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*); 63202ece2SStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, Mat *S); 73202ece2SStefano Zampini 83202ece2SStefano Zampini #undef __FUNCT__ 93202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur" 103202ece2SStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, Mat *S) 113202ece2SStefano Zampini { 123202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 133202ece2SStefano Zampini KSP ksp; 143202ece2SStefano Zampini PC pc; 153202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 163202ece2SStefano Zampini PetscReal fill = 2.0; 173202ece2SStefano Zampini PetscMPIInt size; 183202ece2SStefano Zampini PetscErrorCode ierr; 193202ece2SStefano Zampini 203202ece2SStefano Zampini PetscFunctionBegin; 213202ece2SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr); 223202ece2SStefano Zampini if (size != 1) { 233202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 243202ece2SStefano Zampini } 253202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 263202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 273202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 283202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 293202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 303202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 313202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 323202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 333202ece2SStefano Zampini if (!Bdense) { 343202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 353202ece2SStefano Zampini } else { 363202ece2SStefano Zampini Bd = B; 373202ece2SStefano Zampini } 383202ece2SStefano Zampini 393202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 403202ece2SStefano Zampini Mat fact; 413202ece2SStefano Zampini 423202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 433202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 443202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 453202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 463202ece2SStefano Zampini } else { 473202ece2SStefano Zampini Mat Ainvd; 483202ece2SStefano Zampini 493202ece2SStefano Zampini ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr); 503202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 513202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 523202ece2SStefano Zampini } 533202ece2SStefano Zampini if (!Bdense) { 543202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 553202ece2SStefano Zampini } 563202ece2SStefano Zampini if (!Cdense) { 573202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 583202ece2SStefano Zampini } else { 593202ece2SStefano Zampini Cd = C; 603202ece2SStefano Zampini } 613202ece2SStefano Zampini 623202ece2SStefano Zampini ierr = MatMatMult(Cd, AinvBd, MAT_INITIAL_MATRIX, fill, S);CHKERRQ(ierr); 633202ece2SStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 643202ece2SStefano Zampini if (!Cdense) { 653202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 663202ece2SStefano Zampini } 673202ece2SStefano Zampini 683202ece2SStefano Zampini if (D) { 693202ece2SStefano Zampini Mat Dd; 703202ece2SStefano Zampini PetscBool Ddense; 713202ece2SStefano Zampini 723202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 733202ece2SStefano Zampini if (!Ddense) { 743202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 753202ece2SStefano Zampini } else { 763202ece2SStefano Zampini Dd = D; 773202ece2SStefano Zampini } 783202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 793202ece2SStefano Zampini if (!Ddense) { 803202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 813202ece2SStefano Zampini } 823202ece2SStefano Zampini } else { 833202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 843202ece2SStefano Zampini } 853202ece2SStefano Zampini PetscFunctionReturn(0); 863202ece2SStefano Zampini } 8734a97f8cSStefano Zampini 8834a97f8cSStefano Zampini #undef __FUNCT__ 891580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp" 9008122e43SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool compute_Stilda, PetscBool deluxe, PetscBool use_edges, PetscBool use_faces) 91b1b3d7a2SStefano Zampini { 92b1b3d7a2SStefano Zampini Mat A_II,A_IB,A_BI,A_BB; 93b1b3d7a2SStefano Zampini Mat AE_II,*AE_IE,*AE_EI,*AE_EE; 94883469d8SStefano Zampini Mat S_all,global_schur_subsets,work_mat; 9508122e43SStefano Zampini Mat S_Ej_tilda_all,S_Ej_inv_all; 965db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 975db18549SStefano Zampini IS is_I,*is_subset_B,temp_is; 98883469d8SStefano Zampini PetscInt *nnz,*all_local_idx_G,*all_local_idx_B,*all_local_idx_N; 995db18549SStefano Zampini PetscInt i,subset_size,max_subset_size; 100883469d8SStefano Zampini PetscInt extra,local_size,global_size; 10108122e43SStefano Zampini PetscBool is_symmetric,Stilda_computed; 10208122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 10308122e43SStefano Zampini PetscScalar *work; 104b1b3d7a2SStefano Zampini PetscErrorCode ierr; 105b1b3d7a2SStefano Zampini 106b1b3d7a2SStefano Zampini PetscFunctionBegin; 107b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 108883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 109b7eb3628SStefano Zampini if (compute_Stilda) { 1109bb4a8caSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Computation of Stildas requires MUMPS"); 111b7eb3628SStefano Zampini } 112b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&A_IB,&A_BI,&A_BB);CHKERRQ(ierr); 113b1b3d7a2SStefano Zampini ierr = PetscMalloc4(sub_schurs->n_subs,&is_subset_B, 114b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_IE, 115b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_EI, 116b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_EE);CHKERRQ(ierr); 117*a58a30b4SStefano Zampini } else { 118*a58a30b4SStefano Zampini is_subset_B = NULL; 119*a58a30b4SStefano Zampini AE_IE = NULL; 120*a58a30b4SStefano Zampini AE_EI = NULL; 121*a58a30b4SStefano Zampini AE_EE = NULL; 122b1b3d7a2SStefano Zampini } 123b1b3d7a2SStefano Zampini 124b1b3d7a2SStefano Zampini /* determine interior problems */ 125b96c3477SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 126b1b3d7a2SStefano Zampini if (nlayers >= 0 && xadj != NULL && adjncy != NULL) { /* Interior problems can be different from the original one */ 127b1b3d7a2SStefano Zampini PetscBT touched; 128b1b3d7a2SStefano Zampini const PetscInt* idx_B; 129b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 130b1b3d7a2SStefano Zampini 131b1b3d7a2SStefano Zampini /* get sizes */ 132b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 133b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 134b1b3d7a2SStefano Zampini 135b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 136b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 137b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 138b1b3d7a2SStefano Zampini 139b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 140b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 141b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 142b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 143b1b3d7a2SStefano Zampini } 144b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 145b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 146b1b3d7a2SStefano Zampini 147b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 148b1b3d7a2SStefano Zampini n_local_dofs = n_B; 149b1b3d7a2SStefano Zampini n_prev_added = n_B; 150b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 151b1b3d7a2SStefano Zampini PetscInt n_added; 152b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 153b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 154b1b3d7a2SStefano 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); 155b1b3d7a2SStefano Zampini } 156b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 157b1b3d7a2SStefano Zampini n_prev_added = n_added; 158b1b3d7a2SStefano Zampini n_local_dofs += n_added; 159b1b3d7a2SStefano Zampini if (!n_added) break; 160b1b3d7a2SStefano Zampini } 161b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 162b1b3d7a2SStefano Zampini 163883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 16468270318SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&sub_schurs->is_I_layer);CHKERRQ(ierr); 165b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 16668270318SStefano Zampini ierr = ISSort(sub_schurs->is_I_layer);CHKERRQ(ierr); 167883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 168883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 169b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 170b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 17168270318SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,sub_schurs->is_I_layer,&is_I);CHKERRQ(ierr); 172b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 173b1b3d7a2SStefano Zampini 174b1b3d7a2SStefano Zampini /* II block */ 175b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 176b1b3d7a2SStefano Zampini } 177b1b3d7a2SStefano Zampini } else { 178b1b3d7a2SStefano Zampini PetscInt n_I; 179b1b3d7a2SStefano Zampini 180b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 181b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 18268270318SStefano Zampini sub_schurs->is_I_layer = sub_schurs->is_I; 183b1b3d7a2SStefano Zampini 184b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 185883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 186b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 187b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 188b1b3d7a2SStefano Zampini 189b1b3d7a2SStefano Zampini /* II block is the same */ 190b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 191b1b3d7a2SStefano Zampini AE_II = A_II; 192b1b3d7a2SStefano Zampini } 193b1b3d7a2SStefano Zampini } 194b1b3d7a2SStefano Zampini 195883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 196883469d8SStefano Zampini max_subset_size = 0; 197883469d8SStefano Zampini local_size = 0; 198883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 199883469d8SStefano Zampini PetscInt j = sub_schurs->index_sequential[i]; 200b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[j],&subset_size);CHKERRQ(ierr); 201883469d8SStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 202883469d8SStefano Zampini local_size += subset_size; 203883469d8SStefano Zampini } 204883469d8SStefano Zampini 205883469d8SStefano Zampini /* Work arrays for local indices */ 206883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 207883469d8SStefano Zampini extra = 0; 208883469d8SStefano Zampini if (sub_schurs->use_mumps) { 209883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&extra);CHKERRQ(ierr); 210883469d8SStefano Zampini } 211883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 212883469d8SStefano Zampini if (extra) { 213883469d8SStefano Zampini const PetscInt *idxs; 214883469d8SStefano Zampini ierr = ISGetIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 215883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 216883469d8SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 217883469d8SStefano Zampini } 218883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 219883469d8SStefano Zampini 220883469d8SStefano Zampini /* Get local indices in local numbering */ 221883469d8SStefano Zampini local_size = 0; 222883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 223883469d8SStefano Zampini PetscInt j; 224883469d8SStefano Zampini const PetscInt *idxs; 225883469d8SStefano Zampini 226883469d8SStefano Zampini PetscInt local_problem_index = sub_schurs->index_sequential[i]; 227b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[local_problem_index],&subset_size);CHKERRQ(ierr); 228b96c3477SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 229883469d8SStefano Zampini /* subset indices in local numbering */ 230883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 231b96c3477SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 232883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 233883469d8SStefano Zampini local_size += subset_size; 234883469d8SStefano Zampini } 235883469d8SStefano Zampini 236883469d8SStefano Zampini S_all = NULL; 23708122e43SStefano Zampini ierr = MatIsSymmetric(sub_schurs->A,0.0,&is_symmetric);CHKERRQ(ierr); 238883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 239883469d8SStefano Zampini /* subsets in original and boundary numbering */ 240883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 241b96c3477SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B[i]);CHKERRQ(ierr); 242b1b3d7a2SStefano Zampini } 243b1b3d7a2SStefano Zampini 244b1b3d7a2SStefano Zampini /* EE block */ 245b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 246b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B[i],is_subset_B[i],MAT_INITIAL_MATRIX,&AE_EE[i]);CHKERRQ(ierr); 247b1b3d7a2SStefano Zampini } 248b1b3d7a2SStefano Zampini /* IE block */ 249b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 250b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B[i],MAT_INITIAL_MATRIX,&AE_IE[i]);CHKERRQ(ierr); 251b1b3d7a2SStefano Zampini } 252b1b3d7a2SStefano Zampini /* EI block */ 253b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 254b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B[i],is_I,MAT_INITIAL_MATRIX,&AE_EI[i]);CHKERRQ(ierr); 255b1b3d7a2SStefano Zampini } 256b1b3d7a2SStefano Zampini 257b1b3d7a2SStefano Zampini /* setup Schur complements on subset */ 258b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 259b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 260b1b3d7a2SStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE[i],AE_EI[i],AE_EE[i],&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 261*a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EE[i]);CHKERRQ(ierr); 262*a58a30b4SStefano Zampini ierr = MatDestroy(&AE_IE[i]);CHKERRQ(ierr); 263*a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EI[i]);CHKERRQ(ierr); 264b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 265b1b3d7a2SStefano Zampini KSP ksp; 266b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 267b1b3d7a2SStefano Zampini ierr = MatSchurComplementSetKSP(sub_schurs->S_Ej[i],ksp);CHKERRQ(ierr); 268b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 269b1b3d7a2SStefano Zampini KSP origksp,schurksp; 270b1b3d7a2SStefano Zampini PC origpc,schurpc; 271b1b3d7a2SStefano Zampini KSPType ksp_type; 272b1b3d7a2SStefano Zampini PCType pc_type; 273b1b3d7a2SStefano Zampini PetscInt n_internal; 274b1b3d7a2SStefano Zampini 275b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 276b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S_Ej[i],&schurksp);CHKERRQ(ierr); 277b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 278b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 279b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 280b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 281b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 282b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 283b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 284b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 285b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 286b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 287b1b3d7a2SStefano Zampini if (solver) { 288b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 289b1b3d7a2SStefano Zampini } 290b1b3d7a2SStefano Zampini } 291b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 292b1b3d7a2SStefano Zampini } 293b1b3d7a2SStefano Zampini } 294b1b3d7a2SStefano Zampini /* free */ 295b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 296b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 297b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 298b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_subset_B[i]);CHKERRQ(ierr); 299b1b3d7a2SStefano Zampini } 300b1b3d7a2SStefano Zampini ierr = PetscFree4(is_subset_B,AE_IE,AE_EI,AE_EE);CHKERRQ(ierr); 301883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 302883469d8SStefano Zampini } else { 303883469d8SStefano Zampini Mat A,F; 304883469d8SStefano Zampini IS is_A_all; 305883469d8SStefano Zampini PetscInt *idxs_schur,n_I; 306883469d8SStefano Zampini 307883469d8SStefano Zampini /* get working mat */ 308883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&n_I);CHKERRQ(ierr); 309883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 310b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 311883469d8SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 312883469d8SStefano Zampini 31308122e43SStefano Zampini if (n_I) { 314883469d8SStefano Zampini if (is_symmetric) { 315883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 316883469d8SStefano Zampini } else { 317883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 318883469d8SStefano Zampini } 319883469d8SStefano Zampini 320883469d8SStefano Zampini /* subsets ordered last */ 321883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&idxs_schur);CHKERRQ(ierr); 322883469d8SStefano Zampini for (i=0;i<local_size;i++) { 323883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 324883469d8SStefano Zampini } 325883469d8SStefano Zampini ierr = MatMumpsSetSchurIndices(F,local_size,idxs_schur);CHKERRQ(ierr); 326883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 327883469d8SStefano Zampini 328883469d8SStefano Zampini /* factorization step */ 329883469d8SStefano Zampini if (is_symmetric) { 330883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 331883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 332883469d8SStefano Zampini } else { 333883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 334883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 335883469d8SStefano Zampini } 336883469d8SStefano Zampini 337883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 338883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 339883469d8SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 34008122e43SStefano Zampini } else { 34108122e43SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 34208122e43SStefano Zampini } 34308122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 344883469d8SStefano Zampini #endif 345b1b3d7a2SStefano Zampini } 3465db18549SStefano Zampini 3475db18549SStefano Zampini if (!sub_schurs->n_subs_seq_g) { 3485db18549SStefano Zampini PetscFunctionReturn(0); 3495db18549SStefano Zampini } 3505db18549SStefano Zampini 3515db18549SStefano Zampini /* subset indices in local boundary numbering */ 352b96c3477SStefano Zampini if (!sub_schurs->is_Ej_all) { 353883469d8SStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 3545db18549SStefano Zampini if (subset_size != local_size) { 3555db18549SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 3565db18549SStefano Zampini } 3575db18549SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 358b96c3477SStefano Zampini } 3595db18549SStefano Zampini 3603202ece2SStefano Zampini /* Local matrix of all local Schur on subsets transposed */ 361b96c3477SStefano Zampini if (!sub_schurs->S_Ej_all) { 3625db18549SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3635db18549SStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 3645db18549SStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 3655db18549SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 366b96c3477SStefano Zampini } else { 367b96c3477SStefano Zampini ierr = MatZeroEntries(sub_schurs->S_Ej_all);CHKERRQ(ierr); 368b96c3477SStefano Zampini } 369a1337663SStefano Zampini 370a1337663SStefano Zampini S_Ej_tilda_all = 0; 37108122e43SStefano Zampini S_Ej_inv_all = 0; 372*a58a30b4SStefano Zampini work = NULL; 373*a58a30b4SStefano Zampini pivots = NULL; 374*a58a30b4SStefano Zampini Stilda_computed = PETSC_FALSE; 37508122e43SStefano Zampini if (sub_schurs->n_subs && deluxe) { /* workspace needed only for GETRI */ 37608122e43SStefano Zampini PetscScalar lwork; 37708122e43SStefano Zampini 37808122e43SStefano Zampini B_lwork = -1; 37908122e43SStefano Zampini ierr = PetscBLASIntCast(max_subset_size,&B_N);CHKERRQ(ierr); 38008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38108122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 38208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 38308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 38408122e43SStefano Zampini ierr = PetscBLASIntCast((PetscInt)lwork,&B_lwork);CHKERRQ(ierr); 38508122e43SStefano Zampini ierr = PetscMalloc2(B_lwork,&work,max_subset_size,&pivots);CHKERRQ(ierr); 38608122e43SStefano Zampini } 38708122e43SStefano Zampini 38808122e43SStefano Zampini ierr = PetscBTMemzero(sub_schurs->n_subs,sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 389883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 3905db18549SStefano Zampini PetscScalar *fill_vals; 3915db18549SStefano Zampini PetscInt *dummy_idx; 3925db18549SStefano Zampini 3935db18549SStefano Zampini /* Work arrays */ 3943202ece2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 3953202ece2SStefano Zampini /* Loop on local problems end compute Schur complements explicitly */ 3965db18549SStefano Zampini local_size = 0; 3975db18549SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 3983202ece2SStefano Zampini Mat S_Ej_expl; 3993202ece2SStefano Zampini PetscInt j,lpi = sub_schurs->index_sequential[i]; 4003202ece2SStefano Zampini PetscBool Sdense; 4015db18549SStefano Zampini 4023202ece2SStefano Zampini ierr = PCBDDCComputeExplicitSchur(sub_schurs->S_Ej[lpi],&S_Ej_expl);CHKERRQ(ierr); 403b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[lpi],&subset_size);CHKERRQ(ierr); 4043202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 4053202ece2SStefano Zampini if (Sdense) { 4065db18549SStefano Zampini for (j=0;j<subset_size;j++) { 4075db18549SStefano Zampini dummy_idx[j]=local_size+j; 4085db18549SStefano Zampini } 4093202ece2SStefano Zampini ierr = MatDenseGetArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4105db18549SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,fill_vals,INSERT_VALUES);CHKERRQ(ierr); 4113202ece2SStefano Zampini ierr = MatDenseRestoreArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4123202ece2SStefano Zampini } else { 4133202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 4145db18549SStefano Zampini } 4153202ece2SStefano Zampini local_size += subset_size; 41665d8bf0aSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[lpi]);CHKERRQ(ierr); 41765d8bf0aSStefano Zampini sub_schurs->S_Ej[lpi] = S_Ej_expl; 4183202ece2SStefano Zampini } 4192b973097SStefano Zampini /* Stildas are not computed without mumps */ 4202b973097SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 421b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 4222b973097SStefano Zampini } 4233202ece2SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 424883469d8SStefano Zampini } else { 425883469d8SStefano Zampini PetscInt *dummy_idx,n_all; 426883469d8SStefano Zampini 427a1337663SStefano Zampini if (compute_Stilda) { 428a1337663SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_tilda_all);CHKERRQ(ierr); 429a1337663SStefano Zampini ierr = MatSetSizes(S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 430a1337663SStefano Zampini ierr = MatSetType(S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 431a1337663SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 43208122e43SStefano Zampini if (deluxe) { 43308122e43SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_inv_all);CHKERRQ(ierr); 43408122e43SStefano Zampini ierr = MatSetSizes(S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 43508122e43SStefano Zampini ierr = MatSetType(S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 43608122e43SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 437a1337663SStefano Zampini } 43808122e43SStefano Zampini } 43908122e43SStefano Zampini 44008122e43SStefano Zampini ierr = MatGetSize(S_all,&n_all,NULL);CHKERRQ(ierr); 44108122e43SStefano Zampini local_size = 0; 442883469d8SStefano Zampini /* Work arrays */ 443883469d8SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 444883469d8SStefano Zampini 44508122e43SStefano Zampini Stilda_computed = PETSC_FALSE; 44665d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 44765d8bf0aSStefano Zampini IS is_E; 44865d8bf0aSStefano Zampini PetscScalar *vals; 44965d8bf0aSStefano Zampini PetscInt j; 45065d8bf0aSStefano Zampini 451b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 452a1337663SStefano Zampini for (j=0;j<subset_size;j++) { 453a1337663SStefano Zampini dummy_idx[j]=local_size+j; 454a1337663SStefano Zampini } 45565d8bf0aSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),subset_size,local_size,1,&is_E);CHKERRQ(ierr); 456b96c3477SStefano Zampini if (sub_schurs->S_Ej[i]) { 457b96c3477SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_REUSE_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 458b96c3477SStefano Zampini } else { 459a1337663SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_INITIAL_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 460b96c3477SStefano Zampini } 461a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 462a1337663SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 463a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 464a1337663SStefano Zampini 4656da91a30SStefano Zampini if (compute_Stilda && ((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || use_faces)) { 46608122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,i);CHKERRQ(ierr); 46708122e43SStefano Zampini Stilda_computed = PETSC_TRUE; 46808122e43SStefano Zampini } 46908122e43SStefano Zampini 47008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 47165d8bf0aSStefano Zampini Mat Stilda; 47265d8bf0aSStefano Zampini if (sub_schurs->n_subs == 1) { 47365d8bf0aSStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->S_Ej[i]);CHKERRQ(ierr); 47465d8bf0aSStefano Zampini Stilda = sub_schurs->S_Ej[i]; 47565d8bf0aSStefano Zampini } else { 47665d8bf0aSStefano Zampini KSP ksp; 47765d8bf0aSStefano Zampini PC pc; 47865d8bf0aSStefano Zampini Mat S_EF,S_FE,S_FF,Stilda_impl; 47965d8bf0aSStefano Zampini IS is_F; 4802b973097SStefano Zampini PetscScalar eps=1.e-8; 4812b973097SStefano Zampini PetscBool chop=PETSC_FALSE; 48265d8bf0aSStefano Zampini 48365d8bf0aSStefano Zampini ierr = ISComplement(is_E,0,n_all,&is_F);CHKERRQ(ierr); 48465d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_F,MAT_INITIAL_MATRIX,&S_EF);CHKERRQ(ierr); 48565d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_F,MAT_INITIAL_MATRIX,&S_FF);CHKERRQ(ierr); 48665d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_E,MAT_INITIAL_MATRIX,&S_FE);CHKERRQ(ierr); 487a1337663SStefano Zampini ierr = ISDestroy(&is_F);CHKERRQ(ierr); 4882b973097SStefano Zampini if (chop) { 4892b973097SStefano Zampini ierr = MatChop(S_FF,eps);CHKERRQ(ierr); 4902b973097SStefano Zampini ierr = MatConvert(S_FF,MATAIJ,MAT_REUSE_MATRIX,&S_FF);CHKERRQ(ierr); 4912b973097SStefano Zampini } 492a1337663SStefano Zampini ierr = MatCreateSchurComplement(S_FF,S_FF,S_FE,S_EF,sub_schurs->S_Ej[i],&Stilda_impl);CHKERRQ(ierr); 493a1337663SStefano Zampini ierr = MatDestroy(&S_FF);CHKERRQ(ierr); 494a1337663SStefano Zampini ierr = MatDestroy(&S_FE);CHKERRQ(ierr); 495a1337663SStefano Zampini ierr = MatDestroy(&S_EF);CHKERRQ(ierr); 49665d8bf0aSStefano Zampini ierr = MatSchurComplementGetKSP(Stilda_impl,&ksp);CHKERRQ(ierr); 49765d8bf0aSStefano Zampini ierr = KSPSetType(ksp,KSPPREONLY);CHKERRQ(ierr); 49865d8bf0aSStefano Zampini ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr); 49965d8bf0aSStefano Zampini if (is_symmetric) { 50065d8bf0aSStefano Zampini ierr = PCSetType(pc,PCCHOLESKY);CHKERRQ(ierr); 50165d8bf0aSStefano Zampini } else { 50265d8bf0aSStefano Zampini ierr = PCSetType(pc,PCLU);CHKERRQ(ierr); 50365d8bf0aSStefano Zampini } 5042b973097SStefano Zampini if (!chop) { 50565d8bf0aSStefano Zampini ierr = PCFactorSetUseInPlace(pc,PETSC_TRUE);CHKERRQ(ierr); 5062b973097SStefano Zampini } else { 5072b973097SStefano Zampini ierr = PCFactorSetMatSolverPackage(pc,MATSOLVERMUMPS);CHKERRQ(ierr); 5082b973097SStefano Zampini } 50965d8bf0aSStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 51065d8bf0aSStefano Zampini ierr = PCBDDCComputeExplicitSchur(Stilda_impl,&Stilda);CHKERRQ(ierr); 51165d8bf0aSStefano Zampini ierr = MatDestroy(&Stilda_impl);CHKERRQ(ierr); 51265d8bf0aSStefano Zampini } 5132b973097SStefano Zampini /* 5142b973097SStefano Zampini PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB); 5152b973097SStefano Zampini ierr = MatView(Stilda,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5162b973097SStefano Zampini PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF); 5172b973097SStefano Zampini */ 518b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 5192b973097SStefano Zampini sub_schurs->S_Ej_tilda[i] = Stilda; 52008122e43SStefano Zampini 52108122e43SStefano Zampini 522a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 52308122e43SStefano Zampini if (deluxe) { /* when using deluxe scaling, we need (S_1^-1+S_2^-1)^-1 */ 52408122e43SStefano Zampini PetscScalar *vals2; 52508122e43SStefano Zampini 52608122e43SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 52708122e43SStefano Zampini /* need to be optimized (cholesky) */ 52808122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 52908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 53008122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals,&B_N,pivots,&B_ierr)); 53108122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 53208122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals,&B_N,pivots,work,&B_lwork,&B_ierr)); 53308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 53408122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals2,&B_N,pivots,&B_ierr)); 53508122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 53608122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals2,&B_N,pivots,work,&B_lwork,&B_ierr)); 53708122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 53808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 53908122e43SStefano Zampini ierr = MatSetValues(S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,vals2,INSERT_VALUES);CHKERRQ(ierr); 54008122e43SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 54108122e43SStefano Zampini } 542a1337663SStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 543a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 5442b973097SStefano Zampini } else { 545b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 54665d8bf0aSStefano Zampini } 54765d8bf0aSStefano Zampini ierr = ISDestroy(&is_E);CHKERRQ(ierr); 548883469d8SStefano Zampini local_size += subset_size; 549883469d8SStefano Zampini } 550883469d8SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 5515db18549SStefano Zampini } 552a1337663SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 553a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 5545db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5555db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 556a1337663SStefano Zampini if (compute_Stilda) { 557a1337663SStefano Zampini ierr = MatAssemblyBegin(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 558a1337663SStefano Zampini ierr = MatAssemblyEnd(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55908122e43SStefano Zampini if (deluxe) { 56008122e43SStefano Zampini ierr = MatAssemblyBegin(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 56108122e43SStefano Zampini ierr = MatAssemblyEnd(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 56208122e43SStefano Zampini } 563a1337663SStefano Zampini } 564a1337663SStefano Zampini 5655db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 566883469d8SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)sub_schurs->l2gmap),sub_schurs->l2gmap,local_size,all_local_idx_N+extra,PETSC_NULL,&global_size,&all_local_idx_G);CHKERRQ(ierr); 5675db18549SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,local_size,all_local_idx_G,PETSC_COPY_VALUES,&l2gmap_subsets);CHKERRQ(ierr); 5685db18549SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 5695db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 5705db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 5713927de2eSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 5723927de2eSStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 5733927de2eSStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 5743927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 5753927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 5765db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 577883469d8SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 578883469d8SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->l2gmap),local_size,all_local_idx_G,PETSC_OWN_POINTER,&temp_is);CHKERRQ(ierr); 579b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 580b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 581b96c3477SStefano Zampini } else { 582b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 583b96c3477SStefano Zampini } 58408122e43SStefano Zampini 58508122e43SStefano Zampini if (Stilda_computed) { 586a1337663SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_tilda_all);CHKERRQ(ierr); 587a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 588b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_tilda_all) { 589b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 590b96c3477SStefano Zampini } else { 591b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 592b96c3477SStefano Zampini } 59308122e43SStefano Zampini if (deluxe) { 59408122e43SStefano Zampini PetscInt cum; 59508122e43SStefano Zampini PetscScalar *array,*array2; 59608122e43SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_inv_all);CHKERRQ(ierr); 59708122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 59808122e43SStefano Zampini if (sub_schurs->sum_S_Ej_inv_all) { 59908122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 60008122e43SStefano Zampini } else { 60108122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 60208122e43SStefano Zampini } 60308122e43SStefano Zampini /* invert blocks of sum_S_Ej_inv_all */ 60408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 60508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 60608122e43SStefano Zampini cum = 0; 60708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 60808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 60908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 61008122e43SStefano Zampini /* need to be optimized (cholesky) */ 61108122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 61208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 61308122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 61408122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 61508122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 61608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 61708122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array2+cum,&B_N,pivots,&B_ierr)); 61808122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 61908122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array2+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 62008122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 62108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 62208122e43SStefano Zampini } 62308122e43SStefano Zampini cum += subset_size*subset_size; 62408122e43SStefano Zampini } 62508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 62608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 62708122e43SStefano Zampini } 6283202ece2SStefano Zampini } 6293202ece2SStefano Zampini 63008122e43SStefano Zampini ierr = PetscFree2(work,pivots);CHKERRQ(ierr); 631a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 632a1337663SStefano Zampini ierr = MatDestroy(&S_Ej_tilda_all);CHKERRQ(ierr); 63308122e43SStefano Zampini ierr = MatDestroy(&S_Ej_inv_all);CHKERRQ(ierr); 6343202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 6355db18549SStefano Zampini ierr = ISDestroy(&temp_is);CHKERRQ(ierr); 636b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 637b1b3d7a2SStefano Zampini } 638b1b3d7a2SStefano Zampini 639b1b3d7a2SStefano Zampini #undef __FUNCT__ 640b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 6415db18549SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, Mat A, Mat S, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap, PetscInt seqthreshold) 642b1b3d7a2SStefano Zampini { 6439bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 644b1b3d7a2SStefano Zampini PetscInt *index_sequential,*index_parallel; 645b1b3d7a2SStefano Zampini PetscInt *auxlocal_sequential,*auxlocal_parallel; 646b1b3d7a2SStefano Zampini PetscInt *auxglobal_sequential,*auxglobal_parallel; 647b96c3477SStefano Zampini PetscInt *auxmapping; 648b1b3d7a2SStefano Zampini PetscInt i,max_subset_size; 649b1b3d7a2SStefano Zampini PetscInt n_sequential_problems,n_local_sequential_problems,n_parallel_problems,n_local_parallel_problems; 650b1b3d7a2SStefano Zampini PetscInt n_faces,n_edges,n_all_cc; 651b1b3d7a2SStefano Zampini PetscBool is_sorted; 652b1b3d7a2SStefano Zampini PetscErrorCode ierr; 653b1b3d7a2SStefano Zampini 654b1b3d7a2SStefano Zampini PetscFunctionBegin; 655b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 656b1b3d7a2SStefano Zampini if (!is_sorted) { 657b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 658b1b3d7a2SStefano Zampini } 659b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 660b1b3d7a2SStefano Zampini if (!is_sorted) { 661b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 662b1b3d7a2SStefano Zampini } 663b1b3d7a2SStefano Zampini 664b1b3d7a2SStefano Zampini /* reset any previous data */ 665b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 666b1b3d7a2SStefano Zampini 667b1b3d7a2SStefano Zampini /* get index sets for faces and edges */ 6689bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 669b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 67008122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 67108122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 672b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 673b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 674b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 675b1b3d7a2SStefano Zampini } 676b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 677b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 67808122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 679b1b3d7a2SStefano Zampini } 680b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 681b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 682b1b3d7a2SStefano Zampini 683b1b3d7a2SStefano Zampini /* map interface's subsets */ 684b1b3d7a2SStefano Zampini max_subset_size = 0; 685b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 686b1b3d7a2SStefano Zampini PetscInt subset_size; 687b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 688b1b3d7a2SStefano Zampini max_subset_size = PetscMax(max_subset_size,subset_size); 689b1b3d7a2SStefano Zampini } 690b1b3d7a2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&auxmapping);CHKERRQ(ierr); 691b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&auxlocal_sequential, 692b1b3d7a2SStefano Zampini graph->ncc,&auxlocal_parallel);CHKERRQ(ierr); 693b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&index_sequential, 694b1b3d7a2SStefano Zampini graph->ncc,&index_parallel);CHKERRQ(ierr); 695b1b3d7a2SStefano Zampini 696883469d8SStefano Zampini /* if threshold is negative uses all sequential problems (possibly using MUMPS) */ 697883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 698883469d8SStefano Zampini if (seqthreshold < 0) { 699883469d8SStefano Zampini seqthreshold = max_subset_size; 700883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 701883469d8SStefano Zampini sub_schurs->use_mumps = !!A; 702883469d8SStefano Zampini #endif 703883469d8SStefano Zampini } 704883469d8SStefano Zampini 705b1b3d7a2SStefano Zampini 706b1b3d7a2SStefano Zampini /* determine which problem has to be solved in parallel or sequentially */ 707b1b3d7a2SStefano Zampini n_local_sequential_problems = 0; 708b1b3d7a2SStefano Zampini n_local_parallel_problems = 0; 709b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 710b1b3d7a2SStefano Zampini PetscInt subset_size,j,min_loc = 0; 711b1b3d7a2SStefano Zampini const PetscInt *idxs; 712b1b3d7a2SStefano Zampini 713b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 714b1b3d7a2SStefano Zampini ierr = ISGetIndices(all_cc[i],&idxs);CHKERRQ(ierr); 715b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingApply(graph->l2gmap,subset_size,idxs,auxmapping);CHKERRQ(ierr); 716b1b3d7a2SStefano Zampini for (j=1;j<subset_size;j++) { 717b1b3d7a2SStefano Zampini if (auxmapping[j]<auxmapping[min_loc]) { 718b1b3d7a2SStefano Zampini min_loc = j; 719b1b3d7a2SStefano Zampini } 720b1b3d7a2SStefano Zampini } 721b1b3d7a2SStefano Zampini if (subset_size > seqthreshold) { 722b1b3d7a2SStefano Zampini index_parallel[n_local_parallel_problems] = i; 723b1b3d7a2SStefano Zampini auxlocal_parallel[n_local_parallel_problems] = idxs[min_loc]; 724b1b3d7a2SStefano Zampini n_local_parallel_problems++; 725b1b3d7a2SStefano Zampini } else { 726b1b3d7a2SStefano Zampini index_sequential[n_local_sequential_problems] = i; 727b1b3d7a2SStefano Zampini auxlocal_sequential[n_local_sequential_problems] = idxs[min_loc]; 728b1b3d7a2SStefano Zampini n_local_sequential_problems++; 729b1b3d7a2SStefano Zampini } 730b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(all_cc[i],&idxs);CHKERRQ(ierr); 731b1b3d7a2SStefano Zampini } 732b1b3d7a2SStefano Zampini 733b1b3d7a2SStefano Zampini /* Number parallel problems */ 734b1b3d7a2SStefano Zampini auxglobal_parallel = 0; 735b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_parallel_problems,auxlocal_parallel,PETSC_NULL,&n_parallel_problems,&auxglobal_parallel);CHKERRQ(ierr); 736b1b3d7a2SStefano Zampini 737b1b3d7a2SStefano Zampini /* Number sequential problems */ 738b1b3d7a2SStefano Zampini auxglobal_sequential = 0; 739b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_sequential_problems,auxlocal_sequential,PETSC_NULL,&n_sequential_problems,&auxglobal_sequential);CHKERRQ(ierr); 740b1b3d7a2SStefano Zampini 741b1b3d7a2SStefano Zampini /* update info in sub_schurs */ 742883469d8SStefano Zampini if (sub_schurs->use_mumps && A) { 7431e9c79c2SStefano Zampini ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 7441e9c79c2SStefano Zampini sub_schurs->A = A; 7451e9c79c2SStefano Zampini } 746b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)S);CHKERRQ(ierr); 747b1b3d7a2SStefano Zampini sub_schurs->S = S; 748b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 749b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 750b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 751b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 7525db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 7535db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 7545db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 7555db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 756b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq = n_local_sequential_problems; 757b1b3d7a2SStefano Zampini sub_schurs->n_subs_par = n_local_parallel_problems; 758b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq_g = n_sequential_problems; 759b1b3d7a2SStefano Zampini sub_schurs->n_subs_par_g = n_parallel_problems; 760b1b3d7a2SStefano Zampini sub_schurs->n_subs = sub_schurs->n_subs_seq + sub_schurs->n_subs_par; 761b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 7629bb4a8caSStefano Zampini if (!sub_schurs->use_mumps) { /* for adaptive selection */ 763b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 764b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 765b96c3477SStefano Zampini } 7669bb4a8caSStefano Zampini } 7679bb4a8caSStefano Zampini sub_schurs->is_Ej_com = vertices; 768b1b3d7a2SStefano Zampini sub_schurs->index_sequential = index_sequential; 769b1b3d7a2SStefano Zampini sub_schurs->index_parallel = index_parallel; 770b1b3d7a2SStefano Zampini sub_schurs->auxglobal_sequential = auxglobal_sequential; 771b1b3d7a2SStefano Zampini sub_schurs->auxglobal_parallel = auxglobal_parallel; 772b1b3d7a2SStefano Zampini 773b96c3477SStefano Zampini 774b96c3477SStefano Zampini /* allocate space for schur complements */ 775b96c3477SStefano Zampini ierr = PetscCalloc2(sub_schurs->n_subs,&sub_schurs->S_Ej, 776b96c3477SStefano Zampini sub_schurs->n_subs,&sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 777b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 778b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 77908122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 780b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 781b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 782b96c3477SStefano Zampini 783b1b3d7a2SStefano Zampini /* free workspace */ 784b1b3d7a2SStefano Zampini ierr = PetscFree(auxmapping);CHKERRQ(ierr); 785b1b3d7a2SStefano Zampini ierr = PetscFree2(auxlocal_sequential,auxlocal_parallel);CHKERRQ(ierr); 786b1b3d7a2SStefano Zampini 787b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 788b1b3d7a2SStefano Zampini } 789b1b3d7a2SStefano Zampini 790b1b3d7a2SStefano Zampini #undef __FUNCT__ 79134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 79234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 79334a97f8cSStefano Zampini { 79434a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 79534a97f8cSStefano Zampini PetscErrorCode ierr; 79634a97f8cSStefano Zampini 79734a97f8cSStefano Zampini PetscFunctionBegin; 79834a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 7995ff63025SStefano Zampini schurs_ctx->n_subs = 0; 80034a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 80134a97f8cSStefano Zampini PetscFunctionReturn(0); 80234a97f8cSStefano Zampini } 80334a97f8cSStefano Zampini 80434a97f8cSStefano Zampini #undef __FUNCT__ 80534a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 80634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 80734a97f8cSStefano Zampini { 80834a97f8cSStefano Zampini PetscErrorCode ierr; 80934a97f8cSStefano Zampini 81034a97f8cSStefano Zampini PetscFunctionBegin; 81134a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 81234a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 81334a97f8cSStefano Zampini PetscFunctionReturn(0); 81434a97f8cSStefano Zampini } 81534a97f8cSStefano Zampini 81634a97f8cSStefano Zampini #undef __FUNCT__ 81734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 81834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 81934a97f8cSStefano Zampini { 82034a97f8cSStefano Zampini PetscInt i; 82134a97f8cSStefano Zampini PetscErrorCode ierr; 82234a97f8cSStefano Zampini 82334a97f8cSStefano Zampini PetscFunctionBegin; 8241e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 825b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 826b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 827b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 8285db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 8295db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 83041c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 83141c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 83208122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 833a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 8345db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 8359bb4a8caSStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_com);CHKERRQ(ierr); 83608122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 83708122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 83834a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 839b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 84034a97f8cSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 8412b973097SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 84234a97f8cSStefano Zampini } 84368270318SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 8445ff63025SStefano Zampini if (sub_schurs->n_subs) { 845b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 846b96c3477SStefano Zampini ierr = PetscFree2(sub_schurs->S_Ej,sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 847b1b3d7a2SStefano Zampini ierr = PetscFree2(sub_schurs->index_sequential,sub_schurs->index_parallel);CHKERRQ(ierr); 848b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_sequential);CHKERRQ(ierr); 849b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_parallel);CHKERRQ(ierr); 8505ff63025SStefano Zampini } 85134a97f8cSStefano Zampini sub_schurs->n_subs = 0; 85234a97f8cSStefano Zampini PetscFunctionReturn(0); 85334a97f8cSStefano Zampini } 85434a97f8cSStefano Zampini 85534a97f8cSStefano Zampini #undef __FUNCT__ 85634a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 8572a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 85834a97f8cSStefano Zampini { 85934a97f8cSStefano Zampini PetscInt i,j,n; 86034a97f8cSStefano Zampini PetscErrorCode ierr; 86134a97f8cSStefano Zampini 86234a97f8cSStefano Zampini PetscFunctionBegin; 86334a97f8cSStefano Zampini n = 0; 86434a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 86534a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 86634a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 86734a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 86834a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 86934a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 87034a97f8cSStefano Zampini queue_tip[n] = dof; 87134a97f8cSStefano Zampini n++; 87234a97f8cSStefano Zampini } 87334a97f8cSStefano Zampini } 87434a97f8cSStefano Zampini } 87534a97f8cSStefano Zampini *n_added = n; 87634a97f8cSStefano Zampini PetscFunctionReturn(0); 87734a97f8cSStefano Zampini } 878