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" 90*9552c7c7SStefano 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 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) { 1103dc780c3SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints 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); 117a58a30b4SStefano Zampini } else { 118a58a30b4SStefano Zampini is_subset_B = NULL; 119a58a30b4SStefano Zampini AE_IE = NULL; 120a58a30b4SStefano Zampini AE_EI = NULL; 121a58a30b4SStefano Zampini AE_EE = NULL; 122b1b3d7a2SStefano Zampini } 123b1b3d7a2SStefano Zampini 124b1b3d7a2SStefano Zampini /* determine interior problems */ 125b96c3477SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 1263dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 1273dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 128b1b3d7a2SStefano Zampini PetscBT touched; 129b1b3d7a2SStefano Zampini const PetscInt* idx_B; 130b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 131b1b3d7a2SStefano Zampini 1323dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 1333dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 1343dc780c3SStefano Zampini } 135b1b3d7a2SStefano Zampini /* get sizes */ 136b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 137b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 138b1b3d7a2SStefano Zampini 139b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 140b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 141b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 142b1b3d7a2SStefano Zampini 143b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 144b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 145b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 146b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 147b1b3d7a2SStefano Zampini } 148b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 149b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 150b1b3d7a2SStefano Zampini 151b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 152b1b3d7a2SStefano Zampini n_local_dofs = n_B; 153b1b3d7a2SStefano Zampini n_prev_added = n_B; 154b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 155b1b3d7a2SStefano Zampini PetscInt n_added; 156b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 157b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 158b1b3d7a2SStefano 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); 159b1b3d7a2SStefano Zampini } 160b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 161b1b3d7a2SStefano Zampini n_prev_added = n_added; 162b1b3d7a2SStefano Zampini n_local_dofs += n_added; 163b1b3d7a2SStefano Zampini if (!n_added) break; 164b1b3d7a2SStefano Zampini } 165b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 166b1b3d7a2SStefano Zampini 167883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 16868270318SStefano 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); 169b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 17068270318SStefano Zampini ierr = ISSort(sub_schurs->is_I_layer);CHKERRQ(ierr); 171883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 172883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 173b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 174b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 17568270318SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,sub_schurs->is_I_layer,&is_I);CHKERRQ(ierr); 176b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 177b1b3d7a2SStefano Zampini 178b1b3d7a2SStefano Zampini /* II block */ 179b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 180b1b3d7a2SStefano Zampini } 181b1b3d7a2SStefano Zampini } else { 182b1b3d7a2SStefano Zampini PetscInt n_I; 183b1b3d7a2SStefano Zampini 184b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 185b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 18668270318SStefano Zampini sub_schurs->is_I_layer = sub_schurs->is_I; 187b1b3d7a2SStefano Zampini 188b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 189883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 190b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 191b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 192b1b3d7a2SStefano Zampini 193b1b3d7a2SStefano Zampini /* II block is the same */ 194b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 195b1b3d7a2SStefano Zampini AE_II = A_II; 196b1b3d7a2SStefano Zampini } 197b1b3d7a2SStefano Zampini } 198883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 199883469d8SStefano Zampini max_subset_size = 0; 200883469d8SStefano Zampini local_size = 0; 201883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 202883469d8SStefano Zampini PetscInt j = sub_schurs->index_sequential[i]; 203b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[j],&subset_size);CHKERRQ(ierr); 204883469d8SStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 205883469d8SStefano Zampini local_size += subset_size; 206883469d8SStefano Zampini } 207883469d8SStefano Zampini 208883469d8SStefano Zampini /* Work arrays for local indices */ 209883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 210883469d8SStefano Zampini extra = 0; 211883469d8SStefano Zampini if (sub_schurs->use_mumps) { 212883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&extra);CHKERRQ(ierr); 213883469d8SStefano Zampini } 214883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 215883469d8SStefano Zampini if (extra) { 216883469d8SStefano Zampini const PetscInt *idxs; 217883469d8SStefano Zampini ierr = ISGetIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 218883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 219883469d8SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_I_layer,&idxs);CHKERRQ(ierr); 220883469d8SStefano Zampini } 221883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 222883469d8SStefano Zampini 223883469d8SStefano Zampini /* Get local indices in local numbering */ 224883469d8SStefano Zampini local_size = 0; 225883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 226883469d8SStefano Zampini PetscInt j; 227883469d8SStefano Zampini const PetscInt *idxs; 228883469d8SStefano Zampini 229883469d8SStefano Zampini PetscInt local_problem_index = sub_schurs->index_sequential[i]; 230b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[local_problem_index],&subset_size);CHKERRQ(ierr); 231b96c3477SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 232883469d8SStefano Zampini /* subset indices in local numbering */ 233883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 234b96c3477SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[local_problem_index],&idxs);CHKERRQ(ierr); 235883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 236883469d8SStefano Zampini local_size += subset_size; 237883469d8SStefano Zampini } 238883469d8SStefano Zampini 239883469d8SStefano Zampini S_all = NULL; 240*9552c7c7SStefano Zampini if (sub_schurs->A) { 24108122e43SStefano Zampini ierr = MatIsSymmetric(sub_schurs->A,0.0,&is_symmetric);CHKERRQ(ierr); 242*9552c7c7SStefano Zampini } 2433dc780c3SStefano Zampini if (sub_schurs->n_subs) { 244883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 245883469d8SStefano Zampini /* subsets in original and boundary numbering */ 246883469d8SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 247b96c3477SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B[i]);CHKERRQ(ierr); 248b1b3d7a2SStefano Zampini } 249b1b3d7a2SStefano Zampini 250b1b3d7a2SStefano Zampini /* EE block */ 251b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 252b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B[i],is_subset_B[i],MAT_INITIAL_MATRIX,&AE_EE[i]);CHKERRQ(ierr); 253b1b3d7a2SStefano Zampini } 254b1b3d7a2SStefano Zampini /* IE block */ 255b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 256b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B[i],MAT_INITIAL_MATRIX,&AE_IE[i]);CHKERRQ(ierr); 257b1b3d7a2SStefano Zampini } 258b1b3d7a2SStefano Zampini /* EI block */ 259b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 260b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B[i],is_I,MAT_INITIAL_MATRIX,&AE_EI[i]);CHKERRQ(ierr); 261b1b3d7a2SStefano Zampini } 262b1b3d7a2SStefano Zampini 263b1b3d7a2SStefano Zampini /* setup Schur complements on subset */ 264b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 265b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 266b1b3d7a2SStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE[i],AE_EI[i],AE_EE[i],&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 267a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EE[i]);CHKERRQ(ierr); 268a58a30b4SStefano Zampini ierr = MatDestroy(&AE_IE[i]);CHKERRQ(ierr); 269a58a30b4SStefano Zampini ierr = MatDestroy(&AE_EI[i]);CHKERRQ(ierr); 270b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 271b1b3d7a2SStefano Zampini KSP ksp; 272b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 273b1b3d7a2SStefano Zampini ierr = MatSchurComplementSetKSP(sub_schurs->S_Ej[i],ksp);CHKERRQ(ierr); 274b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 275b1b3d7a2SStefano Zampini KSP origksp,schurksp; 276b1b3d7a2SStefano Zampini PC origpc,schurpc; 277b1b3d7a2SStefano Zampini KSPType ksp_type; 278b1b3d7a2SStefano Zampini PCType pc_type; 279b1b3d7a2SStefano Zampini PetscInt n_internal; 280b1b3d7a2SStefano Zampini 281b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 282b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S_Ej[i],&schurksp);CHKERRQ(ierr); 283b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 284b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 285b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 286b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 287b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 288b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 289b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 290b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 291b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 292b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 293b1b3d7a2SStefano Zampini if (solver) { 294b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 295b1b3d7a2SStefano Zampini } 296b1b3d7a2SStefano Zampini } 297b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 298b1b3d7a2SStefano Zampini } 299b1b3d7a2SStefano Zampini } 300b1b3d7a2SStefano Zampini /* free */ 301b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 302b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 303b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 304b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_subset_B[i]);CHKERRQ(ierr); 305b1b3d7a2SStefano Zampini } 306b1b3d7a2SStefano Zampini ierr = PetscFree4(is_subset_B,AE_IE,AE_EI,AE_EE);CHKERRQ(ierr); 307883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 308883469d8SStefano Zampini } else { 309883469d8SStefano Zampini Mat A,F; 310883469d8SStefano Zampini IS is_A_all; 311883469d8SStefano Zampini PetscInt *idxs_schur,n_I; 312883469d8SStefano Zampini 313883469d8SStefano Zampini /* get working mat */ 314883469d8SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I_layer,&n_I);CHKERRQ(ierr); 315883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 316b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 317883469d8SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 318883469d8SStefano Zampini 31908122e43SStefano Zampini if (n_I) { 320883469d8SStefano Zampini if (is_symmetric) { 321883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 322883469d8SStefano Zampini } else { 323883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 324883469d8SStefano Zampini } 325883469d8SStefano Zampini 326883469d8SStefano Zampini /* subsets ordered last */ 327883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&idxs_schur);CHKERRQ(ierr); 328883469d8SStefano Zampini for (i=0;i<local_size;i++) { 329883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 330883469d8SStefano Zampini } 331883469d8SStefano Zampini ierr = MatMumpsSetSchurIndices(F,local_size,idxs_schur);CHKERRQ(ierr); 332883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 333883469d8SStefano Zampini 334883469d8SStefano Zampini /* factorization step */ 335883469d8SStefano Zampini if (is_symmetric) { 336883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 337883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 338883469d8SStefano Zampini } else { 339883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 340883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 341883469d8SStefano Zampini } 342883469d8SStefano Zampini 343883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 344883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 345883469d8SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 34608122e43SStefano Zampini } else { 34708122e43SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 34808122e43SStefano Zampini } 34908122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 350883469d8SStefano Zampini #endif 351b1b3d7a2SStefano Zampini } 3523dc780c3SStefano Zampini } else { 3533dc780c3SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3543dc780c3SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 3553dc780c3SStefano Zampini } 3565db18549SStefano Zampini 3575db18549SStefano Zampini if (!sub_schurs->n_subs_seq_g) { 3585db18549SStefano Zampini PetscFunctionReturn(0); 3595db18549SStefano Zampini } 3605db18549SStefano Zampini 3615db18549SStefano Zampini /* subset indices in local boundary numbering */ 362b96c3477SStefano Zampini if (!sub_schurs->is_Ej_all) { 363883469d8SStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 3645db18549SStefano Zampini if (subset_size != local_size) { 3655db18549SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 3665db18549SStefano Zampini } 3675db18549SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 368b96c3477SStefano Zampini } 3695db18549SStefano Zampini 3703202ece2SStefano Zampini /* Local matrix of all local Schur on subsets transposed */ 371b96c3477SStefano Zampini if (!sub_schurs->S_Ej_all) { 3725db18549SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3735db18549SStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 3745db18549SStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 3755db18549SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 376b96c3477SStefano Zampini } else { 377b96c3477SStefano Zampini ierr = MatZeroEntries(sub_schurs->S_Ej_all);CHKERRQ(ierr); 378b96c3477SStefano Zampini } 379a1337663SStefano Zampini 380a1337663SStefano Zampini S_Ej_tilda_all = 0; 38108122e43SStefano Zampini S_Ej_inv_all = 0; 382a58a30b4SStefano Zampini work = NULL; 383a58a30b4SStefano Zampini pivots = NULL; 384a58a30b4SStefano Zampini Stilda_computed = PETSC_FALSE; 38508122e43SStefano Zampini if (sub_schurs->n_subs && deluxe) { /* workspace needed only for GETRI */ 38608122e43SStefano Zampini PetscScalar lwork; 38708122e43SStefano Zampini 38808122e43SStefano Zampini B_lwork = -1; 38908122e43SStefano Zampini ierr = PetscBLASIntCast(max_subset_size,&B_N);CHKERRQ(ierr); 39008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39108122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 39208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 39408122e43SStefano Zampini ierr = PetscBLASIntCast((PetscInt)lwork,&B_lwork);CHKERRQ(ierr); 39508122e43SStefano Zampini ierr = PetscMalloc2(B_lwork,&work,max_subset_size,&pivots);CHKERRQ(ierr); 39608122e43SStefano Zampini } 39708122e43SStefano Zampini 39808122e43SStefano Zampini ierr = PetscBTMemzero(sub_schurs->n_subs,sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 399883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 4005db18549SStefano Zampini PetscScalar *fill_vals; 4015db18549SStefano Zampini PetscInt *dummy_idx; 4025db18549SStefano Zampini 4035db18549SStefano Zampini /* Work arrays */ 4043202ece2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 4053202ece2SStefano Zampini /* Loop on local problems end compute Schur complements explicitly */ 4065db18549SStefano Zampini local_size = 0; 4075db18549SStefano Zampini for (i=0;i<sub_schurs->n_subs_seq;i++) { 4083202ece2SStefano Zampini Mat S_Ej_expl; 4093202ece2SStefano Zampini PetscInt j,lpi = sub_schurs->index_sequential[i]; 4103202ece2SStefano Zampini PetscBool Sdense; 4115db18549SStefano Zampini 4123202ece2SStefano Zampini ierr = PCBDDCComputeExplicitSchur(sub_schurs->S_Ej[lpi],&S_Ej_expl);CHKERRQ(ierr); 413b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[lpi],&subset_size);CHKERRQ(ierr); 4143202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 4153202ece2SStefano Zampini if (Sdense) { 4165db18549SStefano Zampini for (j=0;j<subset_size;j++) { 4175db18549SStefano Zampini dummy_idx[j]=local_size+j; 4185db18549SStefano Zampini } 4193202ece2SStefano Zampini ierr = MatDenseGetArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4205db18549SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,fill_vals,INSERT_VALUES);CHKERRQ(ierr); 4213202ece2SStefano Zampini ierr = MatDenseRestoreArray(S_Ej_expl,&fill_vals);CHKERRQ(ierr); 4223202ece2SStefano Zampini } else { 4233202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 4245db18549SStefano Zampini } 4253202ece2SStefano Zampini local_size += subset_size; 42665d8bf0aSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[lpi]);CHKERRQ(ierr); 42765d8bf0aSStefano Zampini sub_schurs->S_Ej[lpi] = S_Ej_expl; 4283202ece2SStefano Zampini } 4292b973097SStefano Zampini /* Stildas are not computed without mumps */ 4302b973097SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 431b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 4322b973097SStefano Zampini } 4333202ece2SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 434883469d8SStefano Zampini } else { 435883469d8SStefano Zampini PetscInt *dummy_idx,n_all; 436883469d8SStefano Zampini 437a1337663SStefano Zampini if (compute_Stilda) { 438a1337663SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_tilda_all);CHKERRQ(ierr); 439a1337663SStefano Zampini ierr = MatSetSizes(S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 440a1337663SStefano Zampini ierr = MatSetType(S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 441a1337663SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 44208122e43SStefano Zampini if (deluxe) { 44308122e43SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_inv_all);CHKERRQ(ierr); 44408122e43SStefano Zampini ierr = MatSetSizes(S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 44508122e43SStefano Zampini ierr = MatSetType(S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 44608122e43SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 447a1337663SStefano Zampini } 44808122e43SStefano Zampini } 44908122e43SStefano Zampini 45008122e43SStefano Zampini ierr = MatGetSize(S_all,&n_all,NULL);CHKERRQ(ierr); 45108122e43SStefano Zampini local_size = 0; 452883469d8SStefano Zampini /* Work arrays */ 453883469d8SStefano Zampini ierr = PetscMalloc1(max_subset_size,&dummy_idx);CHKERRQ(ierr); 454883469d8SStefano Zampini 45508122e43SStefano Zampini Stilda_computed = PETSC_FALSE; 45665d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 45765d8bf0aSStefano Zampini IS is_E; 45865d8bf0aSStefano Zampini PetscScalar *vals; 45965d8bf0aSStefano Zampini PetscInt j; 46065d8bf0aSStefano Zampini 461b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 462a1337663SStefano Zampini for (j=0;j<subset_size;j++) { 463a1337663SStefano Zampini dummy_idx[j]=local_size+j; 464a1337663SStefano Zampini } 46565d8bf0aSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),subset_size,local_size,1,&is_E);CHKERRQ(ierr); 466b96c3477SStefano Zampini if (sub_schurs->S_Ej[i]) { 467b96c3477SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_REUSE_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 468b96c3477SStefano Zampini } else { 469a1337663SStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_E,MAT_INITIAL_MATRIX,&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 470b96c3477SStefano Zampini } 471a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 472a1337663SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 473a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals);CHKERRQ(ierr); 474a1337663SStefano Zampini 4756da91a30SStefano Zampini if (compute_Stilda && ((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || use_faces)) { 47608122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,i);CHKERRQ(ierr); 47708122e43SStefano Zampini Stilda_computed = PETSC_TRUE; 47808122e43SStefano Zampini } 47908122e43SStefano Zampini 48008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 48165d8bf0aSStefano Zampini Mat Stilda; 48265d8bf0aSStefano Zampini if (sub_schurs->n_subs == 1) { 48365d8bf0aSStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->S_Ej[i]);CHKERRQ(ierr); 48465d8bf0aSStefano Zampini Stilda = sub_schurs->S_Ej[i]; 48565d8bf0aSStefano Zampini } else { 48665d8bf0aSStefano Zampini KSP ksp; 48765d8bf0aSStefano Zampini PC pc; 48865d8bf0aSStefano Zampini Mat S_EF,S_FE,S_FF,Stilda_impl; 48965d8bf0aSStefano Zampini IS is_F; 4902b973097SStefano Zampini PetscScalar eps=1.e-8; 4912b973097SStefano Zampini PetscBool chop=PETSC_FALSE; 49265d8bf0aSStefano Zampini 49365d8bf0aSStefano Zampini ierr = ISComplement(is_E,0,n_all,&is_F);CHKERRQ(ierr); 49465d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_E,is_F,MAT_INITIAL_MATRIX,&S_EF);CHKERRQ(ierr); 49565d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_F,MAT_INITIAL_MATRIX,&S_FF);CHKERRQ(ierr); 49665d8bf0aSStefano Zampini ierr = MatGetSubMatrix(S_all,is_F,is_E,MAT_INITIAL_MATRIX,&S_FE);CHKERRQ(ierr); 497a1337663SStefano Zampini ierr = ISDestroy(&is_F);CHKERRQ(ierr); 4982b973097SStefano Zampini if (chop) { 4992b973097SStefano Zampini ierr = MatChop(S_FF,eps);CHKERRQ(ierr); 5002b973097SStefano Zampini ierr = MatConvert(S_FF,MATAIJ,MAT_REUSE_MATRIX,&S_FF);CHKERRQ(ierr); 5012b973097SStefano Zampini } 502a1337663SStefano Zampini ierr = MatCreateSchurComplement(S_FF,S_FF,S_FE,S_EF,sub_schurs->S_Ej[i],&Stilda_impl);CHKERRQ(ierr); 503a1337663SStefano Zampini ierr = MatDestroy(&S_FF);CHKERRQ(ierr); 504a1337663SStefano Zampini ierr = MatDestroy(&S_FE);CHKERRQ(ierr); 505a1337663SStefano Zampini ierr = MatDestroy(&S_EF);CHKERRQ(ierr); 50665d8bf0aSStefano Zampini ierr = MatSchurComplementGetKSP(Stilda_impl,&ksp);CHKERRQ(ierr); 50765d8bf0aSStefano Zampini ierr = KSPSetType(ksp,KSPPREONLY);CHKERRQ(ierr); 50865d8bf0aSStefano Zampini ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr); 50965d8bf0aSStefano Zampini if (is_symmetric) { 51065d8bf0aSStefano Zampini ierr = PCSetType(pc,PCCHOLESKY);CHKERRQ(ierr); 51165d8bf0aSStefano Zampini } else { 51265d8bf0aSStefano Zampini ierr = PCSetType(pc,PCLU);CHKERRQ(ierr); 51365d8bf0aSStefano Zampini } 5142b973097SStefano Zampini if (!chop) { 51565d8bf0aSStefano Zampini ierr = PCFactorSetUseInPlace(pc,PETSC_TRUE);CHKERRQ(ierr); 5162b973097SStefano Zampini } else { 5172b973097SStefano Zampini ierr = PCFactorSetMatSolverPackage(pc,MATSOLVERMUMPS);CHKERRQ(ierr); 5182b973097SStefano Zampini } 51965d8bf0aSStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 52065d8bf0aSStefano Zampini ierr = PCBDDCComputeExplicitSchur(Stilda_impl,&Stilda);CHKERRQ(ierr); 52165d8bf0aSStefano Zampini ierr = MatDestroy(&Stilda_impl);CHKERRQ(ierr); 52265d8bf0aSStefano Zampini } 5232b973097SStefano Zampini /* 5242b973097SStefano Zampini PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB); 5252b973097SStefano Zampini ierr = MatView(Stilda,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5262b973097SStefano Zampini PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF); 5272b973097SStefano Zampini */ 528b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 5292b973097SStefano Zampini sub_schurs->S_Ej_tilda[i] = Stilda; 53008122e43SStefano Zampini 53108122e43SStefano Zampini 532a1337663SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 53308122e43SStefano Zampini if (deluxe) { /* when using deluxe scaling, we need (S_1^-1+S_2^-1)^-1 */ 53408122e43SStefano Zampini PetscScalar *vals2; 53508122e43SStefano Zampini 53608122e43SStefano Zampini ierr = MatDenseGetArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 53708122e43SStefano Zampini /* need to be optimized (cholesky) */ 53808122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 53908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 540*9552c7c7SStefano Zampini if (invert_Stildas) { /* Stildas can be singular */ 54108122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals,&B_N,pivots,&B_ierr)); 54208122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 54308122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals,&B_N,pivots,work,&B_lwork,&B_ierr)); 54408122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 545*9552c7c7SStefano Zampini } 54608122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,vals2,&B_N,pivots,&B_ierr)); 54708122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 54808122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,vals2,&B_N,pivots,work,&B_lwork,&B_ierr)); 54908122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 55008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 55108122e43SStefano Zampini ierr = MatSetValues(S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,vals2,INSERT_VALUES);CHKERRQ(ierr); 55208122e43SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej[i],&vals2);CHKERRQ(ierr); 55308122e43SStefano Zampini } 554a1337663SStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,vals,INSERT_VALUES);CHKERRQ(ierr); 555a1337663SStefano Zampini ierr = MatDenseRestoreArray(sub_schurs->S_Ej_tilda[i],&vals);CHKERRQ(ierr); 5562b973097SStefano Zampini } else { 557b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 55865d8bf0aSStefano Zampini } 55965d8bf0aSStefano Zampini ierr = ISDestroy(&is_E);CHKERRQ(ierr); 560883469d8SStefano Zampini local_size += subset_size; 561883469d8SStefano Zampini } 562883469d8SStefano Zampini ierr = PetscFree(dummy_idx);CHKERRQ(ierr); 5635db18549SStefano Zampini } 564a1337663SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 565a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 5665db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5675db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 568a1337663SStefano Zampini if (compute_Stilda) { 569a1337663SStefano Zampini ierr = MatAssemblyBegin(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 570a1337663SStefano Zampini ierr = MatAssemblyEnd(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57108122e43SStefano Zampini if (deluxe) { 57208122e43SStefano Zampini ierr = MatAssemblyBegin(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57308122e43SStefano Zampini ierr = MatAssemblyEnd(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57408122e43SStefano Zampini } 575a1337663SStefano Zampini } 576a1337663SStefano Zampini 5775db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 578883469d8SStefano 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); 5795db18549SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,local_size,all_local_idx_G,PETSC_COPY_VALUES,&l2gmap_subsets);CHKERRQ(ierr); 5805db18549SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)sub_schurs->l2gmap),1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 5815db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 5825db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 5833927de2eSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 5843927de2eSStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 5853927de2eSStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 5863927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 5873927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 5885db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 589883469d8SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 590883469d8SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->l2gmap),local_size,all_local_idx_G,PETSC_OWN_POINTER,&temp_is);CHKERRQ(ierr); 591b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 592b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 593b96c3477SStefano Zampini } else { 594b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 595b96c3477SStefano Zampini } 59608122e43SStefano Zampini 59708122e43SStefano Zampini if (Stilda_computed) { 598a1337663SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_tilda_all);CHKERRQ(ierr); 599a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 600b96c3477SStefano Zampini if (sub_schurs->sum_S_Ej_tilda_all) { 601b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 602b96c3477SStefano Zampini } else { 603b96c3477SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 604b96c3477SStefano Zampini } 60508122e43SStefano Zampini if (deluxe) { 60608122e43SStefano Zampini PetscInt cum; 60708122e43SStefano Zampini PetscScalar *array,*array2; 60808122e43SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_inv_all);CHKERRQ(ierr); 60908122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 61008122e43SStefano Zampini if (sub_schurs->sum_S_Ej_inv_all) { 61108122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_REUSE_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 61208122e43SStefano Zampini } else { 61308122e43SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,MAT_INITIAL_MATRIX,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 61408122e43SStefano Zampini } 61508122e43SStefano Zampini /* invert blocks of sum_S_Ej_inv_all */ 61608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 61708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 61808122e43SStefano Zampini cum = 0; 61908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 62008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 62108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 62208122e43SStefano Zampini /* need to be optimized (cholesky) */ 62308122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 62408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 62508122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 62608122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 62708122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 62808122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 629*9552c7c7SStefano Zampini if (invert_Stildas) { /* Stildas can be singular */ 63008122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array2+cum,&B_N,pivots,&B_ierr)); 63108122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 63208122e43SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array2+cum,&B_N,pivots,work,&B_lwork,&B_ierr)); 63308122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 634*9552c7c7SStefano Zampini } 63508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 63608122e43SStefano Zampini } 63708122e43SStefano Zampini cum += subset_size*subset_size; 63808122e43SStefano Zampini } 63908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&array);CHKERRQ(ierr); 64008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array2);CHKERRQ(ierr); 64108122e43SStefano Zampini } 6423202ece2SStefano Zampini } 6433202ece2SStefano Zampini 64408122e43SStefano Zampini ierr = PetscFree2(work,pivots);CHKERRQ(ierr); 645a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 646a1337663SStefano Zampini ierr = MatDestroy(&S_Ej_tilda_all);CHKERRQ(ierr); 64708122e43SStefano Zampini ierr = MatDestroy(&S_Ej_inv_all);CHKERRQ(ierr); 6483202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 6495db18549SStefano Zampini ierr = ISDestroy(&temp_is);CHKERRQ(ierr); 650b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 651b1b3d7a2SStefano Zampini } 652b1b3d7a2SStefano Zampini 653b1b3d7a2SStefano Zampini #undef __FUNCT__ 654b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 6555db18549SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, Mat A, Mat S, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap, PetscInt seqthreshold) 656b1b3d7a2SStefano Zampini { 6579bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 658b1b3d7a2SStefano Zampini PetscInt *index_sequential,*index_parallel; 659b1b3d7a2SStefano Zampini PetscInt *auxlocal_sequential,*auxlocal_parallel; 660b1b3d7a2SStefano Zampini PetscInt *auxglobal_sequential,*auxglobal_parallel; 661b96c3477SStefano Zampini PetscInt *auxmapping; 662b1b3d7a2SStefano Zampini PetscInt i,max_subset_size; 663b1b3d7a2SStefano Zampini PetscInt n_sequential_problems,n_local_sequential_problems,n_parallel_problems,n_local_parallel_problems; 664b1b3d7a2SStefano Zampini PetscInt n_faces,n_edges,n_all_cc; 665b1b3d7a2SStefano Zampini PetscBool is_sorted; 666b1b3d7a2SStefano Zampini PetscErrorCode ierr; 667b1b3d7a2SStefano Zampini 668b1b3d7a2SStefano Zampini PetscFunctionBegin; 669b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 670b1b3d7a2SStefano Zampini if (!is_sorted) { 671b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 672b1b3d7a2SStefano Zampini } 673b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 674b1b3d7a2SStefano Zampini if (!is_sorted) { 675b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 676b1b3d7a2SStefano Zampini } 677b1b3d7a2SStefano Zampini 678b1b3d7a2SStefano Zampini /* reset any previous data */ 679b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 680b1b3d7a2SStefano Zampini 681b1b3d7a2SStefano Zampini /* get index sets for faces and edges */ 6829bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 683b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 68408122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 68508122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 686b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 687b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 688b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 689b1b3d7a2SStefano Zampini } 690b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 691b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 69208122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 693b1b3d7a2SStefano Zampini } 694b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 695b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 696b1b3d7a2SStefano Zampini 697b1b3d7a2SStefano Zampini /* map interface's subsets */ 698b1b3d7a2SStefano Zampini max_subset_size = 0; 699b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 700b1b3d7a2SStefano Zampini PetscInt subset_size; 701b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 702b1b3d7a2SStefano Zampini max_subset_size = PetscMax(max_subset_size,subset_size); 703b1b3d7a2SStefano Zampini } 704b1b3d7a2SStefano Zampini ierr = PetscMalloc1(max_subset_size,&auxmapping);CHKERRQ(ierr); 705b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&auxlocal_sequential, 706b1b3d7a2SStefano Zampini graph->ncc,&auxlocal_parallel);CHKERRQ(ierr); 707b1b3d7a2SStefano Zampini ierr = PetscMalloc2(graph->ncc,&index_sequential, 708b1b3d7a2SStefano Zampini graph->ncc,&index_parallel);CHKERRQ(ierr); 709b1b3d7a2SStefano Zampini 710883469d8SStefano Zampini /* if threshold is negative uses all sequential problems (possibly using MUMPS) */ 711883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 712883469d8SStefano Zampini if (seqthreshold < 0) { 713883469d8SStefano Zampini seqthreshold = max_subset_size; 714883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 715883469d8SStefano Zampini sub_schurs->use_mumps = !!A; 716883469d8SStefano Zampini #endif 717883469d8SStefano Zampini } 718883469d8SStefano Zampini 719b1b3d7a2SStefano Zampini 720b1b3d7a2SStefano Zampini /* determine which problem has to be solved in parallel or sequentially */ 721b1b3d7a2SStefano Zampini n_local_sequential_problems = 0; 722b1b3d7a2SStefano Zampini n_local_parallel_problems = 0; 723b1b3d7a2SStefano Zampini for (i=0;i<n_all_cc;i++) { 724b1b3d7a2SStefano Zampini PetscInt subset_size,j,min_loc = 0; 725b1b3d7a2SStefano Zampini const PetscInt *idxs; 726b1b3d7a2SStefano Zampini 727b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(all_cc[i],&subset_size);CHKERRQ(ierr); 728b1b3d7a2SStefano Zampini ierr = ISGetIndices(all_cc[i],&idxs);CHKERRQ(ierr); 729b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingApply(graph->l2gmap,subset_size,idxs,auxmapping);CHKERRQ(ierr); 730b1b3d7a2SStefano Zampini for (j=1;j<subset_size;j++) { 731b1b3d7a2SStefano Zampini if (auxmapping[j]<auxmapping[min_loc]) { 732b1b3d7a2SStefano Zampini min_loc = j; 733b1b3d7a2SStefano Zampini } 734b1b3d7a2SStefano Zampini } 735b1b3d7a2SStefano Zampini if (subset_size > seqthreshold) { 736b1b3d7a2SStefano Zampini index_parallel[n_local_parallel_problems] = i; 737b1b3d7a2SStefano Zampini auxlocal_parallel[n_local_parallel_problems] = idxs[min_loc]; 738b1b3d7a2SStefano Zampini n_local_parallel_problems++; 739b1b3d7a2SStefano Zampini } else { 740b1b3d7a2SStefano Zampini index_sequential[n_local_sequential_problems] = i; 741b1b3d7a2SStefano Zampini auxlocal_sequential[n_local_sequential_problems] = idxs[min_loc]; 742b1b3d7a2SStefano Zampini n_local_sequential_problems++; 743b1b3d7a2SStefano Zampini } 744b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(all_cc[i],&idxs);CHKERRQ(ierr); 745b1b3d7a2SStefano Zampini } 746b1b3d7a2SStefano Zampini 747b1b3d7a2SStefano Zampini /* Number parallel problems */ 748b1b3d7a2SStefano Zampini auxglobal_parallel = 0; 749b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_parallel_problems,auxlocal_parallel,PETSC_NULL,&n_parallel_problems,&auxglobal_parallel);CHKERRQ(ierr); 750b1b3d7a2SStefano Zampini 751b1b3d7a2SStefano Zampini /* Number sequential problems */ 752b1b3d7a2SStefano Zampini auxglobal_sequential = 0; 753b1b3d7a2SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)graph->l2gmap),graph->l2gmap,n_local_sequential_problems,auxlocal_sequential,PETSC_NULL,&n_sequential_problems,&auxglobal_sequential);CHKERRQ(ierr); 754b1b3d7a2SStefano Zampini 755b1b3d7a2SStefano Zampini /* update info in sub_schurs */ 756883469d8SStefano Zampini if (sub_schurs->use_mumps && A) { 7571e9c79c2SStefano Zampini ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 7581e9c79c2SStefano Zampini sub_schurs->A = A; 7591e9c79c2SStefano Zampini } 760b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)S);CHKERRQ(ierr); 761b1b3d7a2SStefano Zampini sub_schurs->S = S; 762b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 763b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 764b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 765b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 7665db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 7675db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 7685db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 7695db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 770b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq = n_local_sequential_problems; 771b1b3d7a2SStefano Zampini sub_schurs->n_subs_par = n_local_parallel_problems; 772b1b3d7a2SStefano Zampini sub_schurs->n_subs_seq_g = n_sequential_problems; 773b1b3d7a2SStefano Zampini sub_schurs->n_subs_par_g = n_parallel_problems; 774b1b3d7a2SStefano Zampini sub_schurs->n_subs = sub_schurs->n_subs_seq + sub_schurs->n_subs_par; 775b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 7769bb4a8caSStefano Zampini if (!sub_schurs->use_mumps) { /* for adaptive selection */ 777b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 778b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 779b96c3477SStefano Zampini } 7809bb4a8caSStefano Zampini } 7819bb4a8caSStefano Zampini sub_schurs->is_Ej_com = vertices; 782b1b3d7a2SStefano Zampini sub_schurs->index_sequential = index_sequential; 783b1b3d7a2SStefano Zampini sub_schurs->index_parallel = index_parallel; 784b1b3d7a2SStefano Zampini sub_schurs->auxglobal_sequential = auxglobal_sequential; 785b1b3d7a2SStefano Zampini sub_schurs->auxglobal_parallel = auxglobal_parallel; 786b1b3d7a2SStefano Zampini 787b96c3477SStefano Zampini 788b96c3477SStefano Zampini /* allocate space for schur complements */ 789b96c3477SStefano Zampini ierr = PetscCalloc2(sub_schurs->n_subs,&sub_schurs->S_Ej, 790b96c3477SStefano Zampini sub_schurs->n_subs,&sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 791b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 792b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 79308122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 794b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 795b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 796b96c3477SStefano Zampini 797b1b3d7a2SStefano Zampini /* free workspace */ 798b1b3d7a2SStefano Zampini ierr = PetscFree(auxmapping);CHKERRQ(ierr); 799b1b3d7a2SStefano Zampini ierr = PetscFree2(auxlocal_sequential,auxlocal_parallel);CHKERRQ(ierr); 800b1b3d7a2SStefano Zampini 801b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 802b1b3d7a2SStefano Zampini } 803b1b3d7a2SStefano Zampini 804b1b3d7a2SStefano Zampini #undef __FUNCT__ 80534a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 80634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 80734a97f8cSStefano Zampini { 80834a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 80934a97f8cSStefano Zampini PetscErrorCode ierr; 81034a97f8cSStefano Zampini 81134a97f8cSStefano Zampini PetscFunctionBegin; 81234a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 8135ff63025SStefano Zampini schurs_ctx->n_subs = 0; 81434a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 81534a97f8cSStefano Zampini PetscFunctionReturn(0); 81634a97f8cSStefano Zampini } 81734a97f8cSStefano Zampini 81834a97f8cSStefano Zampini #undef __FUNCT__ 81934a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 82034a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 82134a97f8cSStefano Zampini { 82234a97f8cSStefano Zampini PetscErrorCode ierr; 82334a97f8cSStefano Zampini 82434a97f8cSStefano Zampini PetscFunctionBegin; 82534a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 82634a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 82734a97f8cSStefano Zampini PetscFunctionReturn(0); 82834a97f8cSStefano Zampini } 82934a97f8cSStefano Zampini 83034a97f8cSStefano Zampini #undef __FUNCT__ 83134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 83234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 83334a97f8cSStefano Zampini { 83434a97f8cSStefano Zampini PetscInt i; 83534a97f8cSStefano Zampini PetscErrorCode ierr; 83634a97f8cSStefano Zampini 83734a97f8cSStefano Zampini PetscFunctionBegin; 8381e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 839b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 840b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 841b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 8425db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 8435db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 84441c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 84541c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 84608122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 847a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 8485db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 8499bb4a8caSStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_com);CHKERRQ(ierr); 85008122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 85108122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 85234a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 853b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 85434a97f8cSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej[i]);CHKERRQ(ierr); 8552b973097SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_tilda[i]);CHKERRQ(ierr); 85634a97f8cSStefano Zampini } 85768270318SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I_layer);CHKERRQ(ierr); 8585ff63025SStefano Zampini if (sub_schurs->n_subs) { 859b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 860b96c3477SStefano Zampini ierr = PetscFree2(sub_schurs->S_Ej,sub_schurs->S_Ej_tilda);CHKERRQ(ierr); 8613dc780c3SStefano Zampini } 862b1b3d7a2SStefano Zampini ierr = PetscFree2(sub_schurs->index_sequential,sub_schurs->index_parallel);CHKERRQ(ierr); 863b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_sequential);CHKERRQ(ierr); 864b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->auxglobal_parallel);CHKERRQ(ierr); 86534a97f8cSStefano Zampini sub_schurs->n_subs = 0; 86634a97f8cSStefano Zampini PetscFunctionReturn(0); 86734a97f8cSStefano Zampini } 86834a97f8cSStefano Zampini 86934a97f8cSStefano Zampini #undef __FUNCT__ 87034a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 8712a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 87234a97f8cSStefano Zampini { 87334a97f8cSStefano Zampini PetscInt i,j,n; 87434a97f8cSStefano Zampini PetscErrorCode ierr; 87534a97f8cSStefano Zampini 87634a97f8cSStefano Zampini PetscFunctionBegin; 87734a97f8cSStefano Zampini n = 0; 87834a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 87934a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 88034a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 88134a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 88234a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 88334a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 88434a97f8cSStefano Zampini queue_tip[n] = dof; 88534a97f8cSStefano Zampini n++; 88634a97f8cSStefano Zampini } 88734a97f8cSStefano Zampini } 88834a97f8cSStefano Zampini } 88934a97f8cSStefano Zampini *n_added = n; 89034a97f8cSStefano Zampini PetscFunctionReturn(0); 89134a97f8cSStefano Zampini } 892