134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h> 234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 308122e43SStefano Zampini #include <petscblaslapack.h> 434a97f8cSStefano Zampini 53202ece2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*); 65ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*); 7d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorDestroy(PC); 8d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC,Vec,Vec); 9d5574798SStefano Zampini 10d5574798SStefano Zampini #undef __FUNCT__ 11d5574798SStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolve" 12d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC pc, Vec rhs, Vec sol) 13d5574798SStefano Zampini { 14d5574798SStefano Zampini PCBDDCMumpsInterior ctx; 15d5574798SStefano Zampini PetscScalar *array,*array_mumps; 16d5574798SStefano Zampini PetscInt ival; 17d5574798SStefano Zampini PetscErrorCode ierr; 18d5574798SStefano Zampini 19d5574798SStefano Zampini PetscFunctionBegin; 20d5574798SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 21d5574798SStefano Zampini ierr = MatMumpsGetIcntl(ctx->F,26,&ival);CHKERRQ(ierr); 22d5574798SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr); 23d5574798SStefano Zampini /* copy rhs into factored matrix workspace (can it be avoided?, MatSolve_MUMPS has another copy b->x internally) */ 24d5574798SStefano Zampini ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 25d5574798SStefano Zampini ierr = VecGetArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 26d5574798SStefano Zampini ierr = PetscMemcpy(array_mumps,array,ctx->n*sizeof(PetscScalar));CHKERRQ(ierr); 27d5574798SStefano Zampini ierr = VecRestoreArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 28d5574798SStefano Zampini ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 29d5574798SStefano Zampini 30d5574798SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 31d5574798SStefano Zampini 32d5574798SStefano Zampini /* get back data to caller worskpace */ 33d5574798SStefano Zampini ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 34d5574798SStefano Zampini ierr = VecGetArray(sol,&array);CHKERRQ(ierr); 35d5574798SStefano Zampini ierr = PetscMemcpy(array,array_mumps,ctx->n*sizeof(PetscScalar));CHKERRQ(ierr); 36d5574798SStefano Zampini ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr); 37d5574798SStefano Zampini ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 38d5574798SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,ival);CHKERRQ(ierr); 39d5574798SStefano Zampini PetscFunctionReturn(0); 40d5574798SStefano Zampini } 41d5574798SStefano Zampini 42d5574798SStefano Zampini #undef __FUNCT__ 43d5574798SStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorDestroy" 44d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorDestroy(PC pc) 45d5574798SStefano Zampini { 46d5574798SStefano Zampini PCBDDCMumpsInterior ctx; 47d5574798SStefano Zampini PetscErrorCode ierr; 48d5574798SStefano Zampini 49d5574798SStefano Zampini PetscFunctionBegin; 50d5574798SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 51d5574798SStefano Zampini ierr = MatDestroy(&ctx->F);CHKERRQ(ierr); 52d5574798SStefano Zampini ierr = VecDestroy(&ctx->sol);CHKERRQ(ierr); 53d5574798SStefano Zampini ierr = VecDestroy(&ctx->rhs);CHKERRQ(ierr); 54d5574798SStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 55d5574798SStefano Zampini PetscFunctionReturn(0); 56d5574798SStefano Zampini } 573202ece2SStefano Zampini 583202ece2SStefano Zampini #undef __FUNCT__ 593202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur" 605ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S) 613202ece2SStefano Zampini { 623202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 633202ece2SStefano Zampini KSP ksp; 643202ece2SStefano Zampini PC pc; 653202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 663202ece2SStefano Zampini PetscReal fill = 2.0; 67f11841e3SStefano Zampini PetscInt n_I; 683202ece2SStefano Zampini PetscMPIInt size; 693202ece2SStefano Zampini PetscErrorCode ierr; 703202ece2SStefano Zampini 713202ece2SStefano Zampini PetscFunctionBegin; 723202ece2SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr); 733202ece2SStefano Zampini if (size != 1) { 743202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 753202ece2SStefano Zampini } 76f11841e3SStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 77f11841e3SStefano Zampini PetscBool Sdense; 78f11841e3SStefano Zampini 79f11841e3SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr); 80f11841e3SStefano Zampini if (!Sdense) { 81f11841e3SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should dense"); 82f11841e3SStefano Zampini } 83f11841e3SStefano Zampini } 843202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 853202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 863202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 873202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 883202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 893202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 903202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 913202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 92f11841e3SStefano Zampini ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr); 93f11841e3SStefano Zampini if (n_I) { 943202ece2SStefano Zampini if (!Bdense) { 953202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 963202ece2SStefano Zampini } else { 973202ece2SStefano Zampini Bd = B; 983202ece2SStefano Zampini } 993202ece2SStefano Zampini 1003202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 1013202ece2SStefano Zampini Mat fact; 1023202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 1033202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 1043202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 1053202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 1063202ece2SStefano Zampini } else { 10707b1e237SStefano Zampini PetscBool ex = PETSC_TRUE; 10807b1e237SStefano Zampini 10907b1e237SStefano Zampini if (ex) { 1103202ece2SStefano Zampini Mat Ainvd; 1113202ece2SStefano Zampini 1123202ece2SStefano Zampini ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr); 1133202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 1143202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 11507b1e237SStefano Zampini } else { 11607b1e237SStefano Zampini Vec sol,rhs; 11707b1e237SStefano Zampini PetscScalar *arrayrhs,*arraysol; 11807b1e237SStefano Zampini PetscInt i,nrhs,n; 11907b1e237SStefano Zampini 12007b1e237SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 12107b1e237SStefano Zampini ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr); 12207b1e237SStefano Zampini ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr); 12307b1e237SStefano Zampini ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr); 12407b1e237SStefano Zampini ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr); 12507b1e237SStefano Zampini ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr); 12607b1e237SStefano Zampini for (i=0;i<nrhs;i++) { 12707b1e237SStefano Zampini ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr); 12807b1e237SStefano Zampini ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr); 12907b1e237SStefano Zampini ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr); 13007b1e237SStefano Zampini ierr = VecResetArray(rhs);CHKERRQ(ierr); 13107b1e237SStefano Zampini ierr = VecResetArray(sol);CHKERRQ(ierr); 13207b1e237SStefano Zampini } 13307b1e237SStefano Zampini ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr); 13407b1e237SStefano Zampini ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr); 13507b1e237SStefano Zampini } 1363202ece2SStefano Zampini } 1375ec10c6aSStefano Zampini if (!Bdense & !issym) { 1383202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 1393202ece2SStefano Zampini } 1405ec10c6aSStefano Zampini 1415ec10c6aSStefano Zampini if (!issym) { 1423202ece2SStefano Zampini if (!Cdense) { 1433202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 1443202ece2SStefano Zampini } else { 1453202ece2SStefano Zampini Cd = C; 1463202ece2SStefano Zampini } 1475ec10c6aSStefano Zampini ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 1483202ece2SStefano Zampini if (!Cdense) { 1493202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 1503202ece2SStefano Zampini } 1515ec10c6aSStefano Zampini } else { 1525ec10c6aSStefano Zampini ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 1535ec10c6aSStefano Zampini if (!Bdense) { 1545ec10c6aSStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 1555ec10c6aSStefano Zampini } 1565ec10c6aSStefano Zampini } 1575ec10c6aSStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 158f11841e3SStefano Zampini } 1593202ece2SStefano Zampini 1603202ece2SStefano Zampini if (D) { 1613202ece2SStefano Zampini Mat Dd; 1623202ece2SStefano Zampini PetscBool Ddense; 1633202ece2SStefano Zampini 1643202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 1653202ece2SStefano Zampini if (!Ddense) { 1663202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 1673202ece2SStefano Zampini } else { 1683202ece2SStefano Zampini Dd = D; 1693202ece2SStefano Zampini } 170f11841e3SStefano Zampini if (n_I) { 1713202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 172f11841e3SStefano Zampini } else { 173f11841e3SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 174f11841e3SStefano Zampini ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr); 175f11841e3SStefano Zampini } else { 176f11841e3SStefano Zampini ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 177f11841e3SStefano Zampini } 178f11841e3SStefano Zampini } 1793202ece2SStefano Zampini if (!Ddense) { 1803202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 1813202ece2SStefano Zampini } 1823202ece2SStefano Zampini } else { 1833202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 1843202ece2SStefano Zampini } 1853202ece2SStefano Zampini PetscFunctionReturn(0); 1863202ece2SStefano Zampini } 18734a97f8cSStefano Zampini 18834a97f8cSStefano Zampini #undef __FUNCT__ 1891580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp" 190a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool faster_deluxe, PetscBool compute_Stilda, PetscBool use_edges, PetscBool use_faces) 191b1b3d7a2SStefano Zampini { 1925a95e1ceSStefano Zampini Mat A_II,A_IB,A_BI,A_BB,AE_II; 193d2627357SStefano Zampini Mat S_all,S_all_inv; 194d2627357SStefano Zampini Mat global_schur_subsets,work_mat; 19508122e43SStefano Zampini Mat S_Ej_tilda_all,S_Ej_inv_all; 1965db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 197*a9b99552SStefano Zampini IS is_I,is_I_layer,temp_is; 198d648f858SStefano Zampini PetscInt *nnz,*all_local_idx_G,*all_local_idx_N; 199eb595f79SStefano Zampini PetscInt *auxnum1,*auxnum2,*all_local_idx_G_rep; 2005a95e1ceSStefano Zampini PetscInt i,subset_size,max_subset_size; 201883469d8SStefano Zampini PetscInt extra,local_size,global_size; 20208122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 20306a4b1faSStefano Zampini PetscScalar *Bwork; 2045a95e1ceSStefano Zampini PetscSubcomm subcomm; 2055a95e1ceSStefano Zampini PetscMPIInt color,rank; 2065a95e1ceSStefano Zampini MPI_Comm comm_n; 207b1b3d7a2SStefano Zampini PetscErrorCode ierr; 208b1b3d7a2SStefano Zampini 209b1b3d7a2SStefano Zampini PetscFunctionBegin; 210a64f4aa4SStefano Zampini /* update info in sub_schurs */ 211a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 212a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 213a64f4aa4SStefano Zampini if (Ain) { 214a64f4aa4SStefano Zampini PetscBool isseqaij; 215a64f4aa4SStefano Zampini 216a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 217a64f4aa4SStefano Zampini if (isseqaij) { 218a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 219a64f4aa4SStefano Zampini sub_schurs->A = Ain; 220a64f4aa4SStefano Zampini } else { /* SeqBAIJ matrices does not support symmetry checking, SeqSBAIJ does not support MatPermute */ 221a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 222a64f4aa4SStefano Zampini } 223a64f4aa4SStefano Zampini } 224a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 225a64f4aa4SStefano Zampini sub_schurs->S = Sin; 226a64f4aa4SStefano Zampini if (sub_schurs->use_mumps) { 227a64f4aa4SStefano Zampini sub_schurs->use_mumps = (PetscBool)(!!sub_schurs->A); 228a64f4aa4SStefano Zampini } 229a64f4aa4SStefano Zampini 2305a95e1ceSStefano Zampini /* preliminary checks */ 2315a95e1ceSStefano Zampini if (!sub_schurs->use_mumps && compute_Stilda) { 2325a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 2335a95e1ceSStefano Zampini } 2345a95e1ceSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 2355a95e1ceSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 2365a95e1ceSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 2375a95e1ceSStefano Zampini if (sub_schurs->A) { 2385a95e1ceSStefano Zampini PetscInt lsize; 2395a95e1ceSStefano Zampini 2405a95e1ceSStefano Zampini ierr = MatGetSize(sub_schurs->A,&lsize,NULL);CHKERRQ(ierr); 2415a95e1ceSStefano Zampini if (lsize) { 2425a95e1ceSStefano Zampini ierr = MatIsHermitian(sub_schurs->A,0.0,&sub_schurs->is_hermitian);CHKERRQ(ierr); 2435a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 2445a95e1ceSStefano Zampini PetscScalar val; 2455a95e1ceSStefano Zampini Vec vec1,vec2; 2465a95e1ceSStefano Zampini 2475a95e1ceSStefano Zampini ierr = MatCreateVecs(sub_schurs->A,&vec1,&vec2);CHKERRQ(ierr); 2485a95e1ceSStefano Zampini ierr = VecSetRandom(vec1,NULL); 2495a95e1ceSStefano Zampini ierr = VecCopy(vec1,vec2);CHKERRQ(ierr); 2505a95e1ceSStefano Zampini ierr = MatMult(sub_schurs->A,vec2,vec1);CHKERRQ(ierr); 2515a95e1ceSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 2525a95e1ceSStefano Zampini if (PetscRealPart(val) > 0. && PetscImaginaryPart(val) == 0.) sub_schurs->is_posdef = PETSC_TRUE; 2535a95e1ceSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 2545a95e1ceSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 2555a95e1ceSStefano Zampini } 2565a95e1ceSStefano Zampini } else { 2575a95e1ceSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 2585a95e1ceSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 2595a95e1ceSStefano Zampini } 2605a95e1ceSStefano Zampini if (compute_Stilda && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 2615a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"General matrix pencils are not currently supported"); 2625a95e1ceSStefano Zampini } 2635a95e1ceSStefano Zampini } 2645a95e1ceSStefano Zampini /* restrict work on active processes */ 2655a95e1ceSStefano Zampini color = 0; 2665a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 2675a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 2685a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 2695a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 2705a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 2715a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 2725a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 2735a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 2745a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 2755a95e1ceSStefano Zampini PetscFunctionReturn(0); 2765a95e1ceSStefano Zampini } 2775a95e1ceSStefano Zampini 278b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 279883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 280a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 281f11841e3SStefano Zampini PetscBool isseqaij; 282a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 283a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 284f11841e3SStefano Zampini if (!isseqaij) { 285a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 286a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 287a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 288a64f4aa4SStefano Zampini } else { 289a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 290a64f4aa4SStefano Zampini A_BB = tA_BB; 291a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 292a64f4aa4SStefano Zampini A_IB = tA_IB; 293a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 294a64f4aa4SStefano Zampini A_BI = tA_BI; 295f11841e3SStefano Zampini } 296a58a30b4SStefano Zampini } else { 2975a95e1ceSStefano Zampini A_II = NULL; 2985a95e1ceSStefano Zampini A_IB = NULL; 2995a95e1ceSStefano Zampini A_BI = NULL; 3005a95e1ceSStefano Zampini A_BB = NULL; 301b1b3d7a2SStefano Zampini } 3025a95e1ceSStefano Zampini S_all = NULL; 3035a95e1ceSStefano Zampini S_all_inv = NULL; 3045a95e1ceSStefano Zampini S_Ej_tilda_all = NULL; 3055a95e1ceSStefano Zampini S_Ej_inv_all = NULL; 306b1b3d7a2SStefano Zampini 307b1b3d7a2SStefano Zampini /* determine interior problems */ 3083dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 3093dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 310b1b3d7a2SStefano Zampini PetscBT touched; 311b1b3d7a2SStefano Zampini const PetscInt* idx_B; 312b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 313b1b3d7a2SStefano Zampini 3143dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 3153dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 3163dc780c3SStefano Zampini } 317b1b3d7a2SStefano Zampini /* get sizes */ 318b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 319b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 320b1b3d7a2SStefano Zampini 321b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 322b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 323b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 324b1b3d7a2SStefano Zampini 325b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 326b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 327b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 328b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 329b1b3d7a2SStefano Zampini } 330b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 331b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 332b1b3d7a2SStefano Zampini 333b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 334b1b3d7a2SStefano Zampini n_local_dofs = n_B; 335b1b3d7a2SStefano Zampini n_prev_added = n_B; 336b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 337b1b3d7a2SStefano Zampini PetscInt n_added; 338b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 339b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 340b1b3d7a2SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+n_B); 341b1b3d7a2SStefano Zampini } 342b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 343b1b3d7a2SStefano Zampini n_prev_added = n_added; 344b1b3d7a2SStefano Zampini n_local_dofs += n_added; 345b1b3d7a2SStefano Zampini if (!n_added) break; 346b1b3d7a2SStefano Zampini } 347b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 348b1b3d7a2SStefano Zampini 349883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 350*a9b99552SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&is_I_layer);CHKERRQ(ierr); 351b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 352*a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 353883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 354883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 355b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 356b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 357*a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 358b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 359b1b3d7a2SStefano Zampini 360b1b3d7a2SStefano Zampini /* II block */ 361b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 362b1b3d7a2SStefano Zampini } 363b1b3d7a2SStefano Zampini } else { 364b1b3d7a2SStefano Zampini PetscInt n_I; 365b1b3d7a2SStefano Zampini 366b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 367b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 368*a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 369b1b3d7a2SStefano Zampini 370b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 371883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 372b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 373b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 374b1b3d7a2SStefano Zampini 375b1b3d7a2SStefano Zampini /* II block is the same */ 376b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 377b1b3d7a2SStefano Zampini AE_II = A_II; 378b1b3d7a2SStefano Zampini } 379b1b3d7a2SStefano Zampini } 3805a95e1ceSStefano Zampini 381883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 3825a95e1ceSStefano Zampini max_subset_size = 0; 383883469d8SStefano Zampini local_size = 0; 3845a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 3855a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 3865a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 387883469d8SStefano Zampini local_size += subset_size; 388883469d8SStefano Zampini } 389883469d8SStefano Zampini 390883469d8SStefano Zampini /* Work arrays for local indices */ 391883469d8SStefano Zampini extra = 0; 392883469d8SStefano Zampini if (sub_schurs->use_mumps) { 393*a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 394883469d8SStefano Zampini } 395883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 396883469d8SStefano Zampini if (extra) { 397883469d8SStefano Zampini const PetscInt *idxs; 398*a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 399883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 400*a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 401883469d8SStefano Zampini } 402883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 403eb595f79SStefano Zampini ierr = PetscMalloc2(sub_schurs->n_subs,&auxnum1,sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 404883469d8SStefano Zampini 405883469d8SStefano Zampini /* Get local indices in local numbering */ 406883469d8SStefano Zampini local_size = 0; 4075a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 408883469d8SStefano Zampini PetscInt j; 409883469d8SStefano Zampini const PetscInt *idxs; 410883469d8SStefano Zampini 4115a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4125a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 413eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 414eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 415eb595f79SStefano Zampini auxnum2[i] = subset_size; 416883469d8SStefano Zampini /* subset indices in local numbering */ 417883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 4185a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 419883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 420883469d8SStefano Zampini local_size += subset_size; 421883469d8SStefano Zampini } 422883469d8SStefano Zampini 4235a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 424d2627357SStefano Zampini Bwork = NULL; 425d2627357SStefano Zampini pivots = NULL; 4265a95e1ceSStefano Zampini if (sub_schurs->n_subs && !sub_schurs->is_hermitian) { 427d2627357SStefano Zampini PetscScalar lwork; 428d2627357SStefano Zampini 429d2627357SStefano Zampini B_lwork = -1; 430d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 431d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 432d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 433d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 434d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 435d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 436d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 437d2627357SStefano Zampini } 438d2627357SStefano Zampini 439d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 440eb595f79SStefano Zampini ierr = PCBDDCSubsetNumbering(comm_n,sub_schurs->l2gmap,sub_schurs->n_subs,auxnum1,auxnum2,&global_size,&all_local_idx_G_rep);CHKERRQ(ierr); 441eb595f79SStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_G);CHKERRQ(ierr); 442eb595f79SStefano Zampini local_size = 0; 443eb595f79SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 444eb595f79SStefano Zampini PetscInt j; 445eb595f79SStefano Zampini for (j=0;j<auxnum2[i];j++) all_local_idx_G[local_size++] = all_local_idx_G_rep[i] + j; 446eb595f79SStefano Zampini } 447eb595f79SStefano Zampini ierr = PetscFree(all_local_idx_G_rep);CHKERRQ(ierr); 448eb595f79SStefano Zampini ierr = PetscFree2(auxnum1,auxnum2);CHKERRQ(ierr); 4495a95e1ceSStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,local_size,all_local_idx_G,PETSC_COPY_VALUES,&l2gmap_subsets);CHKERRQ(ierr); 4505a95e1ceSStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 451d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 452d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 453d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 454d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 4552972d61bSStefano Zampini 4565a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 4575a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 4585a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 4595a95e1ceSStefano Zampini 4605a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 4615a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 4625a95e1ceSStefano Zampini if (subset_size != local_size) { 4635a95e1ceSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 4645a95e1ceSStefano Zampini } 4655a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 466b1b3d7a2SStefano Zampini } 467b1b3d7a2SStefano Zampini 4685a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 4695a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 4705a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4715a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 4725a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 4735a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 474aa83b6aeSStefano Zampini } else { 4755a95e1ceSStefano Zampini ierr = MatZeroEntries(sub_schurs->S_Ej_all);CHKERRQ(ierr); 476aa83b6aeSStefano Zampini } 477b1b3d7a2SStefano Zampini 4785a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 4795a95e1ceSStefano Zampini ierr = PetscBTMemzero(sub_schurs->n_subs,sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 4805a95e1ceSStefano Zampini if (!sub_schurs->use_mumps) { 4815a95e1ceSStefano Zampini Mat S_Ej_expl; 4825a95e1ceSStefano Zampini PetscScalar *work; 4835a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 4845a95e1ceSStefano Zampini PetscBool Sdense; 4855a95e1ceSStefano Zampini 4865a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 4875a95e1ceSStefano Zampini local_size = 0; 488b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4895a95e1ceSStefano Zampini IS is_subset_B; 4905a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 4915a95e1ceSStefano Zampini 4925a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 4935a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 4945a95e1ceSStefano Zampini /* EE block */ 4955a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 4965a95e1ceSStefano Zampini /* IE block */ 4975a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 4985a95e1ceSStefano Zampini /* EI block */ 4995a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 5005a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 5015a95e1ceSStefano Zampini } else { 5025a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 5035a95e1ceSStefano Zampini } 504a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 5055a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 5065a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 5075a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 5085a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 509b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 510b1b3d7a2SStefano Zampini KSP ksp; 511b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 5125a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 513b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 514b1b3d7a2SStefano Zampini KSP origksp,schurksp; 515b1b3d7a2SStefano Zampini PC origpc,schurpc; 516b1b3d7a2SStefano Zampini KSPType ksp_type; 517b1b3d7a2SStefano Zampini PetscInt n_internal; 5185a95e1ceSStefano Zampini PetscBool ispcnone; 519b1b3d7a2SStefano Zampini 520b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 5215a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 522b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 523b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 524b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 525b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 5265a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 5275a95e1ceSStefano Zampini if (!ispcnone) { 5285a95e1ceSStefano Zampini PCType pc_type; 529b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 530b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 5315a95e1ceSStefano Zampini } else { 5325a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 5335a95e1ceSStefano Zampini } 534b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 535b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 536b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 537b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 538b1b3d7a2SStefano Zampini if (solver) { 539b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 540b1b3d7a2SStefano Zampini } 541b1b3d7a2SStefano Zampini } 542b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 543b1b3d7a2SStefano Zampini } 5445a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5455a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 5465a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 5475a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 5485a95e1ceSStefano Zampini if (Sdense) { 5495a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 5505a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 551b1b3d7a2SStefano Zampini } 5525a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 5535a95e1ceSStefano Zampini } else { 5545a95e1ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 5555a95e1ceSStefano Zampini } 5565a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 557a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 5585a95e1ceSStefano Zampini local_size += subset_size; 5595a95e1ceSStefano Zampini } 5605a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 561b1b3d7a2SStefano Zampini /* free */ 562b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 563b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 5645a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 565883469d8SStefano Zampini } else { 566883469d8SStefano Zampini Mat A,F; 567883469d8SStefano Zampini IS is_A_all; 5685a95e1ceSStefano Zampini PetscScalar *work; 569d5574798SStefano Zampini PetscInt *idxs_schur,n_I,n_I_all,*dummy_idx; 570883469d8SStefano Zampini 571883469d8SStefano Zampini /* get working mat */ 572*a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 573883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 574d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(sub_schurs->A,is_A_all,is_A_all,&A);CHKERRQ(ierr); 575883469d8SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 576*a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 577883469d8SStefano Zampini 57808122e43SStefano Zampini if (n_I) { 5799ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 580883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 581883469d8SStefano Zampini } else { 582883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 583883469d8SStefano Zampini } 584883469d8SStefano Zampini 585883469d8SStefano Zampini /* subsets ordered last */ 586883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&idxs_schur);CHKERRQ(ierr); 587883469d8SStefano Zampini for (i=0;i<local_size;i++) { 588883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 589883469d8SStefano Zampini } 5905a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 591883469d8SStefano Zampini ierr = MatMumpsSetSchurIndices(F,local_size,idxs_schur);CHKERRQ(ierr); 5925a95e1ceSStefano Zampini #endif 593883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 594883469d8SStefano Zampini 595883469d8SStefano Zampini /* factorization step */ 5969ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 597883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 598883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 599883469d8SStefano Zampini } else { 600883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 601883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 602883469d8SStefano Zampini } 603883469d8SStefano Zampini 604883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 6055a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 606883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 6075a95e1ceSStefano Zampini #endif 608d5574798SStefano Zampini 609d5574798SStefano Zampini /* we can reuse the interior solver if we are not using the economic version */ 610d5574798SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I_all);CHKERRQ(ierr); 611d5574798SStefano Zampini if (n_I == n_I_all) { 612d5574798SStefano Zampini PCBDDCMumpsInterior msolv_ctx; 613d5574798SStefano Zampini 614a64f4aa4SStefano Zampini ierr = PCDestroy(&sub_schurs->interior_solver);CHKERRQ(ierr); 615d5574798SStefano Zampini ierr = PetscNew(&msolv_ctx);CHKERRQ(ierr); 616d5574798SStefano Zampini msolv_ctx->n = n_I; 617d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 618d5574798SStefano Zampini msolv_ctx->F = F; 619d5574798SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,&msolv_ctx->rhs);CHKERRQ(ierr); 620d5574798SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&sub_schurs->interior_solver);CHKERRQ(ierr); 621d5574798SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 622d5574798SStefano Zampini ierr = PCSetOperators(sub_schurs->interior_solver,A_II,A_II);CHKERRQ(ierr); 623d5574798SStefano Zampini ierr = PCSetType(sub_schurs->interior_solver,PCSHELL);CHKERRQ(ierr); 624d5574798SStefano Zampini ierr = PCShellSetContext(sub_schurs->interior_solver,msolv_ctx);CHKERRQ(ierr); 625d5574798SStefano Zampini ierr = PCShellSetApply(sub_schurs->interior_solver,PCBDDCMumpsInteriorSolve);CHKERRQ(ierr); 626d5574798SStefano Zampini ierr = PCShellSetDestroy(sub_schurs->interior_solver,PCBDDCMumpsInteriorDestroy);CHKERRQ(ierr); 627d5574798SStefano Zampini } 628883469d8SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 62908122e43SStefano Zampini } else { 63008122e43SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 63108122e43SStefano Zampini } 63208122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 633d2627357SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 6345db18549SStefano Zampini 63512d906b1SStefano Zampini if (compute_Stilda) { 636a1337663SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_tilda_all);CHKERRQ(ierr); 637a1337663SStefano Zampini ierr = MatSetSizes(S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 638a1337663SStefano Zampini ierr = MatSetType(S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 639a1337663SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 64008122e43SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_inv_all);CHKERRQ(ierr); 64108122e43SStefano Zampini ierr = MatSetSizes(S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 64208122e43SStefano Zampini ierr = MatSetType(S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 64308122e43SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 64406a4b1faSStefano Zampini 64512d906b1SStefano Zampini /* compute St^-1 */ 64612d906b1SStefano Zampini if (local_size) { /* multilevel guard */ 647d2627357SStefano Zampini PetscScalar *vals; 648f6f667cfSStefano Zampini 6495a95e1ceSStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 650d2627357SStefano Zampini ierr = MatDuplicate(S_all,MAT_COPY_VALUES,&S_all_inv);CHKERRQ(ierr); 651d2627357SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&vals);CHKERRQ(ierr); 652d2627357SStefano Zampini if (!sub_schurs->is_hermitian) { 653d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals,&B_N,pivots,&B_ierr)); 654d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 655d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 656d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 657d2627357SStefano Zampini } else { 658d2627357SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,vals,&B_N,&B_ierr)); 659d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 660d2627357SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,vals,&B_N,&B_ierr)); 661d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 662d2627357SStefano Zampini } 663d2627357SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&vals);CHKERRQ(ierr); 664d2627357SStefano Zampini } 66512d906b1SStefano Zampini } 666d2627357SStefano Zampini 6679087bf02SStefano Zampini /* Work arrays */ 6689087bf02SStefano Zampini if (sub_schurs->n_subs == 1) { 6699087bf02SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 6709087bf02SStefano Zampini } else { 6719087bf02SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 6729087bf02SStefano Zampini } 6739087bf02SStefano Zampini 6745a95e1ceSStefano Zampini local_size = 0; 67565d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6765a95e1ceSStefano Zampini Mat S_Ej; 67765d8bf0aSStefano Zampini IS is_E; 67865d8bf0aSStefano Zampini PetscInt j; 67965d8bf0aSStefano Zampini 6805a95e1ceSStefano Zampini /* get S_E */ 681b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6829087bf02SStefano Zampini if (sub_schurs->n_subs == 1) { 6839087bf02SStefano Zampini ierr = MatDenseGetArray(S_all,&work);CHKERRQ(ierr); 6849087bf02SStefano Zampini S_Ej = NULL; 6859087bf02SStefano Zampini is_E = NULL; 6869087bf02SStefano Zampini } else { 6875a95e1ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,subset_size,local_size,1,&is_E);CHKERRQ(ierr); 6885a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej);CHKERRQ(ierr); 6895a95e1ceSStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_REUSE_MATRIX,&S_Ej);CHKERRQ(ierr); 6909087bf02SStefano Zampini } 6915a95e1ceSStefano Zampini /* insert S_E values */ 692a1337663SStefano Zampini for (j=0;j<subset_size;j++) { 693a1337663SStefano Zampini dummy_idx[j]=local_size+j; 694a1337663SStefano Zampini } 6955a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 696a1337663SStefano Zampini 6975a95e1ceSStefano Zampini /* if adaptivity is requested, invert S_E and insert St_E^-1 blocks */ 698d2627357SStefano Zampini if (compute_Stilda && ((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || (!PetscBTLookup(sub_schurs->is_edge,i) && use_faces))) { 6995a95e1ceSStefano Zampini /* get S_E^-1 */ 70008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 70108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 7022972d61bSStefano Zampini if (!sub_schurs->is_hermitian) { 7035a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 70408122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 7055a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 70608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 7072972d61bSStefano Zampini } else { 7085a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 7092972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 7105a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 7112972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 7122972d61bSStefano Zampini } 71308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 7145a95e1ceSStefano Zampini ierr = MatSetValues(S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 7155a95e1ceSStefano Zampini 7165a95e1ceSStefano Zampini /* get St_E^-1 */ 7179087bf02SStefano Zampini if (sub_schurs->n_subs == 1) { 7189087bf02SStefano Zampini ierr = MatDenseRestoreArray(S_all,&work);CHKERRQ(ierr); 7199087bf02SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&work);CHKERRQ(ierr); 7209087bf02SStefano Zampini } else { 7215a95e1ceSStefano Zampini ierr = MatGetSubMatrix(S_all_inv,is_E,is_E,MAT_REUSE_MATRIX,&S_Ej);CHKERRQ(ierr); 7229087bf02SStefano Zampini } 7235a95e1ceSStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 7249087bf02SStefano Zampini if (sub_schurs->n_subs == 1) { 7259087bf02SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&work);CHKERRQ(ierr); 7269087bf02SStefano Zampini } 7279087bf02SStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,i);CHKERRQ(ierr); 7289087bf02SStefano Zampini } else if (sub_schurs->n_subs == 1) { 7299087bf02SStefano Zampini ierr = MatDenseRestoreArray(S_all,&work);CHKERRQ(ierr); 73008122e43SStefano Zampini } 7315a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 73265d8bf0aSStefano Zampini ierr = ISDestroy(&is_E);CHKERRQ(ierr); 733883469d8SStefano Zampini local_size += subset_size; 734883469d8SStefano Zampini } 7359087bf02SStefano Zampini if (sub_schurs->n_subs == 1) { 7369087bf02SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 7379087bf02SStefano Zampini } else { 7385ec10c6aSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 7395db18549SStefano Zampini } 7409087bf02SStefano Zampini } 741*a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 742a1337663SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 743a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 744d2627357SStefano Zampini ierr = MatDestroy(&S_all_inv);CHKERRQ(ierr); 745a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 746a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 747a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 7485db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7495db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7505a95e1ceSStefano Zampini if (compute_Stilda) { 751a1337663SStefano Zampini ierr = MatAssemblyBegin(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 752a1337663SStefano Zampini ierr = MatAssemblyEnd(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 75308122e43SStefano Zampini ierr = MatAssemblyBegin(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 75408122e43SStefano Zampini ierr = MatAssemblyEnd(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 75508122e43SStefano Zampini } 756a1337663SStefano Zampini 7575db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 7585db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 7593927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 7603927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 7615a95e1ceSStefano Zampini 7625db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 7635a95e1ceSStefano Zampini ierr = ISCreateGeneral(comm_n,local_size,all_local_idx_G,PETSC_OWN_POINTER,&temp_is);CHKERRQ(ierr); 764d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 765d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 76608122e43SStefano Zampini 767ac632422SStefano Zampini /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */ 768ac632422SStefano Zampini if (faster_deluxe) { 7695a95e1ceSStefano Zampini Mat tmpmat; 7705a95e1ceSStefano Zampini PetscScalar *array; 7715a95e1ceSStefano Zampini PetscInt cum; 7725a95e1ceSStefano Zampini 7735a95e1ceSStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 7745a95e1ceSStefano Zampini cum = 0; 7755a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7765a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7775a95e1ceSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 7785a95e1ceSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 7795a95e1ceSStefano Zampini if (!sub_schurs->is_hermitian) { 7805a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 7815a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 7825a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 7835a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 7845a95e1ceSStefano Zampini } else { 7855a95e1ceSStefano Zampini PetscInt j,k; 7865a95e1ceSStefano Zampini 7875a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 7885a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 7895a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 7905a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 7915a95e1ceSStefano Zampini for (j=0;j<B_N;j++) { 7925a95e1ceSStefano Zampini for (k=j+1;k<B_N;k++) { 7935a95e1ceSStefano Zampini array[k*B_N+j+cum] = array[j*B_N+k+cum]; 7945a95e1ceSStefano Zampini } 7955a95e1ceSStefano Zampini } 7965a95e1ceSStefano Zampini } 7975a95e1ceSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 7985a95e1ceSStefano Zampini cum += subset_size*subset_size; 7995a95e1ceSStefano Zampini } 8005a95e1ceSStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 8015a95e1ceSStefano Zampini ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr); 8025a95e1ceSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 803ac632422SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 8045a95e1ceSStefano Zampini sub_schurs->S_Ej_all = tmpmat; 8055a95e1ceSStefano Zampini } 8065a95e1ceSStefano Zampini 807f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 8085a95e1ceSStefano Zampini if (compute_Stilda) { 809a1337663SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_tilda_all);CHKERRQ(ierr); 810a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 811d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 812d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 81308122e43SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_inv_all);CHKERRQ(ierr); 81408122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 815d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 816d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 81708122e43SStefano Zampini } 8183202ece2SStefano Zampini 8195a95e1ceSStefano Zampini /* free workspace */ 82006a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 821a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 822a1337663SStefano Zampini ierr = MatDestroy(&S_Ej_tilda_all);CHKERRQ(ierr); 82308122e43SStefano Zampini ierr = MatDestroy(&S_Ej_inv_all);CHKERRQ(ierr); 8243202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 8255db18549SStefano Zampini ierr = ISDestroy(&temp_is);CHKERRQ(ierr); 8265a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 827b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 828b1b3d7a2SStefano Zampini } 829b1b3d7a2SStefano Zampini 830b1b3d7a2SStefano Zampini #undef __FUNCT__ 831b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 832a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 833b1b3d7a2SStefano Zampini { 8349bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 8355a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 836b1b3d7a2SStefano Zampini PetscBool is_sorted; 837b1b3d7a2SStefano Zampini PetscErrorCode ierr; 838b1b3d7a2SStefano Zampini 839b1b3d7a2SStefano Zampini PetscFunctionBegin; 840b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 841b1b3d7a2SStefano Zampini if (!is_sorted) { 842b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 843b1b3d7a2SStefano Zampini } 844b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 845b1b3d7a2SStefano Zampini if (!is_sorted) { 846b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 847b1b3d7a2SStefano Zampini } 848b1b3d7a2SStefano Zampini 849b1b3d7a2SStefano Zampini /* reset any previous data */ 850b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 851b1b3d7a2SStefano Zampini 8525a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 8539bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 854b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 85508122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 85608122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 857b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 858b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 859b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 860b1b3d7a2SStefano Zampini } 861b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 862b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 86308122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 864b1b3d7a2SStefano Zampini } 865b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 866b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 867b1b3d7a2SStefano Zampini 8685a95e1ceSStefano Zampini /* Determine if MUMPS cane be used */ 869883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 870883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 871a64f4aa4SStefano Zampini sub_schurs->use_mumps = PETSC_TRUE; 872883469d8SStefano Zampini #endif 873b1b3d7a2SStefano Zampini 874b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 875b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 876b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 877b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 8785db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 8795db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 8805db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 8815db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 8825a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 883b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 884a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { /* sort by local ordering mumps is not present */ 885b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 886b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 887b96c3477SStefano Zampini } 8889bb4a8caSStefano Zampini } 8899bb4a8caSStefano Zampini sub_schurs->is_Ej_com = vertices; 890b1b3d7a2SStefano Zampini 891b96c3477SStefano Zampini /* allocate space for schur complements */ 892b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 893b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 89408122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 895b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 896b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 897b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 898b1b3d7a2SStefano Zampini } 899b1b3d7a2SStefano Zampini 900b1b3d7a2SStefano Zampini #undef __FUNCT__ 90134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 90234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 90334a97f8cSStefano Zampini { 90434a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 90534a97f8cSStefano Zampini PetscErrorCode ierr; 90634a97f8cSStefano Zampini 90734a97f8cSStefano Zampini PetscFunctionBegin; 90834a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 9095ff63025SStefano Zampini schurs_ctx->n_subs = 0; 91034a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 91134a97f8cSStefano Zampini PetscFunctionReturn(0); 91234a97f8cSStefano Zampini } 91334a97f8cSStefano Zampini 91434a97f8cSStefano Zampini #undef __FUNCT__ 91534a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 91634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 91734a97f8cSStefano Zampini { 91834a97f8cSStefano Zampini PetscErrorCode ierr; 91934a97f8cSStefano Zampini 92034a97f8cSStefano Zampini PetscFunctionBegin; 92134a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 92234a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 92334a97f8cSStefano Zampini PetscFunctionReturn(0); 92434a97f8cSStefano Zampini } 92534a97f8cSStefano Zampini 92634a97f8cSStefano Zampini #undef __FUNCT__ 92734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 92834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 92934a97f8cSStefano Zampini { 93034a97f8cSStefano Zampini PetscInt i; 93134a97f8cSStefano Zampini PetscErrorCode ierr; 93234a97f8cSStefano Zampini 93334a97f8cSStefano Zampini PetscFunctionBegin; 9341e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 935b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 936b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 937b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 9385db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 9395db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 94041c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 94141c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 94208122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 943a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 9445db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 9459bb4a8caSStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_com);CHKERRQ(ierr); 94608122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 94708122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 94834a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 949b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 95034a97f8cSStefano Zampini } 9515ff63025SStefano Zampini if (sub_schurs->n_subs) { 952b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 9533dc780c3SStefano Zampini } 954d5574798SStefano Zampini ierr = PCDestroy(&sub_schurs->interior_solver);CHKERRQ(ierr); 95534a97f8cSStefano Zampini sub_schurs->n_subs = 0; 95634a97f8cSStefano Zampini PetscFunctionReturn(0); 95734a97f8cSStefano Zampini } 95834a97f8cSStefano Zampini 95934a97f8cSStefano Zampini #undef __FUNCT__ 96034a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 9612a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 96234a97f8cSStefano Zampini { 96334a97f8cSStefano Zampini PetscInt i,j,n; 96434a97f8cSStefano Zampini PetscErrorCode ierr; 96534a97f8cSStefano Zampini 96634a97f8cSStefano Zampini PetscFunctionBegin; 96734a97f8cSStefano Zampini n = 0; 96834a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 96934a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 97034a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 97134a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 97234a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 97334a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 97434a97f8cSStefano Zampini queue_tip[n] = dof; 97534a97f8cSStefano Zampini n++; 97634a97f8cSStefano Zampini } 97734a97f8cSStefano Zampini } 97834a97f8cSStefano Zampini } 97934a97f8cSStefano Zampini *n_added = n; 98034a97f8cSStefano Zampini PetscFunctionReturn(0); 98134a97f8cSStefano Zampini } 982