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 PetscScalar *Smult,*Seigv; 2008122e43SStefano Zampini PetscReal *eigs,thresh; 211b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 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 } 319ab7bb16SStefano Zampini if (!sub_schurs->is_hermitian || !sub_schurs->is_posdef) { 3208122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Not yet implemented"); 3308122e43SStefano Zampini } 3408122e43SStefano Zampini 35fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 36fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 38fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 39fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 40fd14bc51SStefano Zampini } 41fd14bc51SStefano Zampini 4208122e43SStefano Zampini /* max size of subsets */ 4308122e43SStefano Zampini mss = 0; 4408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 4608122e43SStefano Zampini PetscInt subset_size; 4708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4808122e43SStefano Zampini mss = PetscMax(mss,subset_size); 4908122e43SStefano Zampini } 5008122e43SStefano Zampini } 5108122e43SStefano Zampini 5208122e43SStefano Zampini /* min/max and threshold */ 5308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 5418a45a71SStefano Zampini nmin = pcbddc->adaptive_nmin > -1 ? pcbddc->adaptive_nmin : 1; 5508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 5608122e43SStefano Zampini if (pcbddc->adaptive_threshold > 1.0) { 5708122e43SStefano Zampini thresh = 1.0/pcbddc->adaptive_threshold; 5808122e43SStefano Zampini } else { 5908122e43SStefano Zampini thresh = 1.0; 6008122e43SStefano Zampini } 6108122e43SStefano Zampini 6208122e43SStefano Zampini /* allocate lapack workspace */ 6308122e43SStefano Zampini cum = cum2 = 0; 6408122e43SStefano Zampini maxneigs = 0; 6508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 6708122e43SStefano Zampini PetscInt n,subset_size; 6808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6908122e43SStefano Zampini n = PetscMin(subset_size,nmax); 7008122e43SStefano Zampini cum += subset_size*n; 7108122e43SStefano Zampini cum2 += n; 7208122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7308122e43SStefano Zampini } 7408122e43SStefano Zampini } 7508122e43SStefano Zampini 7608122e43SStefano Zampini if (mss) { 779ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7808122e43SStefano Zampini PetscBLASInt B_itype = 1; 7908122e43SStefano Zampini PetscBLASInt B_N = mss; 804c6709b3SStefano Zampini PetscReal zero = 0.0; 814c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8208122e43SStefano Zampini 8308122e43SStefano Zampini B_lwork = -1; 8408122e43SStefano Zampini S = NULL; 8508122e43SStefano Zampini St = NULL; 86a58a30b4SStefano Zampini eigs = NULL; 87a58a30b4SStefano Zampini eigv = NULL; 88a58a30b4SStefano Zampini B_iwork = NULL; 89a58a30b4SStefano Zampini B_ifail = NULL; 9008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9208122e43SStefano 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)); 9308122e43SStefano Zampini #else 9408122e43SStefano 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)); 9508122e43SStefano Zampini #endif 9608122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 9708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 9808122e43SStefano Zampini } else { 9908122e43SStefano Zampini /* TODO */ 10008122e43SStefano Zampini } 10108122e43SStefano Zampini } else { 10208122e43SStefano Zampini lwork = 0; 10308122e43SStefano Zampini } 10408122e43SStefano Zampini 10508122e43SStefano Zampini nv = 0; 10608122e43SStefano Zampini if (sub_schurs->is_Ej_com) { /* complement of subsets, each entry is a vertex */ 10708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&nv);CHKERRQ(ierr); 10808122e43SStefano Zampini } 1094c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 11008122e43SStefano Zampini ierr = PetscMalloc7(mss*mss,&S,mss*mss,&St,mss*mss,&eigv,mss,&eigs, 11108122e43SStefano Zampini B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11308122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 11408122e43SStefano Zampini #endif 11508122e43SStefano Zampini ierr = PetscMalloc2(mss*mss,&Smult,mss*mss,&Seigv);CHKERRQ(ierr); 11608122e43SStefano Zampini ierr = PetscMalloc4(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 11708122e43SStefano Zampini nv+cum2+1,&pcbddc->adaptive_constraints_ptrs, 11808122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 11908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12008122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12108122e43SStefano Zampini 12208122e43SStefano Zampini maxneigs = 0; 12308122e43SStefano Zampini cum = cum2 = cumarray = 0; 12408122e43SStefano Zampini if (sub_schurs->is_Ej_com) { 12508122e43SStefano Zampini const PetscInt *idxs; 12608122e43SStefano Zampini 12708122e43SStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_com,&idxs);CHKERRQ(ierr); 12808122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 12908122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 13108122e43SStefano Zampini pcbddc->adaptive_constraints_ptrs[cum] = cum; 13208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13308122e43SStefano Zampini } 13408122e43SStefano Zampini cum2 = cum; 13508122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_com,&idxs);CHKERRQ(ierr); 13608122e43SStefano Zampini } 13708122e43SStefano Zampini 13808122e43SStefano Zampini if (mss) { /* multilevel */ 13908122e43SStefano Zampini if (pcbddc->use_deluxe_scaling) { 14008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 14108122e43SStefano Zampini } else { 14208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&Sarray);CHKERRQ(ierr); 14308122e43SStefano Zampini } 14408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 14508122e43SStefano Zampini } 14608122e43SStefano Zampini 147fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 148fd14bc51SStefano 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); 149fd14bc51SStefano Zampini } 150fd14bc51SStefano Zampini 15108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15208122e43SStefano Zampini PetscInt j,subset_size; 15308122e43SStefano Zampini 15408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 15508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 15608122e43SStefano Zampini const PetscInt *idxs; 1574c6709b3SStefano Zampini PetscScalar one = 1.0,scalar_zero = 0.0; 1584c6709b3SStefano Zampini PetscReal zero=0.0; 15908122e43SStefano Zampini PetscBLASInt B_N; 16008122e43SStefano Zampini 16108122e43SStefano Zampini /* S should be copied since we need it for deluxe scaling */ 1629ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 16308122e43SStefano Zampini PetscInt j; 16408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 16508122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 16608122e43SStefano Zampini } 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 16808122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 16908122e43SStefano Zampini } 17008122e43SStefano Zampini } else { 17108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17308122e43SStefano Zampini } 17408122e43SStefano Zampini /* is always this the right matrix? */ 17508122e43SStefano Zampini ierr = PetscMemcpy(Smult,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17608122e43SStefano Zampini 17708122e43SStefano Zampini /* we could reuse space already allocated when building sum_S_Ej_tilda_all */ 17808122e43SStefano Zampini /* St = Starray+cumarray; */ 17908122e43SStefano Zampini 1809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 18108122e43SStefano Zampini PetscBLASInt B_itype = 1; 18208122e43SStefano Zampini PetscBLASInt B_IL = 1, B_IU; 1834c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1849552c7c7SStefano Zampini PetscInt nmin_s; 18508122e43SStefano Zampini 18608122e43SStefano Zampini /* ask for eigenvalues lower than thresh */ 187fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 188fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d.\n",i,sub_schurs->n_subs); 189fd14bc51SStefano Zampini } 19008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 19108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 19208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 19308122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 19408122e43SStefano Zampini #else 19508122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 19608122e43SStefano Zampini #endif 19708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 19808122e43SStefano Zampini if (B_ierr) { 19908122e43SStefano Zampini if (B_ierr < 0 ) { 20008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 20108122e43SStefano Zampini } else if (B_ierr <= B_N) { 20208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 20308122e43SStefano Zampini } else { 2049552c7c7SStefano 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); 20508122e43SStefano Zampini } 20608122e43SStefano Zampini } 20708122e43SStefano Zampini 20808122e43SStefano Zampini if (B_neigs > nmax) { 209fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 210fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 211fd14bc51SStefano Zampini } 21208122e43SStefano Zampini B_neigs = nmax; 21308122e43SStefano Zampini } 21408122e43SStefano Zampini 2159552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2169552c7c7SStefano Zampini if (B_neigs < nmin_s) { 21708122e43SStefano Zampini PetscBLASInt B_neigs2; 21808122e43SStefano Zampini 21908122e43SStefano Zampini B_IL = B_neigs + 1; 2209552c7c7SStefano Zampini ierr = PetscBLASIntCast(nmin_s,&B_IU);CHKERRQ(ierr); 221fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 222fd14bc51SStefano 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); 223fd14bc51SStefano Zampini } 2249ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 22508122e43SStefano Zampini PetscInt j; 22608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 22708122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 22808122e43SStefano Zampini } 22908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23008122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23108122e43SStefano Zampini } 23208122e43SStefano Zampini } else { 23308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 23408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 23508122e43SStefano Zampini } 23608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 23808122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&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)); 23908122e43SStefano Zampini #else 24008122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&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)); 24108122e43SStefano Zampini #endif 24208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 24308122e43SStefano Zampini B_neigs += B_neigs2; 24408122e43SStefano Zampini } 24508122e43SStefano Zampini if (B_ierr) { 24608122e43SStefano Zampini if (B_ierr < 0 ) { 24708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 24808122e43SStefano Zampini } else if (B_ierr <= B_N) { 24908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25008122e43SStefano Zampini } else { 2519552c7c7SStefano 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); 25208122e43SStefano Zampini } 25308122e43SStefano Zampini } 254fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 255fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 25608122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 25708122e43SStefano Zampini if (eigs[j] == 0.0) { 258fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 25908122e43SStefano Zampini } else { 260fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1.0/eigs[j]); 261fd14bc51SStefano Zampini } 26208122e43SStefano Zampini } 26308122e43SStefano Zampini } 26408122e43SStefano Zampini } else { 26508122e43SStefano Zampini /* TODO */ 26608122e43SStefano Zampini } 26708122e43SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2681b968477SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 26908122e43SStefano Zampini 27008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2714c6709b3SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&B_N,&B_neigs,&B_N,&one,Smult,&B_N,eigv,&B_N,&scalar_zero,Seigv,&B_N)); 27208122e43SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+cum2,Seigv,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 273fd14bc51SStefano Zampini 274fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2759552c7c7SStefano Zampini PetscInt ii; 2769552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 277fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2789552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 279fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e %1.4e\n",eigv[ii*B_N+j],Seigv[ii*B_N+j]); 2809552c7c7SStefano Zampini } 2819552c7c7SStefano Zampini } 282fd14bc51SStefano Zampini } 28308122e43SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 28408122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 28508122e43SStefano Zampini #if 0 28608122e43SStefano Zampini { 28708122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 28808122e43SStefano Zampini PetscScalar norm; 28908122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 29008122e43SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+cum2,&Blas_one,pcbddc->adaptive_constraints_data+cum2,&Blas_one)); 29108122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 29208122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 29308122e43SStefano Zampini } else { 29408122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 29508122e43SStefano Zampini } 29608122e43SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+cum2,&Blas_one)); 297b1b3d7a2SStefano Zampini } 298b1b3d7a2SStefano Zampini #endif 29908122e43SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+cum2,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 30008122e43SStefano Zampini pcbddc->adaptive_constraints_ptrs[cum++] = cum2; 30108122e43SStefano Zampini cum2 += subset_size; 30208122e43SStefano Zampini } 30308122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 30408122e43SStefano Zampini } 30508122e43SStefano Zampini /* shift for next computation */ 30608122e43SStefano Zampini cumarray += subset_size*subset_size; 30708122e43SStefano Zampini } 308fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 309fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 310fd14bc51SStefano Zampini } 31108122e43SStefano Zampini pcbddc->adaptive_constraints_ptrs[cum] = cum2; 31208122e43SStefano Zampini ierr = PetscFree2(Smult,Seigv);CHKERRQ(ierr); 31308122e43SStefano Zampini 31408122e43SStefano Zampini if (mss) { 31508122e43SStefano Zampini if (pcbddc->use_deluxe_scaling) { 31608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 31708122e43SStefano Zampini } else { 31808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&Sarray);CHKERRQ(ierr); 31908122e43SStefano Zampini } 32008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 32108122e43SStefano Zampini } 32208122e43SStefano Zampini ierr = PetscFree7(S,St,eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 32308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 32508122e43SStefano Zampini #endif 32608122e43SStefano Zampini if (pcbddc->dbg_flag) { 3271b968477SStefano Zampini PetscInt maxneigs_r; 32808122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 32908122e43SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33008122e43SStefano Zampini } 33108122e43SStefano Zampini PetscFunctionReturn(0); 33208122e43SStefano Zampini } 333b1b3d7a2SStefano Zampini 334674ae819SStefano Zampini #undef __FUNCT__ 335c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 336c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 337c8587f34SStefano Zampini { 338c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3398629588bSStefano Zampini PetscScalar *coarse_submat_vals; 340c8587f34SStefano Zampini PetscErrorCode ierr; 341c8587f34SStefano Zampini 342c8587f34SStefano Zampini PetscFunctionBegin; 343f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3445e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 345c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 346c8587f34SStefano Zampini 347684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3480fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 349684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 350c8587f34SStefano Zampini 351c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 352b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 353c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 354c8587f34SStefano Zampini } 355c8587f34SStefano Zampini 3568629588bSStefano Zampini /* 3578629588bSStefano Zampini Setup local correction and local part of coarse basis. 3588629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3598629588bSStefano Zampini */ 36047f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3618629588bSStefano Zampini 3628629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3638629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3648629588bSStefano Zampini 3658629588bSStefano Zampini /* free */ 3668629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 367c8587f34SStefano Zampini PetscFunctionReturn(0); 368c8587f34SStefano Zampini } 369c8587f34SStefano Zampini 370c8587f34SStefano Zampini #undef __FUNCT__ 371674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 372674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 373674ae819SStefano Zampini { 374674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 375674ae819SStefano Zampini PetscErrorCode ierr; 376674ae819SStefano Zampini 377674ae819SStefano Zampini PetscFunctionBegin; 378674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 379674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 380674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 381674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 382785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 383674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 384f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 385f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 386785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 38763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 38863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 389674ae819SStefano Zampini PetscFunctionReturn(0); 390674ae819SStefano Zampini } 391674ae819SStefano Zampini 392674ae819SStefano Zampini #undef __FUNCT__ 393674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 394674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 395674ae819SStefano Zampini { 396674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 397674ae819SStefano Zampini PetscErrorCode ierr; 398674ae819SStefano Zampini 399674ae819SStefano Zampini PetscFunctionBegin; 400b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 401674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 402674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 403674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 404b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 405674ae819SStefano Zampini PetscFunctionReturn(0); 406674ae819SStefano Zampini } 407674ae819SStefano Zampini 408674ae819SStefano Zampini #undef __FUNCT__ 409674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 410674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 411674ae819SStefano Zampini { 412674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 41306656605SStefano Zampini PetscScalar *array; 414674ae819SStefano Zampini PetscErrorCode ierr; 415674ae819SStefano Zampini 416674ae819SStefano Zampini PetscFunctionBegin; 417674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 41906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42158da7f69SStefano Zampini } 422674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 423674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 42415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 42515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 426674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 427674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 42806656605SStefano Zampini if (pcbddc->local_auxmat2) { 42906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 43006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43106656605SStefano Zampini } 432674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43306656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 434674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4368ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 437674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 440f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 441f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 442f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 443f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 444727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 445f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 44670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4476e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 448674ae819SStefano Zampini PetscFunctionReturn(0); 449674ae819SStefano Zampini } 450674ae819SStefano Zampini 451674ae819SStefano Zampini #undef __FUNCT__ 452f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 453f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4546bfb1811SStefano Zampini { 4556bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4566bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4576bfb1811SStefano Zampini VecType impVecType; 458e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4596bfb1811SStefano Zampini PetscErrorCode ierr; 4606bfb1811SStefano Zampini 4616bfb1811SStefano Zampini PetscFunctionBegin; 462f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 463f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 464f4ddd8eeSStefano Zampini } 465e7b262bdSStefano Zampini /* get sizes */ 466e9189074SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_actual_vertices; 467e9189074SStefano Zampini n_R = pcis->n-pcbddc->n_actual_vertices; 4686bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 469e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 470e7b262bdSStefano Zampini /* R nodes */ 471e7b262bdSStefano Zampini old_size = -1; 472e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 473e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 474e7b262bdSStefano Zampini } 475e7b262bdSStefano Zampini if (n_R != old_size) { 476e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 477e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4786bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4796bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4806bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4816bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 482e7b262bdSStefano Zampini } 483e7b262bdSStefano Zampini /* local primal dofs */ 484e7b262bdSStefano Zampini old_size = -1; 485e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 486e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 487e7b262bdSStefano Zampini } 488e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 489e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 491e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4926bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 493e7b262bdSStefano Zampini } 494e7b262bdSStefano Zampini /* local explicit constraints */ 495e7b262bdSStefano Zampini old_size = -1; 496e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 497e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 498e7b262bdSStefano Zampini } 499e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 500e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50283b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50383b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50483b7ccabSStefano Zampini } 5056bfb1811SStefano Zampini PetscFunctionReturn(0); 5066bfb1811SStefano Zampini } 5076bfb1811SStefano Zampini 5086bfb1811SStefano Zampini #undef __FUNCT__ 50947f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51047f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51188ebb749SStefano Zampini { 51225084f0cSStefano Zampini PetscErrorCode ierr; 51325084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51488ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51588ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 51625084f0cSStefano Zampini /* submatrices of local problem */ 51788ebb749SStefano Zampini Mat A_RV,A_VR,A_VV; 51806656605SStefano Zampini /* submatrices of local coarse problem */ 51906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52025084f0cSStefano Zampini /* working matrices */ 52106656605SStefano Zampini Mat C_CR; 52225084f0cSStefano Zampini /* additional working stuff */ 52306656605SStefano Zampini PC pc_R; 52406656605SStefano Zampini Mat F; 52506656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 52606656605SStefano Zampini 52725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 52806656605SStefano Zampini PetscScalar *work; 52906656605SStefano Zampini PetscInt *idx_V_B; 53006656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 532b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53345a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53488ebb749SStefano Zampini MatType impMatType; 53525084f0cSStefano Zampini /* some shortcuts to scalars */ 53606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 53788ebb749SStefano Zampini 53888ebb749SStefano Zampini PetscFunctionBegin; 539e9189074SStefano Zampini /* get number of vertices (corners plus constraints with change of basis) 540e9189074SStefano Zampini pcbddc->n_actual_vertices stores the actual number of vertices, pcbddc->n_vertices the number of corners computed */ 541e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 54288ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54388ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 54488ebb749SStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 54588ebb749SStefano Zampini n_R = pcis->n-n_vertices; 54688ebb749SStefano Zampini 54788ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 54888ebb749SStefano Zampini impMatType = MATSEQDENSE; 54988ebb749SStefano Zampini 55088ebb749SStefano Zampini /* vertices in boundary numbering */ 551785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5525e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->primal_indices_local_idxs,&i,idx_V_B);CHKERRQ(ierr); 55388ebb749SStefano Zampini if (i != n_vertices) { 55422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 55588ebb749SStefano Zampini } 55688ebb749SStefano Zampini 55706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 55806656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 55906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56706656605SStefano Zampini 56806656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 56906656605SStefano Zampini /* allocate workspace */ 57006656605SStefano Zampini n = 0; 57106656605SStefano Zampini if (n_constraints) { 57206656605SStefano Zampini n += n_R*n_constraints; 57306656605SStefano Zampini } 57406656605SStefano Zampini if (n_vertices) { 57506656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57606656605SStefano Zampini } 57706656605SStefano Zampini if (!pcbddc->issym) { 57806656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 57906656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58006656605SStefano Zampini } 58106656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58206656605SStefano Zampini 58306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 58706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 58806656605SStefano Zampini if (isLU || isILU || isCHOL) { 58906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 59006656605SStefano Zampini } else { 59106656605SStefano Zampini F = NULL; 59206656605SStefano Zampini } 59306656605SStefano Zampini 59488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 59588ebb749SStefano Zampini if (n_constraints) { 59606656605SStefano Zampini Mat M1,M2,M3; 59706656605SStefano Zampini IS is_aux; 598f4ddd8eeSStefano Zampini /* see if we can save some allocations */ 599f4ddd8eeSStefano Zampini if (pcbddc->local_auxmat2) { 600f4ddd8eeSStefano Zampini PetscInt on_R,on_constraints; 601f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->local_auxmat2,&on_R,&on_constraints);CHKERRQ(ierr); 602f4ddd8eeSStefano Zampini if (on_R != n_R || on_constraints != n_constraints) { 60306656605SStefano Zampini PetscScalar *marray; 60406656605SStefano Zampini 60506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&marray);CHKERRQ(ierr); 60606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 607f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 608f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 609f4ddd8eeSStefano Zampini } 610f4ddd8eeSStefano Zampini } 61145a1bb75SStefano Zampini /* auxiliary matrices */ 612f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat2) { 61306656605SStefano Zampini PetscScalar *marray; 61406656605SStefano Zampini 61506656605SStefano Zampini ierr = PetscMalloc1(2*n_R*n_constraints,&marray);CHKERRQ(ierr); 61606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,marray,&pcbddc->local_auxmat2);CHKERRQ(ierr); 61706656605SStefano Zampini marray += n_R*n_constraints; 61806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_R,marray,&pcbddc->local_auxmat1);CHKERRQ(ierr); 619f4ddd8eeSStefano Zampini } 62088ebb749SStefano Zampini 62125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6238ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 62425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 62588ebb749SStefano Zampini 62688ebb749SStefano Zampini /* Assemble local_auxmat2 = - A_{RR}^{-1} C^T_{CR} needed by BDDC application */ 62706656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 62906656605SStefano Zampini const PetscScalar *row_cmat_values; 63006656605SStefano Zampini const PetscInt *row_cmat_indices; 63106656605SStefano Zampini PetscInt size_of_constraint,j; 63288ebb749SStefano Zampini 63306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63506656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63606656605SStefano Zampini } 63706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63806656605SStefano Zampini } 63906656605SStefano Zampini if (F) { 64006656605SStefano Zampini Mat B; 64106656605SStefano Zampini 64206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 64306656605SStefano Zampini ierr = MatMatSolve(F,B,pcbddc->local_auxmat2);CHKERRQ(ierr); 64406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64506656605SStefano Zampini } else { 64606656605SStefano Zampini PetscScalar *xarray; 64706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr); 64806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 64906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,xarray+i*n_R);CHKERRQ(ierr); 65106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 65206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65406656605SStefano Zampini } 65506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr); 65606656605SStefano Zampini } 65706656605SStefano Zampini 65806656605SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 65906656605SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 66025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 66206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 66325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 66406656605SStefano Zampini ierr = VecSet(pcbddc->vec1_C,m_one);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 66625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 66725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 66825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 66906656605SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CR} needed by BDDC application in KSP and in preproc */ 670f4ddd8eeSStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_REUSE_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 67106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 67206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 67388ebb749SStefano Zampini } 67488ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 67588ebb749SStefano Zampini if (n_vertices) { 67606656605SStefano Zampini Mat newmat; 67706656605SStefano Zampini IS is_aux; 6783a50541eSStefano Zampini PetscInt ibs,mbs; 6793a50541eSStefano Zampini PetscBool issbaij; 6803a50541eSStefano Zampini 6813a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 6829577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 6839577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 6843a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 6859577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 6869577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 6879577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 6889577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 6899577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 6909577ea80SStefano Zampini } else { 6913a50541eSStefano Zampini /* this is safe */ 6929577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 6933a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 6943a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 6953a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 6963a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 69706656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 6983a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 69903af80d8SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7003a50541eSStefano Zampini } else { 7019577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7029577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7039577ea80SStefano Zampini } 7043a50541eSStefano Zampini } 70525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 70688ebb749SStefano Zampini } 70788ebb749SStefano Zampini 70888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 709f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 71006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 71106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 71206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 71306656605SStefano Zampini } 714f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 71506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 71606656605SStefano Zampini PetscScalar *marray; 71706656605SStefano Zampini 71806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 71906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 720f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 721f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 722f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 723f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 724f4ddd8eeSStefano Zampini } 725f4ddd8eeSStefano Zampini } 72606656605SStefano Zampini 727f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 72806656605SStefano Zampini PetscScalar *marray; 72988ebb749SStefano Zampini 73006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7318eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 73206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 73388ebb749SStefano Zampini } 734b9d89cd5SStefano Zampini if (!pcbddc->issym) { 73506656605SStefano Zampini n *= 2; 736f4ddd8eeSStefano Zampini } 73706656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 73806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 73906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 74006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 74206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74388ebb749SStefano Zampini } 74406656605SStefano Zampini if (!pcbddc->issym) { 74506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7461b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 74806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 74988ebb749SStefano Zampini } 75088ebb749SStefano Zampini } else { 751c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 752c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7531b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 754c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 755c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 756c0553b1fSStefano Zampini } 75788ebb749SStefano Zampini } 75806656605SStefano Zampini } 75906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 76006656605SStefano Zampini /* vertices */ 76106656605SStefano Zampini if (n_vertices) { 76216f15bc4SStefano Zampini 76316f15bc4SStefano Zampini if (n_R) { 76406656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 76506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 76616f15bc4SStefano Zampini PetscScalar *x,*y; 76706656605SStefano Zampini 76806656605SStefano Zampini ierr = PetscMemzero(work,2*n_R*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 76906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 77006656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 77106656605SStefano Zampini if (F) { 77206656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 77306656605SStefano Zampini } else { 77406656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 77506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 77606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 77706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 77806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 77906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 78006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 78106656605SStefano Zampini } 78206656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 78306656605SStefano Zampini } 78406656605SStefano Zampini ierr = MatScale(A_RRmA_RV,m_one);CHKERRQ(ierr); 78506656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 78606656605SStefano Zampini if (n_constraints) { 78706656605SStefano Zampini Mat B; 78806656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,A_RRmA_RV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 78906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 79006656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat2,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 79106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 79206656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 79306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 79406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 79506656605SStefano Zampini } 79606656605SStefano Zampini ierr = MatConvert(A_VR,impMatType,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 79706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 79806656605SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 79906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 80006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 80106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 80206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 80306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 80606656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 80716f15bc4SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 80816f15bc4SStefano Zampini } else { 80916f15bc4SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 81016f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 81116f15bc4SStefano Zampini } 81206656605SStefano Zampini /* coarse basis functions */ 81306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 81416f15bc4SStefano Zampini PetscScalar *y; 81516f15bc4SStefano Zampini 81606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 81706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 81806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 81906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 82206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 82306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 82406656605SStefano Zampini 82506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 82606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 82706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 82806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 83206656605SStefano Zampini } 83306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 83406656605SStefano Zampini } 83516f15bc4SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 83606656605SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 83706656605SStefano Zampini } 83806656605SStefano Zampini 83906656605SStefano Zampini if (n_constraints) { 84006656605SStefano Zampini Mat B; 84106656605SStefano Zampini 84206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat2,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 84606656605SStefano Zampini if (n_vertices) { 84706656605SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_VC);CHKERRQ(ierr); 84806656605SStefano Zampini } 84906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 85006656605SStefano Zampini /* coarse basis functions */ 85106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 85206656605SStefano Zampini PetscScalar *y; 85306656605SStefano Zampini 85406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 85706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86106656605SStefano Zampini 86206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 86506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86906656605SStefano Zampini } 87006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87106656605SStefano Zampini } 87206656605SStefano Zampini } 87306656605SStefano Zampini 87406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 87506656605SStefano Zampini if (!pcbddc->issym) { 87606656605SStefano Zampini Mat B,X; 87706656605SStefano Zampini 87806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work,&B);CHKERRQ(ierr); 87906656605SStefano Zampini 88006656605SStefano Zampini if (n_constraints) { 88116f15bc4SStefano Zampini Mat S_CCT,B_C; 88206656605SStefano Zampini 88306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 88406656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 88616f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 88706656605SStefano Zampini if (n_vertices) { 88816f15bc4SStefano Zampini Mat B_V,S_VCT; 88906656605SStefano Zampini 89006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 89106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 89206656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 89416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 89506656605SStefano Zampini } 89606656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 89706656605SStefano Zampini } 89816f15bc4SStefano Zampini if (n_vertices && n_R) { 89906656605SStefano Zampini Mat A_VRT; 90006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 90106656605SStefano Zampini 90206656605SStefano Zampini if (!n_constraints) { /* if there are no constraints, reset work */ 90306656605SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 90406656605SStefano Zampini } 90506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+pcbddc->local_primal_size*n_R,&A_VRT);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 90806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,work+pcbddc->local_primal_size*n_R,&B_one,work,&B_one)); 90916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 91006656605SStefano Zampini } 91106656605SStefano Zampini 91206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work+pcbddc->local_primal_size*n_R,&X);CHKERRQ(ierr); 91306656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 91406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 91506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini } else { 92206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 92306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 92606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 92806656605SStefano Zampini } 92906656605SStefano Zampini } 93006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 93106656605SStefano Zampini /* coarse basis functions */ 93206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 93306656605SStefano Zampini PetscScalar *y; 93406656605SStefano Zampini 93506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 93706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 93806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 93906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94006656605SStefano Zampini if (i<n_vertices) { 94106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 94206656605SStefano Zampini } 94306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 94506656605SStefano Zampini 94606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 94706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 94806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 95306656605SStefano Zampini } 95406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95506656605SStefano Zampini } 95606656605SStefano Zampini ierr = MatDestroy(&X);CHKERRQ(ierr); 95706656605SStefano Zampini } 95806656605SStefano Zampini 95988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 96306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 96488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 96588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 96688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 96725084f0cSStefano Zampini if (pcbddc->dbg_flag) { 96888ebb749SStefano Zampini Mat coarse_sub_mat; 96925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 97088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 97188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 97288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 97388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 97488ebb749SStefano Zampini PetscReal real_value; 97588ebb749SStefano Zampini 97688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 97788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 97888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 97988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 98088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 98188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 982c0553b1fSStefano Zampini if (unsymmetric_check) { 98388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 98488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 98588ebb749SStefano Zampini } 98688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 98788ebb749SStefano Zampini 98825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 98906656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 99025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 991c0553b1fSStefano Zampini if (unsymmetric_check) { 99288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 99388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 99488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 99588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 99688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 99788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 99888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 99988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 100088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 100188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 100388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 100488ebb749SStefano Zampini } else { 100588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 100688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 100788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 100888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 100988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 101088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 101188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 101288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 101388ebb749SStefano Zampini } 101488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 101588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 101688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 101788ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 101881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 101988ebb749SStefano Zampini ierr = MatNorm(TM1,NORM_INFINITY,&real_value);CHKERRQ(ierr); 102006656605SStefano Zampini 10210fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 102206656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 102325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 102488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 102588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 102688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 102788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 103088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1034c0553b1fSStefano Zampini if (unsymmetric_check) { 103588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 103688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 103788ebb749SStefano Zampini } 103888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 103988ebb749SStefano Zampini } 104006656605SStefano Zampini 104188ebb749SStefano Zampini /* free memory */ 104206656605SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 104388ebb749SStefano Zampini if (n_vertices) { 104488ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 104588ebb749SStefano Zampini } 104688ebb749SStefano Zampini if (n_constraints) { 104788ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 104888ebb749SStefano Zampini } 10498629588bSStefano Zampini /* get back data */ 10508629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 105188ebb749SStefano Zampini PetscFunctionReturn(0); 105288ebb749SStefano Zampini } 105388ebb749SStefano Zampini 105488ebb749SStefano Zampini #undef __FUNCT__ 1055d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1056d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1057d65f70fdSStefano Zampini { 1058d65f70fdSStefano Zampini Mat *work_mat; 1059d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1060d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1061d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1062d65f70fdSStefano Zampini PetscErrorCode ierr; 1063d65f70fdSStefano Zampini 1064d65f70fdSStefano Zampini PetscFunctionBegin; 1065d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1066d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1067d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1068d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1069d65f70fdSStefano Zampini 1070d65f70fdSStefano Zampini if (!rsorted) { 1071d65f70fdSStefano Zampini const PetscInt *idxs; 1072d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1073d65f70fdSStefano Zampini 1074d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1075d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1076d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1077d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1078d65f70fdSStefano Zampini } 1079d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1080d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1081d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1082d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1083d65f70fdSStefano Zampini } 1084d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1085d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1086d65f70fdSStefano Zampini } else { 1087d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1088d65f70fdSStefano Zampini isrow_s = isrow; 1089d65f70fdSStefano Zampini } 1090d65f70fdSStefano Zampini 1091d65f70fdSStefano Zampini if (!csorted) { 1092d65f70fdSStefano Zampini if (isrow == iscol) { 1093d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1094d65f70fdSStefano Zampini iscol_s = isrow_s; 1095d65f70fdSStefano Zampini } else { 1096d65f70fdSStefano Zampini const PetscInt *idxs; 1097d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1098d65f70fdSStefano Zampini 1099d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1100d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1101d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1102d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1103d65f70fdSStefano Zampini } 1104d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1105d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1106d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1107d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1108d65f70fdSStefano Zampini } 1109d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1110d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1111d65f70fdSStefano Zampini } 1112d65f70fdSStefano Zampini } else { 1113d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1114d65f70fdSStefano Zampini iscol_s = iscol; 1115d65f70fdSStefano Zampini } 1116d65f70fdSStefano Zampini 1117d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1118d65f70fdSStefano Zampini 1119d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1120d65f70fdSStefano Zampini Mat new_mat; 1121d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1122d65f70fdSStefano Zampini 1123d65f70fdSStefano Zampini if (!rsorted) { 1124d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1125d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1126d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1127d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1128d65f70fdSStefano Zampini } 1129d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1130d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini } else { 1132d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini } 1134d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini 1136d65f70fdSStefano Zampini if (!csorted) { 1137d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1138d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1139d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1140d65f70fdSStefano Zampini } else { 1141d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1142d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1143d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1144d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1145d65f70fdSStefano Zampini } 1146d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini } 1149d65f70fdSStefano Zampini } else { 1150d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini } 1152d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini 1154d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1155d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1156d65f70fdSStefano Zampini work_mat[0] = new_mat; 1157d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1158d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini } 1160d65f70fdSStefano Zampini 1161d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1162d65f70fdSStefano Zampini *B = work_mat[0]; 1163d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1164d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini PetscFunctionReturn(0); 1167d65f70fdSStefano Zampini } 1168d65f70fdSStefano Zampini 1169d65f70fdSStefano Zampini #undef __FUNCT__ 11705e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 11715e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1172aa0d41d4SStefano Zampini { 1173aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 11745e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1175d65f70fdSStefano Zampini Mat new_mat; 11765e8657edSStefano Zampini IS is_local,is_global; 1177d65f70fdSStefano Zampini PetscInt local_size; 1178d65f70fdSStefano Zampini PetscBool isseqaij; 1179aa0d41d4SStefano Zampini PetscErrorCode ierr; 1180aa0d41d4SStefano Zampini 1181aa0d41d4SStefano Zampini PetscFunctionBegin; 1182aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 11835e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 11845e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1185906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1186906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1187d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1188906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1189906d46d4SStefano Zampini 1190906d46d4SStefano Zampini /* check */ 1191906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1192906d46d4SStefano Zampini Vec x,x_change; 1193906d46d4SStefano Zampini PetscReal error; 1194906d46d4SStefano Zampini 11955e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1196906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 11975e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 11985e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 11995e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12015e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12025e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1203906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1204906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1205906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1206906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1207906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1208906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1209906d46d4SStefano Zampini } 1210906d46d4SStefano Zampini 121122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12129b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 121322d5777bSStefano Zampini if (isseqaij) { 1214d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1215aa0d41d4SStefano Zampini } else { 1216aa0d41d4SStefano Zampini Mat work_mat; 1217aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1219aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1220aa0d41d4SStefano Zampini } 1221d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 122245a1bb75SStefano Zampini /* 122345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 122545a1bb75SStefano Zampini */ 1226d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1227aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1228aa0d41d4SStefano Zampini } 1229aa0d41d4SStefano Zampini 1230aa0d41d4SStefano Zampini #undef __FUNCT__ 1231a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12328ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1233a64d13efSStefano Zampini { 1234a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1235a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12368ce42a96SStefano Zampini IS is_aux1,is_aux2; 1237e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 12383a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 12393a50541eSStefano Zampini PetscInt vbs,bs; 12404641a718SStefano Zampini PetscBT bitmask; 1241a64d13efSStefano Zampini PetscErrorCode ierr; 1242a64d13efSStefano Zampini 1243a64d13efSStefano Zampini PetscFunctionBegin; 1244b23d619eSStefano Zampini /* 1245b23d619eSStefano Zampini No need to setup local scatters if 1246b23d619eSStefano Zampini - primal space is unchanged 1247b23d619eSStefano Zampini AND 1248b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1249b23d619eSStefano Zampini AND 1250b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1251b23d619eSStefano Zampini */ 1252b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1253f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1254f4ddd8eeSStefano Zampini } 1255f4ddd8eeSStefano Zampini /* destroy old objects */ 1256f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1257f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1258f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1259a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1260a64d13efSStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 1261e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 12624641a718SStefano Zampini /* create auxiliary bitmask */ 12634641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 12644641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 1265e9189074SStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->primal_indices_local_idxs[i]);CHKERRQ(ierr); 12664641a718SStefano Zampini } 12673a50541eSStefano Zampini 1268a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 1269854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1270a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 12714641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 1272a64d13efSStefano Zampini idx_R_local[n_R] = i; 1273a64d13efSStefano Zampini n_R++; 1274a64d13efSStefano Zampini } 1275a64d13efSStefano Zampini } 12763a50541eSStefano Zampini 12773a50541eSStefano Zampini /* Block code */ 12783a50541eSStefano Zampini vbs = 1; 12793a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 12803a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 12813a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 12823a50541eSStefano Zampini PetscInt *vary; 12833a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1284785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 12853a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1286e9189074SStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->primal_indices_local_idxs[i]/bs]++; 1287*964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 12883a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 12893a50541eSStefano Zampini is_blocked = PETSC_FALSE; 12903a50541eSStefano Zampini break; 12913a50541eSStefano Zampini } 12923a50541eSStefano Zampini } 12933a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 12943a50541eSStefano Zampini vbs = bs; 12953a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 12963a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 12973a50541eSStefano Zampini } 12983a50541eSStefano Zampini } 12993a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13003a50541eSStefano Zampini } 13013a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 13023a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1303a64d13efSStefano Zampini 1304a64d13efSStefano Zampini /* print some info if requested */ 1305a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1306a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1307a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13080fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1309a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1310a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13113a50541eSStefano 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); 1312a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"pcbddc->n_vertices = %d, pcbddc->n_constraints = %d\n",pcbddc->n_vertices,pcbddc->n_constraints);CHKERRQ(ierr); 1313a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1314a64d13efSStefano Zampini } 1315a64d13efSStefano Zampini 1316a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 13173a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1318854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1319854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1320a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 13214641a718SStefano Zampini for (i=0; i<n_D; i++) { 13224641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 13234641a718SStefano Zampini } 1324a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1325a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13264641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 13274641a718SStefano Zampini aux_array1[j++] = i; 1328a64d13efSStefano Zampini } 1329a64d13efSStefano Zampini } 1330a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1331a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1332a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 13334641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 13344641a718SStefano Zampini aux_array2[j++] = i; 1335a64d13efSStefano Zampini } 1336a64d13efSStefano Zampini } 1337a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1338a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1339a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1340a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1341a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1342a64d13efSStefano Zampini 13438eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1344785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1345a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 13464641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 13474641a718SStefano Zampini aux_array1[j++] = i; 1348a64d13efSStefano Zampini } 1349a64d13efSStefano Zampini } 1350a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1351a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1352a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1353a64d13efSStefano Zampini } 13544641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 13553a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1356a64d13efSStefano Zampini PetscFunctionReturn(0); 1357a64d13efSStefano Zampini } 1358a64d13efSStefano Zampini 1359304d26faSStefano Zampini 1360304d26faSStefano Zampini #undef __FUNCT__ 1361304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1362684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1363304d26faSStefano Zampini { 1364304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1365304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1366304d26faSStefano Zampini PC pc_temp; 1367304d26faSStefano Zampini Mat A_RR; 1368f4ddd8eeSStefano Zampini MatReuse reuse; 1369304d26faSStefano Zampini PetscScalar m_one = -1.0; 1370304d26faSStefano Zampini PetscReal value; 1371af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 13729577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1373304d26faSStefano Zampini PetscErrorCode ierr; 1374e604994aSStefano Zampini /* prefixes stuff */ 1375312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1376e604994aSStefano Zampini size_t len; 1377304d26faSStefano Zampini 1378304d26faSStefano Zampini PetscFunctionBegin; 1379304d26faSStefano Zampini 1380e604994aSStefano Zampini /* compute prefixes */ 1381e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1382e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1383e604994aSStefano Zampini if (!pcbddc->current_level) { 1384e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1385e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1386e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1387e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1388e604994aSStefano Zampini } else { 1389e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1390312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1391e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1392e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1393312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1394312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 139534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 139634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1397e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1398e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1399e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1400e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1401e604994aSStefano Zampini } 1402e604994aSStefano Zampini 1403304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1404684f6988SStefano Zampini if (dirichlet) { 1405*964fefecSStefano Zampini if (pcbddc->issym) { 1406*964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1407*964fefecSStefano Zampini } 1408ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1409*964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1410304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1411304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1412304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1413304d26faSStefano Zampini /* default */ 1414304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1415e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 14169577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1417304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 14189577ea80SStefano Zampini if (issbaij) { 14199577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 14209577ea80SStefano Zampini } else { 1421304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 14229577ea80SStefano Zampini } 1423304d26faSStefano Zampini /* Allow user's customization */ 1424304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1425304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1426304d26faSStefano Zampini } 1427d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1428304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1429304d26faSStefano Zampini if (!n_D) { 1430304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1431304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1432304d26faSStefano Zampini } 1433304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1434304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1435304d26faSStefano Zampini /* set ksp_D into pcis data */ 1436304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1437304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1438304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1439684f6988SStefano Zampini } 1440304d26faSStefano Zampini 1441304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1442684f6988SStefano Zampini A_RR = 0; 1443684f6988SStefano Zampini if (neumann) { 1444f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 14458ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1446f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1447f4ddd8eeSStefano Zampini PetscInt nn_R; 144881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1449f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1450f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1451f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1452f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1453f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1454f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1455f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1456727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1457f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1458f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1459f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1460f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1461f4ddd8eeSStefano Zampini } 1462f4ddd8eeSStefano Zampini } 1463f4ddd8eeSStefano Zampini /* last check */ 1464d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1465f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1466f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1467f4ddd8eeSStefano Zampini } 1468f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1469f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1470f4ddd8eeSStefano Zampini } 1471f4ddd8eeSStefano Zampini /* extract A_RR */ 1472af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1473af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 14743a50541eSStefano Zampini if (ibs != mbs) { 1475af732b37SStefano Zampini Mat newmat; 1476af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1477f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1478af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1479af732b37SStefano Zampini } else { 1480f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1481af732b37SStefano Zampini } 1482*964fefecSStefano Zampini if (pcbddc->issym) { 1483*964fefecSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1484*964fefecSStefano Zampini } 1485f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1486304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1487304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1488304d26faSStefano Zampini /* default */ 1489304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1490e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1491304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 14929577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 14939577ea80SStefano Zampini if (issbaij) { 14949577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 14959577ea80SStefano Zampini } else { 1496304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 14979577ea80SStefano Zampini } 1498304d26faSStefano Zampini /* Allow user's customization */ 1499304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1500304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1501304d26faSStefano Zampini } 1502d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1503304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1504304d26faSStefano Zampini if (!n_R) { 1505304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1506304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1507304d26faSStefano Zampini } 1508304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1509304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1510684f6988SStefano Zampini } 1511304d26faSStefano Zampini 1512304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 15130fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1514684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1515684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1516684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1517684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1518684f6988SStefano Zampini } 1519684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 15200fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 15210fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 15220fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 15230fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 15240fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1525304d26faSStefano Zampini /* need to be adapted? */ 1526b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1527b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1528b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1529304d26faSStefano Zampini /* print info */ 1530304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1531e604994aSStefano 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); 1532304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1533304d26faSStefano Zampini } 1534b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 15358ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1536304d26faSStefano Zampini } 1537684f6988SStefano Zampini } 1538684f6988SStefano Zampini if (neumann) { /* Neumann */ 15390fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 15400fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 15410fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 15420fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 15430fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1544304d26faSStefano Zampini /* need to be adapted? */ 1545b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1546b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1547304d26faSStefano Zampini /* print info */ 1548304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1549e604994aSStefano 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); 1550304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1551304d26faSStefano Zampini } 1552b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 15538ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1554304d26faSStefano Zampini } 15550fccc4e9SStefano Zampini } 1556684f6988SStefano Zampini } 1557304d26faSStefano Zampini /* free Neumann problem's matrix */ 1558304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1559304d26faSStefano Zampini PetscFunctionReturn(0); 1560304d26faSStefano Zampini } 1561304d26faSStefano Zampini 1562304d26faSStefano Zampini #undef __FUNCT__ 1563ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 156420c7b377SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec rhs, Vec sol, Vec work, PetscBool applytranspose) 1565674ae819SStefano Zampini { 1566674ae819SStefano Zampini PetscErrorCode ierr; 1567674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1568674ae819SStefano Zampini 1569674ae819SStefano Zampini PetscFunctionBegin; 157020c7b377SStefano Zampini if (applytranspose) { 1571674ae819SStefano Zampini if (pcbddc->local_auxmat1) { 157220c7b377SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,rhs,work);CHKERRQ(ierr); 157320c7b377SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,work,rhs,rhs);CHKERRQ(ierr); 157420c7b377SStefano Zampini } 157520c7b377SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 157620c7b377SStefano Zampini } else { 157720c7b377SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 157820c7b377SStefano Zampini if (pcbddc->local_auxmat1) { 157920c7b377SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,sol,work);CHKERRQ(ierr); 158020c7b377SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,work,sol,sol);CHKERRQ(ierr); 158120c7b377SStefano Zampini } 1582674ae819SStefano Zampini } 1583674ae819SStefano Zampini PetscFunctionReturn(0); 1584674ae819SStefano Zampini } 1585674ae819SStefano Zampini 1586dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1587674ae819SStefano Zampini #undef __FUNCT__ 1588674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1589dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1590674ae819SStefano Zampini { 1591674ae819SStefano Zampini PetscErrorCode ierr; 1592674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1593674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1594674ae819SStefano Zampini const PetscScalar zero = 0.0; 1595674ae819SStefano Zampini 1596674ae819SStefano Zampini PetscFunctionBegin; 1597dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1598dc359a40SStefano Zampini if (applytranspose) { 1599674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 16008eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1601dc359a40SStefano Zampini } else { 1602674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1603674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 160415aaf578SStefano Zampini } 160512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 160612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 160712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 160812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 160912edc857SStefano Zampini 16109f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 161112edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 161212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1613*964fefecSStefano Zampini Vec rhs,sol; 1614*964fefecSStefano Zampini 1615*964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1616*964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 161712edc857SStefano Zampini if (applytranspose) { 1618*964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 161912edc857SStefano Zampini } else { 1620*964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 162112edc857SStefano Zampini } 162212edc857SStefano Zampini } 1623674ae819SStefano Zampini 1624674ae819SStefano Zampini /* Local solution on R nodes */ 16259f00e9b4SStefano Zampini if (pcis->n) { 1626674ae819SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 1627674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1628674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 16298eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1630674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1631674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1632674ae819SStefano Zampini } 163320c7b377SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcbddc->vec1_R,pcbddc->vec2_R,pcbddc->vec1_C,applytranspose);CHKERRQ(ierr); 1634674ae819SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 1635674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1636674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 16378eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1638674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1639674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1640674ae819SStefano Zampini } 16419f00e9b4SStefano Zampini } 1642674ae819SStefano Zampini 16439f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 16449f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 164512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1646674ae819SStefano Zampini 1647674ae819SStefano Zampini /* Sum contributions from two levels */ 1648dc359a40SStefano Zampini if (applytranspose) { 1649dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1650dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1651dc359a40SStefano Zampini } else { 1652674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 16538eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1654dc359a40SStefano Zampini } 1655674ae819SStefano Zampini PetscFunctionReturn(0); 1656674ae819SStefano Zampini } 1657674ae819SStefano Zampini 1658674ae819SStefano Zampini #undef __FUNCT__ 1659674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 166012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1661674ae819SStefano Zampini { 1662674ae819SStefano Zampini PetscErrorCode ierr; 1663674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 166458da7f69SStefano Zampini PetscScalar *array; 166512edc857SStefano Zampini Vec from,to; 1666674ae819SStefano Zampini 1667674ae819SStefano Zampini PetscFunctionBegin; 166812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 166912edc857SStefano Zampini from = pcbddc->coarse_vec; 167012edc857SStefano Zampini to = pcbddc->vec1_P; 167112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 167212edc857SStefano Zampini Vec tvec; 167358da7f69SStefano Zampini 167458da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 167558da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 167612edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 167758da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 167858da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 167958da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 168012edc857SStefano Zampini } 168112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 168212edc857SStefano Zampini from = pcbddc->vec1_P; 168312edc857SStefano Zampini to = pcbddc->coarse_vec; 168412edc857SStefano Zampini } 168512edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1686674ae819SStefano Zampini PetscFunctionReturn(0); 1687674ae819SStefano Zampini } 1688674ae819SStefano Zampini 1689674ae819SStefano Zampini #undef __FUNCT__ 1690674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 169112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1692674ae819SStefano Zampini { 1693674ae819SStefano Zampini PetscErrorCode ierr; 1694674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 169558da7f69SStefano Zampini PetscScalar *array; 169612edc857SStefano Zampini Vec from,to; 1697674ae819SStefano Zampini 1698674ae819SStefano Zampini PetscFunctionBegin; 169912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 170012edc857SStefano Zampini from = pcbddc->coarse_vec; 170112edc857SStefano Zampini to = pcbddc->vec1_P; 170212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 170312edc857SStefano Zampini from = pcbddc->vec1_P; 170412edc857SStefano Zampini to = pcbddc->coarse_vec; 170512edc857SStefano Zampini } 170612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 170712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 170812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 170912edc857SStefano Zampini Vec tvec; 171058da7f69SStefano Zampini 171112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 171258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 171358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 171458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 171558da7f69SStefano Zampini } 171658da7f69SStefano Zampini } else { 171758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 171858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 171912edc857SStefano Zampini } 172012edc857SStefano Zampini } 1721674ae819SStefano Zampini PetscFunctionReturn(0); 1722674ae819SStefano Zampini } 1723674ae819SStefano Zampini 1724984c4197SStefano Zampini /* uncomment for testing purposes */ 1725984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1726674ae819SStefano Zampini #undef __FUNCT__ 1727674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1728674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1729674ae819SStefano Zampini { 1730674ae819SStefano Zampini PetscErrorCode ierr; 1731674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1732674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1733674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1734984c4197SStefano Zampini /* one and zero */ 1735984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1736984c4197SStefano Zampini /* space to store constraints and their local indices */ 1737984c4197SStefano Zampini PetscScalar *temp_quadrature_constraint; 1738984c4197SStefano Zampini PetscInt *temp_indices,*temp_indices_to_constraint,*temp_indices_to_constraint_B; 1739984c4197SStefano Zampini /* iterators */ 1740984c4197SStefano Zampini PetscInt i,j,k,total_counts,temp_start_ptr; 1741984c4197SStefano Zampini /* BLAS integers */ 1742e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1743e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1744c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1745727cdba6SStefano Zampini /* reuse */ 1746727cdba6SStefano Zampini PetscInt olocal_primal_size; 1747727cdba6SStefano Zampini PetscInt *oprimal_indices_local_idxs; 1748984c4197SStefano Zampini /* change of basis */ 1749984c4197SStefano Zampini PetscInt *aux_primal_numbering,*aux_primal_minloc,*global_indices; 1750a717540cSStefano Zampini PetscBool boolforchange,qr_needed; 1751a717540cSStefano Zampini PetscBT touched,change_basis,qr_needed_idx; 1752984c4197SStefano Zampini /* auxiliary stuff */ 175339e2fb2aSStefano Zampini PetscInt *nnz,*is_indices,*aux_primal_numbering_B; 1754a58a30b4SStefano Zampini PetscInt ncc,*gidxs=NULL,*permutation=NULL,*temp_indices_to_constraint_work=NULL; 1755a58a30b4SStefano Zampini PetscScalar *temp_quadrature_constraint_work=NULL; 1756984c4197SStefano Zampini /* some quantities */ 175745a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1758a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1759984c4197SStefano Zampini 1760674ae819SStefano Zampini PetscFunctionBegin; 17618e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 17628e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 17638e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1764cf5a6209SStefano Zampini 1765cf5a6209SStefano Zampini /* print some info */ 1766cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1767cf5a6209SStefano Zampini IS vertices; 1768cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1769cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1770cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1771cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1772cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1773cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1774cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1775fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1776fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1777cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1778cf5a6209SStefano Zampini } 1779cf5a6209SStefano Zampini 1780cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 1781cf5a6209SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges,*used_IS; 1782cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1783cf5a6209SStefano Zampini const Vec *nearnullvecs; 1784cf5a6209SStefano Zampini Vec *localnearnullsp; 1785cf5a6209SStefano Zampini PetscScalar *array; 1786cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1787cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1788cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1789cf5a6209SStefano Zampini PetscBool skip_lapack; 1790cf5a6209SStefano Zampini PetscScalar *work; 1791cf5a6209SStefano Zampini PetscReal *singular_vals; 1792cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1793cf5a6209SStefano Zampini PetscReal *rwork; 1794cf5a6209SStefano Zampini #endif 1795cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1796cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1797cf5a6209SStefano Zampini #else 1798*964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1799*964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1800cf5a6209SStefano Zampini #endif 1801cf5a6209SStefano Zampini 1802674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1803d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1804d06fc5fdSStefano Zampini /* free unneeded index sets */ 1805d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1806d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1807674ae819SStefano Zampini } 1808d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1809d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1810d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1811d06fc5fdSStefano Zampini } 1812d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1813d06fc5fdSStefano Zampini n_ISForEdges = 0; 1814d06fc5fdSStefano Zampini } 1815d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1816d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1817d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1818d06fc5fdSStefano Zampini } 1819d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1820d06fc5fdSStefano Zampini n_ISForFaces = 0; 1821d06fc5fdSStefano Zampini } 1822d06fc5fdSStefano Zampini /* HACKS (the following two blocks of code) */ 1823b9b85e73SStefano Zampini if (!ISForVertices && pcbddc->NullSpace && !pcbddc->user_ChangeOfBasisMatrix) { 1824b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1825d06fc5fdSStefano Zampini if (!ISForEdges) { 1826d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1827d06fc5fdSStefano Zampini } 1828b8ffe317SStefano Zampini } 182998a51de6SStefano Zampini if (pcbddc->NullSpace) { 183098a51de6SStefano Zampini /* use_change_of_basis should be consistent among processors */ 1831d06fc5fdSStefano Zampini PetscBool tbool[2],gbool[2]; 1832d06fc5fdSStefano Zampini tbool [0] = pcbddc->use_change_of_basis; 1833d06fc5fdSStefano Zampini tbool [1] = pcbddc->use_change_on_faces; 1834d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1835d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1836d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 183798a51de6SStefano Zampini } 183808122e43SStefano Zampini 1839674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1840674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1841674ae819SStefano Zampini if (nearnullsp) { 1842674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1843f4ddd8eeSStefano Zampini /* remove any stored info */ 1844f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1845f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1846f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1847f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1848f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1849473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1850f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1851f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1852f4ddd8eeSStefano Zampini } 1853984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1854984c4197SStefano Zampini nnsp_size = 0; 1855674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1856674ae819SStefano Zampini } 1857984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1858984c4197SStefano Zampini max_constraints = nnsp_size; 1859984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1860984c4197SStefano Zampini 1861674ae819SStefano Zampini /* 1862674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 1863674ae819SStefano Zampini - temp_indices will contain start index of each constraint stored as follows 1864911cabfeSStefano Zampini - temp_indices_to_constraint [temp_indices[i],...,temp_indices[i+1]-1] will contain the indices (in local numbering) on which the constraint acts 1865911cabfeSStefano Zampini - temp_indices_to_constraint_B[temp_indices[i],...,temp_indices[i+1]-1] will contain the indices (in boundary numbering) on which the constraint acts 1866911cabfeSStefano Zampini - temp_quadrature_constraint [temp_indices[i],...,temp_indices[i+1]-1] will contain the scalars representing the constraint itself 1867674ae819SStefano Zampini */ 1868674ae819SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 1869984c4197SStefano Zampini total_counts *= max_constraints; 1870674ae819SStefano Zampini n_vertices = 0; 1871674ae819SStefano Zampini if (ISForVertices) { 1872674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1873674ae819SStefano Zampini } 1874674ae819SStefano Zampini total_counts += n_vertices; 1875854ce69bSBarry Smith ierr = PetscMalloc1(total_counts+1,&temp_indices);CHKERRQ(ierr); 18764641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 1877674ae819SStefano Zampini total_counts = 0; 1878674ae819SStefano Zampini max_size_of_constraint = 0; 1879674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 1880674ae819SStefano Zampini if (i<n_ISForEdges) { 1881674ae819SStefano Zampini used_IS = &ISForEdges[i]; 1882674ae819SStefano Zampini } else { 1883674ae819SStefano Zampini used_IS = &ISForFaces[i-n_ISForEdges]; 1884674ae819SStefano Zampini } 1885674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&j);CHKERRQ(ierr); 1886674ae819SStefano Zampini total_counts += j; 1887674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1888674ae819SStefano Zampini } 1889984c4197SStefano Zampini total_counts *= max_constraints; 1890674ae819SStefano Zampini total_counts += n_vertices; 1891906d46d4SStefano Zampini ierr = PetscMalloc3(total_counts,&temp_quadrature_constraint,total_counts,&temp_indices_to_constraint,total_counts,&temp_indices_to_constraint_B);CHKERRQ(ierr); 1892984c4197SStefano Zampini /* get local part of global near null space vectors */ 1893785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1894984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1895984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1896984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1897984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1898984c4197SStefano Zampini } 1899674ae819SStefano Zampini 1900242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1901242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1902a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1903242a89d7SStefano Zampini 19048f1c130eSStefano Zampini /* allocate some auxiliary stuff */ 19058f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 19068f1c130eSStefano Zampini ierr = PetscMalloc4(max_size_of_constraint,&gidxs,max_size_of_constraint,&permutation,max_size_of_constraint,&temp_indices_to_constraint_work,max_size_of_constraint,&temp_quadrature_constraint_work);CHKERRQ(ierr); 1907f9208b2eSJed Brown } else { 1908f9208b2eSJed Brown gidxs = NULL; 1909f9208b2eSJed Brown permutation = NULL; 1910f9208b2eSJed Brown temp_indices_to_constraint_work = NULL; 1911f9208b2eSJed Brown temp_quadrature_constraint_work = NULL; 19128f1c130eSStefano Zampini } 19138f1c130eSStefano Zampini 1914984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1915a773dcb8SStefano Zampini if (!skip_lapack) { 1916674ae819SStefano Zampini PetscScalar temp_work; 1917911cabfeSStefano Zampini 1918674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1919984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1920785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 1921785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 1922785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 1923674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1924785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 1925674ae819SStefano Zampini #endif 1926674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1927c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 1928c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 1929674ae819SStefano Zampini lwork = -1; 1930674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1931674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1932c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 1933674ae819SStefano Zampini #else 1934c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 1935674ae819SStefano Zampini #endif 1936674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1937984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 1938674ae819SStefano Zampini #else /* on missing GESVD */ 1939674ae819SStefano Zampini /* SVD */ 1940674ae819SStefano Zampini PetscInt max_n,min_n; 1941674ae819SStefano Zampini max_n = max_size_of_constraint; 1942984c4197SStefano Zampini min_n = max_constraints; 1943984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 1944674ae819SStefano Zampini min_n = max_size_of_constraint; 1945984c4197SStefano Zampini max_n = max_constraints; 1946674ae819SStefano Zampini } 1947785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 1948674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1949785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 1950674ae819SStefano Zampini #endif 1951674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1952674ae819SStefano Zampini lwork = -1; 1953e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 1954e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 1955b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 1956674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1957674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1958cf5a6209SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr)); 1959674ae819SStefano Zampini #else 1960cf5a6209SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr)); 1961674ae819SStefano Zampini #endif 1962674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1963984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 1964984c4197SStefano Zampini #endif /* on missing GESVD */ 1965674ae819SStefano Zampini /* Allocate optimal workspace */ 1966674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 1967854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 1968674ae819SStefano Zampini } 1969674ae819SStefano Zampini /* Now we can loop on constraining sets */ 1970674ae819SStefano Zampini total_counts = 0; 1971674ae819SStefano Zampini temp_indices[0] = 0; 1972674ae819SStefano Zampini /* vertices */ 1973674ae819SStefano Zampini if (ISForVertices) { 1974674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1975674ae819SStefano Zampini if (nnsp_has_cnst) { /* consider all vertices */ 197639e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 1977674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1978674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1979674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1980674ae819SStefano Zampini total_counts++; 1981674ae819SStefano Zampini } 1982674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 1983984c4197SStefano Zampini PetscBool used_vertex; 1984674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1985674ae819SStefano Zampini used_vertex = PETSC_FALSE; 1986674ae819SStefano Zampini k = 0; 1987674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 1988984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1989984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 1990674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i]; 1991674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1992674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1993674ae819SStefano Zampini total_counts++; 1994674ae819SStefano Zampini used_vertex = PETSC_TRUE; 1995674ae819SStefano Zampini } 1996984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1997674ae819SStefano Zampini k++; 1998674ae819SStefano Zampini } 1999674ae819SStefano Zampini } 2000674ae819SStefano Zampini } 2001674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2002674ae819SStefano Zampini n_vertices = total_counts; 2003674ae819SStefano Zampini } 2004984c4197SStefano Zampini 2005674ae819SStefano Zampini /* edges and faces */ 2006911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 2007911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 2008911cabfeSStefano Zampini used_IS = &ISForEdges[ncc]; 2009984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2010674ae819SStefano Zampini } else { 2011911cabfeSStefano Zampini used_IS = &ISForFaces[ncc-n_ISForEdges]; 2012984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2013674ae819SStefano Zampini } 2014674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 2015674ae819SStefano Zampini temp_start_ptr = total_counts; /* need to know the starting index of constraints stored */ 2016674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&size_of_constraint);CHKERRQ(ierr); 2017674ae819SStefano Zampini ierr = ISGetIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2018984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2019984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2020674ae819SStefano Zampini if (nnsp_has_cnst) { 20215b08dc53SStefano Zampini PetscScalar quad_value; 2022674ae819SStefano Zampini temp_constraints++; 2023a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2024674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2025a773dcb8SStefano Zampini } else { 2026a773dcb8SStefano Zampini quad_value = 1.0; 2027a773dcb8SStefano Zampini } 202839e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2029674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 2030674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=quad_value; 2031674ae819SStefano Zampini } 203208122e43SStefano Zampini /* sort by global ordering if using lapack subroutines (not needed!) */ 20338f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 2034911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 2035911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2036911cabfeSStefano Zampini permutation[j]=j; 2037911cabfeSStefano Zampini } 2038911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 2039911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 204008122e43SStefano Zampini if (permutation[j]!=j) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP,"This should not happen"); 204108122e43SStefano Zampini } 204208122e43SStefano Zampini for (j=0;j<size_of_constraint;j++) { 2043911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 2044911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 2045911cabfeSStefano Zampini } 2046911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2047911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 2048911cabfeSStefano Zampini } 2049674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 2050674ae819SStefano Zampini total_counts++; 2051674ae819SStefano Zampini } 2052674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2053984c4197SStefano Zampini PetscReal real_value; 2054984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 205539e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2056674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 2057984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=array[is_indices[j]]; 2058674ae819SStefano Zampini } 2059984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2060984c4197SStefano Zampini /* check if array is null on the connected component */ 2061e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2062e310c8b4SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_one)); 20635b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2064911cabfeSStefano Zampini /* sort by global ordering if using lapack subroutines */ 20658f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 2066911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 2067911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2068911cabfeSStefano Zampini permutation[j]=j; 2069911cabfeSStefano Zampini } 2070911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 2071911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2072911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 2073911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 2074911cabfeSStefano Zampini } 2075911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 2076911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 2077911cabfeSStefano Zampini } 2078674ae819SStefano Zampini temp_constraints++; 2079674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 2080674ae819SStefano Zampini total_counts++; 2081674ae819SStefano Zampini } 2082674ae819SStefano Zampini } 2083674ae819SStefano Zampini ierr = ISRestoreIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 208445a1bb75SStefano Zampini valid_constraints = temp_constraints; 2085eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2086a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 2087a773dcb8SStefano Zampini PetscScalar norm; 2088a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2089a773dcb8SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one)); 2090a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 2091a773dcb8SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one)); 2092a773dcb8SStefano Zampini } else { /* perform SVD */ 2093984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 2094674ae819SStefano Zampini 2095674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2096984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2097984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2098984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2099984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2100984c4197SStefano Zampini from that computed using LAPACKgesvd 2101984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2102984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2103984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2104674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2105e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2106984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2107674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2108674ae819SStefano Zampini for (k=0;k<j+1;k++) { 2109cf5a6209SStefano Zampini PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k]=BLASdot_(&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr+k]],&Blas_one,&temp_quadrature_constraint[temp_indices[temp_start_ptr+j]],&Blas_one)); 2110674ae819SStefano Zampini } 2111674ae819SStefano Zampini } 2112e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2113e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2114e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2115674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2116c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2117674ae819SStefano Zampini #else 2118c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2119674ae819SStefano Zampini #endif 2120674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2121984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2122984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2123674ae819SStefano Zampini j = 0; 2124984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2125674ae819SStefano Zampini total_counts = total_counts-j; 212645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2127e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2128c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2129c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2130c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2131c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2132c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2133c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2134674ae819SStefano Zampini if (j<temp_constraints) { 2135984c4197SStefano Zampini PetscInt ii; 2136984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k]=1.0/PetscSqrtReal(singular_vals[k]); 2137674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2138c4303822SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC)); 2139674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2140984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2141674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 2142984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[temp_start_ptr+k]+ii]=singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii]; 2143674ae819SStefano Zampini } 2144674ae819SStefano Zampini } 2145674ae819SStefano Zampini } 2146674ae819SStefano Zampini #else /* on missing GESVD */ 2147e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2148e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2149b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2150674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2151674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2152cf5a6209SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr)); 2153674ae819SStefano Zampini #else 2154cf5a6209SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr)); 2155674ae819SStefano Zampini #endif 2156984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2157674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2158984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2159e310c8b4SStefano Zampini k = temp_constraints; 2160e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2161674ae819SStefano Zampini j = 0; 2162e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 216345a1bb75SStefano Zampini valid_constraints = k-j; 2164911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2165984c4197SStefano Zampini #endif /* on missing GESVD */ 2166674ae819SStefano Zampini } 2167a773dcb8SStefano Zampini } 216845a1bb75SStefano Zampini /* setting change_of_basis flag is safe now */ 216945a1bb75SStefano Zampini if (boolforchange) { 217045a1bb75SStefano Zampini for (j=0;j<valid_constraints;j++) { 217145a1bb75SStefano Zampini PetscBTSet(change_basis,total_counts-j-1); 217245a1bb75SStefano Zampini } 217345a1bb75SStefano Zampini } 2174674ae819SStefano Zampini } 2175984c4197SStefano Zampini /* free workspace */ 21768f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 2177911cabfeSStefano Zampini ierr = PetscFree4(gidxs,permutation,temp_indices_to_constraint_work,temp_quadrature_constraint_work);CHKERRQ(ierr); 21788f1c130eSStefano Zampini } 21798f1c130eSStefano Zampini if (!skip_lapack) { 2180984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2181984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2182984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2183984c4197SStefano Zampini #endif 2184984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2185984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2186984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2187984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2188984c4197SStefano Zampini #endif 2189984c4197SStefano Zampini } 2190984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2191984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2192984c4197SStefano Zampini } 2193984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2194cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2195cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2196cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2197cf5a6209SStefano Zampini } 2198cf5a6209SStefano Zampini if (n_ISForFaces) { 2199cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2200cf5a6209SStefano Zampini } 2201cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2202cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2203cf5a6209SStefano Zampini } 2204cf5a6209SStefano Zampini if (n_ISForEdges) { 2205cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2206cf5a6209SStefano Zampini } 2207cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 220808122e43SStefano Zampini } else { 220908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 221008122e43SStefano Zampini PetscInt cum = 0; 221108122e43SStefano Zampini 221208122e43SStefano Zampini total_counts = 0; 221308122e43SStefano Zampini n_vertices = 0; 221408122e43SStefano Zampini if (sub_schurs->is_Ej_com) { 221508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&n_vertices);CHKERRQ(ierr); 221608122e43SStefano Zampini } 221708122e43SStefano Zampini max_constraints = 0; 221808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 221908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 222008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 222108122e43SStefano Zampini } 222208122e43SStefano Zampini temp_indices = pcbddc->adaptive_constraints_ptrs; 222308122e43SStefano Zampini temp_indices_to_constraint = pcbddc->adaptive_constraints_idxs; 222408122e43SStefano Zampini temp_quadrature_constraint = pcbddc->adaptive_constraints_data; 222508122e43SStefano Zampini 222608122e43SStefano Zampini #if 0 222708122e43SStefano Zampini printf("Found %d totals\n",total_counts); 222808122e43SStefano Zampini for (i=0;i<total_counts;i++) { 222908122e43SStefano Zampini printf("const %d, start %d",i,temp_indices[i]); 223008122e43SStefano Zampini printf(" end %d:\n",temp_indices[i+1]); 223108122e43SStefano Zampini for (j=temp_indices[i];j<temp_indices[i+1];j++) { 223208122e43SStefano Zampini printf(" idxs %d",temp_indices_to_constraint[j]); 223308122e43SStefano Zampini printf(" data %1.2e\n",temp_quadrature_constraint[j]); 223408122e43SStefano Zampini } 223508122e43SStefano Zampini } 22361b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 22371b968477SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i+n_vertices]); 22381b968477SStefano Zampini } 22391b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 22401b968477SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] sub %d, edge %d, n %d\n",PetscGlobalRank,i,PetscBTLookup(sub_schurs->is_edge,i),pcbddc->adaptive_constraints_n[i+n_vertices]); 22411b968477SStefano Zampini } 224208122e43SStefano Zampini #endif 224308122e43SStefano Zampini 224408122e43SStefano Zampini for (i=0;i<total_counts;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,temp_indices[i+1]-temp_indices[i]); 224508122e43SStefano Zampini ierr = PetscMalloc1(temp_indices[total_counts],&temp_indices_to_constraint_B);CHKERRQ(ierr); 224608122e43SStefano Zampini /* Change of basis */ 224708122e43SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 224808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 2249bbb9e6c6SStefano Zampini cum = n_vertices; 225008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 225108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2252bbb9e6c6SStefano Zampini for (j=0;j<pcbddc->adaptive_constraints_n[i+n_vertices];j++) { 225308122e43SStefano Zampini ierr = PetscBTSet(change_basis,cum+j);CHKERRQ(ierr); 225408122e43SStefano Zampini } 225508122e43SStefano Zampini } 2256bbb9e6c6SStefano Zampini cum += pcbddc->adaptive_constraints_n[i+n_vertices]; 225708122e43SStefano Zampini } 225808122e43SStefano Zampini } 225908122e43SStefano Zampini } 226008122e43SStefano Zampini 226108122e43SStefano Zampini /* map temp_indices_to_constraint in boundary numbering */ 226208122e43SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,temp_indices[total_counts],temp_indices_to_constraint,&i,temp_indices_to_constraint_B);CHKERRQ(ierr); 226308122e43SStefano Zampini if (i != temp_indices[total_counts]) { 226408122e43SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",temp_indices[total_counts],i); 226508122e43SStefano Zampini } 2266984c4197SStefano Zampini 2267727cdba6SStefano Zampini /* set quantities in pcbddc data structure and store previous primal size */ 2268984c4197SStefano Zampini /* n_vertices defines the number of subdomain corners in the primal space */ 2269674ae819SStefano Zampini /* n_constraints defines the number of averages (they can be point primal dofs if change of basis is requested) */ 2270727cdba6SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 2271984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2272674ae819SStefano Zampini pcbddc->n_vertices = n_vertices; 2273984c4197SStefano Zampini pcbddc->n_constraints = pcbddc->local_primal_size-pcbddc->n_vertices; 2274674ae819SStefano Zampini 2275674ae819SStefano Zampini /* Create constraint matrix */ 2276674ae819SStefano Zampini /* The constraint matrix is used to compute the l2g map of primal dofs */ 2277674ae819SStefano Zampini /* so we need to set it up properly either with or without change of basis */ 2278674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 227916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2280984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2281984c4197SStefano Zampini /* array to compute a local numbering of constraints : vertices first then constraints */ 2282785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_numbering);CHKERRQ(ierr); 2283984c4197SStefano Zampini /* array to select the proper local node (of minimum index with respect to global ordering) when changing the basis */ 2284984c4197SStefano Zampini /* note: it should not be needed since IS for faces and edges are already sorted by global ordering when analyzing the graph but... just in case */ 2285785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_minloc);CHKERRQ(ierr); 2286984c4197SStefano Zampini /* auxiliary stuff for basis change */ 2287785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&global_indices);CHKERRQ(ierr); 22884641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&touched);CHKERRQ(ierr); 2289984c4197SStefano Zampini 2290984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2291984c4197SStefano Zampini total_primal_vertices=0; 2292984c4197SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 2293674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 2294984c4197SStefano Zampini if (size_of_constraint == 1) { 22954641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]]);CHKERRQ(ierr); 2296984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]]; 2297984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=0; 2298984c4197SStefano Zampini total_primal_vertices++; 22994641a718SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { /* Same procedure used in PCBDDCGetPrimalConstraintsLocalIdx */ 2300984c4197SStefano Zampini PetscInt min_loc,min_index; 2301984c4197SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],global_indices);CHKERRQ(ierr); 2302984c4197SStefano Zampini /* find first untouched local node */ 2303674ae819SStefano Zampini k = 0; 23044641a718SStefano Zampini while (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) k++; 2305984c4197SStefano Zampini min_index = global_indices[k]; 2306984c4197SStefano Zampini min_loc = k; 2307984c4197SStefano Zampini /* search the minimum among global nodes already untouched on the cc */ 2308984c4197SStefano Zampini for (k=1;k<size_of_constraint;k++) { 2309984c4197SStefano Zampini /* there can be more than one constraint on a single connected component */ 23104641a718SStefano Zampini if (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k]) && min_index > global_indices[k]) { 2311984c4197SStefano Zampini min_index = global_indices[k]; 2312984c4197SStefano Zampini min_loc = k; 2313674ae819SStefano Zampini } 2314674ae819SStefano Zampini } 23154641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]+min_loc]);CHKERRQ(ierr); 2316984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]+min_loc]; 2317984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=min_loc; 2318984c4197SStefano Zampini total_primal_vertices++; 2319984c4197SStefano Zampini } 2320984c4197SStefano Zampini } 2321a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2322a717540cSStefano Zampini qr_needed = PETSC_FALSE; 2323a717540cSStefano Zampini ierr = PetscBTCreate(pcbddc->local_primal_size,&qr_needed_idx);CHKERRQ(ierr); 2324a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 2325a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 2326fa434743SStefano Zampini if (!pcbddc->use_qr_single) { 2327a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 2328a717540cSStefano Zampini j = 0; 2329a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 2330a717540cSStefano Zampini if (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) { 2331a717540cSStefano Zampini j++; 2332a717540cSStefano Zampini } 2333a717540cSStefano Zampini } 2334a717540cSStefano Zampini /* found more than one primal dof on the cc */ 2335a717540cSStefano Zampini if (j > 1) { 2336a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2337a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2338a717540cSStefano Zampini } 2339fa434743SStefano Zampini } else { 2340fa434743SStefano Zampini PetscBTSet(qr_needed_idx,i); 2341fa434743SStefano Zampini qr_needed = PETSC_TRUE; 2342fa434743SStefano Zampini } 2343a717540cSStefano Zampini } 2344a717540cSStefano Zampini } 2345984c4197SStefano Zampini /* free workspace */ 2346984c4197SStefano Zampini ierr = PetscFree(global_indices);CHKERRQ(ierr); 234745a1bb75SStefano Zampini 2348674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 2349a717540cSStefano Zampini ierr = PetscSortInt(total_primal_vertices,aux_primal_numbering);CHKERRQ(ierr); 2350984c4197SStefano Zampini 2351984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 2352785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2353984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i]=1; 2354984c4197SStefano Zampini j=total_primal_vertices; 2355984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 23564641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2357674ae819SStefano Zampini nnz[j]=temp_indices[i+1]-temp_indices[i]; 2358674ae819SStefano Zampini j++; 2359674ae819SStefano Zampini } 2360674ae819SStefano Zampini } 2361674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2362674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2363674ae819SStefano Zampini /* set values in constraint matrix */ 2364984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 2365984c4197SStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,aux_primal_numbering[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2366674ae819SStefano Zampini } 2367984c4197SStefano Zampini total_counts = total_primal_vertices; 2368984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 23694641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2370674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 2371674ae819SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&total_counts,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],&temp_quadrature_constraint[temp_indices[i]],INSERT_VALUES);CHKERRQ(ierr); 2372674ae819SStefano Zampini total_counts++; 2373674ae819SStefano Zampini } 2374674ae819SStefano Zampini } 2375674ae819SStefano Zampini /* assembling */ 2376674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2377674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2378984c4197SStefano Zampini /* 237945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2380984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2381984c4197SStefano Zampini */ 2382674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2383674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2384026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2385984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2386026de310SStefano Zampini /* iterator on aux_primal_minloc (ordered as read from nearnullspace: vertices, edges and then constraints) */ 2387026de310SStefano Zampini PetscInt primal_counter; 2388984c4197SStefano Zampini /* working stuff for GEQRF */ 238981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2390984c4197SStefano Zampini PetscBLASInt lqr_work; 2391984c4197SStefano Zampini /* working stuff for UNGQR */ 2392984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2393984c4197SStefano Zampini PetscBLASInt lgqr_work; 2394984c4197SStefano Zampini /* working stuff for TRTRS */ 2395984c4197SStefano Zampini PetscScalar *trs_rhs; 23963f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2397984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2398984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2399984c4197SStefano Zampini PetscScalar *start_vals; 2400984c4197SStefano Zampini /* working stuff for values insertion */ 24014641a718SStefano Zampini PetscBT is_primal; 2402906d46d4SStefano Zampini /* matrix sizes */ 2403906d46d4SStefano Zampini PetscInt global_size,local_size; 2404906d46d4SStefano Zampini /* temporary change of basis */ 2405906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2406cf5a6209SStefano Zampini /* extra space for debugging */ 2407cf5a6209SStefano Zampini PetscScalar *dbg_work; 2408984c4197SStefano Zampini 2409906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2410906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 241116f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2412bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2413906d46d4SStefano Zampini /* nonzeros for local mat */ 2414bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2415bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 2416a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 2417a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 2418a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 2419a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 2420bbb9e6c6SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint[temp_indices[i]+j]] = size_of_constraint; 2421a717540cSStefano Zampini } else { 2422bbb9e6c6SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint[temp_indices[i]+j]] = 2; 2423a717540cSStefano Zampini /* get local primal index on the cc */ 2424a717540cSStefano Zampini j = 0; 2425a717540cSStefano Zampini while (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+j])) j++; 2426bbb9e6c6SStefano Zampini nnz[temp_indices_to_constraint[temp_indices[i]+j]] = size_of_constraint; 2427a717540cSStefano Zampini } 2428a717540cSStefano Zampini } 2429a717540cSStefano Zampini } 2430906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2431bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2432a717540cSStefano Zampini /* Set initial identity in the matrix */ 2433bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2434906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2435a717540cSStefano Zampini } 2436a717540cSStefano Zampini 2437a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2438a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2439a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2440a717540cSStefano Zampini } 2441a717540cSStefano Zampini 2442a717540cSStefano Zampini 2443a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2444a717540cSStefano Zampini /* 2445a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2446a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2447a717540cSStefano Zampini 2448a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2449a717540cSStefano Zampini 2450a6b551f4SStefano 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) 2451a6b551f4SStefano Zampini 2452a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2453a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2454a717540cSStefano Zampini | ... | 2455a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2456a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2457a717540cSStefano Zampini 2458a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2459a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2460a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2461a6b551f4SStefano Zampini 2462a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2463a717540cSStefano Zampini */ 2464a717540cSStefano Zampini if (qr_needed) { 2465984c4197SStefano Zampini /* space to store Q */ 2466854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2467984c4197SStefano Zampini /* first we issue queries for optimal work */ 24683f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 24693f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 24703f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2471984c4197SStefano Zampini lqr_work = -1; 24723f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2473984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2474984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2475785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2476984c4197SStefano Zampini lgqr_work = -1; 24773f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 24783f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 24793f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 24803f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 24813f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 24823f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2483984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2484984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2485785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2486984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2487785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2488984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2489785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2490a717540cSStefano Zampini /* allocating workspace for check */ 2491a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2492cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2493a717540cSStefano Zampini } 2494a717540cSStefano Zampini } 2495984c4197SStefano Zampini /* array to store whether a node is primal or not */ 24964641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2497473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 24985e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,aux_primal_numbering,&i,aux_primal_numbering_B);CHKERRQ(ierr); 249939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 250039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 25014641a718SStefano Zampini } 250239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 250339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 250439e2fb2aSStefano Zampini } 250539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2506984c4197SStefano Zampini 2507a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 2508026de310SStefano Zampini /* -> using implicit ordering contained in temp_indices data */ 2509026de310SStefano Zampini total_counts = pcbddc->n_vertices; 2510026de310SStefano Zampini primal_counter = total_counts; 2511026de310SStefano Zampini while (total_counts<pcbddc->local_primal_size) { 2512026de310SStefano Zampini primal_dofs = 1; 25134641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2514026de310SStefano Zampini /* get all constraints with same support: if more then one constraint is present on the cc then surely indices are stored contiguosly */ 2515bbb9e6c6SStefano Zampini while (total_counts+primal_dofs < pcbddc->local_primal_size && temp_indices_to_constraint[temp_indices[total_counts]] == temp_indices_to_constraint[temp_indices[total_counts+primal_dofs]]) { 2516026de310SStefano Zampini primal_dofs++; 2517674ae819SStefano Zampini } 2518984c4197SStefano Zampini /* get constraint info */ 2519026de310SStefano Zampini size_of_constraint = temp_indices[total_counts+1]-temp_indices[total_counts]; 2520984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2521984c4197SStefano Zampini 2522984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2523a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d to %d (incl) need a change of basis (size %d)\n",total_counts,total_counts+primal_dofs-1,size_of_constraint);CHKERRQ(ierr); 2524674ae819SStefano Zampini } 2525984c4197SStefano Zampini 2526fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2527a717540cSStefano Zampini 2528a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2529a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2530cf5a6209SStefano Zampini ierr = PetscMemcpy(dbg_work,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2531a717540cSStefano Zampini } 2532984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 2533026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2534984c4197SStefano Zampini 2535984c4197SStefano Zampini /* compute QR decomposition of constraints */ 25363f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25373f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2539674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25403f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2541984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2542674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2543984c4197SStefano Zampini 2544984c4197SStefano Zampini /* explictly compute R^-T */ 2545984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2546984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 25473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25483f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 25493f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2551984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25523f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2553984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2554984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2555984c4197SStefano Zampini 2556a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 25573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 25603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2561984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25623f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2563984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2564984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2565984c4197SStefano Zampini 2566984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2567984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2568984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 25693f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 25703f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 25713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 25723f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 25733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 25743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2575984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2576c4303822SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_LDC)); 2577984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2578026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2579984c4197SStefano Zampini 2580984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 2581bbb9e6c6SStefano Zampini start_rows = &temp_indices_to_constraint[temp_indices[total_counts]]; 2582984c4197SStefano Zampini /* insert cols for primal dofs */ 2583984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2584984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 2585bbb9e6c6SStefano Zampini start_cols = &temp_indices_to_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter+j]]; 2586906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2587984c4197SStefano Zampini } 2588984c4197SStefano Zampini /* insert cols for dual dofs */ 2589984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 25904641a718SStefano Zampini if (!PetscBTLookup(is_primal,temp_indices_to_constraint_B[temp_indices[total_counts]+k])) { 2591984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 2592bbb9e6c6SStefano Zampini start_cols = &temp_indices_to_constraint[temp_indices[total_counts]+k]; 2593906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2594984c4197SStefano Zampini j++; 2595674ae819SStefano Zampini } 2596674ae819SStefano Zampini } 2597984c4197SStefano Zampini 2598984c4197SStefano Zampini /* check change of basis */ 2599984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2600984c4197SStefano Zampini PetscInt ii,jj; 2601984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2602c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2603c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2604c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2605c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2606c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2607c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2608984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2609cf5a6209SStefano 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)); 2610984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2611984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2612984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2613cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2614cf5a6209SStefano 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; 2615674ae819SStefano Zampini } 2616674ae819SStefano Zampini } 2617984c4197SStefano Zampini if (!valid_qr) { 261822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2619984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2620984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2621cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2622cf5a6209SStefano 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])); 2623674ae819SStefano Zampini } 2624cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2625cf5a6209SStefano 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])); 2626984c4197SStefano Zampini } 2627984c4197SStefano Zampini } 2628984c4197SStefano Zampini } 2629674ae819SStefano Zampini } else { 263022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2631674ae819SStefano Zampini } 2632674ae819SStefano Zampini } 2633a717540cSStefano Zampini } else { /* simple transformation block */ 2634a717540cSStefano Zampini PetscInt row,col; 2635a6b551f4SStefano Zampini PetscScalar val,norm; 2636a6b551f4SStefano Zampini 2637a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2638a6b551f4SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one)); 2639a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2640bbb9e6c6SStefano Zampini PetscInt row_B = temp_indices_to_constraint_B[temp_indices[total_counts]+j]; 2641bbb9e6c6SStefano Zampini row = temp_indices_to_constraint[temp_indices[total_counts]+j]; 2642bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 2643bbb9e6c6SStefano Zampini col = temp_indices_to_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2644906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 2645a6b551f4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,temp_quadrature_constraint[temp_indices[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2646a717540cSStefano Zampini } else { 2647a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 2648bbb9e6c6SStefano Zampini col = temp_indices_to_constraint[temp_indices[total_counts]+k]; 2649a717540cSStefano Zampini if (row != col) { 2650a717540cSStefano Zampini val = -temp_quadrature_constraint[temp_indices[total_counts]+k]/temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2651a717540cSStefano Zampini } else { 2652a6b551f4SStefano Zampini val = temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]/norm; 2653a717540cSStefano Zampini } 2654906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2655a717540cSStefano Zampini } 2656a717540cSStefano Zampini } 2657a717540cSStefano Zampini } 265898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 265922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2660a717540cSStefano Zampini } 2661674ae819SStefano Zampini } 2662026de310SStefano Zampini /* increment primal counter */ 2663026de310SStefano Zampini primal_counter += primal_dofs; 2664984c4197SStefano Zampini } else { 2665984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2666026de310SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,temp_indices[total_counts+1]-temp_indices[total_counts]);CHKERRQ(ierr); 2667674ae819SStefano Zampini } 2668674ae819SStefano Zampini } 2669026de310SStefano Zampini /* increment constraint counter total_counts */ 2670026de310SStefano Zampini total_counts += primal_dofs; 2671674ae819SStefano Zampini } 2672a717540cSStefano Zampini 2673a717540cSStefano Zampini /* free workspace */ 2674a717540cSStefano Zampini if (qr_needed) { 2675984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2676cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2677984c4197SStefano Zampini } 2678984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2679984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2680984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2681984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2682984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2683674ae819SStefano Zampini } 2684a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2685906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2686906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2687906d46d4SStefano Zampini 2688906d46d4SStefano Zampini /* assembling of global change of variable */ 2689bbb9e6c6SStefano Zampini { 2690bbb9e6c6SStefano Zampini Mat tmat; 269116f15bc4SStefano Zampini PetscInt bs; 269216f15bc4SStefano Zampini 2693906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2694906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2695bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2696bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2697bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2698bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 269916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 270016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2701906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2702bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2703bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2704bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2705bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2706bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2707bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2708bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2709bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2710bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2711906d46d4SStefano Zampini } 2712906d46d4SStefano Zampini /* check */ 2713906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2714906d46d4SStefano Zampini PetscReal error; 2715906d46d4SStefano Zampini Vec x,x_change; 2716906d46d4SStefano Zampini 2717906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2718906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2719906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2720906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2721bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2722bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2723bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2724bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2725bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2726906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2727906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2728906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2729906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2730bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2731906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2732906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2733906d46d4SStefano Zampini } 2734b96c3477SStefano Zampini 2735b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2736b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2737b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2738b96c3477SStefano Zampini if (sub_schurs->n_subs_par_g) { 2739b7eb3628SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Change of basis with deluxe scaling and parallel problems still needs to be implemented"); 2740b96c3477SStefano Zampini } 2741b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2742b96c3477SStefano Zampini Mat S_1,S_2,tmat; 2743bbb9e6c6SStefano Zampini IS is_all_N; 2744bbb9e6c6SStefano Zampini 2745bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 2746d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr); 2747bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2748b96c3477SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_1);CHKERRQ(ierr); 2749b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2750b96c3477SStefano Zampini sub_schurs->S_Ej_all = S_1; 2751b96c3477SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_2);CHKERRQ(ierr); 2752b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2753b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = S_2; 2754b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2755b96c3477SStefano Zampini } 2756b96c3477SStefano Zampini } 2757906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2758906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2759b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2760b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2761b9b85e73SStefano Zampini } 2762906d46d4SStefano Zampini 2763906d46d4SStefano Zampini /* set up change of basis context */ 2764906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2765906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2766906d46d4SStefano Zampini 2767906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2768906d46d4SStefano Zampini PetscInt global_size,local_size; 2769906d46d4SStefano Zampini 2770906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2771906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2772906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2773906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2774906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2775906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2776906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2777906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2778906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2779906d46d4SStefano Zampini } else { 2780906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2781906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2782906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2783906d46d4SStefano Zampini } 2784906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2785906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2786906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2787906d46d4SStefano Zampini } else { 2788906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2789906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2790906d46d4SStefano Zampini } 2791906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2792906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2793906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2794906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2795b9b85e73SStefano Zampini } 2796a717540cSStefano Zampini 2797727cdba6SStefano Zampini /* get indices in local ordering for vertices and constraints */ 2798727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { /* if this is true, I need to check if a new primal space has been introduced */ 2799473ba861SJed Brown ierr = PetscMalloc1(olocal_primal_size,&oprimal_indices_local_idxs);CHKERRQ(ierr); 2800727cdba6SStefano Zampini ierr = PetscMemcpy(oprimal_indices_local_idxs,pcbddc->primal_indices_local_idxs,olocal_primal_size*sizeof(PetscInt));CHKERRQ(ierr); 2801727cdba6SStefano Zampini } 2802727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2803f347579bSStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2804473ba861SJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2805727cdba6SStefano Zampini ierr = PCBDDCGetPrimalVerticesLocalIdx(pc,&i,&aux_primal_numbering);CHKERRQ(ierr); 2806727cdba6SStefano Zampini ierr = PetscMemcpy(pcbddc->primal_indices_local_idxs,aux_primal_numbering,i*sizeof(PetscInt));CHKERRQ(ierr); 2807727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2808727cdba6SStefano Zampini ierr = PCBDDCGetPrimalConstraintsLocalIdx(pc,&j,&aux_primal_numbering);CHKERRQ(ierr); 2809727cdba6SStefano Zampini ierr = PetscMemcpy(&pcbddc->primal_indices_local_idxs[i],aux_primal_numbering,j*sizeof(PetscInt));CHKERRQ(ierr); 2810727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2811e9189074SStefano Zampini /* set quantities in PCBDDC data struct */ 2812e9189074SStefano Zampini pcbddc->n_actual_vertices = i; 2813727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2814727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2815727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2816727cdba6SStefano Zampini ierr = PetscMemcmp(pcbddc->primal_indices_local_idxs,oprimal_indices_local_idxs,olocal_primal_size,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2817c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2818727cdba6SStefano Zampini ierr = PetscFree(oprimal_indices_local_idxs);CHKERRQ(ierr); 2819727cdba6SStefano Zampini } 2820727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2821727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2822727cdba6SStefano Zampini 2823a717540cSStefano Zampini /* flush dbg viewer */ 2824b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2825b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2826b8ffe317SStefano Zampini } 2827a717540cSStefano Zampini 2828e310c8b4SStefano Zampini /* free workspace */ 2829a717540cSStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 2830a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 2831984c4197SStefano Zampini ierr = PetscFree(aux_primal_minloc);CHKERRQ(ierr); 28324641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 283308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 2834906d46d4SStefano Zampini ierr = PetscFree(temp_indices);CHKERRQ(ierr); 2835906d46d4SStefano Zampini ierr = PetscFree3(temp_quadrature_constraint,temp_indices_to_constraint,temp_indices_to_constraint_B);CHKERRQ(ierr); 283608122e43SStefano Zampini } else { 283708122e43SStefano Zampini ierr = PetscFree4(pcbddc->adaptive_constraints_n, 283808122e43SStefano Zampini pcbddc->adaptive_constraints_ptrs, 283908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 284008122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 284108122e43SStefano Zampini ierr = PetscFree(temp_indices_to_constraint_B);CHKERRQ(ierr); 284208122e43SStefano Zampini } 2843674ae819SStefano Zampini PetscFunctionReturn(0); 2844674ae819SStefano Zampini } 2845674ae819SStefano Zampini 2846674ae819SStefano Zampini #undef __FUNCT__ 2847674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2848674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2849674ae819SStefano Zampini { 2850674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2851674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2852674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 285363602bcaSStefano Zampini PetscInt ierr,i,vertex_size; 2854674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2855674ae819SStefano Zampini 2856674ae819SStefano Zampini PetscFunctionBegin; 28578e61c736SStefano Zampini /* Reset previously computed graph */ 28588e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2859674ae819SStefano Zampini /* Init local Graph struct */ 2860674ae819SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping);CHKERRQ(ierr); 2861674ae819SStefano Zampini 2862575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2863575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2864575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2865575ad6abSStefano Zampini } 28669577ea80SStefano Zampini 2867674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 28684d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 28694d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 28704d379d7bSStefano Zampini PetscInt nvtxs; 2871674ae819SStefano Zampini 28724d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 28732fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 28742fffb893SStefano Zampini 28752fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 28762fffb893SStefano Zampini if (flg_row) { 28774d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 2878b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 28792fffb893SStefano Zampini } 28802fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 28812fffb893SStefano Zampini } else if (pcbddc->current_level) { /* just compute subdomain's connected components for coarser levels */ 28824d379d7bSStefano Zampini IS is_dummy; 28834d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 28844d379d7bSStefano Zampini PetscInt j,sum; 28854d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 28864d379d7bSStefano Zampini const PetscInt *idxs; 28874d379d7bSStefano Zampini PCBDDCGraph graph; 28884d379d7bSStefano Zampini PetscBT is_on_boundary; 28894d379d7bSStefano Zampini 28904d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 28914d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 28924d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 28934d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 28944d379d7bSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy);CHKERRQ(ierr); 28954d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 28964d379d7bSStefano Zampini graph->xadj = xadj; 28974d379d7bSStefano Zampini graph->adjncy = adjncy; 28984d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 28994d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 29004d379d7bSStefano Zampini 29014d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 29024d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains\n",PetscGlobalRank,graph->ncc);CHKERRQ(ierr); 29034d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29044d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 29054d379d7bSStefano Zampini } 29064d379d7bSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29074d379d7bSStefano Zampini } 29084d379d7bSStefano Zampini 29094d379d7bSStefano Zampini ierr = PetscBTCreate(nvtxs,&is_on_boundary);CHKERRQ(ierr); 29104d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29114d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 29124d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 29134d379d7bSStefano Zampini } 29144d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 29154d379d7bSStefano Zampini 29164d379d7bSStefano Zampini ierr = PetscCalloc1(nvtxs+1,&cxadj);CHKERRQ(ierr); 29174d379d7bSStefano Zampini sum = 0; 29184d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29194d379d7bSStefano Zampini PetscInt sizecc = 0; 29204d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29214d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29224d379d7bSStefano Zampini sizecc++; 29234d379d7bSStefano Zampini } 29244d379d7bSStefano Zampini } 29254d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29264d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29274d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 29284d379d7bSStefano Zampini } 29294d379d7bSStefano Zampini } 29304d379d7bSStefano Zampini sum += sizecc*sizecc; 29314d379d7bSStefano Zampini } 29324d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 29334d379d7bSStefano Zampini sum = 0; 29344d379d7bSStefano Zampini for (i=0;i<nvtxs;i++) { 29354d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 29364d379d7bSStefano Zampini cxadj[i] = sum; 29374d379d7bSStefano Zampini sum += temp; 29384d379d7bSStefano Zampini } 29394d379d7bSStefano Zampini cxadj[nvtxs] = sum; 29404d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 29414d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 29424d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 29434d379d7bSStefano Zampini PetscInt k,sizecc = 0; 29444d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 29454d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 29464d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 29474d379d7bSStefano Zampini sizecc++; 29484d379d7bSStefano Zampini } 29494d379d7bSStefano Zampini } 29504d379d7bSStefano Zampini } 29514d379d7bSStefano Zampini } 29524d379d7bSStefano Zampini } 29534d379d7bSStefano Zampini if (nvtxs) { 29544d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 29554d379d7bSStefano Zampini } else { 29564d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 29574d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 29584d379d7bSStefano Zampini } 29594d379d7bSStefano Zampini graph->xadj = 0; 29604d379d7bSStefano Zampini graph->adjncy = 0; 29614d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 29624d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 29634d379d7bSStefano Zampini } 2964674ae819SStefano Zampini } 2965674ae819SStefano Zampini 296663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 2967674ae819SStefano Zampini vertex_size = 1; 296863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 296963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 297095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 297163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 2972a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 297363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 2974674ae819SStefano Zampini } 297563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 297663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 297763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 2978674ae819SStefano Zampini } 297963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 2980674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 298163602bcaSStefano Zampini } else { 298263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 298363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 2984854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 298563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 298663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 298763602bcaSStefano Zampini } 298863602bcaSStefano Zampini } 2989674ae819SStefano Zampini } 2990674ae819SStefano Zampini 2991674ae819SStefano Zampini /* Setup of Graph */ 2992785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 2993a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 2994785d1243SStefano Zampini } 2995785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 2996a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2997785d1243SStefano Zampini } 299863602bcaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices); 2999674ae819SStefano Zampini 3000674ae819SStefano Zampini /* Graph's connected components analysis */ 3001674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3002674ae819SStefano Zampini 3003674ae819SStefano Zampini /* print some info to stdout */ 3004674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3005e49050b4SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer); 3006674ae819SStefano Zampini } 3007fb180af4SStefano Zampini 3008fb180af4SStefano Zampini /* mark topography has done */ 3009fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3010674ae819SStefano Zampini PetscFunctionReturn(0); 3011674ae819SStefano Zampini } 3012674ae819SStefano Zampini 3013674ae819SStefano Zampini #undef __FUNCT__ 3014674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalVerticesLocalIdx" 3015f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalVerticesLocalIdx(PC pc, PetscInt *n_vertices, PetscInt **vertices_idx) 3016674ae819SStefano Zampini { 3017674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 3018674ae819SStefano Zampini PetscInt *vertices,*row_cmat_indices,n,i,size_of_constraint,local_primal_size; 3019674ae819SStefano Zampini PetscErrorCode ierr; 3020674ae819SStefano Zampini 3021674ae819SStefano Zampini PetscFunctionBegin; 3022674ae819SStefano Zampini n = 0; 3023674ae819SStefano Zampini vertices = 0; 3024674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 3025674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&i);CHKERRQ(ierr); 3026b120a5c6SStefano Zampini for (i=0;i<local_primal_size;i++) { 3027b120a5c6SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 3028b120a5c6SStefano Zampini if (size_of_constraint == 1) n++; 3029b120a5c6SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 3030b120a5c6SStefano Zampini } 3031811e8ca2SStefano Zampini if (vertices_idx) { 3032785e854fSJed Brown ierr = PetscMalloc1(n,&vertices);CHKERRQ(ierr); 3033b120a5c6SStefano Zampini n = 0; 3034674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 3035674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 3036674ae819SStefano Zampini if (size_of_constraint == 1) { 3037674ae819SStefano Zampini vertices[n++]=row_cmat_indices[0]; 3038674ae819SStefano Zampini } 3039674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 3040674ae819SStefano Zampini } 3041674ae819SStefano Zampini } 3042811e8ca2SStefano Zampini } 3043674ae819SStefano Zampini *n_vertices = n; 3044811e8ca2SStefano Zampini if (vertices_idx) *vertices_idx = vertices; 3045674ae819SStefano Zampini PetscFunctionReturn(0); 3046674ae819SStefano Zampini } 3047674ae819SStefano Zampini 3048674ae819SStefano Zampini #undef __FUNCT__ 3049674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalConstraintsLocalIdx" 3050f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalConstraintsLocalIdx(PC pc, PetscInt *n_constraints, PetscInt **constraints_idx) 3051674ae819SStefano Zampini { 3052674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 3053674ae819SStefano Zampini PetscInt *constraints_index,*row_cmat_indices,*row_cmat_global_indices; 3054674ae819SStefano Zampini PetscInt n,i,j,size_of_constraint,local_primal_size,local_size,max_size_of_constraint,min_index,min_loc; 30554641a718SStefano Zampini PetscBT touched; 3056674ae819SStefano Zampini PetscErrorCode ierr; 3057674ae819SStefano Zampini 3058f34684f1SStefano Zampini /* This function assumes that the number of local constraints per connected component 3059f34684f1SStefano Zampini is not greater than the number of nodes defined for the connected component 3060f34684f1SStefano Zampini (otherwise we will surely have linear dependence between constraints and thus a singular coarse problem) */ 3061674ae819SStefano Zampini PetscFunctionBegin; 3062674ae819SStefano Zampini n = 0; 3063674ae819SStefano Zampini constraints_index = 0; 3064674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 3065674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&local_size);CHKERRQ(ierr); 3066674ae819SStefano Zampini max_size_of_constraint = 0; 3067674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 3068674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 3069674ae819SStefano Zampini if (size_of_constraint > 1) { 3070674ae819SStefano Zampini n++; 3071674ae819SStefano Zampini } 3072674ae819SStefano Zampini max_size_of_constraint = PetscMax(size_of_constraint,max_size_of_constraint); 3073674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 3074674ae819SStefano Zampini } 3075811e8ca2SStefano Zampini if (constraints_idx) { 3076785e854fSJed Brown ierr = PetscMalloc1(n,&constraints_index);CHKERRQ(ierr); 3077785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&row_cmat_global_indices);CHKERRQ(ierr); 30784641a718SStefano Zampini ierr = PetscBTCreate(local_size,&touched);CHKERRQ(ierr); 3079674ae819SStefano Zampini n = 0; 3080674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 3081674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 3082674ae819SStefano Zampini if (size_of_constraint > 1) { 3083674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,row_cmat_indices,row_cmat_global_indices);CHKERRQ(ierr); 308482d3d8afSStefano Zampini /* find first untouched local node */ 308582d3d8afSStefano Zampini j = 0; 30864641a718SStefano Zampini while (PetscBTLookup(touched,row_cmat_indices[j])) j++; 308782d3d8afSStefano Zampini min_index = row_cmat_global_indices[j]; 308882d3d8afSStefano Zampini min_loc = j; 308982d3d8afSStefano Zampini /* search the minimum among nodes not yet touched on the connected component 309082d3d8afSStefano Zampini since there can be more than one constraint on a single cc */ 3091674ae819SStefano Zampini for (j=1;j<size_of_constraint;j++) { 30924641a718SStefano Zampini if (!PetscBTLookup(touched,row_cmat_indices[j]) && min_index > row_cmat_global_indices[j]) { 3093674ae819SStefano Zampini min_index = row_cmat_global_indices[j]; 3094674ae819SStefano Zampini min_loc = j; 3095674ae819SStefano Zampini } 3096674ae819SStefano Zampini } 30974641a718SStefano Zampini ierr = PetscBTSet(touched,row_cmat_indices[min_loc]);CHKERRQ(ierr); 3098674ae819SStefano Zampini constraints_index[n++] = row_cmat_indices[min_loc]; 3099674ae819SStefano Zampini } 3100674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 3101674ae819SStefano Zampini } 31024641a718SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 3103674ae819SStefano Zampini ierr = PetscFree(row_cmat_global_indices);CHKERRQ(ierr); 3104811e8ca2SStefano Zampini } 3105811e8ca2SStefano Zampini } 3106674ae819SStefano Zampini *n_constraints = n; 3107811e8ca2SStefano Zampini if (constraints_idx) *constraints_idx = constraints_index; 3108674ae819SStefano Zampini PetscFunctionReturn(0); 3109674ae819SStefano Zampini } 3110674ae819SStefano Zampini 3111674ae819SStefano Zampini #undef __FUNCT__ 3112674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3113674ae819SStefano 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[]) 3114674ae819SStefano Zampini { 3115674ae819SStefano Zampini Vec local_vec,global_vec; 3116674ae819SStefano Zampini IS seqis,paris; 3117674ae819SStefano Zampini VecScatter scatter_ctx; 3118674ae819SStefano Zampini PetscScalar *array; 3119674ae819SStefano Zampini PetscInt *temp_global_dofs; 3120674ae819SStefano Zampini PetscScalar globalsum; 3121674ae819SStefano Zampini PetscInt i,j,s; 3122674ae819SStefano Zampini PetscInt nlocals,first_index,old_index,max_local; 3123674ae819SStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm,max_global; 3124674ae819SStefano Zampini PetscMPIInt *dof_sizes,*dof_displs; 3125674ae819SStefano Zampini PetscBool first_found; 3126674ae819SStefano Zampini PetscErrorCode ierr; 3127674ae819SStefano Zampini 3128674ae819SStefano Zampini PetscFunctionBegin; 3129674ae819SStefano Zampini /* mpi buffers */ 3130b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3131b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3132674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 3133785e854fSJed Brown ierr = PetscMalloc1(j,&dof_sizes);CHKERRQ(ierr); 3134785e854fSJed Brown ierr = PetscMalloc1(j,&dof_displs);CHKERRQ(ierr); 3135674ae819SStefano Zampini /* get maximum size of subset */ 3136785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3137674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3138674ae819SStefano Zampini max_local = 0; 3139b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3140674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3141674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3142674ae819SStefano Zampini } 3143674ae819SStefano Zampini } 3144b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3145674ae819SStefano Zampini max_global++; 3146674ae819SStefano Zampini max_local = 0; 3147b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3148674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3149674ae819SStefano Zampini max_local = local_dofs[i]; 3150674ae819SStefano Zampini } 3151674ae819SStefano Zampini } 3152674ae819SStefano Zampini max_local++; 3153674ae819SStefano Zampini /* allocate workspace */ 3154674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3155674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3156674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3157674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3158674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3159674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3160674ae819SStefano Zampini /* create scatter */ 3161674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3162674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3163674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3164674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3165674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3166674ae819SStefano Zampini /* init array */ 3167674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3168674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3169674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3170674ae819SStefano Zampini if (local_dofs_mult) { 3171674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3172674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3173674ae819SStefano Zampini } 3174674ae819SStefano Zampini } else { 3175674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3176674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3177674ae819SStefano Zampini } 3178674ae819SStefano Zampini } 3179674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3180674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3181674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3182674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3183674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 31845b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3185674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3186674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3187674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3188674ae819SStefano Zampini nlocals = 0; 3189674ae819SStefano Zampini first_index = -1; 3190674ae819SStefano Zampini first_found = PETSC_FALSE; 3191674ae819SStefano Zampini for (i=0;i<s;i++) { 3192b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3193674ae819SStefano Zampini first_found = PETSC_TRUE; 3194674ae819SStefano Zampini first_index = i; 3195674ae819SStefano Zampini } 31965b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3197674ae819SStefano Zampini } 3198674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3199674ae819SStefano Zampini if (!rank_prec_comm) { 3200674ae819SStefano Zampini dof_displs[0]=0; 3201674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3202674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3203674ae819SStefano Zampini } 3204674ae819SStefano Zampini } 3205674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3206674ae819SStefano Zampini if (first_found) { 3207674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3208674ae819SStefano Zampini old_index = first_index; 3209674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3210b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3211674ae819SStefano Zampini array[i] += array[old_index]; 3212674ae819SStefano Zampini old_index = i; 3213674ae819SStefano Zampini } 3214674ae819SStefano Zampini } 3215674ae819SStefano Zampini } 3216674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3218674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3219674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3220674ae819SStefano Zampini /* get global ordering of local dofs */ 32212b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3222674ae819SStefano Zampini if (local_dofs_mult) { 3223674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32245b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3225674ae819SStefano Zampini } 3226674ae819SStefano Zampini } else { 3227674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32285b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3229674ae819SStefano Zampini } 3230674ae819SStefano Zampini } 32312b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3232674ae819SStefano Zampini /* free workspace */ 3233674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3234674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3235674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 3236674ae819SStefano Zampini ierr = PetscFree(dof_sizes);CHKERRQ(ierr); 3237674ae819SStefano Zampini ierr = PetscFree(dof_displs);CHKERRQ(ierr); 3238674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3239674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3240674ae819SStefano Zampini PetscFunctionReturn(0); 3241674ae819SStefano Zampini } 32429a7d3425SStefano Zampini 32439a7d3425SStefano Zampini #undef __FUNCT__ 32449a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32459a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32469a7d3425SStefano Zampini { 32479a7d3425SStefano Zampini PetscInt i,j; 32489a7d3425SStefano Zampini PetscScalar *alphas; 32499a7d3425SStefano Zampini PetscErrorCode ierr; 32509a7d3425SStefano Zampini 32519a7d3425SStefano Zampini PetscFunctionBegin; 32529a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3253785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 32549a7d3425SStefano Zampini for (i=0;i<n;i++) { 32559a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 32569a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 32579a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 32589a7d3425SStefano Zampini } 32599a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 32609a7d3425SStefano Zampini PetscFunctionReturn(0); 32619a7d3425SStefano Zampini } 32629a7d3425SStefano Zampini 3263e7931f94SStefano Zampini #undef __FUNCT__ 326470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 326528143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3266e7931f94SStefano Zampini { 3267e7931f94SStefano Zampini Mat subdomain_adj; 3268e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3269e7931f94SStefano Zampini MatPartitioning partitioner; 3270e7931f94SStefano Zampini Mat_IS *matis; 3271e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3272e7931f94SStefano Zampini PetscInt prank; 3273e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3274e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3275e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 32763837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3277e7931f94SStefano Zampini PetscErrorCode ierr; 32782b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3279e7931f94SStefano Zampini PetscSubcomm subcomm; 3280a57a6d2fSStefano Zampini 3281e7931f94SStefano Zampini PetscFunctionBegin; 32822b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 32832b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 32842b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3285e7931f94SStefano Zampini 3286e7931f94SStefano Zampini /* Get info on mapping */ 3287e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3288e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3289e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3290e7931f94SStefano Zampini 3291e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3292785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3293e7931f94SStefano Zampini xadj[0] = 0; 3294e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3295785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3296785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3297e7931f94SStefano Zampini 32982b510759SStefano Zampini if (threshold) { 3299d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33002b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3301d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3302d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3303d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3304d023bfaeSStefano Zampini xadj_count++; 3305e7931f94SStefano Zampini } 3306e7931f94SStefano Zampini } 3307d023bfaeSStefano Zampini xadj[1] = xadj_count; 3308c8587f34SStefano Zampini } else { 3309e7931f94SStefano Zampini if (xadj[1]) { 3310e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3311e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3312c8587f34SStefano Zampini } 3313e7931f94SStefano Zampini } 3314e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3315e7931f94SStefano Zampini if (use_square) { 3316e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3317e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3318e7931f94SStefano Zampini } 3319e7931f94SStefano Zampini } 3320e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3321e7931f94SStefano Zampini 33223837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3323e7931f94SStefano Zampini 3324e7931f94SStefano Zampini /* 3325e7931f94SStefano Zampini Restrict work on active processes only. 3326e7931f94SStefano Zampini */ 3327e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3328e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3329e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33302b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3331d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3332e7931f94SStefano Zampini if (color) { 3333e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3334e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3335e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3336c8587f34SStefano Zampini } else { 333728143c3dSStefano Zampini PetscInt coarsening_ratio; 3338306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3339785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3340e7931f94SStefano Zampini prank = rank; 3341306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33428002ef2cSStefano Zampini /* 3343e7931f94SStefano Zampini for (i=0;i<size;i++) { 3344e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3345c8587f34SStefano Zampini } 33468002ef2cSStefano Zampini */ 3347e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3348e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3349c8587f34SStefano Zampini } 3350e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3351306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 335222b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3353e7931f94SStefano Zampini 3354e7931f94SStefano Zampini /* Partition */ 3355306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3356e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3357e7931f94SStefano Zampini if (use_vwgt) { 33583837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3359e7931f94SStefano Zampini v_wgt[0] = local_size; 3360e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3361c8587f34SStefano Zampini } 336228143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 336328143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 336428143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3365e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3366e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 336722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3368e7931f94SStefano Zampini 3369e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 337028143c3dSStefano Zampini if (contiguous) { 337128143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 337228143c3dSStefano Zampini } else { 337384ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 337428143c3dSStefano Zampini } 3375e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3376e7931f94SStefano Zampini /* clean up */ 3377e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3378e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3379e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3380e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3381e7931f94SStefano Zampini } 3382e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3383e7931f94SStefano Zampini 3384e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3385e7931f94SStefano Zampini i = 1; 3386e7931f94SStefano Zampini if (color) i=0; 3387e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3388e7931f94SStefano Zampini 3389e7931f94SStefano Zampini /* get back IS */ 3390e7931f94SStefano Zampini *is_sends = ranks_send_to; 3391e7931f94SStefano Zampini PetscFunctionReturn(0); 3392e7931f94SStefano Zampini } 3393e7931f94SStefano Zampini 3394e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3395e7931f94SStefano Zampini 3396e7931f94SStefano Zampini #undef __FUNCT__ 3397e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 339828143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3399e7931f94SStefano Zampini { 340070cf5478SStefano Zampini Mat local_mat; 3401e7931f94SStefano Zampini Mat_IS *matis; 3402e7931f94SStefano Zampini IS is_sends_internal; 34039d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 340428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34059d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3406e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3407e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3408e7931f94SStefano Zampini const PetscInt* is_indices; 3409e7931f94SStefano Zampini MatType new_local_type; 3410e7931f94SStefano Zampini /* buffers */ 3411e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 341228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34139d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3414e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3415e7931f94SStefano Zampini /* MPI */ 341628143c3dSStefano Zampini MPI_Comm comm,comm_n; 341728143c3dSStefano Zampini PetscSubcomm subcomm; 3418e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 341928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 342028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 342128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 342228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 342328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3424e7931f94SStefano Zampini PetscErrorCode ierr; 3425e7931f94SStefano Zampini 3426e7931f94SStefano Zampini PetscFunctionBegin; 342728143c3dSStefano Zampini /* TODO: add missing checks */ 342828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 342928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 343028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 343128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3432e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 343328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3434e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3435e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3436e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3437e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3438e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 343928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 344070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 344170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 344228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 344370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 344470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 344570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 344670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 344770cf5478SStefano Zampini } 3448e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3449e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3450e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3451e7931f94SStefano Zampini if (!is_sends) { 345228143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 345328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 345428143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3455c8587f34SStefano Zampini } else { 3456e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3457e7931f94SStefano Zampini is_sends_internal = is_sends; 3458c8587f34SStefano Zampini } 3459e7931f94SStefano Zampini 3460e7931f94SStefano Zampini /* get pointer of MATIS data */ 3461e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3462e7931f94SStefano Zampini 3463e7931f94SStefano Zampini /* get comm */ 3464a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3465e7931f94SStefano Zampini 3466e7931f94SStefano Zampini /* compute number of sends */ 3467e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3468e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3469e7931f94SStefano Zampini 3470e7931f94SStefano Zampini /* compute number of receives */ 3471e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3472785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3473e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3474e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3475e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3476e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3477e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3478e7931f94SStefano Zampini 347928143c3dSStefano Zampini /* restrict comm if requested */ 348028143c3dSStefano Zampini subcomm = 0; 348128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 348228143c3dSStefano Zampini if (restrict_comm) { 3483779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3484779c1cceSStefano Zampini 348528143c3dSStefano Zampini color = 0; 3486779c1cceSStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */ 348728143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 348828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 348928143c3dSStefano Zampini /* check if reuse has been requested */ 349028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 349128143c3dSStefano Zampini if (*mat_n) { 349228143c3dSStefano Zampini PetscMPIInt subcommsize2; 349328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 349428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 349528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 349628143c3dSStefano Zampini } else { 349728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 349828143c3dSStefano Zampini } 349928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3500779c1cceSStefano Zampini PetscMPIInt rank; 3501779c1cceSStefano Zampini 3502779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 350328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 350428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 350528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3506306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 350728143c3dSStefano Zampini } 350828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 350928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 351028143c3dSStefano Zampini } else { 351128143c3dSStefano Zampini comm_n = comm; 351228143c3dSStefano Zampini } 351328143c3dSStefano Zampini 3514e7931f94SStefano Zampini /* prepare send/receive buffers */ 3515785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3516e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3517785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3518e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 351928143c3dSStefano Zampini if (nis) { 3520854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 352128143c3dSStefano Zampini } 3522e7931f94SStefano Zampini 352328143c3dSStefano Zampini /* Get data from local matrices */ 3524e7931f94SStefano Zampini if (!isdense) { 3525a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3526e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3527e7931f94SStefano Zampini /* 3528e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3529e7931f94SStefano Zampini send_buffer_idxs should contain: 3530e7931f94SStefano Zampini - MatType_PRIVATE type 3531e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3532e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3533e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3534e7931f94SStefano Zampini */ 3535e7931f94SStefano Zampini } else { 3536e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3537e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3538854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3539e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3540e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3541e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3542e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3543e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3544e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3545e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3546e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3547e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3548c8587f34SStefano Zampini } 3549c8587f34SStefano Zampini } 3550e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 355128143c3dSStefano Zampini /* additional is (if any) */ 355228143c3dSStefano Zampini if (nis) { 355328143c3dSStefano Zampini PetscMPIInt psum; 355428143c3dSStefano Zampini PetscInt j; 355528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 355628143c3dSStefano Zampini PetscInt plen; 355728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 355828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 355928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 356028143c3dSStefano Zampini } 3561854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 356228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 356328143c3dSStefano Zampini PetscInt plen; 356428143c3dSStefano Zampini const PetscInt *is_array_idxs; 356528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 356628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 356728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 356828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 356928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 357028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 357128143c3dSStefano Zampini } 357228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 357328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 357428143c3dSStefano Zampini } 357528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 357628143c3dSStefano Zampini } 357728143c3dSStefano Zampini 3578e7931f94SStefano Zampini buf_size_idxs = 0; 3579e7931f94SStefano Zampini buf_size_vals = 0; 358028143c3dSStefano Zampini buf_size_idxs_is = 0; 3581e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3582e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3583e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 358428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3585e7931f94SStefano Zampini } 3586785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3587785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 358895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3589e7931f94SStefano Zampini 3590e7931f94SStefano Zampini /* get new tags for clean communications */ 3591e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3592e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 359328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3594e7931f94SStefano Zampini 3595e7931f94SStefano Zampini /* allocate for requests */ 3596785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3597785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 359895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3599785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3600785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 360195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3602e7931f94SStefano Zampini 3603e7931f94SStefano Zampini /* communications */ 3604e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3605e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 360628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3607e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3608e7931f94SStefano Zampini source_dest = onodes[i]; 3609e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3610e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3611e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3612e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 361328143c3dSStefano Zampini if (nis) { 361428143c3dSStefano 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); 361528143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 361628143c3dSStefano Zampini } 3617e7931f94SStefano Zampini } 3618e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3619e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3620e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3621e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 362228143c3dSStefano Zampini if (nis) { 362328143c3dSStefano 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); 362428143c3dSStefano Zampini } 3625e7931f94SStefano Zampini } 3626e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3627e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3628e7931f94SStefano Zampini 3629e7931f94SStefano Zampini /* assemble new l2g map */ 3630e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3631e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36329d30be91SStefano Zampini new_local_rows = 0; 3633e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36349d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3635e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3636e7931f94SStefano Zampini } 36379d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3638e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36399d30be91SStefano Zampini new_local_rows = 0; 3640e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36419d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36429d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3643e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3644e7931f94SStefano Zampini } 36459d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36469d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3647e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3648e7931f94SStefano Zampini 3649e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3650e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3651e7931f94SStefano 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) */ 3652e7931f94SStefano Zampini if (n_recvs) { 365328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3654e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3655e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3656e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3657e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3658e7931f94SStefano Zampini break; 3659e7931f94SStefano Zampini } 3660e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3661e7931f94SStefano Zampini } 3662e7931f94SStefano Zampini switch (new_local_type_private) { 366328143c3dSStefano Zampini case MATDENSE_PRIVATE: 366428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3665e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3666e7931f94SStefano Zampini bs = 1; 366728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 366828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 366928143c3dSStefano Zampini bs = 1; 367028143c3dSStefano Zampini } 3671e7931f94SStefano Zampini break; 3672e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3673e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3674e7931f94SStefano Zampini bs = 1; 3675e7931f94SStefano Zampini break; 3676e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3677e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3678e7931f94SStefano Zampini break; 3679e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3680e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3681e7931f94SStefano Zampini break; 3682e7931f94SStefano Zampini default: 36839d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3684e7931f94SStefano Zampini break; 3685e7931f94SStefano Zampini } 368628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 368728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 368828143c3dSStefano Zampini bs = 1; 3689e7931f94SStefano Zampini } 3690e7931f94SStefano Zampini 369170cf5478SStefano Zampini /* create MATIS object if needed */ 369270cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3693e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 369428143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 369570cf5478SStefano Zampini } else { 369670cf5478SStefano Zampini /* it also destroys the local matrices */ 369770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 369870cf5478SStefano Zampini } 369970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3700e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37019d30be91SStefano Zampini 37029d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37039d30be91SStefano Zampini 37049d30be91SStefano Zampini /* Global to local map of received indices */ 37059d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37069d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37079d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37089d30be91SStefano Zampini 37099d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37109d30be91SStefano Zampini buf_size_idxs = 0; 37119d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37129d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37139d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37149d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37159d30be91SStefano Zampini } 37169d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37179d30be91SStefano Zampini 37189d30be91SStefano Zampini /* set preallocation */ 37199d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37209d30be91SStefano Zampini if (!newisdense) { 37219d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37229d30be91SStefano Zampini 37239d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37249d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37259d30be91SStefano Zampini if (n_recvs) { 37269d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37279d30be91SStefano Zampini } 37289d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37299d30be91SStefano Zampini PetscInt j; 37309d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37319d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37329d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37339d30be91SStefano Zampini } 37349d30be91SStefano Zampini } else { 37359d30be91SStefano Zampini /* TODO */ 37369d30be91SStefano Zampini } 37379d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37389d30be91SStefano Zampini } 37399d30be91SStefano Zampini if (new_local_nnz) { 37409d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37419d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37429d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37439d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37449d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37459d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37469d30be91SStefano Zampini } else { 37479d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37489d30be91SStefano Zampini } 37499d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 37509d30be91SStefano Zampini } else { 37519d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37529d30be91SStefano Zampini } 3753e7931f94SStefano Zampini 3754e7931f94SStefano Zampini /* set values */ 3755e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 37569d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3757e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3758e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3759e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 37609d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3761e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3762e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3763e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 376428143c3dSStefano Zampini } else { 376528143c3dSStefano Zampini /* TODO */ 3766e7931f94SStefano Zampini } 3767e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3768e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3769e7931f94SStefano Zampini } 3770e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3771e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 377270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 377370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 37749d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 37759d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3776e7931f94SStefano Zampini 3777dfd14d43SStefano Zampini #if 0 377828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3779e7931f94SStefano Zampini Vec lvec,rvec; 3780e7931f94SStefano Zampini PetscReal infty_error; 3781e7931f94SStefano Zampini 37822a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3783e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3784e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3785e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 378670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3787e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3788e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3789e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3790e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3791e7931f94SStefano Zampini } 379228143c3dSStefano Zampini #endif 3793e7931f94SStefano Zampini 379428143c3dSStefano Zampini /* assemble new additional is (if any) */ 379528143c3dSStefano Zampini if (nis) { 379628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 379728143c3dSStefano Zampini 379828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3799854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 380028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 380128143c3dSStefano Zampini psum = 0; 380228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 380328143c3dSStefano Zampini for (j=0;j<nis;j++) { 380428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 380528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 380628143c3dSStefano Zampini psum += plen; 380728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 380828143c3dSStefano Zampini } 380928143c3dSStefano Zampini } 3810854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3811854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 381228143c3dSStefano Zampini for (i=1;i<nis;i++) { 381328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 381428143c3dSStefano Zampini } 381528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 381628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 381728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 381828143c3dSStefano Zampini for (j=0;j<nis;j++) { 381928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 382028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 382128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 382228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 382328143c3dSStefano Zampini } 382428143c3dSStefano Zampini } 382528143c3dSStefano Zampini for (i=0;i<nis;i++) { 382628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 382728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 382828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 382928143c3dSStefano Zampini } 383028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 383128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 383228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 383328143c3dSStefano Zampini } 3834e7931f94SStefano Zampini /* free workspace */ 383528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3836e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3837e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3838e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3839e7931f94SStefano Zampini if (isdense) { 3840e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3841e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3842e7931f94SStefano Zampini } else { 3843e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3844e7931f94SStefano Zampini } 384528143c3dSStefano Zampini if (nis) { 384628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 384728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 384828143c3dSStefano Zampini } 3849e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3850e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 385128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3852e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3853e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 385428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3855e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3856e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3857e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3858e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3859e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 386028143c3dSStefano Zampini if (nis) { 386128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 386228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 386328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 386428143c3dSStefano Zampini } 386528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 386628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 386728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 386828143c3dSStefano Zampini for (i=0;i<nis;i++) { 386928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 387028143c3dSStefano Zampini } 387128143c3dSStefano Zampini } 3872e7931f94SStefano Zampini PetscFunctionReturn(0); 3873e7931f94SStefano Zampini } 3874a57a6d2fSStefano Zampini 387512edc857SStefano Zampini /* temporary hack into ksp private data structure */ 387612edc857SStefano Zampini #include <petsc-private/kspimpl.h> 387712edc857SStefano Zampini 3878c8587f34SStefano Zampini #undef __FUNCT__ 3879c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3880c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3881c8587f34SStefano Zampini { 3882c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3883c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 388420a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 38859881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 388620a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 38876e683305SStefano Zampini IS coarse_is,*isarray; 38886e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 38896e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3890f9eb5b7dSStefano Zampini PC pc_temp; 3891c8587f34SStefano Zampini PCType coarse_pc_type; 3892c8587f34SStefano Zampini KSPType coarse_ksp_type; 3893f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 38944f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 38956e683305SStefano Zampini Mat t_coarse_mat_is; 38966e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 38976e683305SStefano Zampini PetscMPIInt all_procs; 389874e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 389968457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 390022bc73bbSStefano Zampini PetscScalar *array; 39019881197aSStefano Zampini PetscErrorCode ierr; 3902fdc09c96SStefano Zampini 3903c8587f34SStefano Zampini PetscFunctionBegin; 3904c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 390568457ee5SStefano 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 */ 3906fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39075a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3908fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3909f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3910f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3911f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3912fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3913fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3914727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3915fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3916fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3917fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3918f4ddd8eeSStefano Zampini } 3919fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3920fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3921f4ddd8eeSStefano Zampini } 392270cf5478SStefano Zampini /* reset any subassembling information */ 392370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39246e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39256e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3926fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3927f4ddd8eeSStefano Zampini } 3928c8587f34SStefano Zampini 39296e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39302b510759SStefano Zampini im_active = !!(pcis->n); 39312b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39326e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39336e683305SStefano Zampini void_procs = all_procs-active_procs; 39346e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 393574e2c79eSStefano Zampini redist = PETSC_FALSE; 393622bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39376e683305SStefano Zampini csin_ml = PETSC_TRUE; 39386e683305SStefano Zampini ncoarse_ml = void_procs; 3939779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3940779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 394118a45a71SStefano Zampini csin_ds = PETSC_TRUE; 394218a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 394318a45a71SStefano Zampini redist = PETSC_TRUE; 394418a45a71SStefano Zampini } else { 39456e683305SStefano Zampini csin_ds = PETSC_TRUE; 3946779c1cceSStefano Zampini ncoarse_ds = active_procs; 3947779c1cceSStefano Zampini redist = PETSC_TRUE; 394818a45a71SStefano Zampini } 39496e683305SStefano Zampini } else { 39506e683305SStefano Zampini csin_ml = PETSC_FALSE; 39516e683305SStefano Zampini ncoarse_ml = all_procs; 39526e683305SStefano Zampini if (void_procs) { 39536e683305SStefano Zampini csin_ds = PETSC_TRUE; 39546e683305SStefano Zampini ncoarse_ds = void_procs; 39556e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 39566e683305SStefano Zampini } else { 3957779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 395874e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 395974e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 396074e2c79eSStefano Zampini redist = PETSC_TRUE; 396174e2c79eSStefano Zampini } else { 39626e683305SStefano Zampini csin_ds = PETSC_FALSE; 39636e683305SStefano Zampini ncoarse_ds = all_procs; 39646e683305SStefano Zampini } 39656e683305SStefano Zampini } 396674e2c79eSStefano Zampini } 39676e683305SStefano Zampini 39686e683305SStefano Zampini /* 39696e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 39706e683305SStefano Zampini - we have not exceeded the number of levels requested 39716e683305SStefano Zampini - we can actually subassemble the active processes 39726e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 39736e683305SStefano Zampini */ 39746e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 39756e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 39766e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 39776e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 39786e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3979f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 39802b510759SStefano Zampini } else { 3981f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3982c8587f34SStefano Zampini } 3983c8587f34SStefano Zampini } 39846e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 39856e683305SStefano Zampini if (multilevel_allowed) { 39866e683305SStefano Zampini ncoarse = ncoarse_ml; 39876e683305SStefano Zampini csin = csin_ml; 398858da7f69SStefano Zampini redist = PETSC_FALSE; 39896e683305SStefano Zampini } else { 39906e683305SStefano Zampini ncoarse = ncoarse_ds; 39916e683305SStefano Zampini csin = csin_ds; 39926e683305SStefano Zampini } 3993e7931f94SStefano Zampini 3994abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3995abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3996abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3997abbbba34SStefano Zampini 3998abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 399922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 400022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 400122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 400222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4003b9b85e73SStefano Zampini #if 0 4004b9b85e73SStefano Zampini { 4005b9b85e73SStefano Zampini PetscViewer viewer; 4006b9b85e73SStefano Zampini char filename[256]; 4007b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4008b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4009b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4010b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4011b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4012b9b85e73SStefano Zampini } 4013b9b85e73SStefano Zampini #endif 40146e683305SStefano 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); 40156e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40166e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40176e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4018abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4019abbbba34SStefano Zampini 40206e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40216e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40226e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40236e683305SStefano Zampini const PetscInt *idxs; 40246e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40256e683305SStefano Zampini 40266e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40270be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40286e683305SStefano Zampini /* allocate space for temporary storage */ 4029854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4030854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40316e683305SStefano Zampini /* allocate for IS array */ 40326e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40336e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40346e683305SStefano Zampini nis = nisdofs + nisneu; 4035854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40366e683305SStefano Zampini /* dofs splitting */ 40376e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40386e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40396e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40406e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40416e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40426e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40436e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40446e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40456e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40466e683305SStefano Zampini } 40476e683305SStefano Zampini /* neumann boundaries */ 40486e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 40496e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 40506e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 40516e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40526e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40536e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40546e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40556e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 40566e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 40576e683305SStefano Zampini } 40586e683305SStefano Zampini /* free memory */ 40596e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 40606e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 40616e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 40626e683305SStefano Zampini } else { 40636e683305SStefano Zampini nis = 0; 40646e683305SStefano Zampini nisdofs = 0; 40656e683305SStefano Zampini nisneu = 0; 40666e683305SStefano Zampini isarray = NULL; 40676e683305SStefano Zampini } 40686e683305SStefano Zampini /* destroy no longer needed map */ 40696e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 40706e683305SStefano Zampini 40716e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 40726e683305SStefano Zampini coarse_mat_is = NULL; 40736e683305SStefano Zampini if (csin) { 40746e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 407574e2c79eSStefano Zampini if (redist) { 407674e2c79eSStefano Zampini PetscMPIInt rank; 4077779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 407874e2c79eSStefano Zampini 407974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 408058da7f69SStefano Zampini spc = active_procs/ncoarse; 408158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4082779c1cceSStefano Zampini if (im_active) { 4083779c1cceSStefano Zampini destsize = 1; 408474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 408574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 408674e2c79eSStefano Zampini } else { 408774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 408874e2c79eSStefano Zampini } 408974e2c79eSStefano Zampini } else { 4090779c1cceSStefano Zampini destsize = 0; 40916e683305SStefano Zampini } 4092779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4093779c1cceSStefano Zampini } else if (csin_type_simple) { 40946e683305SStefano Zampini PetscMPIInt rank; 40956e683305SStefano Zampini PetscInt issize,isidx; 4096779c1cceSStefano Zampini 40976e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 40986e683305SStefano Zampini if (im_active) { 40996e683305SStefano Zampini issize = 1; 41006e683305SStefano Zampini isidx = (PetscInt)rank; 41016e683305SStefano Zampini } else { 41026e683305SStefano Zampini issize = 0; 41036e683305SStefano Zampini isidx = -1; 41046e683305SStefano Zampini } 41056e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4106779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41076e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41086e683305SStefano Zampini } 4109779c1cceSStefano Zampini 4110779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4111779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4112779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4113779c1cceSStefano Zampini PetscInt *coarse_candidates; 4114779c1cceSStefano Zampini const PetscInt* tisindices; 4115779c1cceSStefano Zampini 4116779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4117779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4118779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4119779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4120779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4121779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4122779c1cceSStefano Zampini } 4123779c1cceSStefano Zampini } 4124779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4125779c1cceSStefano Zampini 4126779c1cceSStefano Zampini 41276e683305SStefano Zampini if (pcbddc->dbg_flag) { 41286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41296e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41306e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4132779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41346e683305SStefano Zampini } 41356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41366e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41376e683305SStefano Zampini } 41386e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41396e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41406e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4141854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41426e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41436e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41446e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41456e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41466e683305SStefano Zampini } 41476e683305SStefano Zampini if (pcbddc->dbg_flag) { 41486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 41506e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41516e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41526e683305SStefano Zampini } 4153779c1cceSStefano Zampini } 41546e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 41556e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 41566e683305SStefano Zampini } else { 41576e683305SStefano Zampini if (pcbddc->dbg_flag) { 41586e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 41606e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41616e683305SStefano Zampini } 41626e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 41636e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 41646e683305SStefano Zampini } 41656e683305SStefano Zampini 41666e683305SStefano Zampini /* create local to global scatters for coarse problem */ 416768457ee5SStefano Zampini if (compute_vecs) { 41686e683305SStefano Zampini PetscInt lrows; 41696e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 41706e683305SStefano Zampini if (coarse_mat_is) { 41716e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 41726e683305SStefano Zampini } else { 41736e683305SStefano Zampini lrows = 0; 41746e683305SStefano Zampini } 41756e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 41766e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 41776e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 41786e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41796e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 41806e683305SStefano Zampini } 41816e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4183c8587f34SStefano Zampini 4184f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4185f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4186f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4187f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4188f9eb5b7dSStefano Zampini } else { 4189f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4190f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4191c8587f34SStefano Zampini } 4192c8587f34SStefano Zampini 41936e683305SStefano Zampini /* print some info if requested */ 41946e683305SStefano Zampini if (pcbddc->dbg_flag) { 41956e683305SStefano Zampini if (!multilevel_allowed) { 41966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41976e683305SStefano Zampini if (multilevel_requested) { 41986e683305SStefano 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); 41996e683305SStefano Zampini } else if (pcbddc->max_levels) { 42006e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42016e683305SStefano Zampini } 42026e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42036e683305SStefano Zampini } 42046e683305SStefano Zampini } 42056e683305SStefano Zampini 4206f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42076e683305SStefano Zampini if (coarse_mat_is) { 42086e683305SStefano Zampini MatReuse coarse_mat_reuse; 42096a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42106e683305SStefano Zampini if (pcbddc->dbg_flag) { 42116e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42126e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42136e683305SStefano Zampini } 4214f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4215312be037SStefano Zampini char prefix[256],str_level[16]; 4216e604994aSStefano Zampini size_t len; 42176e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4218c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4219f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42205f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4221c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42226e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4223c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4224c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4225e604994aSStefano Zampini /* prefix */ 4226e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4227e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4228e604994aSStefano Zampini if (!pcbddc->current_level) { 4229e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4230e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4231c8587f34SStefano Zampini } else { 4232e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4233312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4234312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 423534d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4236312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4237e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4238e604994aSStefano Zampini } 4239e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 4240f9eb5b7dSStefano Zampini /* allow user customization */ 4241f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 42427e0def11SStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4243312be037SStefano Zampini } 4244f9eb5b7dSStefano Zampini 4245f9eb5b7dSStefano Zampini /* get some info after set from options */ 4246f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4247f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4248f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 42494f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 42506e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4251f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4252f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4253f9eb5b7dSStefano Zampini } 42544f3a063dSStefano Zampini if (isredundant) { 42554f3a063dSStefano Zampini KSP inner_ksp; 42564f3a063dSStefano Zampini PC inner_pc; 42574f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 42584f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 42594f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 42604f3a063dSStefano Zampini } 4261f9eb5b7dSStefano Zampini 42626e683305SStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 4263f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 4264f9eb5b7dSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 4265f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 42666e683305SStefano Zampini if (nisdofs) { 42676e683305SStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 42686e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 42696e683305SStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 42701035eff8SStefano Zampini } 42711035eff8SStefano Zampini } 42726e683305SStefano Zampini if (nisneu) { 42736e683305SStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 42746e683305SStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 42751035eff8SStefano Zampini } 4276fdc09c96SStefano Zampini 4277f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4278fa7f1dd8SStefano Zampini if (coarse_reuse) { 427981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4280fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 42816e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4282fa7f1dd8SStefano Zampini } else { 42836e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4284fa7f1dd8SStefano Zampini } 4285c8587f34SStefano Zampini if (isbddc || isnn) { 428622bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 428770cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 428828143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 428922b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 42906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 42926e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 42936e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 429422b6e8a2SStefano Zampini } 429570cf5478SStefano Zampini } 42966e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 429770cf5478SStefano Zampini } else { 429822bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 429922bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 430022bc73bbSStefano Zampini } 430122bc73bbSStefano Zampini } else { 43022e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4303c8587f34SStefano Zampini } 4304c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4305c8587f34SStefano Zampini 4306c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4307b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 43085a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4309c8587f34SStefano Zampini 43106e683305SStefano Zampini /* set operators */ 43115f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43126e683305SStefano Zampini if (pcbddc->dbg_flag) { 43136e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43146e683305SStefano Zampini } 43156e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43166e683305SStefano Zampini coarse_mat = 0; 43176e683305SStefano Zampini } 43186e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4319b9b85e73SStefano Zampini #if 0 4320b9b85e73SStefano Zampini { 4321b9b85e73SStefano Zampini PetscViewer viewer; 4322b9b85e73SStefano Zampini char filename[256]; 4323b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4324b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4325b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4326b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4327b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4328b9b85e73SStefano Zampini } 4329b9b85e73SStefano Zampini #endif 4330c8587f34SStefano Zampini 4331c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4332c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4333c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 433498a51de6SStefano Zampini } 433598a51de6SStefano Zampini 433698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 433798a51de6SStefano Zampini Vec crhs,csol; 433898a51de6SStefano Zampini PetscBool ispreonly; 433998a51de6SStefano Zampini if (CoarseNullSpace) { 4340c8587f34SStefano Zampini if (isbddc) { 4341c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4342c8587f34SStefano Zampini } else { 4343c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4344c8587f34SStefano Zampini } 4345c8587f34SStefano Zampini } 4346f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4347f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4348f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4349f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 43506e683305SStefano Zampini /* hack */ 4351f347579bSStefano Zampini if (!csol) { 43522a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4353f9eb5b7dSStefano Zampini } 4354f347579bSStefano Zampini if (!crhs) { 43552a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4356f347579bSStefano Zampini } 4357cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4358cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 43596e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4360c8587f34SStefano Zampini KSP check_ksp; 43612b510759SStefano Zampini KSPType check_ksp_type; 4362c8587f34SStefano Zampini PC check_pc; 43636e683305SStefano Zampini Vec check_vec,coarse_vec; 43646a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 43652b510759SStefano Zampini PetscInt its; 43666e683305SStefano Zampini PetscBool compute_eigs; 43676e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 43686e683305SStefano Zampini PetscInt neigs; 43698e185a42SStefano Zampini const char *prefix; 4370c8587f34SStefano Zampini 43712b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 43726e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 437323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4374f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 43752b510759SStefano Zampini if (ispreonly) { 43762b510759SStefano Zampini check_ksp_type = KSPPREONLY; 43776e683305SStefano Zampini compute_eigs = PETSC_FALSE; 43782b510759SStefano Zampini } else { 4379cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 43806e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4381c8587f34SStefano Zampini } 4382c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 43836e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 43846e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 43856e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4386a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4387a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4388a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4389a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4390c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4391c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4392c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4393c8587f34SStefano Zampini /* create random vec */ 43946e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 43956e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4396c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4397c8587f34SStefano Zampini if (CoarseNullSpace) { 4398c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4399c8587f34SStefano Zampini } 44006e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4401c8587f34SStefano Zampini /* solve coarse problem */ 44026e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4403c8587f34SStefano Zampini if (CoarseNullSpace) { 44046e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4405c8587f34SStefano Zampini } 4406cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44076e683305SStefano Zampini if (compute_eigs) { 4408854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4409854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44106e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44116e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44126e683305SStefano Zampini lambda_min = eigs_r[0]; 44136e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44146e683305SStefano Zampini if (lambda_max>lambda_min) { 4415cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4416cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4417cbcc2c2aSStefano Zampini } 4418c8587f34SStefano Zampini } 4419c8587f34SStefano Zampini } 4420cbcc2c2aSStefano Zampini 4421c8587f34SStefano Zampini /* check coarse problem residual error */ 44226e683305SStefano Zampini if (pcbddc->dbg_flag) { 44236e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44246e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44256e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4426c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44276e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4429c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4430779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44316e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44326e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44356e683305SStefano Zampini if (compute_eigs) { 44366e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4437deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4438c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44396e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 44406e683305SStefano 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); 44416e683305SStefano Zampini for (i=0;i<neigs;i++) { 44426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4443c8587f34SStefano Zampini } 44446e683305SStefano Zampini } 44456e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 44466e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44476e683305SStefano Zampini } 4448c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 44496e683305SStefano Zampini if (compute_eigs) { 44506e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 44516e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4452c8587f34SStefano Zampini } 44536e683305SStefano Zampini } 44546e683305SStefano Zampini } 4455cbcc2c2aSStefano Zampini /* print additional info */ 4456cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 44576e683305SStefano Zampini /* waits until all processes reaches this point */ 44586e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4459cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4460cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4461cbcc2c2aSStefano Zampini } 4462cbcc2c2aSStefano Zampini 44632b510759SStefano Zampini /* free memory */ 4464c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4465fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4466c8587f34SStefano Zampini PetscFunctionReturn(0); 4467c8587f34SStefano Zampini } 4468674ae819SStefano Zampini 4469f34684f1SStefano Zampini #undef __FUNCT__ 4470f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4471f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4472f34684f1SStefano Zampini { 4473f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4474f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4475f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4476727cdba6SStefano Zampini PetscInt i,coarse_size; 4477727cdba6SStefano Zampini PetscInt *local_primal_indices; 4478f34684f1SStefano Zampini PetscErrorCode ierr; 4479f34684f1SStefano Zampini 4480f34684f1SStefano Zampini PetscFunctionBegin; 4481f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 448289c96988SStefano Zampini if (!pcbddc->primal_indices_local_idxs && pcbddc->local_primal_size) { 448389c96988SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Local primal indices have not been created"); 4484727cdba6SStefano Zampini } 4485727cdba6SStefano 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); 4486f34684f1SStefano Zampini 4487f34684f1SStefano Zampini /* check numbering */ 4488f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4489f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4490b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4491f34684f1SStefano Zampini 4492f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4493f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4494f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 44950fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4496f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4497f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4498727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4499f34684f1SStefano Zampini } 4500f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4501f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4502f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4503f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4504f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4505f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4506f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4507f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4508f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4509f34684f1SStefano Zampini if (array[i] == 1.0) { 4510b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4511f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4512f34684f1SStefano Zampini } 4513f34684f1SStefano Zampini } 4514b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4515f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4516f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4517f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4518f34684f1SStefano Zampini } 4519f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4520f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4521f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4522f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4523f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4524f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4525b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4526f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4527f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4528f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4529f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4530727cdba6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%d]=%d (%d)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i]); 4531f34684f1SStefano Zampini } 4532f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4533f34684f1SStefano Zampini } 4534f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4535b9b85e73SStefano Zampini if (set_error_reduced) { 4536b9b85e73SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4537b9b85e73SStefano Zampini } 4538f34684f1SStefano Zampini } 4539f34684f1SStefano Zampini /* get back data */ 4540f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4541f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4542674ae819SStefano Zampini PetscFunctionReturn(0); 4543674ae819SStefano Zampini } 4544674ae819SStefano Zampini 4545e456f2a8SStefano Zampini #undef __FUNCT__ 4546e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4547a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4548e456f2a8SStefano Zampini { 4549e456f2a8SStefano Zampini IS localis_t; 4550a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4551e456f2a8SStefano Zampini PetscScalar *vals; 4552e456f2a8SStefano Zampini PetscErrorCode ierr; 4553e456f2a8SStefano Zampini 4554e456f2a8SStefano Zampini PetscFunctionBegin; 4555a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4556e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4557854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4558e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4559e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4560a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4561a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 45621035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4563a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 45641035eff8SStefano Zampini } 4565a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4566e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4567e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4568a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4569a7dc3881SStefano Zampini /* now compute set in local ordering */ 4570a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4571a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4572a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4573a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4574a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4575ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4576e456f2a8SStefano Zampini lsize++; 4577e456f2a8SStefano Zampini } 4578e456f2a8SStefano Zampini } 4579854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4580a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4581ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4582e456f2a8SStefano Zampini idxs[lsize++] = i; 4583e456f2a8SStefano Zampini } 4584e456f2a8SStefano Zampini } 4585a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4586a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4587e456f2a8SStefano Zampini *localis = localis_t; 4588e456f2a8SStefano Zampini PetscFunctionReturn(0); 4589e456f2a8SStefano Zampini } 4590906d46d4SStefano Zampini 4591906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4592906d46d4SStefano Zampini #undef __FUNCT__ 4593906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4594906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4595906d46d4SStefano Zampini { 4596906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4597906d46d4SStefano Zampini PetscErrorCode ierr; 4598906d46d4SStefano Zampini 4599906d46d4SStefano Zampini PetscFunctionBegin; 4600906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4601906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4602906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4603906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4604906d46d4SStefano Zampini PetscFunctionReturn(0); 4605906d46d4SStefano Zampini } 4606906d46d4SStefano Zampini 4607906d46d4SStefano Zampini #undef __FUNCT__ 4608906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4609906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4610906d46d4SStefano Zampini { 4611906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4612906d46d4SStefano Zampini PetscErrorCode ierr; 4613906d46d4SStefano Zampini 4614906d46d4SStefano Zampini PetscFunctionBegin; 4615906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4616906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4617906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4618906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4619906d46d4SStefano Zampini PetscFunctionReturn(0); 4620906d46d4SStefano Zampini } 4621b96c3477SStefano Zampini 4622b96c3477SStefano Zampini #undef __FUNCT__ 4623b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 462408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4625b96c3477SStefano Zampini { 4626b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4627b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4628b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4629b96c3477SStefano Zampini PetscBool free_used_adj; 4630b96c3477SStefano Zampini PetscErrorCode ierr; 4631b96c3477SStefano Zampini 4632b96c3477SStefano Zampini PetscFunctionBegin; 4633b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4634b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 463508122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4636b96c3477SStefano Zampini used_xadj = NULL; 4637b96c3477SStefano Zampini used_adjncy = NULL; 4638b96c3477SStefano Zampini } else { 463908122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 464008122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 464108122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 464208122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4643b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4644b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4645b96c3477SStefano Zampini } else { 46462fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4647b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4648b96c3477SStefano Zampini PetscInt nvtxs; 4649b96c3477SStefano Zampini 46502fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 46512fffb893SStefano Zampini if (flg_row) { 4652b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4653b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4654b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4655b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 46562fffb893SStefano Zampini } else { 46572fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 46582fffb893SStefano Zampini used_xadj = NULL; 46592fffb893SStefano Zampini used_adjncy = NULL; 46602fffb893SStefano Zampini } 46612fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4662b96c3477SStefano Zampini } 4663b96c3477SStefano Zampini } 46649552c7c7SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->adaptive_selection,pcbddc->use_deluxe_scaling,pcbddc->adaptive_invert_Stildas,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4665b96c3477SStefano Zampini 4666b96c3477SStefano Zampini /* free adjacency */ 4667b96c3477SStefano Zampini if (free_used_adj) { 4668b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4669b96c3477SStefano Zampini } 4670b96c3477SStefano Zampini PetscFunctionReturn(0); 4671b96c3477SStefano Zampini } 4672b96c3477SStefano Zampini 4673b96c3477SStefano Zampini #undef __FUNCT__ 4674b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 467508122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4676b96c3477SStefano Zampini { 4677b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4678b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4679b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4680b96c3477SStefano Zampini PCBDDCGraph graph; 4681b96c3477SStefano Zampini Mat S_j; 4682b96c3477SStefano Zampini PetscErrorCode ierr; 4683b96c3477SStefano Zampini 4684b96c3477SStefano Zampini PetscFunctionBegin; 4685b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 468608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4687b96c3477SStefano Zampini IS verticesIS; 4688b96c3477SStefano Zampini 4689b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4690b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 4691b96c3477SStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap);CHKERRQ(ierr); 4692b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4693b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4694b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4695b96c3477SStefano Zampini /* 4696b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4697b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4698b96c3477SStefano Zampini } 4699b96c3477SStefano Zampini */ 4700b96c3477SStefano Zampini } else { 4701b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4702b96c3477SStefano Zampini } 4703b96c3477SStefano Zampini 4704b96c3477SStefano Zampini /* Create Schur complement matrix */ 4705b96c3477SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4706b96c3477SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4707b96c3477SStefano Zampini 4708b96c3477SStefano Zampini /* sub_schurs init */ 470908122e43SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcbddc->local_mat,S_j,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_threshold);CHKERRQ(ierr); 4710b96c3477SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4711b96c3477SStefano Zampini /* free graph struct */ 471208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4713b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4714b96c3477SStefano Zampini } 4715b96c3477SStefano Zampini PetscFunctionReturn(0); 4716b96c3477SStefano Zampini } 4717