490906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4491906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4492906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4493906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4494906d46d4SStefano Zampini PetscFunctionReturn(0); 4495906d46d4SStefano Zampini } 4496906d46d4SStefano Zampini 4497906d46d4SStefano Zampini #undef __FUNCT__ 4498906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4499906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4500906d46d4SStefano Zampini { 4501906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4502906d46d4SStefano Zampini PetscErrorCode ierr; 4503906d46d4SStefano Zampini 4504906d46d4SStefano Zampini PetscFunctionBegin; 4505906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4506906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4507906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4508906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4509906d46d4SStefano Zampini PetscFunctionReturn(0); 4510906d46d4SStefano Zampini } 4511b96c3477SStefano Zampini 4512b96c3477SStefano Zampini #undef __FUNCT__ 4513b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 451408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4515b96c3477SStefano Zampini { 4516a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4517b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4518b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4519a64f4aa4SStefano Zampini Mat S_j; 4520b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4521b96c3477SStefano Zampini PetscBool free_used_adj; 4522b96c3477SStefano Zampini PetscErrorCode ierr; 4523b96c3477SStefano Zampini 4524b96c3477SStefano Zampini PetscFunctionBegin; 4525b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4526b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 452708122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4528b96c3477SStefano Zampini used_xadj = NULL; 4529b96c3477SStefano Zampini used_adjncy = NULL; 4530b96c3477SStefano Zampini } else { 453108122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 453208122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 453308122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 453408122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4535b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4536b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4537b96c3477SStefano Zampini } else { 45382fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4539b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4540b96c3477SStefano Zampini PetscInt nvtxs; 4541b96c3477SStefano Zampini 45422fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 45432fffb893SStefano Zampini if (flg_row) { 4544b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4545b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4546b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4547b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 45482fffb893SStefano Zampini } else { 45492fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 45502fffb893SStefano Zampini used_xadj = NULL; 45512fffb893SStefano Zampini used_adjncy = NULL; 45522fffb893SStefano Zampini } 45532fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4554b96c3477SStefano Zampini } 4555b96c3477SStefano Zampini } 4556d5574798SStefano Zampini 4557d5574798SStefano Zampini /* setup sub_schurs data */ 4558a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4559a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4560a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4561a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4562a64f4aa4SStefano 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); 4563a64f4aa4SStefano Zampini } else { 4564a64f4aa4SStefano 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); 4565a64f4aa4SStefano Zampini } 4566a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4567b96c3477SStefano Zampini 4568b96c3477SStefano Zampini /* free adjacency */ 4569b96c3477SStefano Zampini if (free_used_adj) { 4570b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4571b96c3477SStefano Zampini } 4572b96c3477SStefano Zampini PetscFunctionReturn(0); 4573b96c3477SStefano Zampini } 4574b96c3477SStefano Zampini 4575b96c3477SStefano Zampini #undef __FUNCT__ 4576b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 457708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4578b96c3477SStefano Zampini { 4579b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4580b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4581b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4582b96c3477SStefano Zampini PCBDDCGraph graph; 4583b96c3477SStefano Zampini PetscErrorCode ierr; 4584b96c3477SStefano Zampini 4585b96c3477SStefano Zampini PetscFunctionBegin; 4586b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 458708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4588b96c3477SStefano Zampini IS verticesIS; 4589b96c3477SStefano Zampini 4590b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4591b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 45927fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4593b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4594b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4595b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4596b96c3477SStefano Zampini /* 4597b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4598b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4599b96c3477SStefano Zampini } 4600b96c3477SStefano Zampini */ 4601b96c3477SStefano Zampini } else { 4602b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4603b96c3477SStefano Zampini } 4604b96c3477SStefano Zampini 4605b96c3477SStefano Zampini /* sub_schurs init */ 4606a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4607a64f4aa4SStefano Zampini 4608b96c3477SStefano Zampini /* free graph struct */ 460908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4610b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4611b96c3477SStefano Zampini } 4612b96c3477SStefano Zampini PetscFunctionReturn(0); 4613b96c3477SStefano Zampini } 4614