1674ae819SStefano Zampini #include "bddc.h" 2674ae819SStefano Zampini #include "bddcprivate.h" 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5674ae819SStefano Zampini #undef __FUNCT__ 6674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 7674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 8674ae819SStefano Zampini { 9674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 10674ae819SStefano Zampini PetscInt i; 11674ae819SStefano Zampini PetscErrorCode ierr; 12674ae819SStefano Zampini 13674ae819SStefano Zampini PetscFunctionBegin; 14674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 15674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 16674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 17674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 18674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 19674ae819SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 20674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 21674ae819SStefano Zampini } 22674ae819SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 23674ae819SStefano Zampini PetscFunctionReturn(0); 24674ae819SStefano Zampini } 25674ae819SStefano Zampini 26674ae819SStefano Zampini #undef __FUNCT__ 27674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 28674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 29674ae819SStefano Zampini { 30674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 31674ae819SStefano Zampini PetscErrorCode ierr; 32674ae819SStefano Zampini 33674ae819SStefano Zampini PetscFunctionBegin; 34674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 35674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 36674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 37674ae819SStefano Zampini PetscFunctionReturn(0); 38674ae819SStefano Zampini } 39674ae819SStefano Zampini 40674ae819SStefano Zampini #undef __FUNCT__ 41674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 42674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 43674ae819SStefano Zampini { 44674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 45674ae819SStefano Zampini PetscErrorCode ierr; 46674ae819SStefano Zampini 47674ae819SStefano Zampini PetscFunctionBegin; 48674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->temp_solution);CHKERRQ(ierr); 49674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->original_rhs);CHKERRQ(ierr); 50674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 51674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 52674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_rhs);CHKERRQ(ierr); 53674ae819SStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 54674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_mat);CHKERRQ(ierr); 55674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 56674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 5715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 5815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 59674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 60674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 61674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 62674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 63674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 64674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 65674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec4_D);CHKERRQ(ierr); 66674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 67674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 68674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 69674ae819SStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 70674ae819SStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 71674ae819SStefano Zampini ierr = PetscFree(pcbddc->local_primal_indices);CHKERRQ(ierr); 72674ae819SStefano Zampini ierr = PetscFree(pcbddc->replicated_local_primal_indices);CHKERRQ(ierr); 73674ae819SStefano Zampini ierr = PetscFree(pcbddc->replicated_local_primal_values);CHKERRQ(ierr); 74674ae819SStefano Zampini ierr = PetscFree(pcbddc->local_primal_displacements);CHKERRQ(ierr); 75674ae819SStefano Zampini ierr = PetscFree(pcbddc->local_primal_sizes);CHKERRQ(ierr); 76674ae819SStefano Zampini PetscFunctionReturn(0); 77674ae819SStefano Zampini } 78674ae819SStefano Zampini 79674ae819SStefano Zampini #undef __FUNCT__ 80674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSolveSaddlePoint" 81674ae819SStefano Zampini static PetscErrorCode PCBDDCSolveSaddlePoint(PC pc) 82674ae819SStefano Zampini { 83674ae819SStefano Zampini PetscErrorCode ierr; 84674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 85674ae819SStefano Zampini 86674ae819SStefano Zampini PetscFunctionBegin; 87674ae819SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 88674ae819SStefano Zampini if (pcbddc->local_auxmat1) { 89674ae819SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,pcbddc->vec2_R,pcbddc->vec1_C);CHKERRQ(ierr); 90674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 91674ae819SStefano Zampini } 92674ae819SStefano Zampini PetscFunctionReturn(0); 93674ae819SStefano Zampini } 94674ae819SStefano Zampini 95674ae819SStefano Zampini #undef __FUNCT__ 96674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 97674ae819SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc) 98674ae819SStefano Zampini { 99674ae819SStefano Zampini PetscErrorCode ierr; 100674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 101674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 102674ae819SStefano Zampini const PetscScalar zero = 0.0; 103674ae819SStefano Zampini 104674ae819SStefano Zampini PetscFunctionBegin; 10515aaf578SStefano Zampini /* Application of PHI^T (or PSI^T) */ 10615aaf578SStefano Zampini if (pcbddc->coarse_psi_B) { 10715aaf578SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 10815aaf578SStefano Zampini if (pcbddc->inexact_prec_type) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 10915aaf578SStefano Zampini } else { 110674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 111674ae819SStefano Zampini if (pcbddc->inexact_prec_type) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 11215aaf578SStefano Zampini } 113674ae819SStefano Zampini /* Scatter data of coarse_rhs */ 114674ae819SStefano Zampini if (pcbddc->coarse_rhs) { ierr = VecSet(pcbddc->coarse_rhs,zero);CHKERRQ(ierr); } 115674ae819SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,pcbddc->vec1_P,pcbddc->coarse_rhs,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116674ae819SStefano Zampini 117674ae819SStefano Zampini /* Local solution on R nodes */ 118674ae819SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 119674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 120674ae819SStefano Zampini ierr = VecScatterEnd (pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 121674ae819SStefano Zampini if (pcbddc->inexact_prec_type) { 122674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 123674ae819SStefano Zampini ierr = VecScatterEnd (pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 124674ae819SStefano Zampini } 125674ae819SStefano Zampini ierr = PCBDDCSolveSaddlePoint(pc);CHKERRQ(ierr); 126674ae819SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 127674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 128674ae819SStefano Zampini ierr = VecScatterEnd (pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 129674ae819SStefano Zampini if (pcbddc->inexact_prec_type) { 130674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 131674ae819SStefano Zampini ierr = VecScatterEnd (pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 132674ae819SStefano Zampini } 133674ae819SStefano Zampini 134674ae819SStefano Zampini /* Coarse solution */ 135674ae819SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,pcbddc->vec1_P,pcbddc->coarse_rhs,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 136674ae819SStefano Zampini if (pcbddc->coarse_rhs) { /* TODO remove null space when doing multilevel */ 137674ae819SStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,pcbddc->coarse_rhs,pcbddc->coarse_vec);CHKERRQ(ierr); 138674ae819SStefano Zampini } 139674ae819SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 140674ae819SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd (pc,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 141674ae819SStefano Zampini 142674ae819SStefano Zampini /* Sum contributions from two levels */ 143674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 144674ae819SStefano Zampini if (pcbddc->inexact_prec_type) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 145674ae819SStefano Zampini PetscFunctionReturn(0); 146674ae819SStefano Zampini } 147674ae819SStefano Zampini 148674ae819SStefano Zampini #undef __FUNCT__ 149674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 150674ae819SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,Vec vec_from, Vec vec_to, InsertMode imode, ScatterMode smode) 151674ae819SStefano Zampini { 152674ae819SStefano Zampini PetscErrorCode ierr; 153674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 154674ae819SStefano Zampini 155674ae819SStefano Zampini PetscFunctionBegin; 156674ae819SStefano Zampini switch (pcbddc->coarse_communications_type) { 157674ae819SStefano Zampini case SCATTERS_BDDC: 158674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,vec_from,vec_to,imode,smode);CHKERRQ(ierr); 159674ae819SStefano Zampini break; 160674ae819SStefano Zampini case GATHERS_BDDC: 161674ae819SStefano Zampini break; 162674ae819SStefano Zampini } 163674ae819SStefano Zampini PetscFunctionReturn(0); 164674ae819SStefano Zampini } 165674ae819SStefano Zampini 166674ae819SStefano Zampini #undef __FUNCT__ 167674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 168674ae819SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc,Vec vec_from, Vec vec_to, InsertMode imode, ScatterMode smode) 169674ae819SStefano Zampini { 170674ae819SStefano Zampini PetscErrorCode ierr; 171674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 172674ae819SStefano Zampini PetscScalar* array_to; 173674ae819SStefano Zampini PetscScalar* array_from; 174674ae819SStefano Zampini MPI_Comm comm; 175674ae819SStefano Zampini PetscInt i; 176674ae819SStefano Zampini 177674ae819SStefano Zampini PetscFunctionBegin; 178674ae819SStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 179674ae819SStefano Zampini switch (pcbddc->coarse_communications_type) { 180674ae819SStefano Zampini case SCATTERS_BDDC: 181674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,vec_from,vec_to,imode,smode);CHKERRQ(ierr); 182674ae819SStefano Zampini break; 183674ae819SStefano Zampini case GATHERS_BDDC: 184674ae819SStefano Zampini if (vec_from) { 185674ae819SStefano Zampini ierr = VecGetArray(vec_from,&array_from);CHKERRQ(ierr); 186674ae819SStefano Zampini } 187674ae819SStefano Zampini if (vec_to) { 188674ae819SStefano Zampini ierr = VecGetArray(vec_to,&array_to);CHKERRQ(ierr); 189674ae819SStefano Zampini } 190674ae819SStefano Zampini switch(pcbddc->coarse_problem_type){ 191674ae819SStefano Zampini case SEQUENTIAL_BDDC: 192674ae819SStefano Zampini if (smode == SCATTER_FORWARD) { 193674ae819SStefano Zampini ierr = MPI_Gatherv(&array_from[0],pcbddc->local_primal_size,MPIU_SCALAR,&pcbddc->replicated_local_primal_values[0],pcbddc->local_primal_sizes,pcbddc->local_primal_displacements,MPIU_SCALAR,0,comm);CHKERRQ(ierr); 194674ae819SStefano Zampini if (vec_to) { 195674ae819SStefano Zampini if (imode == ADD_VALUES) { 196674ae819SStefano Zampini for (i=0;i<pcbddc->replicated_primal_size;i++) { 197674ae819SStefano Zampini array_to[pcbddc->replicated_local_primal_indices[i]]+=pcbddc->replicated_local_primal_values[i]; 198674ae819SStefano Zampini } 199674ae819SStefano Zampini } else { 200674ae819SStefano Zampini for (i=0;i<pcbddc->replicated_primal_size;i++) { 201674ae819SStefano Zampini array_to[pcbddc->replicated_local_primal_indices[i]]=pcbddc->replicated_local_primal_values[i]; 202674ae819SStefano Zampini } 203674ae819SStefano Zampini } 204674ae819SStefano Zampini } 205674ae819SStefano Zampini } else { 206674ae819SStefano Zampini if (vec_from) { 207674ae819SStefano Zampini if (imode == ADD_VALUES) { 208674ae819SStefano Zampini MPI_Comm vec_from_comm; 209674ae819SStefano Zampini ierr = PetscObjectGetComm((PetscObject)(vec_from),&vec_from_comm);CHKERRQ(ierr); 210674ae819SStefano Zampini SETERRQ2(vec_from_comm,PETSC_ERR_SUP,"Unsupported insert mode ADD_VALUES for SCATTER_REVERSE in %s for case %d\n",__FUNCT__,pcbddc->coarse_problem_type); 211674ae819SStefano Zampini } 212674ae819SStefano Zampini for (i=0;i<pcbddc->replicated_primal_size;i++) { 213674ae819SStefano Zampini pcbddc->replicated_local_primal_values[i]=array_from[pcbddc->replicated_local_primal_indices[i]]; 214674ae819SStefano Zampini } 215674ae819SStefano Zampini } 216674ae819SStefano Zampini ierr = MPI_Scatterv(&pcbddc->replicated_local_primal_values[0],pcbddc->local_primal_sizes,pcbddc->local_primal_displacements,MPIU_SCALAR,&array_to[0],pcbddc->local_primal_size,MPIU_SCALAR,0,comm);CHKERRQ(ierr); 217674ae819SStefano Zampini } 218674ae819SStefano Zampini break; 219674ae819SStefano Zampini case REPLICATED_BDDC: 220674ae819SStefano Zampini if (smode == SCATTER_FORWARD) { 221674ae819SStefano Zampini ierr = MPI_Allgatherv(&array_from[0],pcbddc->local_primal_size,MPIU_SCALAR,&pcbddc->replicated_local_primal_values[0],pcbddc->local_primal_sizes,pcbddc->local_primal_displacements,MPIU_SCALAR,comm);CHKERRQ(ierr); 222674ae819SStefano Zampini if (imode == ADD_VALUES) { 223674ae819SStefano Zampini for (i=0;i<pcbddc->replicated_primal_size;i++) { 224674ae819SStefano Zampini array_to[pcbddc->replicated_local_primal_indices[i]]+=pcbddc->replicated_local_primal_values[i]; 225674ae819SStefano Zampini } 226674ae819SStefano Zampini } else { 227674ae819SStefano Zampini for (i=0;i<pcbddc->replicated_primal_size;i++) { 228674ae819SStefano Zampini array_to[pcbddc->replicated_local_primal_indices[i]]=pcbddc->replicated_local_primal_values[i]; 229674ae819SStefano Zampini } 230674ae819SStefano Zampini } 231674ae819SStefano Zampini } else { /* no communications needed for SCATTER_REVERSE since needed data is already present */ 232674ae819SStefano Zampini if (imode == ADD_VALUES) { 233674ae819SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 234674ae819SStefano Zampini array_to[i]+=array_from[pcbddc->local_primal_indices[i]]; 235674ae819SStefano Zampini } 236674ae819SStefano Zampini } else { 237674ae819SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 238674ae819SStefano Zampini array_to[i]=array_from[pcbddc->local_primal_indices[i]]; 239674ae819SStefano Zampini } 240674ae819SStefano Zampini } 241674ae819SStefano Zampini } 242674ae819SStefano Zampini break; 243674ae819SStefano Zampini case MULTILEVEL_BDDC: 244674ae819SStefano Zampini break; 245674ae819SStefano Zampini case PARALLEL_BDDC: 246674ae819SStefano Zampini break; 247674ae819SStefano Zampini } 248674ae819SStefano Zampini if (vec_from) { 249674ae819SStefano Zampini ierr = VecRestoreArray(vec_from,&array_from);CHKERRQ(ierr); 250674ae819SStefano Zampini } 251674ae819SStefano Zampini if (vec_to) { 252674ae819SStefano Zampini ierr = VecRestoreArray(vec_to,&array_to);CHKERRQ(ierr); 253674ae819SStefano Zampini } 254674ae819SStefano Zampini break; 255674ae819SStefano Zampini } 256674ae819SStefano Zampini PetscFunctionReturn(0); 257674ae819SStefano Zampini } 258674ae819SStefano Zampini 259984c4197SStefano Zampini /* uncomment for testing purposes */ 260984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 261674ae819SStefano Zampini #undef __FUNCT__ 262674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 263674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 264674ae819SStefano Zampini { 265674ae819SStefano Zampini PetscErrorCode ierr; 266674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 267674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 268674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 269984c4197SStefano Zampini /* constraint and (optionally) change of basis matrix implemented as SeqAIJ */ 270674ae819SStefano Zampini MatType impMatType=MATSEQAIJ; 271984c4197SStefano Zampini /* one and zero */ 272984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 273984c4197SStefano Zampini /* space to store constraints and their local indices */ 274984c4197SStefano Zampini PetscScalar *temp_quadrature_constraint; 275984c4197SStefano Zampini PetscInt *temp_indices,*temp_indices_to_constraint,*temp_indices_to_constraint_B; 276984c4197SStefano Zampini /* iterators */ 277984c4197SStefano Zampini PetscInt i,j,k,total_counts,temp_start_ptr; 278984c4197SStefano Zampini /* stuff to store connected components stored in pcbddc->mat_graph */ 279984c4197SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges,*used_IS; 280984c4197SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges; 281984c4197SStefano Zampini PetscBool get_faces,get_edges,get_vertices; 282984c4197SStefano Zampini /* near null space stuff */ 283674ae819SStefano Zampini MatNullSpace nearnullsp; 284674ae819SStefano Zampini const Vec *nearnullvecs; 285674ae819SStefano Zampini Vec *localnearnullsp; 286984c4197SStefano Zampini PetscBool nnsp_has_cnst; 287984c4197SStefano Zampini PetscInt nnsp_size; 288984c4197SStefano Zampini PetscScalar *array; 289984c4197SStefano Zampini /* BLAS integers */ 290984c4197SStefano Zampini PetscBLASInt Bs,Bt,lwork,lierr,Bone=1; 291c4303822SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K; 292c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 293984c4197SStefano Zampini /* LAPACK working arrays for SVD or POD */ 294242a89d7SStefano Zampini PetscBool skip_lapack; 295984c4197SStefano Zampini PetscScalar *work; 296984c4197SStefano Zampini PetscReal *singular_vals; 297984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 298984c4197SStefano Zampini PetscReal *rwork; 299674ae819SStefano Zampini #endif 300984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 301610854a6SStefano Zampini PetscBLASInt Bone_2=1; 302984c4197SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 303b7d8b9f8SStefano Zampini #else 304b7d8b9f8SStefano Zampini PetscBLASInt dummy_int_1=1,dummy_int_2=1; 305b7d8b9f8SStefano Zampini PetscScalar dummy_scalar_1=0.0,dummy_scalar_2=0.0; 306984c4197SStefano Zampini #endif 307984c4197SStefano Zampini /* change of basis */ 308984c4197SStefano Zampini PetscInt *aux_primal_numbering,*aux_primal_minloc,*global_indices; 309984c4197SStefano Zampini PetscBool boolforchange,*change_basis,*touched; 310984c4197SStefano Zampini /* auxiliary stuff */ 311984c4197SStefano Zampini PetscInt *nnz,*is_indices,*local_to_B; 312984c4197SStefano Zampini /* some quantities */ 313984c4197SStefano Zampini PetscInt n_vertices,total_primal_vertices; 314984c4197SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint,max_constraints,temp_constraints; 315984c4197SStefano Zampini 316674ae819SStefano Zampini 317674ae819SStefano Zampini PetscFunctionBegin; 318674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 319674ae819SStefano Zampini get_faces = PETSC_TRUE; 320674ae819SStefano Zampini get_edges = PETSC_TRUE; 321674ae819SStefano Zampini get_vertices = PETSC_TRUE; 322674ae819SStefano Zampini if (pcbddc->vertices_flag) { 323674ae819SStefano Zampini get_faces = PETSC_FALSE; 324674ae819SStefano Zampini get_edges = PETSC_FALSE; 325674ae819SStefano Zampini } 326674ae819SStefano Zampini if (pcbddc->constraints_flag) { 327674ae819SStefano Zampini get_vertices = PETSC_FALSE; 328674ae819SStefano Zampini } 329674ae819SStefano Zampini if (pcbddc->faces_flag) { 330674ae819SStefano Zampini get_edges = PETSC_FALSE; 331674ae819SStefano Zampini } 332674ae819SStefano Zampini if (pcbddc->edges_flag) { 333674ae819SStefano Zampini get_faces = PETSC_FALSE; 334674ae819SStefano Zampini } 335674ae819SStefano Zampini /* default */ 336674ae819SStefano Zampini if (!get_faces && !get_edges && !get_vertices) { 337674ae819SStefano Zampini get_vertices = PETSC_TRUE; 338674ae819SStefano Zampini } 339674ae819SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,get_faces,get_edges,get_vertices,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices); 340984c4197SStefano Zampini /* print some info */ 341674ae819SStefano Zampini if (pcbddc->dbg_flag) { 342674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 343674ae819SStefano Zampini i = 0; 344674ae819SStefano Zampini if (ISForVertices) { 345674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&i);CHKERRQ(ierr); 346674ae819SStefano Zampini } 347674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices\n",PetscGlobalRank,i);CHKERRQ(ierr); 348674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges\n",PetscGlobalRank,n_ISForEdges);CHKERRQ(ierr); 34915aaf578SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces\n",PetscGlobalRank,n_ISForFaces);CHKERRQ(ierr); 350674ae819SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 351674ae819SStefano Zampini } 352674ae819SStefano Zampini /* check if near null space is attached to global mat */ 353674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 354674ae819SStefano Zampini if (nearnullsp) { 355674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 356984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 357984c4197SStefano Zampini nnsp_size = 0; 358674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 359674ae819SStefano Zampini } 360984c4197SStefano Zampini /* get max number of constraints on a single cc */ 361984c4197SStefano Zampini max_constraints = nnsp_size; 362984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 363984c4197SStefano Zampini 364674ae819SStefano Zampini /* 365674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 366674ae819SStefano Zampini - temp_indices will contain start index of each constraint stored as follows 367674ae819SStefano Zampini - temp_indices_to_constraint [temp_indices[i],...,temp[indices[i+1]-1] will contain the indices (in local numbering) on which the constraint acts 368674ae819SStefano Zampini - temp_indices_to_constraint_B[temp_indices[i],...,temp[indices[i+1]-1] will contain the indices (in boundary numbering) on which the constraint acts 369674ae819SStefano Zampini - temp_quadrature_constraint [temp_indices[i],...,temp[indices[i+1]-1] will contain the scalars representing the constraint itself 370674ae819SStefano Zampini */ 371674ae819SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 372984c4197SStefano Zampini total_counts *= max_constraints; 373674ae819SStefano Zampini n_vertices = 0; 374674ae819SStefano Zampini if (ISForVertices) { 375674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 376674ae819SStefano Zampini } 377674ae819SStefano Zampini total_counts += n_vertices; 378674ae819SStefano Zampini ierr = PetscMalloc((total_counts+1)*sizeof(PetscInt),&temp_indices);CHKERRQ(ierr); 379674ae819SStefano Zampini ierr = PetscMalloc((total_counts+1)*sizeof(PetscBool),&change_basis);CHKERRQ(ierr); 380674ae819SStefano Zampini total_counts = 0; 381674ae819SStefano Zampini max_size_of_constraint = 0; 382674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 383674ae819SStefano Zampini if (i<n_ISForEdges) { 384674ae819SStefano Zampini used_IS = &ISForEdges[i]; 385674ae819SStefano Zampini } else { 386674ae819SStefano Zampini used_IS = &ISForFaces[i-n_ISForEdges]; 387674ae819SStefano Zampini } 388674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&j);CHKERRQ(ierr); 389674ae819SStefano Zampini total_counts += j; 390674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 391674ae819SStefano Zampini } 392984c4197SStefano Zampini total_counts *= max_constraints; 393674ae819SStefano Zampini total_counts += n_vertices; 394674ae819SStefano Zampini ierr = PetscMalloc(total_counts*sizeof(PetscScalar),&temp_quadrature_constraint);CHKERRQ(ierr); 395674ae819SStefano Zampini ierr = PetscMalloc(total_counts*sizeof(PetscInt),&temp_indices_to_constraint);CHKERRQ(ierr); 396674ae819SStefano Zampini ierr = PetscMalloc(total_counts*sizeof(PetscInt),&temp_indices_to_constraint_B);CHKERRQ(ierr); 397984c4197SStefano Zampini /* local to boundary numbering */ 398674ae819SStefano Zampini ierr = PetscMalloc(pcis->n*sizeof(PetscInt),&local_to_B);CHKERRQ(ierr); 399674ae819SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 400984c4197SStefano Zampini for (i=0;i<pcis->n;i++) local_to_B[i]=-1; 401984c4197SStefano Zampini for (i=0;i<pcis->n_B;i++) local_to_B[is_indices[i]]=i; 402674ae819SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 403984c4197SStefano Zampini /* get local part of global near null space vectors */ 404984c4197SStefano Zampini ierr = PetscMalloc(nnsp_size*sizeof(Vec),&localnearnullsp);CHKERRQ(ierr); 405984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 406984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 407984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 409984c4197SStefano Zampini } 410674ae819SStefano Zampini 411242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 412242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 413242a89d7SStefano Zampini if (n_ISForFaces+n_ISForEdges) skip_lapack = PETSC_FALSE; 414242a89d7SStefano Zampini 415984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 416242a89d7SStefano Zampini if (!pcbddc->use_nnsp_true && !skip_lapack) { 417674ae819SStefano Zampini PetscScalar temp_work; 418674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 419984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 420984c4197SStefano Zampini ierr = PetscMalloc(max_constraints*max_constraints*sizeof(PetscScalar),&correlation_mat);CHKERRQ(ierr); 421984c4197SStefano Zampini ierr = PetscMalloc(max_constraints*sizeof(PetscReal),&singular_vals);CHKERRQ(ierr); 422984c4197SStefano Zampini ierr = PetscMalloc(max_size_of_constraint*max_constraints*sizeof(PetscScalar),&temp_basis);CHKERRQ(ierr); 423674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 424984c4197SStefano Zampini ierr = PetscMalloc(3*max_constraints*sizeof(PetscReal),&rwork);CHKERRQ(ierr); 425674ae819SStefano Zampini #endif 426674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 427c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 428c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 429674ae819SStefano Zampini lwork = -1; 430674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 431674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 432c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 433674ae819SStefano Zampini #else 434c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 435674ae819SStefano Zampini #endif 436674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 437984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 438674ae819SStefano Zampini #else /* on missing GESVD */ 439674ae819SStefano Zampini /* SVD */ 440674ae819SStefano Zampini PetscInt max_n,min_n; 441674ae819SStefano Zampini max_n = max_size_of_constraint; 442984c4197SStefano Zampini min_n = max_constraints; 443984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 444674ae819SStefano Zampini min_n = max_size_of_constraint; 445984c4197SStefano Zampini max_n = max_constraints; 446674ae819SStefano Zampini } 447674ae819SStefano Zampini ierr = PetscMalloc(min_n*sizeof(PetscReal),&singular_vals);CHKERRQ(ierr); 448674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 449674ae819SStefano Zampini ierr = PetscMalloc(5*min_n*sizeof(PetscReal),&rwork);CHKERRQ(ierr); 450674ae819SStefano Zampini #endif 451674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 452674ae819SStefano Zampini lwork = -1; 453674ae819SStefano Zampini ierr = PetscBLASIntCast(max_n,&Bs);CHKERRQ(ierr); 454674ae819SStefano Zampini ierr = PetscBLASIntCast(min_n,&Bt);CHKERRQ(ierr); 455b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 456674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 457674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 458b7d8b9f8SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Bs,&Bt,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar_1,&dummy_int_1,&dummy_scalar_2,&dummy_int_2,&temp_work,&lwork,&lierr)); 459674ae819SStefano Zampini #else 460b7d8b9f8SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Bs,&Bt,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar_1,&dummy_int_1,&dummy_scalar_2,&dummy_int_2,&temp_work,&lwork,rwork,&lierr)); 461674ae819SStefano Zampini #endif 462674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 463984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 464984c4197SStefano Zampini #endif /* on missing GESVD */ 465674ae819SStefano Zampini /* Allocate optimal workspace */ 466674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 467984c4197SStefano Zampini ierr = PetscMalloc((PetscInt)lwork*sizeof(PetscScalar),&work);CHKERRQ(ierr); 468674ae819SStefano Zampini } 469674ae819SStefano Zampini /* Now we can loop on constraining sets */ 470674ae819SStefano Zampini total_counts = 0; 471674ae819SStefano Zampini temp_indices[0] = 0; 472674ae819SStefano Zampini /* vertices */ 473674ae819SStefano Zampini if (ISForVertices) { 474674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 475674ae819SStefano Zampini if (nnsp_has_cnst) { /* consider all vertices */ 476674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 477674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i]; 478674ae819SStefano Zampini temp_indices_to_constraint_B[temp_indices[total_counts]]=local_to_B[is_indices[i]]; 479674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 480674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 481674ae819SStefano Zampini change_basis[total_counts]=PETSC_FALSE; 482674ae819SStefano Zampini total_counts++; 483674ae819SStefano Zampini } 484674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 485984c4197SStefano Zampini PetscBool used_vertex; 486674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 487674ae819SStefano Zampini used_vertex = PETSC_FALSE; 488674ae819SStefano Zampini k = 0; 489674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 490984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 491984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 492674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i]; 493674ae819SStefano Zampini temp_indices_to_constraint_B[temp_indices[total_counts]]=local_to_B[is_indices[i]]; 494674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 495674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 496674ae819SStefano Zampini change_basis[total_counts]=PETSC_FALSE; 497674ae819SStefano Zampini total_counts++; 498674ae819SStefano Zampini used_vertex = PETSC_TRUE; 499674ae819SStefano Zampini } 500984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 501674ae819SStefano Zampini k++; 502674ae819SStefano Zampini } 503674ae819SStefano Zampini } 504674ae819SStefano Zampini } 505674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 506674ae819SStefano Zampini n_vertices = total_counts; 507674ae819SStefano Zampini } 508984c4197SStefano Zampini 509674ae819SStefano Zampini /* edges and faces */ 510674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 511674ae819SStefano Zampini if (i<n_ISForEdges) { 512674ae819SStefano Zampini used_IS = &ISForEdges[i]; 513984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 514674ae819SStefano Zampini } else { 515674ae819SStefano Zampini used_IS = &ISForFaces[i-n_ISForEdges]; 516984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 517674ae819SStefano Zampini } 518674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 519674ae819SStefano Zampini temp_start_ptr = total_counts; /* need to know the starting index of constraints stored */ 520674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&size_of_constraint);CHKERRQ(ierr); 521674ae819SStefano Zampini ierr = ISGetIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 522984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 523984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 524674ae819SStefano Zampini if (nnsp_has_cnst) { 5255b08dc53SStefano Zampini PetscScalar quad_value; 526674ae819SStefano Zampini temp_constraints++; 527674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 528674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 529674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]+j]=is_indices[j]; 530674ae819SStefano Zampini temp_indices_to_constraint_B[temp_indices[total_counts]+j]=local_to_B[is_indices[j]]; 531674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=quad_value; 532674ae819SStefano Zampini } 533674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 53451b0f6cfSStefano Zampini change_basis[total_counts]=boolforchange; 535674ae819SStefano Zampini total_counts++; 536674ae819SStefano Zampini } 537674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 538984c4197SStefano Zampini PetscReal real_value; 539984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 540674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 541674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]+j]=is_indices[j]; 542674ae819SStefano Zampini temp_indices_to_constraint_B[temp_indices[total_counts]+j]=local_to_B[is_indices[j]]; 543984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=array[is_indices[j]]; 544674ae819SStefano Zampini } 545984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 546984c4197SStefano Zampini /* check if array is null on the connected component */ 547674ae819SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Bs);CHKERRQ(ierr); 5485b08dc53SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Bs,&temp_quadrature_constraint[temp_indices[total_counts]],&Bone)); 5495b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 550674ae819SStefano Zampini temp_constraints++; 551674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 55251b0f6cfSStefano Zampini change_basis[total_counts]=boolforchange; 553674ae819SStefano Zampini total_counts++; 554674ae819SStefano Zampini } 555674ae819SStefano Zampini } 556674ae819SStefano Zampini ierr = ISRestoreIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 557984c4197SStefano Zampini /* perform SVD on the constraints if use_nnsp_true has not be requested by the user */ 558984c4197SStefano Zampini if (!pcbddc->use_nnsp_true) { 559984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 560984c4197SStefano Zampini 561674ae819SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Bs);CHKERRQ(ierr); 562674ae819SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Bt);CHKERRQ(ierr); 563674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 564984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 565984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 566984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 567984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 568984c4197SStefano Zampini from that computed using LAPACKgesvd 569984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 570984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 571c8244a33SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 572c8244a33SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 573984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 574674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 575984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 576674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 577674ae819SStefano Zampini for (k=0;k<j+1;k++) { 578610854a6SStefano Zampini PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k]=BLASdot_(&Bs,&temp_quadrature_constraint[temp_indices[temp_start_ptr+k]],&Bone,&temp_quadrature_constraint[temp_indices[temp_start_ptr+j]],&Bone_2)); 579674ae819SStefano Zampini } 580674ae819SStefano Zampini } 581674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 582c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 583674ae819SStefano Zampini #else 584c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 585674ae819SStefano Zampini #endif 586674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 587984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 588984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 589674ae819SStefano Zampini j=0; 590984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 591674ae819SStefano Zampini total_counts=total_counts-j; 592c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 593c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 594c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 595c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 596c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 597c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 598674ae819SStefano Zampini if (j<temp_constraints) { 599984c4197SStefano Zampini PetscInt ii; 600984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k]=1.0/PetscSqrtReal(singular_vals[k]); 601674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 602c4303822SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC)); 603674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 604984c4197SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 605984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 606674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 607984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[temp_start_ptr+k]+ii]=singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii]; 608674ae819SStefano Zampini } 609674ae819SStefano Zampini } 610674ae819SStefano Zampini } 611674ae819SStefano Zampini #else /* on missing GESVD */ 612b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 613674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 614674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 615b7d8b9f8SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Bs,&Bt,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar_1,&dummy_int_1,&dummy_scalar_2,&dummy_int_2,work,&lwork,&lierr)); 616674ae819SStefano Zampini #else 617b7d8b9f8SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Bs,&Bt,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar_1,&dummy_int_1,&dummy_scalar_2,&dummy_int_2,work,&lwork,rwork,&lierr)); 618674ae819SStefano Zampini #endif 619984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 620674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 621984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 622b7d8b9f8SStefano Zampini PetscInt min_n = temp_constraints; 623b7d8b9f8SStefano Zampini if (min_n > size_of_constraint) min_n = size_of_constraint; 624674ae819SStefano Zampini j = 0; 625674ae819SStefano Zampini while (j < min_n && singular_vals[min_n-j-1] < tol) j++; 626984c4197SStefano Zampini total_counts = total_counts-temp_constraints+min_n-j; 627984c4197SStefano Zampini #endif /* on missing GESVD */ 628674ae819SStefano Zampini } 629674ae819SStefano Zampini } 630674ae819SStefano Zampini /* free index sets of faces, edges and vertices */ 631674ae819SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 632674ae819SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 633674ae819SStefano Zampini } 634674ae819SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 635674ae819SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 636674ae819SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 637674ae819SStefano Zampini } 638674ae819SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 639674ae819SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 640674ae819SStefano Zampini 641984c4197SStefano Zampini /* free workspace */ 642242a89d7SStefano Zampini if (!pcbddc->use_nnsp_true && !skip_lapack) { 643984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 644984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 645984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 646984c4197SStefano Zampini #endif 647984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 648984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 649984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 650984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 651984c4197SStefano Zampini #endif 652984c4197SStefano Zampini } 653984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 654984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 655984c4197SStefano Zampini } 656984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 657984c4197SStefano Zampini 658674ae819SStefano Zampini /* set quantities in pcbddc data structure */ 659984c4197SStefano Zampini /* n_vertices defines the number of subdomain corners in the primal space */ 660674ae819SStefano Zampini /* n_constraints defines the number of averages (they can be point primal dofs if change of basis is requested) */ 661984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 662674ae819SStefano Zampini pcbddc->n_vertices = n_vertices; 663984c4197SStefano Zampini pcbddc->n_constraints = pcbddc->local_primal_size-pcbddc->n_vertices; 664674ae819SStefano Zampini 665674ae819SStefano Zampini /* Create constraint matrix */ 666674ae819SStefano Zampini /* The constraint matrix is used to compute the l2g map of primal dofs */ 667674ae819SStefano Zampini /* so we need to set it up properly either with or without change of basis */ 668674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 669674ae819SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,impMatType);CHKERRQ(ierr); 670984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 671984c4197SStefano Zampini /* array to compute a local numbering of constraints : vertices first then constraints */ 672984c4197SStefano Zampini ierr = PetscMalloc(pcbddc->local_primal_size*sizeof(PetscInt),&aux_primal_numbering);CHKERRQ(ierr); 673984c4197SStefano Zampini /* array to select the proper local node (of minimum index with respect to global ordering) when changing the basis */ 674984c4197SStefano Zampini /* note: it should not be needed since IS for faces and edges are already sorted by global ordering when analyzing the graph but... just in case */ 675984c4197SStefano Zampini ierr = PetscMalloc(pcbddc->local_primal_size*sizeof(PetscInt),&aux_primal_minloc);CHKERRQ(ierr); 676984c4197SStefano Zampini /* auxiliary stuff for basis change */ 677984c4197SStefano Zampini ierr = PetscMalloc(max_size_of_constraint*sizeof(PetscInt),&global_indices);CHKERRQ(ierr); 678984c4197SStefano Zampini ierr = PetscMalloc(pcis->n_B*sizeof(PetscBool),&touched);CHKERRQ(ierr); 679984c4197SStefano Zampini ierr = PetscMemzero(touched,pcis->n_B*sizeof(PetscBool));CHKERRQ(ierr); 680984c4197SStefano Zampini 681984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 682984c4197SStefano Zampini total_primal_vertices=0; 683984c4197SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 684674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 685984c4197SStefano Zampini if (size_of_constraint == 1) { 686984c4197SStefano Zampini touched[temp_indices_to_constraint_B[temp_indices[i]]]=PETSC_TRUE; 687984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]]; 688984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=0; 689984c4197SStefano Zampini total_primal_vertices++; 690984c4197SStefano Zampini } else if (change_basis[i]) { /* Same procedure used in PCBDDCGetPrimalConstraintsLocalIdx */ 691984c4197SStefano Zampini PetscInt min_loc,min_index; 692984c4197SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],global_indices);CHKERRQ(ierr); 693984c4197SStefano Zampini /* find first untouched local node */ 694674ae819SStefano Zampini k = 0; 695984c4197SStefano Zampini while (touched[temp_indices_to_constraint_B[temp_indices[i]+k]]) k++; 696984c4197SStefano Zampini min_index = global_indices[k]; 697984c4197SStefano Zampini min_loc = k; 698984c4197SStefano Zampini /* search the minimum among global nodes already untouched on the cc */ 699984c4197SStefano Zampini for (k=1;k<size_of_constraint;k++) { 700984c4197SStefano Zampini /* there can be more than one constraint on a single connected component */ 701984c4197SStefano Zampini if (min_index > global_indices[k] && !touched[temp_indices_to_constraint_B[temp_indices[i]+k]]) { 702984c4197SStefano Zampini min_index = global_indices[k]; 703984c4197SStefano Zampini min_loc = k; 704674ae819SStefano Zampini } 705674ae819SStefano Zampini } 706984c4197SStefano Zampini touched[temp_indices_to_constraint_B[temp_indices[i]+min_loc]] = PETSC_TRUE; 707984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]+min_loc]; 708984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=min_loc; 709984c4197SStefano Zampini total_primal_vertices++; 710984c4197SStefano Zampini } 711984c4197SStefano Zampini } 712984c4197SStefano Zampini /* free workspace */ 713984c4197SStefano Zampini ierr = PetscFree(global_indices);CHKERRQ(ierr); 714984c4197SStefano Zampini ierr = PetscFree(touched);CHKERRQ(ierr); 715674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 716984c4197SStefano Zampini ierr = PetscSortInt(total_primal_vertices,aux_primal_numbering); 717984c4197SStefano Zampini 718984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 719984c4197SStefano Zampini ierr = PetscMalloc(pcbddc->local_primal_size*sizeof(PetscInt),&nnz);CHKERRQ(ierr); 720984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i]=1; 721984c4197SStefano Zampini j=total_primal_vertices; 722984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 723674ae819SStefano Zampini if (!change_basis[i]) { 724674ae819SStefano Zampini nnz[j]=temp_indices[i+1]-temp_indices[i]; 725674ae819SStefano Zampini j++; 726674ae819SStefano Zampini } 727674ae819SStefano Zampini } 728674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 729674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 730674ae819SStefano Zampini /* set values in constraint matrix */ 731984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 732984c4197SStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,aux_primal_numbering[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 733674ae819SStefano Zampini } 734984c4197SStefano Zampini total_counts = total_primal_vertices; 735984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 736674ae819SStefano Zampini if (!change_basis[i]) { 737674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 738674ae819SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&total_counts,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],&temp_quadrature_constraint[temp_indices[i]],INSERT_VALUES);CHKERRQ(ierr); 739674ae819SStefano Zampini total_counts++; 740674ae819SStefano Zampini } 741674ae819SStefano Zampini } 742674ae819SStefano Zampini /* assembling */ 743674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 744674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 745984c4197SStefano Zampini /* 746984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 747984c4197SStefano Zampini */ 748674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 749674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 750984c4197SStefano Zampini PetscBool qr_needed = PETSC_FALSE; 751984c4197SStefano Zampini /* change of basis acts on local interfaces -> dimension is n_B x n_B */ 752674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 753674ae819SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,impMatType);CHKERRQ(ierr); 754674ae819SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,pcis->n_B,pcis->n_B,pcis->n_B,pcis->n_B);CHKERRQ(ierr); 755674ae819SStefano Zampini /* work arrays */ 756674ae819SStefano Zampini ierr = PetscMalloc(pcis->n_B*sizeof(PetscInt),&nnz);CHKERRQ(ierr); 757984c4197SStefano Zampini for (i=0;i<pcis->n_B;i++) nnz[i]=1; 758984c4197SStefano Zampini /* nonzeros per row */ 759984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 760674ae819SStefano Zampini if (change_basis[i]) { 761984c4197SStefano Zampini qr_needed = PETSC_TRUE; 762674ae819SStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 763984c4197SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = size_of_constraint; 764674ae819SStefano Zampini } 765674ae819SStefano Zampini } 766674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 767674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 768674ae819SStefano Zampini /* Set initial identity in the matrix */ 769674ae819SStefano Zampini for (i=0;i<pcis->n_B;i++) { 770674ae819SStefano Zampini ierr = MatSetValue(pcbddc->ChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 771674ae819SStefano Zampini } 772984c4197SStefano Zampini 773674ae819SStefano Zampini /* Now we loop on the constraints which need a change of basis */ 774674ae819SStefano Zampini /* Change of basis matrix is evaluated as the FIRST APPROACH in */ 775984c4197SStefano Zampini /* Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) */ 776984c4197SStefano Zampini /* Change of basis matrix T computed via QR decomposition of constraints */ 777984c4197SStefano Zampini if (qr_needed) { 778026de310SStefano Zampini /* dual and primal dofs on a single cc */ 779984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 780026de310SStefano Zampini /* iterator on aux_primal_minloc (ordered as read from nearnullspace: vertices, edges and then constraints) */ 781026de310SStefano Zampini PetscInt primal_counter; 782984c4197SStefano Zampini /* working stuff for GEQRF */ 783984c4197SStefano Zampini PetscScalar *qr_basis,*qr_tau,*qr_work,lqr_work_t; 784984c4197SStefano Zampini PetscBLASInt lqr_work; 785984c4197SStefano Zampini /* working stuff for UNGQR */ 786984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 787984c4197SStefano Zampini PetscBLASInt lgqr_work; 788984c4197SStefano Zampini /* working stuff for TRTRS */ 789984c4197SStefano Zampini PetscScalar *trs_rhs; 790*3f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 791984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 792984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 793984c4197SStefano Zampini PetscScalar *start_vals; 794984c4197SStefano Zampini /* working stuff for values insertion */ 795984c4197SStefano Zampini PetscBool *is_primal; 796984c4197SStefano Zampini 797984c4197SStefano Zampini /* space to store Q */ 798984c4197SStefano Zampini ierr = PetscMalloc((max_size_of_constraint)*(max_size_of_constraint)*sizeof(PetscScalar),&qr_basis);CHKERRQ(ierr); 799984c4197SStefano Zampini /* first we issue queries for optimal work */ 800*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 801*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 802*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 803984c4197SStefano Zampini lqr_work = -1; 804*3f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 805984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 806984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 807984c4197SStefano Zampini ierr = PetscMalloc((PetscInt)PetscRealPart(lqr_work_t)*sizeof(*qr_work),&qr_work);CHKERRQ(ierr); 808984c4197SStefano Zampini lgqr_work = -1; 809*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 810*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 811*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 812*3f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 813*3f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 814*3f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 815984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 816984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 817984c4197SStefano Zampini ierr = PetscMalloc((PetscInt)PetscRealPart(lgqr_work_t)*sizeof(*gqr_work),&gqr_work);CHKERRQ(ierr); 818984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 819984c4197SStefano Zampini ierr = PetscMalloc(max_constraints*sizeof(*qr_tau),&qr_tau);CHKERRQ(ierr); 820984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 821984c4197SStefano Zampini ierr = PetscMalloc(max_constraints*max_constraints*sizeof(*trs_rhs),&trs_rhs);CHKERRQ(ierr); 822984c4197SStefano Zampini /* array to store whether a node is primal or not */ 823984c4197SStefano Zampini ierr = PetscMalloc(pcis->n_B*sizeof(*is_primal),&is_primal);CHKERRQ(ierr); 824984c4197SStefano Zampini ierr = PetscMemzero(is_primal,pcis->n_B*sizeof(*is_primal));CHKERRQ(ierr); 825984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) is_primal[local_to_B[aux_primal_numbering[i]]] = PETSC_TRUE; 826984c4197SStefano Zampini 827984c4197SStefano Zampini /* allocating workspace for check */ 828984c4197SStefano Zampini if (pcbddc->dbg_flag) { 829984c4197SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 830984c4197SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 831984c4197SStefano Zampini ierr = PetscMalloc(max_size_of_constraint*(max_constraints+max_size_of_constraint)*sizeof(*work),&work);CHKERRQ(ierr); 832674ae819SStefano Zampini } 833984c4197SStefano Zampini 834026de310SStefano Zampini /* loop on constraints and see whether or not they need a change of basis */ 835026de310SStefano Zampini /* -> using implicit ordering contained in temp_indices data */ 836026de310SStefano Zampini total_counts = pcbddc->n_vertices; 837026de310SStefano Zampini primal_counter = total_counts; 838026de310SStefano Zampini while (total_counts<pcbddc->local_primal_size) { 839026de310SStefano Zampini primal_dofs = 1; 840026de310SStefano Zampini if (change_basis[total_counts]) { 841026de310SStefano Zampini /* get all constraints with same support: if more then one constraint is present on the cc then surely indices are stored contiguosly */ 842026de310SStefano Zampini while (total_counts+primal_dofs < pcbddc->local_primal_size && temp_indices_to_constraint_B[temp_indices[total_counts]] == temp_indices_to_constraint_B[temp_indices[total_counts+primal_dofs]]) { 843026de310SStefano Zampini primal_dofs++; 844674ae819SStefano Zampini } 845984c4197SStefano Zampini /* get constraint info */ 846026de310SStefano Zampini size_of_constraint = temp_indices[total_counts+1]-temp_indices[total_counts]; 847984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 848984c4197SStefano Zampini 849984c4197SStefano Zampini /* copy quadrature constraints for change of basis check */ 850984c4197SStefano Zampini if (pcbddc->dbg_flag) { 851026de310SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d to %d need a change of basis (size %d)\n",total_counts,total_counts+primal_dofs,size_of_constraint);CHKERRQ(ierr); 852026de310SStefano Zampini ierr = PetscMemcpy(work,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 853674ae819SStefano Zampini } 854984c4197SStefano Zampini 855984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 856026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 857984c4197SStefano Zampini 858984c4197SStefano Zampini /* compute QR decomposition of constraints */ 859*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 860*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 861*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 862674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 863*3f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 864984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 865674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 866984c4197SStefano Zampini 867984c4197SStefano Zampini /* explictly compute R^-T */ 868984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 869984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 870*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 871*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 872*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 873*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 874984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 875*3f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 876984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 877984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 878984c4197SStefano Zampini 879984c4197SStefano Zampini /* explcitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 880*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 881*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 882*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 883*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 884984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 885*3f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 886984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 887984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 888984c4197SStefano Zampini 889984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 890984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 891984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 892*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 893*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 894*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 895*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 896*3f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 897*3f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 898984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 899c4303822SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_LDC)); 900984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 901026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 902984c4197SStefano Zampini 903984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 904026de310SStefano Zampini start_rows = &temp_indices_to_constraint_B[temp_indices[total_counts]]; 905984c4197SStefano Zampini /* insert cols for primal dofs */ 906984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 907984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 908026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+aux_primal_minloc[primal_counter+j]]; 909984c4197SStefano Zampini ierr = MatSetValues(pcbddc->ChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 910984c4197SStefano Zampini } 911984c4197SStefano Zampini /* insert cols for dual dofs */ 912984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 913026de310SStefano Zampini if (!is_primal[temp_indices_to_constraint_B[temp_indices[total_counts]+k]]) { 914984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 915026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+k]; 916984c4197SStefano Zampini ierr = MatSetValues(pcbddc->ChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 917984c4197SStefano Zampini j++; 918674ae819SStefano Zampini } 919674ae819SStefano Zampini } 920984c4197SStefano Zampini 921984c4197SStefano Zampini /* check change of basis */ 922984c4197SStefano Zampini if (pcbddc->dbg_flag) { 923984c4197SStefano Zampini PetscInt ii,jj; 924984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 925c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 926c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 927c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 928c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 929c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 930c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 931984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 932c4303822SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("T","N",&Blas_M,&Blas_N,&Blas_K,&one,work,&Blas_LDA,qr_basis,&Blas_LDB,&zero,&work[size_of_constraint*primal_dofs],&Blas_LDC)); 933984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 934984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 935984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 936984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 937984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE; 938674ae819SStefano Zampini } 939674ae819SStefano Zampini } 940984c4197SStefano Zampini if (!valid_qr) { 941984c4197SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n",PetscGlobalRank);CHKERRQ(ierr); 942984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 943984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 944984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 945984c4197SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not orthogonal to constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 946674ae819SStefano Zampini } 947984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 948984c4197SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not unitary w.r.t constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 949984c4197SStefano Zampini } 950984c4197SStefano Zampini } 951984c4197SStefano Zampini } 952674ae819SStefano Zampini } else { 953984c4197SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n",PetscGlobalRank);CHKERRQ(ierr); 954674ae819SStefano Zampini } 955674ae819SStefano Zampini } 956026de310SStefano Zampini /* increment primal counter */ 957026de310SStefano Zampini primal_counter += primal_dofs; 958984c4197SStefano Zampini } else { 959984c4197SStefano Zampini if (pcbddc->dbg_flag) { 960026de310SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,temp_indices[total_counts+1]-temp_indices[total_counts]);CHKERRQ(ierr); 961674ae819SStefano Zampini } 962674ae819SStefano Zampini } 963026de310SStefano Zampini /* increment constraint counter total_counts */ 964026de310SStefano Zampini total_counts += primal_dofs; 965674ae819SStefano Zampini } 966984c4197SStefano Zampini if (pcbddc->dbg_flag) { 967984c4197SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 968984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 969984c4197SStefano Zampini } 970984c4197SStefano Zampini /* free workspace */ 971984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 972984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 973984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 974984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 975984c4197SStefano Zampini ierr = PetscFree(is_primal);CHKERRQ(ierr); 976984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 977674ae819SStefano Zampini } 978674ae819SStefano Zampini /* assembling */ 979674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 980674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 981984c4197SStefano Zampini /* 982984c4197SStefano Zampini ierr = MatView(pcbddc->ChangeOfBasisMatrix,(PetscViewer)0);CHKERRQ(ierr); 983984c4197SStefano Zampini */ 984674ae819SStefano Zampini } 985674ae819SStefano Zampini /* free workspace no longer needed */ 986984c4197SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 987984c4197SStefano Zampini ierr = PetscFree(aux_primal_minloc);CHKERRQ(ierr); 988674ae819SStefano Zampini ierr = PetscFree(temp_indices);CHKERRQ(ierr); 989674ae819SStefano Zampini ierr = PetscFree(change_basis);CHKERRQ(ierr); 990674ae819SStefano Zampini ierr = PetscFree(temp_indices_to_constraint);CHKERRQ(ierr); 991674ae819SStefano Zampini ierr = PetscFree(temp_indices_to_constraint_B);CHKERRQ(ierr); 992674ae819SStefano Zampini ierr = PetscFree(local_to_B);CHKERRQ(ierr); 993674ae819SStefano Zampini ierr = PetscFree(temp_quadrature_constraint);CHKERRQ(ierr); 994674ae819SStefano Zampini PetscFunctionReturn(0); 995674ae819SStefano Zampini } 996674ae819SStefano Zampini 997674ae819SStefano Zampini #undef __FUNCT__ 998674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 999674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 1000674ae819SStefano Zampini { 1001674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1002674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1003674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1004674ae819SStefano Zampini PetscInt bs,ierr,i,vertex_size; 1005674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 1006674ae819SStefano Zampini 1007674ae819SStefano Zampini PetscFunctionBegin; 1008674ae819SStefano Zampini /* Init local Graph struct */ 1009674ae819SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping);CHKERRQ(ierr); 1010674ae819SStefano Zampini 1011575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 1012575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 1013575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1014575ad6abSStefano Zampini } 1015674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 1016674ae819SStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 1017674ae819SStefano Zampini Mat mat_adj; 1018674ae819SStefano Zampini const PetscInt *xadj,*adjncy; 1019674ae819SStefano Zampini PetscBool flg_row=PETSC_TRUE; 1020674ae819SStefano Zampini 1021674ae819SStefano Zampini ierr = MatConvert(matis->A,MATMPIADJ,MAT_INITIAL_MATRIX,&mat_adj);CHKERRQ(ierr); 1022674ae819SStefano Zampini ierr = MatGetRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&i,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1023674ae819SStefano Zampini if (!flg_row) { 1024674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatGetRowIJ called in %s\n",__FUNCT__); 1025674ae819SStefano Zampini } 1026674ae819SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,i,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 1027674ae819SStefano Zampini ierr = MatRestoreRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&i,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1028674ae819SStefano Zampini if (!flg_row) { 1029674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__); 1030674ae819SStefano Zampini } 1031674ae819SStefano Zampini ierr = MatDestroy(&mat_adj);CHKERRQ(ierr); 1032674ae819SStefano Zampini } 1033674ae819SStefano Zampini 1034674ae819SStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting */ 1035674ae819SStefano Zampini vertex_size = 1; 1036674ae819SStefano Zampini if (!pcbddc->n_ISForDofs) { 1037674ae819SStefano Zampini IS *custom_ISForDofs; 1038674ae819SStefano Zampini 1039674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&bs);CHKERRQ(ierr); 1040674ae819SStefano Zampini ierr = PetscMalloc(bs*sizeof(IS),&custom_ISForDofs);CHKERRQ(ierr); 1041674ae819SStefano Zampini for (i=0;i<bs;i++) { 1042674ae819SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n/bs,i,bs,&custom_ISForDofs[i]);CHKERRQ(ierr); 1043674ae819SStefano Zampini } 1044674ae819SStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,bs,custom_ISForDofs);CHKERRQ(ierr); 1045674ae819SStefano Zampini /* remove my references to IS objects */ 1046674ae819SStefano Zampini for (i=0;i<bs;i++) { 1047674ae819SStefano Zampini ierr = ISDestroy(&custom_ISForDofs[i]);CHKERRQ(ierr); 1048674ae819SStefano Zampini } 1049674ae819SStefano Zampini ierr = PetscFree(custom_ISForDofs);CHKERRQ(ierr); 1050674ae819SStefano Zampini } else { /* mat block size as vertex size (used for elasticity) */ 1051674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 1052674ae819SStefano Zampini } 1053674ae819SStefano Zampini 1054674ae819SStefano Zampini /* Setup of Graph */ 1055674ae819SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundaries,pcbddc->DirichletBoundaries,pcbddc->n_ISForDofs,pcbddc->ISForDofs,pcbddc->user_primal_vertices); 1056674ae819SStefano Zampini 1057674ae819SStefano Zampini /* Graph's connected components analysis */ 1058674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 1059674ae819SStefano Zampini 1060674ae819SStefano Zampini /* print some info to stdout */ 1061674ae819SStefano Zampini if (pcbddc->dbg_flag) { 1062e49050b4SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer); 1063674ae819SStefano Zampini } 1064674ae819SStefano Zampini PetscFunctionReturn(0); 1065674ae819SStefano Zampini } 1066674ae819SStefano Zampini 1067674ae819SStefano Zampini #undef __FUNCT__ 1068674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalVerticesLocalIdx" 1069674ae819SStefano Zampini PetscErrorCode PCBDDCGetPrimalVerticesLocalIdx(PC pc, PetscInt *n_vertices, PetscInt *vertices_idx[]) 1070674ae819SStefano Zampini { 1071674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 1072674ae819SStefano Zampini PetscInt *vertices,*row_cmat_indices,n,i,size_of_constraint,local_primal_size; 1073674ae819SStefano Zampini PetscErrorCode ierr; 1074674ae819SStefano Zampini 1075674ae819SStefano Zampini PetscFunctionBegin; 1076674ae819SStefano Zampini n = 0; 1077674ae819SStefano Zampini vertices = 0; 1078674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 1079674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&i);CHKERRQ(ierr); 1080b120a5c6SStefano Zampini for (i=0;i<local_primal_size;i++) { 1081b120a5c6SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 1082b120a5c6SStefano Zampini if (size_of_constraint == 1) n++; 1083b120a5c6SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 1084b120a5c6SStefano Zampini } 1085b120a5c6SStefano Zampini ierr = PetscMalloc(n*sizeof(PetscInt),&vertices);CHKERRQ(ierr); 1086b120a5c6SStefano Zampini n = 0; 1087674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 1088674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 1089674ae819SStefano Zampini if (size_of_constraint == 1) { 1090674ae819SStefano Zampini vertices[n++]=row_cmat_indices[0]; 1091674ae819SStefano Zampini } 1092674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 1093674ae819SStefano Zampini } 1094674ae819SStefano Zampini } 1095674ae819SStefano Zampini *n_vertices = n; 1096674ae819SStefano Zampini *vertices_idx = vertices; 1097674ae819SStefano Zampini PetscFunctionReturn(0); 1098674ae819SStefano Zampini } 1099674ae819SStefano Zampini 1100674ae819SStefano Zampini #undef __FUNCT__ 1101674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalConstraintsLocalIdx" 1102674ae819SStefano Zampini PetscErrorCode PCBDDCGetPrimalConstraintsLocalIdx(PC pc, PetscInt *n_constraints, PetscInt *constraints_idx[]) 1103674ae819SStefano Zampini { 1104674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 1105674ae819SStefano Zampini PetscInt *constraints_index,*row_cmat_indices,*row_cmat_global_indices; 1106674ae819SStefano Zampini PetscInt n,i,j,size_of_constraint,local_primal_size,local_size,max_size_of_constraint,min_index,min_loc; 1107674ae819SStefano Zampini PetscBool *touched; 1108674ae819SStefano Zampini PetscErrorCode ierr; 1109674ae819SStefano Zampini 1110674ae819SStefano Zampini PetscFunctionBegin; 1111674ae819SStefano Zampini n = 0; 1112674ae819SStefano Zampini constraints_index = 0; 1113674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 1114674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&local_size);CHKERRQ(ierr); 1115674ae819SStefano Zampini max_size_of_constraint = 0; 1116674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 1117674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 1118674ae819SStefano Zampini if (size_of_constraint > 1) { 1119674ae819SStefano Zampini n++; 1120674ae819SStefano Zampini } 1121674ae819SStefano Zampini max_size_of_constraint = PetscMax(size_of_constraint,max_size_of_constraint); 1122674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 1123674ae819SStefano Zampini } 1124674ae819SStefano Zampini ierr = PetscMalloc(n*sizeof(PetscInt),&constraints_index);CHKERRQ(ierr); 1125674ae819SStefano Zampini ierr = PetscMalloc(max_size_of_constraint*sizeof(PetscInt),&row_cmat_global_indices);CHKERRQ(ierr); 1126674ae819SStefano Zampini ierr = PetscMalloc(local_size*sizeof(PetscBool),&touched);CHKERRQ(ierr); 1127674ae819SStefano Zampini ierr = PetscMemzero(touched,local_size*sizeof(PetscBool));CHKERRQ(ierr); 1128674ae819SStefano Zampini n = 0; 1129674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 1130674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 1131674ae819SStefano Zampini if (size_of_constraint > 1) { 1132674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,row_cmat_indices,row_cmat_global_indices);CHKERRQ(ierr); 1133674ae819SStefano Zampini min_index = row_cmat_global_indices[0]; 1134674ae819SStefano Zampini min_loc = 0; 1135674ae819SStefano Zampini for (j=1;j<size_of_constraint;j++) { 1136674ae819SStefano Zampini /* there can be more than one constraint on a single connected component */ 1137674ae819SStefano Zampini if (min_index > row_cmat_global_indices[j] && !touched[row_cmat_indices[j]]) { 1138674ae819SStefano Zampini min_index = row_cmat_global_indices[j]; 1139674ae819SStefano Zampini min_loc = j; 1140674ae819SStefano Zampini } 1141674ae819SStefano Zampini } 1142674ae819SStefano Zampini touched[row_cmat_indices[min_loc]] = PETSC_TRUE; 1143674ae819SStefano Zampini constraints_index[n++] = row_cmat_indices[min_loc]; 1144674ae819SStefano Zampini } 1145674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 1146674ae819SStefano Zampini } 1147674ae819SStefano Zampini } 1148674ae819SStefano Zampini ierr = PetscFree(touched);CHKERRQ(ierr); 1149674ae819SStefano Zampini ierr = PetscFree(row_cmat_global_indices);CHKERRQ(ierr); 1150674ae819SStefano Zampini *n_constraints = n; 1151674ae819SStefano Zampini *constraints_idx = constraints_index; 1152674ae819SStefano Zampini PetscFunctionReturn(0); 1153674ae819SStefano Zampini } 1154674ae819SStefano Zampini 1155674ae819SStefano Zampini /* the next two functions has been adapted from pcis.c */ 1156674ae819SStefano Zampini #undef __FUNCT__ 1157674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplySchur" 1158674ae819SStefano Zampini PetscErrorCode PCBDDCApplySchur(PC pc, Vec v, Vec vec1_B, Vec vec2_B, Vec vec1_D, Vec vec2_D) 1159674ae819SStefano Zampini { 1160674ae819SStefano Zampini PetscErrorCode ierr; 1161674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 1162674ae819SStefano Zampini 1163674ae819SStefano Zampini PetscFunctionBegin; 1164674ae819SStefano Zampini if (!vec2_B) { vec2_B = v; } 1165674ae819SStefano Zampini ierr = MatMult(pcis->A_BB,v,vec1_B);CHKERRQ(ierr); 1166674ae819SStefano Zampini ierr = MatMult(pcis->A_IB,v,vec1_D);CHKERRQ(ierr); 1167674ae819SStefano Zampini ierr = KSPSolve(pcis->ksp_D,vec1_D,vec2_D);CHKERRQ(ierr); 1168674ae819SStefano Zampini ierr = MatMult(pcis->A_BI,vec2_D,vec2_B);CHKERRQ(ierr); 1169674ae819SStefano Zampini ierr = VecAXPY(vec1_B,-1.0,vec2_B);CHKERRQ(ierr); 1170674ae819SStefano Zampini PetscFunctionReturn(0); 1171674ae819SStefano Zampini } 1172674ae819SStefano Zampini 1173674ae819SStefano Zampini #undef __FUNCT__ 1174674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplySchurTranspose" 1175674ae819SStefano Zampini PetscErrorCode PCBDDCApplySchurTranspose(PC pc, Vec v, Vec vec1_B, Vec vec2_B, Vec vec1_D, Vec vec2_D) 1176674ae819SStefano Zampini { 1177674ae819SStefano Zampini PetscErrorCode ierr; 1178674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 1179674ae819SStefano Zampini 1180674ae819SStefano Zampini PetscFunctionBegin; 1181674ae819SStefano Zampini if (!vec2_B) { vec2_B = v; } 1182674ae819SStefano Zampini ierr = MatMultTranspose(pcis->A_BB,v,vec1_B);CHKERRQ(ierr); 1183674ae819SStefano Zampini ierr = MatMultTranspose(pcis->A_BI,v,vec1_D);CHKERRQ(ierr); 1184674ae819SStefano Zampini ierr = KSPSolveTranspose(pcis->ksp_D,vec1_D,vec2_D);CHKERRQ(ierr); 1185674ae819SStefano Zampini ierr = MatMultTranspose(pcis->A_IB,vec2_D,vec2_B);CHKERRQ(ierr); 1186674ae819SStefano Zampini ierr = VecAXPY(vec1_B,-1.0,vec2_B);CHKERRQ(ierr); 1187674ae819SStefano Zampini PetscFunctionReturn(0); 1188674ae819SStefano Zampini } 1189674ae819SStefano Zampini 1190674ae819SStefano Zampini #undef __FUNCT__ 1191674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 1192674ae819SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(MPI_Comm comm,ISLocalToGlobalMapping l2gmap, PetscInt n_local_dofs, PetscInt local_dofs[], PetscInt local_dofs_mult[], PetscInt* n_global_subset, PetscInt* global_numbering_subset[]) 1193674ae819SStefano Zampini { 1194674ae819SStefano Zampini Vec local_vec,global_vec; 1195674ae819SStefano Zampini IS seqis,paris; 1196674ae819SStefano Zampini VecScatter scatter_ctx; 1197674ae819SStefano Zampini PetscScalar *array; 1198674ae819SStefano Zampini PetscInt *temp_global_dofs; 1199674ae819SStefano Zampini PetscScalar globalsum; 1200674ae819SStefano Zampini PetscInt i,j,s; 1201674ae819SStefano Zampini PetscInt nlocals,first_index,old_index,max_local; 1202674ae819SStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm,max_global; 1203674ae819SStefano Zampini PetscMPIInt *dof_sizes,*dof_displs; 1204674ae819SStefano Zampini PetscBool first_found; 1205674ae819SStefano Zampini PetscErrorCode ierr; 1206674ae819SStefano Zampini 1207674ae819SStefano Zampini PetscFunctionBegin; 1208674ae819SStefano Zampini /* mpi buffers */ 1209674ae819SStefano Zampini MPI_Comm_size(comm,&size_prec_comm); 1210674ae819SStefano Zampini MPI_Comm_rank(comm,&rank_prec_comm); 1211674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 1212674ae819SStefano Zampini ierr = PetscMalloc(j*sizeof(*dof_sizes),&dof_sizes);CHKERRQ(ierr); 1213674ae819SStefano Zampini ierr = PetscMalloc(j*sizeof(*dof_displs),&dof_displs);CHKERRQ(ierr); 1214674ae819SStefano Zampini /* get maximum size of subset */ 1215674ae819SStefano Zampini ierr = PetscMalloc(n_local_dofs*sizeof(PetscInt),&temp_global_dofs);CHKERRQ(ierr); 1216674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 1217674ae819SStefano Zampini max_local = 0; 1218674ae819SStefano Zampini if (n_local_dofs) { 1219674ae819SStefano Zampini max_local = temp_global_dofs[0]; 1220674ae819SStefano Zampini for (i=1;i<n_local_dofs;i++) { 1221674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 1222674ae819SStefano Zampini max_local = temp_global_dofs[i]; 1223674ae819SStefano Zampini } 1224674ae819SStefano Zampini } 1225674ae819SStefano Zampini } 1226674ae819SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm); 1227674ae819SStefano Zampini max_global++; 1228674ae819SStefano Zampini max_local = 0; 1229674ae819SStefano Zampini if (n_local_dofs) { 1230674ae819SStefano Zampini max_local = local_dofs[0]; 1231674ae819SStefano Zampini for (i=1;i<n_local_dofs;i++) { 1232674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 1233674ae819SStefano Zampini max_local = local_dofs[i]; 1234674ae819SStefano Zampini } 1235674ae819SStefano Zampini } 1236674ae819SStefano Zampini } 1237674ae819SStefano Zampini max_local++; 1238674ae819SStefano Zampini /* allocate workspace */ 1239674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 1240674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 1241674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 1242674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 1243674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 1244674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 1245674ae819SStefano Zampini /* create scatter */ 1246674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 1247674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 1248674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 1249674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 1250674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 1251674ae819SStefano Zampini /* init array */ 1252674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 1253674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 1254674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 1255674ae819SStefano Zampini if (local_dofs_mult) { 1256674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 1257674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 1258674ae819SStefano Zampini } 1259674ae819SStefano Zampini } else { 1260674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 1261674ae819SStefano Zampini array[local_dofs[i]]=1.0; 1262674ae819SStefano Zampini } 1263674ae819SStefano Zampini } 1264674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 1265674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 1266674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1267674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1268674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 12695b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 1270674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 1271674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 1272674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 1273674ae819SStefano Zampini nlocals = 0; 1274674ae819SStefano Zampini first_index = -1; 1275674ae819SStefano Zampini first_found = PETSC_FALSE; 1276674ae819SStefano Zampini for (i=0;i<s;i++) { 12775b08dc53SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.0) { 1278674ae819SStefano Zampini first_found = PETSC_TRUE; 1279674ae819SStefano Zampini first_index = i; 1280674ae819SStefano Zampini } 12815b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 1282674ae819SStefano Zampini } 1283674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 1284674ae819SStefano Zampini if (!rank_prec_comm) { 1285674ae819SStefano Zampini dof_displs[0]=0; 1286674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 1287674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 1288674ae819SStefano Zampini } 1289674ae819SStefano Zampini } 1290674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 1291674ae819SStefano Zampini if (first_found) { 1292674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 1293674ae819SStefano Zampini old_index = first_index; 1294674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 12955b08dc53SStefano Zampini if (PetscRealPart(array[i]) > 0.0) { 1296674ae819SStefano Zampini array[i] += array[old_index]; 1297674ae819SStefano Zampini old_index = i; 1298674ae819SStefano Zampini } 1299674ae819SStefano Zampini } 1300674ae819SStefano Zampini } 1301674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 1302674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 1303674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1304674ae819SStefano Zampini ierr = VecScatterEnd (scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1305674ae819SStefano Zampini /* get global ordering of local dofs */ 1306674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 1307674ae819SStefano Zampini if (local_dofs_mult) { 1308674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 13095b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 1310674ae819SStefano Zampini } 1311674ae819SStefano Zampini } else { 1312674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 13135b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 1314674ae819SStefano Zampini } 1315674ae819SStefano Zampini } 1316674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 1317674ae819SStefano Zampini /* free workspace */ 1318674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1319674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1320674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1321674ae819SStefano Zampini ierr = PetscFree(dof_sizes);CHKERRQ(ierr); 1322674ae819SStefano Zampini ierr = PetscFree(dof_displs);CHKERRQ(ierr); 1323674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 1324674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 1325674ae819SStefano Zampini PetscFunctionReturn(0); 1326674ae819SStefano Zampini } 1327