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" 909552c7c7SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool compute_Stilda, PetscBool deluxe, PetscBool invert_Stildas, 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 PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 10208122e43SStefano Zampini PetscScalar *work; 103b1b3d7a2SStefano Zampini PetscErrorCode ierr; 104b1b3d7a2SStefano Zampini 105b1b3d7a2SStefano Zampini PetscFunctionBegin; 106b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 107883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 108b7eb3628SStefano Zampini if (compute_Stilda) { 1093dc780c3SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 110b7eb3628SStefano Zampini } 111b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&A_IB,&A_BI,&A_BB);CHKERRQ(ierr); 112b1b3d7a2SStefano Zampini ierr = PetscMalloc4(sub_schurs->n_subs,&is_subset_B, 113b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_IE, 114b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_EI, 115b1b3d7a2SStefano Zampini sub_schurs->n_subs,&AE_EE);CHKERRQ(ierr); 116a58a30b4SStefano Zampini } else { 117a58a30b4SStefano Zampini is_subset_B = NULL; 118a58a30b4SStefano Zampini AE_IE = NULL; 119a58a30b4SStefano Zampini AE_EI = NULL; 120a58a30b4SStefano Zampini AE_EE = NULL; 121b1b3d7a2SStefano Zampini } 122b1b3d7a2SStefano Zampini 123b1b3d7a2SStefano Zampini /* determine interior problems */ 124b96c3477SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 1253dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 1263dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* 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 1313dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 1323dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 1333dc780c3SStefano Zampini } 134b1b3d7a2SStefano Zampini /* get sizes */ 135b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 136b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 137b1b3d7a2SStefano Zampini 138b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 139b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 140b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 141b1b3d7a2SStefano Zampini 142b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 143b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 144b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 145b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 146b1b3d7a2SStefano Zampini } 147b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 148b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 149b1b3d7a2SStefano Zampini 150b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 151b1b3d7a2SStefano Zampini n_local_dofs = n_B; 152b1b3d7a2SStefano Zampini n_prev_added = n_B; 153b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 154b1b3d7a2SStefano Zampini PetscInt n_added; 155b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 156b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 157b1b3d7a2SStefano 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); 158b1b3d7a2SStefano Zampini } 159b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 160b1b3d7a2SStefano Zampini n_prev_added = n_added; 161b1b3d7a2SStefano Zampini n_local_dofs += n_added; 162b1b3d7a2SStefano Zampini if (!n_added) break; 163b1b3d7a2SStefano Zampini } 164b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 165b1b3d7a2SStefano Zampini 166883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 16768270318SStefano 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); 168b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 16968270318SStefano Zampini ierr = ISSort(sub_schurs->is_I_layer);CHKERRQ(ierr); 170883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 171883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 172b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 173b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 17468270318SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,sub_schurs->is_I_layer,&is_I);CHKERRQ(ierr); 175b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 176b1b3d7a2SStefano Zampini 177b1b3d7a2SStefano Zampini /* II block */ 178b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 179b1b3d7a2SStefano Zampini } 180b1b3d7a2SStefano Zampini } else { 181b1b3d7a2SStefano Zampini PetscInt n_I; 182b1b3d7a2SStefano Zampini 183b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 184b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 18568270318SStefano Zampini sub_schurs->is_I_layer = sub_schurs->is_I; 186b1b3d7a2SStefano Zampini 187b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 188883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 189b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 190b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 191b1b3d7a2SStefano Zampini 192b1b3d7a2SStefano Zampini /* II block is the same */ 193b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 194b1b3d7a2SStefano Zampini AE_II = A_II; 195b1b3d7a2SStefano Zampini } 196b1b3d7a2SStefano Zampini } 197883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 198883469d8SStefano Zampini max_subset_size = 0; 199883469d8SStefano Zampini local_size = 0; 200883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 201883469d8SStefano Zampini PetscInt j = sub_schurs->index_sequential[i]; 202b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[j],&subset_size);CHKERRQ(ierr); 203883469d8SStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 204883469d8SStefano Zampini local_size += subset_size; 205883469d8SStefano Zampini } 206883469d8SStefano Zampini 207883469d8SStefano Zampini /* Work arrays for local indices */ 208883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 209883469d8SStefano Zampini extra = 0; 210883469d8SStefano Zampini if (sub_schurs->use_mumps) { 211883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&extra);CHKERRQ(ierr); 212883469d8SStefano Zampini } 213883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 214883469d8SStefano Zampini if (extra) { 215883469d8SStefano Zampini const PetscInt *idxs; 216883469d8SStefano Zampini ierr = ISGetIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 217883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 218883469d8SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 219883469d8SStefano Zampini } 220883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 221883469d8SStefano Zampini 222883469d8SStefano Zampini /* Get local indices in local numbering */ 223883469d8SStefano Zampini local_size = 0; 224883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 225883469d8SStefano Zampini PetscInt j; 226883469d8SStefano Zampini const PetscInt *idxs; 227883469d8SStefano Zampini 228883469d8SStefano Zampini PetscInt local_problem_index = sub_schurs->index_sequential[i]; 229b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[local_problem_index],&subset_size);CHKERRQ(ierr); 230b96c3477SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 231883469d8SStefano Zampini /* subset indices in local numbering */ 232883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 233b96c3477SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 234883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 235883469d8SStefano Zampini local_size += subset_size; 236883469d8SStefano Zampini } 237883469d8SStefano Zampini 238883469d8SStefano Zampini S_all = NULL; 239*9ab7bb16SStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 240*9ab7bb16SStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 2419552c7c7SStefano Zampini if (sub_schurs->A) { 242*9ab7bb16SStefano Zampini ierr = MatIsHermitian(sub_schurs->A,0.0,&sub_schurs->is_hermitian);CHKERRQ(ierr); 243*9ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 244*9ab7bb16SStefano Zampini PetscScalar val; 245*9ab7bb16SStefano Zampini Vec vec1,vec2; 246*9ab7bb16SStefano Zampini 247*9ab7bb16SStefano Zampini ierr = MatCreateVecs(sub_schurs->A,&vec1,&vec2);CHKERRQ(ierr); 248*9ab7bb16SStefano Zampini ierr = VecSetRandom(vec1,NULL); 249*9ab7bb16SStefano Zampini ierr = VecCopy(vec1,vec2);CHKERRQ(ierr); 250*9ab7bb16SStefano Zampini ierr = MatMult(sub_schurs->A,vec2,vec1);CHKERRQ(ierr); 251*9ab7bb16SStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 252*9ab7bb16SStefano Zampini if (PetscRealPart(val) > 0. && PetscImaginaryPart(val) == 0.) sub_schurs->is_posdef = PETSC_TRUE; 253*9ab7bb16SStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 254*9ab7bb16SStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 255*9ab7bb16SStefano Zampini } 2569552c7c7SStefano Zampini } 2573dc780c3SStefano Zampini if (sub_schurs->n_subs) { 258883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 259883469d8SStefano Zampini /* subsets in original and boundary numbering */ 260883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 261b96c3477SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B[i]);CHKERRQ(ierr); 262b1b3d7a2SStefano Zampini } 263b1b3d7a2SStefano Zampini 264b1b3d7a2SStefano Zampini /* EE block */ 265b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 266b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B[i],is_subset_B[i],MAT_INITIAL_MATRIX,&AE_EE[i]);CHKERRQ(ierr); 267b1b3d7a2SStefano Zampini } 268b1b3d7a2SStefano Zampini /* IE block */ 269b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 270b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B[i],MAT_INITIAL_MATRIX,&AE_IE[i]);CHKERRQ(ierr); 271b1b3d7a2SStefano Zampini } 272b1b3d7a2SStefano Zampini /* EI block */ 273b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 274b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B[i],is_I,MAT_INITIAL_MATRIX,&AE_EI[i]);CHKERRQ(ierr); 275b1b3d7a2SStefano Zampini } 276b1b3d7a2SStefano Zampini 277b1b3d7a2SStefano Zampini /* setup Schur complements on subset */ 278b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 279b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 280b1b3d7a2SStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE[i],AE_EI[i],AE_EE[i],&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 281a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EE[i]);CHKERRQ(ierr); 282a58a30b4SStefano Zampini ierr = MatDestroy(&AE_IE[i]);CHKERRQ(ierr); 283a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EI[i]);CHKERRQ(ierr); 284b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 285b1b3d7a2SStefano Zampini KSP ksp; 286b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 287b1b3d7a2SStefano Zampini ierr = MatSchurComplementSetKSP(sub_schurs->S_Ej[i],ksp);CHKERRQ(ierr); 288b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 289b1b3d7a2SStefano Zampini KSP origksp,schurksp; 290b1b3d7a2SStefano Zampini PC origpc,schurpc; 291b1b3d7a2SStefano Zampini KSPType ksp_type; 292b1b3d7a2SStefano Zampini PCType pc_type; 293b1b3d7a2SStefano Zampini PetscInt n_internal; 294b1b3d7a2SStefano Zampini 295b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 296b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S_Ej[i],&schurksp);CHKERRQ(ierr); 297b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 298b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 299b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 300b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 301b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 302b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 303b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 304b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 305b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 306b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 307b1b3d7a2SStefano Zampini if (solver) { 308b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 309b1b3d7a2SStefano Zampini } 310b1b3d7a2SStefano Zampini } 311b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 312b1b3d7a2SStefano Zampini } 313b1b3d7a2SStefano Zampini } 314b1b3d7a2SStefano Zampini /* free */ 315b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 316b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 317b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 318b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_subset_B[i]);CHKERRQ(ierr); 319b1b3d7a2SStefano Zampini } 320b1b3d7a2SStefano Zampini ierr = PetscFree4(is_subset_B,AE_IE,AE_EI,AE_EE);CHKERRQ(ierr); 321883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 322883469d8SStefano Zampini } else { 323883469d8SStefano Zampini Mat A,F; 324883469d8SStefano Zampini IS is_A_all; 325883469d8SStefano Zampini PetscInt *idxs_schur,n_I; 326883469d8SStefano Zampini 327883469d8SStefano Zampini /* get working mat */ 328883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&n_I);CHKERRQ(ierr); 329883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 330b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 331883469d8SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 332883469d8SStefano Zampini 33308122e43SStefano Zampini if (n_I) { 334*9ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 335883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 336883469d8SStefano Zampini } else { 337883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 338883469d8SStefano Zampini } 339883469d8SStefano Zampini 340883469d8SStefano Zampini /* subsets ordered last */ 341883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&idxs_schur);CHKERRQ(ierr); 342883469d8SStefano Zampini for (i=0;i<local_size;i++) { 343883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 344883469d8SStefano Zampini } 345883469d8SStefano Zampini ierr = MatMumpsSetSchurIndices(F,local_size,idxs_schur);CHKERRQ(ierr); 346883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 347883469d8SStefano Zampini 348883469d8SStefano Zampini /* factorization step */ 349*9ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 350883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 351883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 352883469d8SStefano Zampini } else { 353883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 354883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 355883469d8SStefano Zampini } 356883469d8SStefano Zampini 357883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 358883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 359883469d8SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 36008122e43SStefano Zampini } else { 36108122e43SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 36208122e43SStefano Zampini } 36308122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 364883469d8SStefano Zampini #endif 365b1b3d7a2SStefano Zampini } 3663dc780c3SStefano Zampini } else { 3673dc780c3SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3683dc780c3SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 3693dc780c3SStefano Zampini } 3705db18549SStefano Zampini 3715db18549SStefano Zampini if (!sub_schurs->n_subs_seq_g) { 3725db18549SStefano Zampini PetscFunctionReturn(0); 3735db18549SStefano Zampini } 3745db18549SStefano Zampini 3755db18549SStefano Zampini /* subset indices in local boundary numbering */ 376b96c3477SStefano Zampini if (!sub_schurs->is_Ej_all) { 377883469d8SStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 3785db18549SStefano Zampini if (subset_size != local_size) { 3795db18549SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 3805db18549SStefano Zampini } 3815db18549SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 382b96c3477SStefano Zampini } 3835db18549SStefano Zampini 3843202ece2SStefano Zampini /* Local matrix of all local Schur on subsets transposed */ 385b96c3477SStefano Zampini if (!sub_schurs->S_Ej_all) { 3865db18549SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3875db18549SStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 3885db18549SStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 3895db18549SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 390b96c3477SStefano Zampini } else { 391b96c3477SStefano Zampini ierr = MatZeroEntries(sub_schurs->S_Ej_all);CHKERRQ(ierr); 392b96c3477SStefano Zampini } 393a1337663SStefano Zampini 394a1337663SStefano Zampini S_Ej_tilda_all = 0; 39508122e43SStefano Zampini S_Ej_inv_all = 0; 396a58a30b4SStefano Zampini work = NULL; 397a58a30b4SStefano Zampini pivots = NULL; 39808122e43SStefano Zampini if (sub_schurs->n_subs && deluxe) { /* workspace needed only for GETRI */ 39908122e43SStefano Zampini PetscScalar lwork; 40008122e43SStefano Zampini 40108122e43SStefano Zampini B_lwork = -1; 40208122e43SStefano Zampini ierr = PetscBLASIntCast(max_subset_size,&B_N);CHKERRQ(ierr); 40308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40408122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 40508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 40708122e43SStefano Zampini ierr = PetscBLASIntCast((PetscInt)lwork,&B_lwork);CHKERRQ(ierr); 40808122e43SStefano Zampini ierr = PetscMalloc2(B_lwork,&work,max_subset_size,&pivots);CHKERRQ(ierr); 40908122e43SStefano Zampini } 41008122e43SStefano Zampini 41108122e43SStefano Zampini ierr = PetscBTMemzero(sub_schurs->n_subs,sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 412883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 4135db18549SStefano Zampini PetscScalar *fill_vals; 4145db18549SStefano Zampini PetscInt *dummy_idx; 4155db18549SStefano Zampini 4165db18549SStefano Zampini /* Work arrays */ 4173202ece2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 4183202ece2SStefano Zampini /* Loop on local problems end compute Schur complements explicitly */ 4195db18549SStefano Zampini local_size = 0; 4205db18549SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 4213202ece2SStefano Zampini Mat S_Ej_expl; 4223202ece2SStefano Zampini PetscInt j,lpi = sub_schurs->index_sequential[i]; 4233202ece2SStefano Zampini PetscBool Sdense; 4245db18549SStefano Zampini 4253202ece2SStefano Zampini ierr = PCBDDCComputeExplicitSchur(sub_schurs->S_Ej[lpi],&S_Ej_expl);CHKERRQ(ierr); 426b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[lpi],&subset_size);CHKERRQ(ierr); 4273202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 4283202ece2SStefano Zampini if (Sdense) { 4295db18549SStefano Zampini for (j=0;j<subset_size;j++) { 4305db18549SStefano Zampini dummy_idx[j]=local_size+j; 4315db18549SStefano Zampini } 4323202ece2SStefano Zampini ierr = MatDenseGetArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4335db18549SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,fill_vals,INSERT_VALUES);CHKERRQ(ierr); 4343202ece2SStefano Zampini ierr = MatDenseRestoreArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4353202ece2SStefano Zampini } else { 4363202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 4375db18549SStefano Zampini } 4383202ece2SStefano Zampini local_size += subset_size; 43965d8bf0aSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[lpi]);CHKERRQ(ierr); 44065d8bf0aSStefano Zampini sub_schurs->S_Ej[lpi] = S_Ej_expl; 4413202ece2SStefano Zampini } 4422b973097SStefano Zampini /* Stildas are not computed without mumps */ 4432b973097SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 444b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 4452b973097SStefano Zampini } 4463202ece2SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 447883469d8SStefano Zampini } else { 448883469d8SStefano Zampini PetscInt *dummy_idx,n_all; 449883469d8SStefano Zampini 450a1337663SStefano Zampini if (compute_Stilda) { 451a1337663SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_tilda_all);CHKERRQ(ierr); 452a1337663SStefano Zampini ierr = MatSetSizes(S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 453a1337663SStefano Zampini ierr = MatSetType(S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 454a1337663SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 45508122e43SStefano Zampini if (deluxe) { 45608122e43SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_inv_all);CHKERRQ(ierr); 45708122e43SStefano Zampini ierr = MatSetSizes(S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 45808122e43SStefano Zampini ierr = MatSetType(S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 45908122e43SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 460a1337663SStefano Zampini } 46108122e43SStefano Zampini } 46208122e43SStefano Zampini 4638275c6a8SStefano Zampini if (S_all) { /* multilevel */ 46408122e43SStefano Zampini ierr = MatGetSize(S_all,&n_all,NULL);CHKERRQ(ierr); 4658275c6a8SStefano Zampini } else { 4668275c6a8SStefano Zampini n_all = 0; 4678275c6a8SStefano Zampini } 46808122e43SStefano Zampini local_size = 0; 469883469d8SStefano Zampini /* Work arrays */ 470883469d8SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 471883469d8SStefano Zampini 47265d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 47365d8bf0aSStefano Zampini IS is_E; 47465d8bf0aSStefano Zampini PetscScalar *vals; 47565d8bf0aSStefano Zampini PetscInt j; 47665d8bf0aSStefano Zampini 477b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 478a1337663SStefano Zampini for (j=0;j<subset_size;j++) { 479a1337663SStefano Zampini dummy_idx[j]=local_size+j; 480a1337663SStefano Zampini } 48165d8bf0aSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),subset_size,local_size,1,&is_E);CHKERRQ(ierr); 482b96c3477SStefano Zampini if (sub_schurs->S_Ej[i]) { 483b96c3477SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_REUSE_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 484b96c3477SStefano Zampini } else { 485a1337663SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_INITIAL_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 486b96c3477SStefano Zampini } 487a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 488a1337663SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 489a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 490a1337663SStefano Zampini 4916da91a30SStefano Zampini if (compute_Stilda && ((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || use_faces)) { 49208122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,i);CHKERRQ(ierr); 49308122e43SStefano Zampini } 49408122e43SStefano Zampini 49508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 49665d8bf0aSStefano Zampini Mat Stilda; 49765d8bf0aSStefano Zampini if (sub_schurs->n_subs == 1) { 49865d8bf0aSStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->S_Ej[i]);CHKERRQ(ierr); 49965d8bf0aSStefano Zampini Stilda = sub_schurs->S_Ej[i]; 50065d8bf0aSStefano Zampini } else { 50165d8bf0aSStefano Zampini KSP ksp; 50265d8bf0aSStefano Zampini PC pc; 50365d8bf0aSStefano Zampini Mat S_EF,S_FE,S_FF,Stilda_impl; 50465d8bf0aSStefano Zampini IS is_F; 5052b973097SStefano Zampini PetscScalar eps=1.e-8; 5062b973097SStefano Zampini PetscBool chop=PETSC_FALSE; 50765d8bf0aSStefano Zampini 50865d8bf0aSStefano Zampini ierr = ISComplement(is_E,0,n_all,&is_F);CHKERRQ(ierr); 50965d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_F,MAT_INITIAL_MATRIX,&S_EF);CHKERRQ(ierr); 51065d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_F,MAT_INITIAL_MATRIX,&S_FF);CHKERRQ(ierr); 51165d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_E,MAT_INITIAL_MATRIX,&S_FE);CHKERRQ(ierr); 512a1337663SStefano Zampini ierr = ISDestroy(&is_F);CHKERRQ(ierr); 5132b973097SStefano Zampini if (chop) { 5142b973097SStefano Zampini ierr = MatChop(S_FF,eps);CHKERRQ(ierr); 5152b973097SStefano Zampini ierr = MatConvert(S_FF,MATAIJ,MAT_REUSE_MATRIX,&S_FF);CHKERRQ(ierr); 5162b973097SStefano Zampini } 517a1337663SStefano Zampini ierr = MatCreateSchurComplement(S_FF,S_FF,S_FE,S_EF,sub_schurs->S_Ej[i],&Stilda_impl);CHKERRQ(ierr); 518a1337663SStefano Zampini ierr = MatDestroy(&S_FF);CHKERRQ(ierr); 519a1337663SStefano Zampini ierr = MatDestroy(&S_FE);CHKERRQ(ierr); 520a1337663SStefano Zampini ierr = MatDestroy(&S_EF);CHKERRQ(ierr); 52165d8bf0aSStefano Zampini ierr = MatSchurComplementGetKSP(Stilda_impl,&ksp);CHKERRQ(ierr); 52265d8bf0aSStefano Zampini ierr = KSPSetType(ksp,KSPPREONLY);CHKERRQ(ierr); 52365d8bf0aSStefano Zampini ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr); 524*9ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 52565d8bf0aSStefano Zampini ierr = PCSetType(pc,PCCHOLESKY);CHKERRQ(ierr); 52665d8bf0aSStefano Zampini } else { 52765d8bf0aSStefano Zampini ierr = PCSetType(pc,PCLU);CHKERRQ(ierr); 52865d8bf0aSStefano Zampini } 5292b973097SStefano Zampini if (!chop) { 53065d8bf0aSStefano Zampini ierr = PCFactorSetUseInPlace(pc,PETSC_TRUE);CHKERRQ(ierr); 5312b973097SStefano Zampini } else { 5322b973097SStefano Zampini ierr = PCFactorSetMatSolverPackage(pc,MATSOLVERMUMPS);CHKERRQ(ierr); 5332b973097SStefano Zampini } 53465d8bf0aSStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 53565d8bf0aSStefano Zampini ierr = PCBDDCComputeExplicitSchur(Stilda_impl,&Stilda);CHKERRQ(ierr); 53665d8bf0aSStefano Zampini ierr = MatDestroy(&Stilda_impl);CHKERRQ(ierr); 53765d8bf0aSStefano Zampini } 5382b973097SStefano Zampini /* 5392b973097SStefano Zampini PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB); 5402b973097SStefano Zampini ierr = MatView(Stilda,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5412b973097SStefano Zampini PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF); 5422b973097SStefano Zampini */ 543b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 5442b973097SStefano Zampini sub_schurs->S_Ej_tilda[i] = Stilda; 54508122e43SStefano Zampini 546a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 54708122e43SStefano Zampini if (deluxe) { /* when using deluxe scaling, we need (S_1^-1+S_2^-1)^-1 */ 54808122e43SStefano Zampini PetscScalar *vals2; 54908122e43SStefano Zampini 55008122e43SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 55108122e43SStefano Zampini /* need to be optimized (cholesky) */ 55208122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 55308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5549552c7c7SStefano Zampini if (invert_Stildas) { /* Stildas can be singular */ 55508122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals,&B_N,pivots,&B_ierr)); 55608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 55708122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals,&B_N,pivots,work,&B_lwork,&B_ierr)); 55808122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 5599552c7c7SStefano Zampini } 56008122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals2,&B_N,pivots,&B_ierr)); 56108122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 56208122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals2,&B_N,pivots,work,&B_lwork,&B_ierr)); 56308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 56408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 56508122e43SStefano Zampini ierr = MatSetValues(S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,vals2,INSERT_VALUES);CHKERRQ(ierr); 56608122e43SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 56708122e43SStefano Zampini } 568a1337663SStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 569a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 5702b973097SStefano Zampini } else { 571b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 57265d8bf0aSStefano Zampini } 57365d8bf0aSStefano Zampini ierr = ISDestroy(&is_E);CHKERRQ(ierr); 574883469d8SStefano Zampini local_size += subset_size; 575883469d8SStefano Zampini } 576883469d8SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 5775db18549SStefano Zampini } 578a1337663SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 579a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 5805db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5815db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5828275c6a8SStefano Zampini if (S_Ej_tilda_all) { 583a1337663SStefano Zampini ierr = MatAssemblyBegin(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 584a1337663SStefano Zampini ierr = MatAssemblyEnd(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5858275c6a8SStefano Zampini } 5868275c6a8SStefano Zampini if (S_Ej_inv_all) { 58708122e43SStefano Zampini ierr = MatAssemblyBegin(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 58808122e43SStefano Zampini ierr = MatAssemblyEnd(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 58908122e43SStefano Zampini } 590a1337663SStefano Zampini 5915db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 592883469d8SStefano 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); 5935db18549SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,local_size,all_local_idx_G,PETSC_COPY_VALUES,&l2gmap_subsets);CHKERRQ(ierr); 5945db18549SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 5955db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 5965db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 5973927de2eSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 5983927de2eSStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 5993927de2eSStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 6003927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 6013927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 6025db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 603883469d8SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 604883469d8SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->l2gmap),local_size,all_local_idx_G,PETSC_OWN_POINTER,&temp_is);CHKERRQ(ierr); 605b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 606b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 607b96c3477SStefano Zampini } else { 608b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 609b96c3477SStefano Zampini } 61008122e43SStefano Zampini 6118275c6a8SStefano Zampini if (S_Ej_tilda_all) { 612a1337663SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_tilda_all);CHKERRQ(ierr); 613a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 614b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_tilda_all) { 615b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 616b96c3477SStefano Zampini } else { 617b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 618b96c3477SStefano Zampini } 6198275c6a8SStefano Zampini } 6208275c6a8SStefano Zampini if (S_Ej_inv_all) { 62108122e43SStefano Zampini PetscInt cum; 62208122e43SStefano Zampini PetscScalar *array,*array2; 62308122e43SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_inv_all);CHKERRQ(ierr); 62408122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 62508122e43SStefano Zampini if (sub_schurs->sum_S_Ej_inv_all) { 62608122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 62708122e43SStefano Zampini } else { 62808122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 62908122e43SStefano Zampini } 63008122e43SStefano Zampini /* invert blocks of sum_S_Ej_inv_all */ 63108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 63208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 63308122e43SStefano Zampini cum = 0; 63408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 63508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 63608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 63708122e43SStefano Zampini /* need to be optimized (cholesky) */ 63808122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 63908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 64008122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 64108122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 64208122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 64308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 6449552c7c7SStefano Zampini if (invert_Stildas) { /* Stildas can be singular */ 64508122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array2+cum,&B_N,pivots,&B_ierr)); 64608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 64708122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array2+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 64808122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 6499552c7c7SStefano Zampini } 65008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 65108122e43SStefano Zampini } 65208122e43SStefano Zampini cum += subset_size*subset_size; 65308122e43SStefano Zampini } 65408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 65508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 65608122e43SStefano Zampini } 6573202ece2SStefano Zampini 65808122e43SStefano Zampini ierr = PetscFree2(work,pivots);CHKERRQ(ierr); 659a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 660a1337663SStefano Zampini ierr = MatDestroy(&S_Ej_tilda_all);CHKERRQ(ierr); 66108122e43SStefano Zampini ierr = MatDestroy(&S_Ej_inv_all);CHKERRQ(ierr); 6623202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 6635db18549SStefano Zampini ierr = ISDestroy(&temp_is);CHKERRQ(ierr); 664b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 665b1b3d7a2SStefano Zampini } 666b1b3d7a2SStefano Zampini 667b1b3d7a2SStefano Zampini #undef __FUNCT__ 668b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 6695db18549SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, Mat A, Mat S, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap, PetscInt seqthreshold) 670b1b3d7a2SStefano Zampini { 6719bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 672b1b3d7a2SStefano Zampini PetscInt *index_sequential,*index_parallel; 673b1b3d7a2SStefano Zampini PetscInt *auxlocal_sequential,*auxlocal_parallel; 674b1b3d7a2SStefano Zampini PetscInt *auxglobal_sequential,*auxglobal_parallel; 675b96c3477SStefano Zampini PetscInt *auxmapping; 676b1b3d7a2SStefano Zampini PetscInt i,max_subset_size; 677b1b3d7a2SStefano Zampini PetscInt n_sequential_problems,n_local_sequential_problems,n_parallel_problems,n_local_parallel_problems; 678b1b3d7a2SStefano Zampini PetscInt n_faces,n_edges,n_all_cc; 679b1b3d7a2SStefano Zampini PetscBool is_sorted; 680b1b3d7a2SStefano Zampini PetscErrorCode ierr; 681b1b3d7a2SStefano Zampini 682b1b3d7a2SStefano Zampini PetscFunctionBegin; 683b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 684b1b3d7a2SStefano Zampini if (!is_sorted) { 685b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 686b1b3d7a2SStefano Zampini } 687b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 688b1b3d7a2SStefano Zampini if (!is_sorted) { 689b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 690b1b3d7a2SStefano Zampini } 691b1b3d7a2SStefano Zampini 692b1b3d7a2SStefano Zampini /* reset any previous data */ 693b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 694b1b3d7a2SStefano Zampini 695b1b3d7a2SStefano Zampini /* get index sets for faces and edges */ 6969bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 697b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 69808122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 69908122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 700b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 701b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 702b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 703b1b3d7a2SStefano Zampini } 704b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 705b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 70608122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 707b1b3d7a2SStefano Zampini } 708b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 709b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 710b1b3d7a2SStefano Zampini 711b1b3d7a2SStefano Zampini /* map interface's subsets */ 712b1b3d7a2SStefano Zampini max_subset_size = 0; 713b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 714b1b3d7a2SStefano Zampini PetscInt subset_size; 715b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 716b1b3d7a2SStefano Zampini max_subset_size = PetscMax(max_subset_size,subset_size); 717b1b3d7a2SStefano Zampini } 718b1b3d7a2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&auxmapping);CHKERRQ(ierr); 719b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&auxlocal_sequential, 720b1b3d7a2SStefano Zampini graph->ncc,&auxlocal_parallel);CHKERRQ(ierr); 721b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&index_sequential, 722b1b3d7a2SStefano Zampini graph->ncc,&index_parallel);CHKERRQ(ierr); 723b1b3d7a2SStefano Zampini 724883469d8SStefano Zampini /* if threshold is negative uses all sequential problems (possibly using MUMPS) */ 725883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 726883469d8SStefano Zampini if (seqthreshold < 0) { 727883469d8SStefano Zampini seqthreshold = max_subset_size; 728883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 729883469d8SStefano Zampini sub_schurs->use_mumps = !!A; 730883469d8SStefano Zampini #endif 731883469d8SStefano Zampini } 732883469d8SStefano Zampini 733b1b3d7a2SStefano Zampini 734b1b3d7a2SStefano Zampini /* determine which problem has to be solved in parallel or sequentially */ 735b1b3d7a2SStefano Zampini n_local_sequential_problems = 0; 736b1b3d7a2SStefano Zampini n_local_parallel_problems = 0; 737b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 738b1b3d7a2SStefano Zampini PetscInt subset_size,j,min_loc = 0; 739b1b3d7a2SStefano Zampini const PetscInt *idxs; 740b1b3d7a2SStefano Zampini 741b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 742b1b3d7a2SStefano Zampini ierr = ISGetIndices(all_cc[i],&idxs);CHKERRQ(ierr); 743b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingApply(graph->l2gmap,subset_size,idxs,auxmapping);CHKERRQ(ierr); 744b1b3d7a2SStefano Zampini for (j=1;j<subset_size;j++) { 745b1b3d7a2SStefano Zampini if (auxmapping[j]<auxmapping[min_loc]) { 746b1b3d7a2SStefano Zampini min_loc = j; 747b1b3d7a2SStefano Zampini } 748b1b3d7a2SStefano Zampini } 749b1b3d7a2SStefano Zampini if (subset_size > seqthreshold) { 750b1b3d7a2SStefano Zampini index_parallel[n_local_parallel_problems] = i; 751b1b3d7a2SStefano Zampini auxlocal_parallel[n_local_parallel_problems] = idxs[min_loc]; 752b1b3d7a2SStefano Zampini n_local_parallel_problems++; 753b1b3d7a2SStefano Zampini } else { 754b1b3d7a2SStefano Zampini index_sequential[n_local_sequential_problems] = i; 755b1b3d7a2SStefano Zampini auxlocal_sequential[n_local_sequential_problems] = idxs[min_loc]; 756b1b3d7a2SStefano Zampini n_local_sequential_problems++; 757b1b3d7a2SStefano Zampini } 758b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(all_cc[i],&idxs);CHKERRQ(ierr); 759b1b3d7a2SStefano Zampini } 760b1b3d7a2SStefano Zampini 761b1b3d7a2SStefano Zampini /* Number parallel problems */ 762b1b3d7a2SStefano Zampini auxglobal_parallel = 0; 763b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_parallel_problems,auxlocal_parallel,PETSC_NULL,&n_parallel_problems,&auxglobal_parallel);CHKERRQ(ierr); 764b1b3d7a2SStefano Zampini 765b1b3d7a2SStefano Zampini /* Number sequential problems */ 766b1b3d7a2SStefano Zampini auxglobal_sequential = 0; 767b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_sequential_problems,auxlocal_sequential,PETSC_NULL,&n_sequential_problems,&auxglobal_sequential);CHKERRQ(ierr); 768b1b3d7a2SStefano Zampini 769b1b3d7a2SStefano Zampini /* update info in sub_schurs */ 770883469d8SStefano Zampini if (sub_schurs->use_mumps && A) { 771*9ab7bb16SStefano Zampini PetscBool isseqaij; 772*9ab7bb16SStefano Zampini 773*9ab7bb16SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 774*9ab7bb16SStefano Zampini if (isseqaij) { 7751e9c79c2SStefano Zampini ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 7761e9c79c2SStefano Zampini sub_schurs->A = A; 777*9ab7bb16SStefano Zampini } else { /* SeqBAIJ matrices does not support symmetry checking, SeqSBAIJ does not support MatPermute */ 778*9ab7bb16SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 779*9ab7bb16SStefano Zampini } 7801e9c79c2SStefano Zampini } 781b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)S);CHKERRQ(ierr); 782b1b3d7a2SStefano Zampini sub_schurs->S = S; 783b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 784b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 785b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 786b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 7875db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 7885db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 7895db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 7905db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 791b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq = n_local_sequential_problems; 792b1b3d7a2SStefano Zampini sub_schurs->n_subs_par = n_local_parallel_problems; 793b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq_g = n_sequential_problems; 794b1b3d7a2SStefano Zampini sub_schurs->n_subs_par_g = n_parallel_problems; 795b1b3d7a2SStefano Zampini sub_schurs->n_subs = sub_schurs->n_subs_seq + sub_schurs->n_subs_par; 796b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 7979bb4a8caSStefano Zampini if (!sub_schurs->use_mumps) { /* for adaptive selection */ 798b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 799b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 800b96c3477SStefano Zampini } 8019bb4a8caSStefano Zampini } 8029bb4a8caSStefano Zampini sub_schurs->is_Ej_com = vertices; 803b1b3d7a2SStefano Zampini sub_schurs->index_sequential = index_sequential; 804b1b3d7a2SStefano Zampini sub_schurs->index_parallel = index_parallel; 805b1b3d7a2SStefano Zampini sub_schurs->auxglobal_sequential = auxglobal_sequential; 806b1b3d7a2SStefano Zampini sub_schurs->auxglobal_parallel = auxglobal_parallel; 807b1b3d7a2SStefano Zampini 808b96c3477SStefano Zampini 809b96c3477SStefano Zampini /* allocate space for schur complements */ 810b96c3477SStefano Zampini ierr = PetscCalloc2(sub_schurs->n_subs,&sub_schurs->S_Ej, 811b96c3477SStefano Zampini sub_schurs->n_subs,&sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 812b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 813b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 81408122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 815b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 816b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 817b96c3477SStefano Zampini 818b1b3d7a2SStefano Zampini /* free workspace */ 819b1b3d7a2SStefano Zampini ierr = PetscFree(auxmapping);CHKERRQ(ierr); 820b1b3d7a2SStefano Zampini ierr = PetscFree2(auxlocal_sequential,auxlocal_parallel);CHKERRQ(ierr); 821b1b3d7a2SStefano Zampini 822b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 823b1b3d7a2SStefano Zampini } 824b1b3d7a2SStefano Zampini 825b1b3d7a2SStefano Zampini #undef __FUNCT__ 82634a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 82734a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 82834a97f8cSStefano Zampini { 82934a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 83034a97f8cSStefano Zampini PetscErrorCode ierr; 83134a97f8cSStefano Zampini 83234a97f8cSStefano Zampini PetscFunctionBegin; 83334a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 8345ff63025SStefano Zampini schurs_ctx->n_subs = 0; 83534a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 83634a97f8cSStefano Zampini PetscFunctionReturn(0); 83734a97f8cSStefano Zampini } 83834a97f8cSStefano Zampini 83934a97f8cSStefano Zampini #undef __FUNCT__ 84034a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 84134a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 84234a97f8cSStefano Zampini { 84334a97f8cSStefano Zampini PetscErrorCode ierr; 84434a97f8cSStefano Zampini 84534a97f8cSStefano Zampini PetscFunctionBegin; 84634a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 84734a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 84834a97f8cSStefano Zampini PetscFunctionReturn(0); 84934a97f8cSStefano Zampini } 85034a97f8cSStefano Zampini 85134a97f8cSStefano Zampini #undef __FUNCT__ 85234a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 85334a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 85434a97f8cSStefano Zampini { 85534a97f8cSStefano Zampini PetscInt i; 85634a97f8cSStefano Zampini PetscErrorCode ierr; 85734a97f8cSStefano Zampini 85834a97f8cSStefano Zampini PetscFunctionBegin; 8591e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 860b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 861b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 862b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 8635db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 8645db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 86541c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 86641c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 86708122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 868a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 8695db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 8709bb4a8caSStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_com);CHKERRQ(ierr); 87108122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 87208122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 87334a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 874b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 87534a97f8cSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 8762b973097SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 87734a97f8cSStefano Zampini } 87868270318SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 8795ff63025SStefano Zampini if (sub_schurs->n_subs) { 880b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 881b96c3477SStefano Zampini ierr = PetscFree2(sub_schurs->S_Ej,sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 8823dc780c3SStefano Zampini } 883b1b3d7a2SStefano Zampini ierr = PetscFree2(sub_schurs->index_sequential,sub_schurs->index_parallel);CHKERRQ(ierr); 884b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_sequential);CHKERRQ(ierr); 885b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_parallel);CHKERRQ(ierr); 88634a97f8cSStefano Zampini sub_schurs->n_subs = 0; 88734a97f8cSStefano Zampini PetscFunctionReturn(0); 88834a97f8cSStefano Zampini } 88934a97f8cSStefano Zampini 89034a97f8cSStefano Zampini #undef __FUNCT__ 89134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 8922a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 89334a97f8cSStefano Zampini { 89434a97f8cSStefano Zampini PetscInt i,j,n; 89534a97f8cSStefano Zampini PetscErrorCode ierr; 89634a97f8cSStefano Zampini 89734a97f8cSStefano Zampini PetscFunctionBegin; 89834a97f8cSStefano Zampini n = 0; 89934a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 90034a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 90134a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 90234a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 90334a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 90434a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 90534a97f8cSStefano Zampini queue_tip[n] = dof; 90634a97f8cSStefano Zampini n++; 90734a97f8cSStefano Zampini } 90834a97f8cSStefano Zampini } 90934a97f8cSStefano Zampini } 91034a97f8cSStefano Zampini *n_added = n; 91134a97f8cSStefano Zampini PetscFunctionReturn(0); 91234a97f8cSStefano Zampini } 913