1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8674ae819SStefano Zampini #undef __FUNCT__ 9c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 10c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 11c8587f34SStefano Zampini { 12c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 138629588bSStefano Zampini PetscScalar *coarse_submat_vals; 14c8587f34SStefano Zampini PetscErrorCode ierr; 15c8587f34SStefano Zampini 16c8587f34SStefano Zampini PetscFunctionBegin; 17c8587f34SStefano Zampini /* Compute matrix after change of basis and extract local submatrices */ 18c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalMatrices(pc);CHKERRQ(ierr); 19c8587f34SStefano Zampini 20f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 21f4ddd8eeSStefano Zampini /* PCBDDCSetUpLocalWorkVectors and PCBDDCSetUpLocalMatrices should be called first! */ 22c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 23c8587f34SStefano Zampini 24c8587f34SStefano Zampini /* Setup local solvers ksp_D and ksp_R */ 250fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 26c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc);CHKERRQ(ierr); 27c8587f34SStefano Zampini 28c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 29b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 30c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 31c8587f34SStefano Zampini } 32c8587f34SStefano Zampini 338629588bSStefano Zampini /* 348629588bSStefano Zampini Setup local correction and local part of coarse basis. 358629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 368629588bSStefano Zampini */ 3747f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 388629588bSStefano Zampini 398629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 408629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 418629588bSStefano Zampini 428629588bSStefano Zampini /* free */ 438629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 44c8587f34SStefano Zampini PetscFunctionReturn(0); 45c8587f34SStefano Zampini } 46c8587f34SStefano Zampini 47c8587f34SStefano Zampini #undef __FUNCT__ 48674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 49674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 50674ae819SStefano Zampini { 51674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 52674ae819SStefano Zampini PetscErrorCode ierr; 53674ae819SStefano Zampini 54674ae819SStefano Zampini PetscFunctionBegin; 55674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 56674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 57674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 58674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 59785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 60674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 61f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 62f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 63785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 6463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 6563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 66674ae819SStefano Zampini PetscFunctionReturn(0); 67674ae819SStefano Zampini } 68674ae819SStefano Zampini 69674ae819SStefano Zampini #undef __FUNCT__ 70674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 71674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 72674ae819SStefano Zampini { 73674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 74674ae819SStefano Zampini PetscErrorCode ierr; 75674ae819SStefano Zampini 76674ae819SStefano Zampini PetscFunctionBegin; 77b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 78674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 79674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 80674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 81*5ff63025SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 82674ae819SStefano Zampini PetscFunctionReturn(0); 83674ae819SStefano Zampini } 84674ae819SStefano Zampini 85674ae819SStefano Zampini #undef __FUNCT__ 86674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 87674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 88674ae819SStefano Zampini { 89674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 90674ae819SStefano Zampini PetscErrorCode ierr; 91674ae819SStefano Zampini 92674ae819SStefano Zampini PetscFunctionBegin; 93674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 94674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_rhs);CHKERRQ(ierr); 95674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 96674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 9715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 9815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 99674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 100674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 101674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 102674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 103674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 104674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 1058ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 106674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 107674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 108674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 109f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 110f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 111f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 112f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 113727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 114f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 11570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 1166e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 117674ae819SStefano Zampini PetscFunctionReturn(0); 118674ae819SStefano Zampini } 119674ae819SStefano Zampini 120674ae819SStefano Zampini #undef __FUNCT__ 121f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 122f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 1236bfb1811SStefano Zampini { 1246bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1256bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1266bfb1811SStefano Zampini VecType impVecType; 127e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 1286bfb1811SStefano Zampini PetscErrorCode ierr; 1296bfb1811SStefano Zampini 1306bfb1811SStefano Zampini PetscFunctionBegin; 131f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 132f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 133f4ddd8eeSStefano Zampini } 134e7b262bdSStefano Zampini /* get sizes */ 135e9189074SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_actual_vertices; 136e9189074SStefano Zampini n_R = pcis->n-pcbddc->n_actual_vertices; 1376bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 138e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 139e7b262bdSStefano Zampini /* R nodes */ 140e7b262bdSStefano Zampini old_size = -1; 141e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 142e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 143e7b262bdSStefano Zampini } 144e7b262bdSStefano Zampini if (n_R != old_size) { 145e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 146e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 1476bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 1486bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 1496bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 1506bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 151e7b262bdSStefano Zampini } 152e7b262bdSStefano Zampini /* local primal dofs */ 153e7b262bdSStefano Zampini old_size = -1; 154e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 155e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 156e7b262bdSStefano Zampini } 157e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 158e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 15983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 160e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 1616bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 162e7b262bdSStefano Zampini } 163e7b262bdSStefano Zampini /* local explicit constraints */ 164e7b262bdSStefano Zampini old_size = -1; 165e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 166e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 167e7b262bdSStefano Zampini } 168e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 169e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 17083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 17183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 17283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 17383b7ccabSStefano Zampini } 1746bfb1811SStefano Zampini PetscFunctionReturn(0); 1756bfb1811SStefano Zampini } 1766bfb1811SStefano Zampini 1776bfb1811SStefano Zampini #undef __FUNCT__ 17847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 17947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 18088ebb749SStefano Zampini { 18125084f0cSStefano Zampini PetscErrorCode ierr; 18225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 18388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 18488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 18525084f0cSStefano Zampini /* submatrices of local problem */ 18688ebb749SStefano Zampini Mat A_RV,A_VR,A_VV; 18725084f0cSStefano Zampini /* working matrices */ 18825084f0cSStefano Zampini Mat M1,M2,M3,C_CR; 18925084f0cSStefano Zampini /* working vectors */ 19025084f0cSStefano Zampini Vec vec1_C,vec2_C,vec1_V,vec2_V; 19125084f0cSStefano Zampini /* additional working stuff */ 19225084f0cSStefano Zampini IS is_aux; 19325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 19425084f0cSStefano Zampini const PetscScalar *array,*row_cmat_values; 19525084f0cSStefano Zampini const PetscInt *row_cmat_indices,*idx_R_local; 196e9189074SStefano Zampini PetscInt *idx_V_B,*auxindices; 19725084f0cSStefano Zampini PetscInt n_vertices,n_constraints,size_of_constraint; 19825084f0cSStefano Zampini PetscInt i,j,n_R,n_D,n_B; 199b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 20045a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 20188ebb749SStefano Zampini MatType impMatType; 20225084f0cSStefano Zampini /* some shortcuts to scalars */ 20325084f0cSStefano Zampini PetscScalar zero=0.0,one=1.0,m_one=-1.0; 20425084f0cSStefano Zampini /* for debugging purposes */ 20588ebb749SStefano Zampini PetscReal *coarsefunctions_errors,*constraints_errors; 20688ebb749SStefano Zampini 20788ebb749SStefano Zampini PetscFunctionBegin; 208e9189074SStefano Zampini /* get number of vertices (corners plus constraints with change of basis) 209e9189074SStefano Zampini pcbddc->n_actual_vertices stores the actual number of vertices, pcbddc->n_vertices the number of corners computed */ 210e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 21188ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 21288ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 21388ebb749SStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 21488ebb749SStefano Zampini n_R = pcis->n-n_vertices; 21588ebb749SStefano Zampini 21688ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 21788ebb749SStefano Zampini impMatType = MATSEQDENSE; 21888ebb749SStefano Zampini 21988ebb749SStefano Zampini /* Allocating some extra storage just to be safe */ 22088ebb749SStefano Zampini ierr = PetscMalloc (pcis->n*sizeof(PetscInt),&auxindices);CHKERRQ(ierr); 22188ebb749SStefano Zampini for (i=0;i<pcis->n;i++) auxindices[i]=i; 22288ebb749SStefano Zampini 22388ebb749SStefano Zampini /* vertices in boundary numbering */ 224785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 225e9189074SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcbddc->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->primal_indices_local_idxs,&i,idx_V_B);CHKERRQ(ierr); 22688ebb749SStefano Zampini if (i != n_vertices) { 22722d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 22888ebb749SStefano Zampini } 22988ebb749SStefano Zampini 23088ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 23188ebb749SStefano Zampini if (n_constraints) { 232f4ddd8eeSStefano Zampini /* see if we can save some allocations */ 233f4ddd8eeSStefano Zampini if (pcbddc->local_auxmat2) { 234f4ddd8eeSStefano Zampini PetscInt on_R,on_constraints; 235f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->local_auxmat2,&on_R,&on_constraints);CHKERRQ(ierr); 236f4ddd8eeSStefano Zampini if (on_R != n_R || on_constraints != n_constraints) { 237f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 238f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 239f4ddd8eeSStefano Zampini } 240f4ddd8eeSStefano Zampini } 24145a1bb75SStefano Zampini /* work vectors */ 24245a1bb75SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_C,&vec1_C);CHKERRQ(ierr); 24345a1bb75SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_C,&vec2_C);CHKERRQ(ierr); 24445a1bb75SStefano Zampini /* auxiliary matrices */ 245f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat2) { 24688ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->local_auxmat2);CHKERRQ(ierr); 24725084f0cSStefano Zampini ierr = MatSetSizes(pcbddc->local_auxmat2,n_R,n_constraints,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 24888ebb749SStefano Zampini ierr = MatSetType(pcbddc->local_auxmat2,impMatType);CHKERRQ(ierr); 24925084f0cSStefano Zampini ierr = MatSetUp(pcbddc->local_auxmat2);CHKERRQ(ierr); 250f4ddd8eeSStefano Zampini } 25188ebb749SStefano Zampini 25225084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 25325084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 2548ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 25525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 25688ebb749SStefano Zampini 25788ebb749SStefano Zampini /* Assemble local_auxmat2 = - A_{RR}^{-1} C^T_{CR} needed by BDDC application */ 25888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 25988ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 26088ebb749SStefano Zampini /* Get row of constraint matrix in R numbering */ 26125084f0cSStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 26225084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_R,size_of_constraint,row_cmat_indices,row_cmat_values,INSERT_VALUES);CHKERRQ(ierr); 26325084f0cSStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 26425084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_R);CHKERRQ(ierr); 26525084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_R);CHKERRQ(ierr); 26688ebb749SStefano Zampini /* Solve for row of constraint matrix in R numbering */ 26788ebb749SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 26825084f0cSStefano Zampini /* Set values in local_auxmat2 */ 26925084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec2_R,&array);CHKERRQ(ierr); 27088ebb749SStefano Zampini ierr = MatSetValues(pcbddc->local_auxmat2,n_R,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 27125084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec2_R,&array);CHKERRQ(ierr); 27288ebb749SStefano Zampini } 27388ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_auxmat2,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27488ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_auxmat2,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27525084f0cSStefano Zampini ierr = MatScale(pcbddc->local_auxmat2,m_one);CHKERRQ(ierr); 27688ebb749SStefano Zampini 27788ebb749SStefano Zampini /* Assemble explicitly M1 = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} needed in preproc */ 27825084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 27925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 28088ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&M1);CHKERRQ(ierr); 28188ebb749SStefano Zampini ierr = MatSetSizes(M1,n_constraints,n_constraints,n_constraints,n_constraints);CHKERRQ(ierr); 28288ebb749SStefano Zampini ierr = MatSetType(M1,impMatType);CHKERRQ(ierr); 28325084f0cSStefano Zampini ierr = MatSetUp(M1);CHKERRQ(ierr); 28425084f0cSStefano Zampini ierr = MatDuplicate(M1,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 28525084f0cSStefano Zampini ierr = MatZeroEntries(M2);CHKERRQ(ierr); 28625084f0cSStefano Zampini ierr = VecSet(vec1_C,m_one);CHKERRQ(ierr); 28725084f0cSStefano Zampini ierr = MatDiagonalSet(M2,vec1_C,INSERT_VALUES);CHKERRQ(ierr); 28825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 28925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 29025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 29188ebb749SStefano Zampini /* Assemble local_auxmat1 = M1*C_{CR} needed by BDDC application in KSP and in preproc */ 292f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat1) { 29388ebb749SStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 294f4ddd8eeSStefano Zampini } else { 295f4ddd8eeSStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_REUSE_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 296f4ddd8eeSStefano Zampini } 29788ebb749SStefano Zampini } 29888ebb749SStefano Zampini 29988ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 30088ebb749SStefano Zampini if (n_vertices) { 3013a50541eSStefano Zampini PetscInt ibs,mbs; 3023a50541eSStefano Zampini PetscBool issbaij; 303af732b37SStefano Zampini Mat newmat; 3043a50541eSStefano Zampini 3053a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 3069577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3079577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 3083a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 3099577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 3109577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3119577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3129577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 3139577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 3149577ea80SStefano Zampini } else { 3153a50541eSStefano Zampini /* this is safe */ 3169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 3173a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3183a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 3193a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 32045a1bb75SStefano Zampini /* which of the two approaches is faster? */ 3213a50541eSStefano Zampini /* ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3223a50541eSStefano Zampini ierr = MatCreateTranspose(A_RV,&A_VR);CHKERRQ(ierr);*/ 3233a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3243a50541eSStefano Zampini ierr = MatCreateTranspose(A_VR,&A_RV);CHKERRQ(ierr); 3253a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 3263a50541eSStefano Zampini } else { 3279577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3289577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3299577ea80SStefano Zampini } 3303a50541eSStefano Zampini } 3312a7a6963SBarry Smith ierr = MatCreateVecs(A_RV,&vec1_V,NULL);CHKERRQ(ierr); 33245a1bb75SStefano Zampini ierr = VecDuplicate(vec1_V,&vec2_V);CHKERRQ(ierr); 33325084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 33488ebb749SStefano Zampini } 33588ebb749SStefano Zampini 33688ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 337f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 338f4ddd8eeSStefano Zampini PetscInt on_B,on_primal; 339f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 340f4ddd8eeSStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size) { 341f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 342f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 343f4ddd8eeSStefano Zampini } 344f4ddd8eeSStefano Zampini } 345f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_D) { 346f4ddd8eeSStefano Zampini PetscInt on_D,on_primal; 347f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,&on_primal);CHKERRQ(ierr); 348f4ddd8eeSStefano Zampini if (on_D != n_D || on_primal != pcbddc->local_primal_size) { 349f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 350f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 351f4ddd8eeSStefano Zampini } 352f4ddd8eeSStefano Zampini } 353f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 35488ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 35588ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_phi_B,n_B,pcbddc->local_primal_size,n_B,pcbddc->local_primal_size);CHKERRQ(ierr); 35688ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_phi_B,impMatType);CHKERRQ(ierr); 35725084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_phi_B);CHKERRQ(ierr); 358f4ddd8eeSStefano Zampini } 359f4ddd8eeSStefano Zampini if ( (pcbddc->switch_static || pcbddc->dbg_flag) && !pcbddc->coarse_phi_D ) { 36088ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 36188ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_phi_D,n_D,pcbddc->local_primal_size,n_D,pcbddc->local_primal_size);CHKERRQ(ierr); 36288ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_phi_D,impMatType);CHKERRQ(ierr); 36325084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_phi_D);CHKERRQ(ierr); 36488ebb749SStefano Zampini } 36588ebb749SStefano Zampini 36625084f0cSStefano Zampini if (pcbddc->dbg_flag) { 3678ce42a96SStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,&idx_R_local);CHKERRQ(ierr); 368785e854fSJed Brown ierr = PetscMalloc1(2*pcbddc->local_primal_size,&coarsefunctions_errors);CHKERRQ(ierr); 369785e854fSJed Brown ierr = PetscMalloc1(2*pcbddc->local_primal_size,&constraints_errors);CHKERRQ(ierr); 37088ebb749SStefano Zampini } 37188ebb749SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 372854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 37388ebb749SStefano Zampini 37488ebb749SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 37525084f0cSStefano Zampini 37625084f0cSStefano Zampini /* vertices */ 37788ebb749SStefano Zampini for (i=0;i<n_vertices;i++) { 37845a1bb75SStefano Zampini /* this should not be needed, but MatMult_BAIJ is broken when using compressed row routines */ 379f4ddd8eeSStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); /* TODO: REMOVE IT */ 38088ebb749SStefano Zampini ierr = VecSet(vec1_V,zero);CHKERRQ(ierr); 38188ebb749SStefano Zampini ierr = VecSetValue(vec1_V,i,one,INSERT_VALUES);CHKERRQ(ierr); 38288ebb749SStefano Zampini ierr = VecAssemblyBegin(vec1_V);CHKERRQ(ierr); 38388ebb749SStefano Zampini ierr = VecAssemblyEnd(vec1_V);CHKERRQ(ierr); 38425084f0cSStefano Zampini /* simplified solution of saddle point problem with null rhs on constraints multipliers */ 38588ebb749SStefano Zampini ierr = MatMult(A_RV,vec1_V,pcbddc->vec1_R);CHKERRQ(ierr); 38688ebb749SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 38788ebb749SStefano Zampini ierr = VecScale(pcbddc->vec1_R,m_one);CHKERRQ(ierr); 38888ebb749SStefano Zampini if (n_constraints) { 38988ebb749SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,pcbddc->vec1_R,vec1_C);CHKERRQ(ierr); 39088ebb749SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 39188ebb749SStefano Zampini ierr = VecScale(vec1_C,m_one);CHKERRQ(ierr); 39288ebb749SStefano Zampini } 39388ebb749SStefano Zampini ierr = MatMult(A_VR,pcbddc->vec1_R,vec2_V);CHKERRQ(ierr); 39488ebb749SStefano Zampini ierr = MatMultAdd(A_VV,vec1_V,vec2_V,vec2_V);CHKERRQ(ierr); 39588ebb749SStefano Zampini 39688ebb749SStefano Zampini /* Set values in coarse basis function and subdomain part of coarse_mat */ 39788ebb749SStefano Zampini /* coarse basis functions */ 39888ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 39988ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40088ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40125084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 40288ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_B,n_B,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 40325084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 40488ebb749SStefano Zampini ierr = MatSetValue(pcbddc->coarse_phi_B,idx_V_B[i],i,one,INSERT_VALUES);CHKERRQ(ierr); 4058eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 40688ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40788ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40825084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 40988ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_D,n_D,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 41025084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 41188ebb749SStefano Zampini } 41225084f0cSStefano Zampini /* subdomain contribution to coarse matrix. WARNING -> column major ordering */ 41325084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 41425084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[i*pcbddc->local_primal_size],array,n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 41525084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 41688ebb749SStefano Zampini if (n_constraints) { 41725084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 41825084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[i*pcbddc->local_primal_size+n_vertices],array,n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 41925084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 42088ebb749SStefano Zampini } 42188ebb749SStefano Zampini 42225084f0cSStefano Zampini /* check */ 42325084f0cSStefano Zampini if (pcbddc->dbg_flag) { 42425084f0cSStefano Zampini /* assemble subdomain vector on local nodes */ 42588ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 42625084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 4279f00e9b4SStefano Zampini if (n_R) { 42825084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 4299f00e9b4SStefano Zampini } 43025084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 431e9189074SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],one,INSERT_VALUES);CHKERRQ(ierr); 43225084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 43325084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 43488ebb749SStefano Zampini /* assemble subdomain vector of lagrange multipliers (i.e. primal nodes) */ 43588ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 43625084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 43725084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_vertices,auxindices,array,INSERT_VALUES);CHKERRQ(ierr); 43825084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 43988ebb749SStefano Zampini if (n_constraints) { 44025084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 44125084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_constraints,&auxindices[n_vertices],array,INSERT_VALUES);CHKERRQ(ierr); 44225084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 44388ebb749SStefano Zampini } 44425084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 44525084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 44688ebb749SStefano Zampini ierr = VecScale(pcbddc->vec1_P,m_one);CHKERRQ(ierr); 44788ebb749SStefano Zampini /* check saddle point solution */ 44888ebb749SStefano Zampini ierr = MatMult(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 44988ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 45088ebb749SStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[i]);CHKERRQ(ierr); 45188ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 45225084f0cSStefano Zampini /* shift by the identity matrix */ 45325084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,i,m_one,ADD_VALUES);CHKERRQ(ierr); 45425084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 45525084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 45688ebb749SStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[i]);CHKERRQ(ierr); 45788ebb749SStefano Zampini } 45888ebb749SStefano Zampini } 45988ebb749SStefano Zampini 46025084f0cSStefano Zampini /* constraints */ 46188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 46288ebb749SStefano Zampini ierr = VecSet(vec2_C,zero);CHKERRQ(ierr); 46388ebb749SStefano Zampini ierr = VecSetValue(vec2_C,i,m_one,INSERT_VALUES);CHKERRQ(ierr); 46488ebb749SStefano Zampini ierr = VecAssemblyBegin(vec2_C);CHKERRQ(ierr); 46588ebb749SStefano Zampini ierr = VecAssemblyEnd(vec2_C);CHKERRQ(ierr); 46625084f0cSStefano Zampini /* simplified solution of saddle point problem with null rhs on vertices multipliers */ 46788ebb749SStefano Zampini ierr = MatMult(M1,vec2_C,vec1_C);CHKERRQ(ierr); 46888ebb749SStefano Zampini ierr = MatMult(pcbddc->local_auxmat2,vec1_C,pcbddc->vec1_R);CHKERRQ(ierr); 46988ebb749SStefano Zampini ierr = VecScale(vec1_C,m_one);CHKERRQ(ierr); 47025084f0cSStefano Zampini if (n_vertices) { 47125084f0cSStefano Zampini ierr = MatMult(A_VR,pcbddc->vec1_R,vec2_V);CHKERRQ(ierr); 47225084f0cSStefano Zampini } 47388ebb749SStefano Zampini /* Set values in coarse basis function and subdomain part of coarse_mat */ 47488ebb749SStefano Zampini /* coarse basis functions */ 47525084f0cSStefano Zampini j = i+n_vertices; /* don't touch this! */ 47688ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 47788ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 47888ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 47925084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 48025084f0cSStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_B,n_B,auxindices,1,&j,array,INSERT_VALUES);CHKERRQ(ierr); 48125084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 4828eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 48388ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 48488ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 48525084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 48625084f0cSStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_D,n_D,auxindices,1,&j,array,INSERT_VALUES);CHKERRQ(ierr); 48725084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 48888ebb749SStefano Zampini } 48925084f0cSStefano Zampini /* subdomain contribution to coarse matrix. WARNING -> column major ordering */ 49088ebb749SStefano Zampini if (n_vertices) { 49125084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 49225084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[j*pcbddc->local_primal_size],array,n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 49325084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 49488ebb749SStefano Zampini } 49525084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 49625084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices],array,n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 49725084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 49888ebb749SStefano Zampini 49925084f0cSStefano Zampini if (pcbddc->dbg_flag) { 50088ebb749SStefano Zampini /* assemble subdomain vector on nodes */ 50188ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 50225084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 5039f00e9b4SStefano Zampini if (n_R) { 50425084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 5059f00e9b4SStefano Zampini } 50625084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 50725084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 50825084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 50988ebb749SStefano Zampini /* assemble subdomain vector of lagrange multipliers */ 51088ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 51188ebb749SStefano Zampini if (n_vertices) { 51225084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 51325084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_vertices,auxindices,array,INSERT_VALUES);CHKERRQ(ierr); 51425084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 51588ebb749SStefano Zampini } 51625084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 51725084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_constraints,&auxindices[n_vertices],array,INSERT_VALUES);CHKERRQ(ierr); 51825084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 51925084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 52025084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 52125084f0cSStefano Zampini ierr = VecScale(pcbddc->vec1_P,m_one);CHKERRQ(ierr); 52288ebb749SStefano Zampini /* check saddle point solution */ 52388ebb749SStefano Zampini ierr = MatMult(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 52488ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 52525084f0cSStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[j]);CHKERRQ(ierr); 52688ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 52725084f0cSStefano Zampini /* shift by the identity matrix */ 52825084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,j,m_one,ADD_VALUES);CHKERRQ(ierr); 52925084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 53025084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 53125084f0cSStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[j]);CHKERRQ(ierr); 53288ebb749SStefano Zampini } 53388ebb749SStefano Zampini } 53425084f0cSStefano Zampini /* call assembling routines for local coarse basis */ 53588ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_phi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53688ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_phi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 53888ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_phi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53988ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_phi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54088ebb749SStefano Zampini } 54125084f0cSStefano Zampini 54288ebb749SStefano Zampini /* compute other basis functions for non-symmetric problems */ 54351cb09feSStefano Zampini /*ierr = MatIsSymmetric(pc->pmat,0.,&pcbddc->issym);CHKERRQ(ierr);*/ 54451cb09feSStefano Zampini { /* this is a temporary workaround since seqbaij matrices does not have support for symmetry checking */ 54551cb09feSStefano Zampini PetscBool setsym; 54651cb09feSStefano Zampini ierr = MatIsSymmetricKnown(pc->pmat,&setsym,&pcbddc->issym);CHKERRQ(ierr); 54751cb09feSStefano Zampini if (!setsym) pcbddc->issym = PETSC_FALSE; 54851cb09feSStefano Zampini } 54951cb09feSStefano Zampini 550b9d89cd5SStefano Zampini if (!pcbddc->issym) { 551f4ddd8eeSStefano Zampini if (!pcbddc->coarse_psi_B) { 55288ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 55388ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_psi_B,n_B,pcbddc->local_primal_size,n_B,pcbddc->local_primal_size);CHKERRQ(ierr); 55488ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_psi_B,impMatType);CHKERRQ(ierr); 55525084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_psi_B);CHKERRQ(ierr); 556f4ddd8eeSStefano Zampini } 557f4ddd8eeSStefano Zampini if ( (pcbddc->switch_static || pcbddc->dbg_flag) && !pcbddc->coarse_psi_D) { 55888ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 55988ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_psi_D,n_D,pcbddc->local_primal_size,n_D,pcbddc->local_primal_size);CHKERRQ(ierr); 56088ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_psi_D,impMatType);CHKERRQ(ierr); 56125084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_psi_D);CHKERRQ(ierr); 56288ebb749SStefano Zampini } 56388ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 56488ebb749SStefano Zampini if (n_constraints) { 56588ebb749SStefano Zampini ierr = VecSet(vec1_C,zero);CHKERRQ(ierr); 56688ebb749SStefano Zampini for (j=0;j<n_constraints;j++) { 56725084f0cSStefano Zampini ierr = VecSetValue(vec1_C,j,coarse_submat_vals[(j+n_vertices)*pcbddc->local_primal_size+i],INSERT_VALUES);CHKERRQ(ierr); 56888ebb749SStefano Zampini } 56925084f0cSStefano Zampini ierr = VecAssemblyBegin(vec1_C);CHKERRQ(ierr); 57025084f0cSStefano Zampini ierr = VecAssemblyEnd(vec1_C);CHKERRQ(ierr); 57188ebb749SStefano Zampini } 57288ebb749SStefano Zampini if (i<n_vertices) { 57388ebb749SStefano Zampini ierr = VecSet(vec1_V,zero);CHKERRQ(ierr); 57488ebb749SStefano Zampini ierr = VecSetValue(vec1_V,i,m_one,INSERT_VALUES);CHKERRQ(ierr); 57588ebb749SStefano Zampini ierr = VecAssemblyBegin(vec1_V);CHKERRQ(ierr); 57688ebb749SStefano Zampini ierr = VecAssemblyEnd(vec1_V);CHKERRQ(ierr); 57788ebb749SStefano Zampini ierr = MatMultTranspose(A_VR,vec1_V,pcbddc->vec1_R);CHKERRQ(ierr); 57888ebb749SStefano Zampini if (n_constraints) { 57988ebb749SStefano Zampini ierr = MatMultTransposeAdd(C_CR,vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 58088ebb749SStefano Zampini } 58188ebb749SStefano Zampini } else { 58288ebb749SStefano Zampini ierr = MatMultTranspose(C_CR,vec1_C,pcbddc->vec1_R);CHKERRQ(ierr); 58388ebb749SStefano Zampini } 58488ebb749SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 58588ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 58688ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58788ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58825084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 58988ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_psi_B,n_B,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 59025084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 59188ebb749SStefano Zampini if (i<n_vertices) { 59288ebb749SStefano Zampini ierr = MatSetValue(pcbddc->coarse_psi_B,idx_V_B[i],i,one,INSERT_VALUES);CHKERRQ(ierr); 59388ebb749SStefano Zampini } 5948eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 59588ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 59688ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 59725084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 59888ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_psi_D,n_D,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 59925084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 60088ebb749SStefano Zampini } 60188ebb749SStefano Zampini 60225084f0cSStefano Zampini if (pcbddc->dbg_flag) { 60388ebb749SStefano Zampini /* assemble subdomain vector on nodes */ 60488ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 60525084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 6069f00e9b4SStefano Zampini if (n_R) { 60725084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 6089f00e9b4SStefano Zampini } 60925084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 61025084f0cSStefano Zampini if (i<n_vertices) { 611e9189074SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],one,INSERT_VALUES);CHKERRQ(ierr); 61288ebb749SStefano Zampini } 61325084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 61425084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 61525084f0cSStefano Zampini /* assemble subdomain vector of lagrange multipliers */ 61625084f0cSStefano Zampini for (j=0;j<pcbddc->local_primal_size;j++) { 61725084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,j,-coarse_submat_vals[j*pcbddc->local_primal_size+i],INSERT_VALUES);CHKERRQ(ierr); 61825084f0cSStefano Zampini } 61925084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 62025084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 62188ebb749SStefano Zampini /* check saddle point solution */ 62288ebb749SStefano Zampini ierr = MatMultTranspose(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 62388ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 62488ebb749SStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[i+pcbddc->local_primal_size]);CHKERRQ(ierr); 62588ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 62625084f0cSStefano Zampini /* shift by the identity matrix */ 62725084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,i,m_one,ADD_VALUES);CHKERRQ(ierr); 62825084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 62925084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 63088ebb749SStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[i+pcbddc->local_primal_size]);CHKERRQ(ierr); 63188ebb749SStefano Zampini } 63288ebb749SStefano Zampini } 63388ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_psi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63488ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_psi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6358eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 63688ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_psi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63788ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_psi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63888ebb749SStefano Zampini } 639c0553b1fSStefano Zampini unsymmetric_check = PETSC_TRUE; 640c0553b1fSStefano Zampini } else { /* take references to already computed coarse basis */ 641c0553b1fSStefano Zampini unsymmetric_check = PETSC_FALSE; 642c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 643c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 644c0553b1fSStefano Zampini if (pcbddc->coarse_phi_D) { 645c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 646c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 647c0553b1fSStefano Zampini } 64888ebb749SStefano Zampini } 64988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 65088ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 65188ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 65288ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 65325084f0cSStefano Zampini if (pcbddc->dbg_flag) { 65488ebb749SStefano Zampini Mat coarse_sub_mat; 65525084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 65688ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 65788ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 65888ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 65988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 66088ebb749SStefano Zampini PetscReal real_value; 66188ebb749SStefano Zampini 66288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 66388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 66488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 66588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 66688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 66788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 668c0553b1fSStefano Zampini if (unsymmetric_check) { 66988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 67088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 67188ebb749SStefano Zampini } 67288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 67388ebb749SStefano Zampini 67425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 67525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat and local basis functions\n");CHKERRQ(ierr); 67625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 677c0553b1fSStefano Zampini if (unsymmetric_check) { 67888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 67988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 68088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 68288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 68388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 68588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 68688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 68888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 68988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 69088ebb749SStefano Zampini } else { 69188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 69288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 69388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 69488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 69588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 69688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 69788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 69888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 69988ebb749SStefano Zampini } 70088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70388ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 70481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70588ebb749SStefano Zampini ierr = MatNorm(TM1,NORM_INFINITY,&real_value);CHKERRQ(ierr); 7060fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 70725084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"----------------------------------\n");CHKERRQ(ierr); 70825084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d \n",PetscGlobalRank);CHKERRQ(ierr); 70925084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"matrix error = % 1.14e\n",real_value);CHKERRQ(ierr); 71025084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"coarse functions (phi) errors\n");CHKERRQ(ierr); 71188ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 71225084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i,coarsefunctions_errors[i]);CHKERRQ(ierr); 71388ebb749SStefano Zampini } 71425084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"constraints (phi) errors\n");CHKERRQ(ierr); 71588ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 71625084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i,constraints_errors[i]);CHKERRQ(ierr); 71788ebb749SStefano Zampini } 718c0553b1fSStefano Zampini if (unsymmetric_check) { 71925084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"coarse functions (psi) errors\n");CHKERRQ(ierr); 72088ebb749SStefano Zampini for (i=pcbddc->local_primal_size;i<2*pcbddc->local_primal_size;i++) { 72125084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i-pcbddc->local_primal_size,coarsefunctions_errors[i]);CHKERRQ(ierr); 72288ebb749SStefano Zampini } 72325084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"constraints (psi) errors\n");CHKERRQ(ierr); 72488ebb749SStefano Zampini for (i=pcbddc->local_primal_size;i<2*pcbddc->local_primal_size;i++) { 72525084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i-pcbddc->local_primal_size,constraints_errors[i]);CHKERRQ(ierr); 72688ebb749SStefano Zampini } 72788ebb749SStefano Zampini } 72825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 72988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 73088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 73188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 73288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 73388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 73488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 73588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 73688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 73788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 73888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 739c0553b1fSStefano Zampini if (unsymmetric_check) { 74088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 74188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 74288ebb749SStefano Zampini } 74388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 7448ce42a96SStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,&idx_R_local);CHKERRQ(ierr); 74588ebb749SStefano Zampini ierr = PetscFree(coarsefunctions_errors);CHKERRQ(ierr); 74688ebb749SStefano Zampini ierr = PetscFree(constraints_errors);CHKERRQ(ierr); 74788ebb749SStefano Zampini } 74888ebb749SStefano Zampini /* free memory */ 74988ebb749SStefano Zampini if (n_vertices) { 75088ebb749SStefano Zampini ierr = VecDestroy(&vec1_V);CHKERRQ(ierr); 75188ebb749SStefano Zampini ierr = VecDestroy(&vec2_V);CHKERRQ(ierr); 75288ebb749SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 75388ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 75488ebb749SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 75588ebb749SStefano Zampini } 75688ebb749SStefano Zampini if (n_constraints) { 75788ebb749SStefano Zampini ierr = VecDestroy(&vec1_C);CHKERRQ(ierr); 75888ebb749SStefano Zampini ierr = VecDestroy(&vec2_C);CHKERRQ(ierr); 75988ebb749SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 76088ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 76188ebb749SStefano Zampini } 76288ebb749SStefano Zampini ierr = PetscFree(auxindices);CHKERRQ(ierr); 7638629588bSStefano Zampini /* get back data */ 7648629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 76588ebb749SStefano Zampini PetscFunctionReturn(0); 76688ebb749SStefano Zampini } 76788ebb749SStefano Zampini 76888ebb749SStefano Zampini #undef __FUNCT__ 769aa0d41d4SStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalMatrices" 770aa0d41d4SStefano Zampini PetscErrorCode PCBDDCSetUpLocalMatrices(PC pc) 771aa0d41d4SStefano Zampini { 772aa0d41d4SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 773aa0d41d4SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 774aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 77522d5777bSStefano Zampini PetscBool issbaij,isseqaij; 776aa0d41d4SStefano Zampini /* manage repeated solves */ 777aa0d41d4SStefano Zampini MatReuse reuse; 778aa0d41d4SStefano Zampini PetscErrorCode ierr; 779aa0d41d4SStefano Zampini 780aa0d41d4SStefano Zampini PetscFunctionBegin; 781b9b85e73SStefano Zampini if ( (pcbddc->use_change_of_basis && !pcbddc->ChangeOfBasisMatrix) || (pcbddc->user_ChangeOfBasisMatrix && !pcbddc->ChangeOfBasisMatrix) ) { 782f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Change of basis matrix has not been created"); 783f4ddd8eeSStefano Zampini } 784aa0d41d4SStefano Zampini /* get mat flags */ 785aa0d41d4SStefano Zampini reuse = MAT_INITIAL_MATRIX; 786aa0d41d4SStefano Zampini if (pc->setupcalled) { 787d1e9a80fSBarry Smith if (pc->flag == SAME_NONZERO_PATTERN) { 788aa0d41d4SStefano Zampini reuse = MAT_REUSE_MATRIX; 789aa0d41d4SStefano Zampini } else { 790aa0d41d4SStefano Zampini reuse = MAT_INITIAL_MATRIX; 791aa0d41d4SStefano Zampini } 792aa0d41d4SStefano Zampini } 793aa0d41d4SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 794aa0d41d4SStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 795aa0d41d4SStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 796aa0d41d4SStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 797aa0d41d4SStefano Zampini ierr = MatDestroy(&pcis->A_BB);CHKERRQ(ierr); 798aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 799aa0d41d4SStefano Zampini } 800aa0d41d4SStefano Zampini 801aa0d41d4SStefano Zampini /* transform local matrices if needed */ 802b9b85e73SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 803906d46d4SStefano Zampini Mat *change_mat_all; 804906d46d4SStefano Zampini IS is_local,is_global; 805906d46d4SStefano Zampini PetscBool sorted; 806906d46d4SStefano Zampini PetscInt *idxs_perm; 807aa0d41d4SStefano Zampini 808906d46d4SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_local);CHKERRQ(ierr); 809906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 810906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 811906d46d4SStefano Zampini ierr = ISSorted(is_global,&sorted);CHKERRQ(ierr); 812906d46d4SStefano Zampini if (!sorted) { 813906d46d4SStefano Zampini const PetscInt *idxs; 814906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 815aa0d41d4SStefano Zampini 816906d46d4SStefano Zampini ierr = PetscMalloc1(pcis->n,&idxs_perm);CHKERRQ(ierr); 817906d46d4SStefano Zampini ierr = PetscMalloc1(pcis->n,&idxs_sorted);CHKERRQ(ierr); 818906d46d4SStefano Zampini for (i=0;i<pcis->n;i++) { 819906d46d4SStefano Zampini idxs_perm[i] = i; 820aa0d41d4SStefano Zampini } 821906d46d4SStefano Zampini ierr = ISGetIndices(is_global,&idxs);CHKERRQ(ierr); 822906d46d4SStefano Zampini ierr = PetscSortIntWithPermutation(pcis->n,idxs,idxs_perm);CHKERRQ(ierr); 823906d46d4SStefano Zampini for (i=0;i<pcis->n;i++) { 824906d46d4SStefano Zampini idxs_sorted[i] = idxs[idxs_perm[i]]; 825aa0d41d4SStefano Zampini } 826906d46d4SStefano Zampini ierr = ISRestoreIndices(is_global,&idxs);CHKERRQ(ierr); 827906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 828906d46d4SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcis->n,idxs_sorted,PETSC_OWN_POINTER,&is_global);CHKERRQ(ierr); 829aa0d41d4SStefano Zampini } 830906d46d4SStefano Zampini 831906d46d4SStefano Zampini /* get change of basis on the whole set of local dofs */ 832906d46d4SStefano Zampini ierr = MatGetSubMatrices(pcbddc->ChangeOfBasisMatrix,1,&is_global,&is_global,MAT_INITIAL_MATRIX,&change_mat_all);CHKERRQ(ierr); 833906d46d4SStefano Zampini 834906d46d4SStefano Zampini if (!sorted) { 835906d46d4SStefano Zampini Mat new_mat; 836906d46d4SStefano Zampini IS is_perm; 837906d46d4SStefano Zampini PetscInt *idxs,i; 838906d46d4SStefano Zampini 839906d46d4SStefano Zampini ierr = PetscMalloc1(pcis->n,&idxs);CHKERRQ(ierr); 840906d46d4SStefano Zampini for (i=0;i<pcis->n;i++) { 841906d46d4SStefano Zampini idxs[idxs_perm[i]] = i; 842906d46d4SStefano Zampini } 843906d46d4SStefano Zampini ierr = PetscFree(idxs_perm);CHKERRQ(ierr); 844906d46d4SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcis->n,idxs,PETSC_OWN_POINTER,&is_perm);CHKERRQ(ierr); 845906d46d4SStefano Zampini ierr = ISSetPermutation(is_perm);CHKERRQ(ierr); 846906d46d4SStefano Zampini ierr = MatPermute(change_mat_all[0],is_perm,is_perm,&new_mat);CHKERRQ(ierr); 847906d46d4SStefano Zampini ierr = MatDestroy(&change_mat_all[0]);CHKERRQ(ierr); 848906d46d4SStefano Zampini change_mat_all[0] = new_mat; 849906d46d4SStefano Zampini ierr = ISDestroy(&is_perm);CHKERRQ(ierr); 850906d46d4SStefano Zampini } 851906d46d4SStefano Zampini 852906d46d4SStefano Zampini /* check */ 853906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 854906d46d4SStefano Zampini Vec x,x_change; 855906d46d4SStefano Zampini PetscReal error; 856906d46d4SStefano Zampini 857906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 858906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 859906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 860906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 861906d46d4SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 862906d46d4SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 863906d46d4SStefano Zampini ierr = MatMult(change_mat_all[0],pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 864906d46d4SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 865906d46d4SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 866906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 867906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 868906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 869906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 870906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 871906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 872906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 873906d46d4SStefano Zampini } 874906d46d4SStefano Zampini 87522d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 8769b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 87722d5777bSStefano Zampini if (isseqaij) { 878906d46d4SStefano Zampini ierr = MatPtAP(matis->A,change_mat_all[0],reuse,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 879aa0d41d4SStefano Zampini } else { 880aa0d41d4SStefano Zampini Mat work_mat; 881aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 882906d46d4SStefano Zampini ierr = MatPtAP(work_mat,change_mat_all[0],reuse,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 883aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 884aa0d41d4SStefano Zampini } 88545a1bb75SStefano Zampini /* 88645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 88745a1bb75SStefano Zampini ierr = MatView(change_mat_all,(PetscViewer)0);CHKERRQ(ierr); 88845a1bb75SStefano Zampini */ 889906d46d4SStefano Zampini ierr = MatDestroyMatrices(1,&change_mat_all);CHKERRQ(ierr); 890906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 891aa0d41d4SStefano Zampini } else { 892aa0d41d4SStefano Zampini /* without change of basis, the local matrix is unchanged */ 893aa0d41d4SStefano Zampini if (!pcbddc->local_mat) { 894aa0d41d4SStefano Zampini ierr = PetscObjectReference((PetscObject)matis->A);CHKERRQ(ierr); 895aa0d41d4SStefano Zampini pcbddc->local_mat = matis->A; 896aa0d41d4SStefano Zampini } 897aa0d41d4SStefano Zampini } 898aa0d41d4SStefano Zampini 899aa0d41d4SStefano Zampini /* get submatrices */ 900aa0d41d4SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcis->is_I_local,pcis->is_I_local,reuse,&pcis->A_II);CHKERRQ(ierr); 9019577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcis->is_B_local,pcis->is_B_local,reuse,&pcis->A_BB);CHKERRQ(ierr); 9029577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 9039577ea80SStefano Zampini if (!issbaij) { 904aa0d41d4SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcis->is_I_local,pcis->is_B_local,reuse,&pcis->A_IB);CHKERRQ(ierr); 905aa0d41d4SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcis->is_B_local,pcis->is_I_local,reuse,&pcis->A_BI);CHKERRQ(ierr); 9069577ea80SStefano Zampini } else { 9079577ea80SStefano Zampini Mat newmat; 9089577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 9099577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcis->is_I_local,pcis->is_B_local,reuse,&pcis->A_IB);CHKERRQ(ierr); 9109577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcis->is_B_local,pcis->is_I_local,reuse,&pcis->A_BI);CHKERRQ(ierr); 9119577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 9129577ea80SStefano Zampini } 913aa0d41d4SStefano Zampini PetscFunctionReturn(0); 914aa0d41d4SStefano Zampini } 915aa0d41d4SStefano Zampini 916aa0d41d4SStefano Zampini #undef __FUNCT__ 917a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 9188ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 919a64d13efSStefano Zampini { 920a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 921a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 9228ce42a96SStefano Zampini IS is_aux1,is_aux2; 923e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 9243a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 9253a50541eSStefano Zampini PetscInt vbs,bs; 9264641a718SStefano Zampini PetscBT bitmask; 927a64d13efSStefano Zampini PetscErrorCode ierr; 928a64d13efSStefano Zampini 929a64d13efSStefano Zampini PetscFunctionBegin; 930b23d619eSStefano Zampini /* 931b23d619eSStefano Zampini No need to setup local scatters if 932b23d619eSStefano Zampini - primal space is unchanged 933b23d619eSStefano Zampini AND 934b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 935b23d619eSStefano Zampini AND 936b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 937b23d619eSStefano Zampini */ 938b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 939f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 940f4ddd8eeSStefano Zampini } 941f4ddd8eeSStefano Zampini /* destroy old objects */ 942f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 943f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 944f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 945a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 946a64d13efSStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 947e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 9484641a718SStefano Zampini /* create auxiliary bitmask */ 9494641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 9504641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 951e9189074SStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->primal_indices_local_idxs[i]);CHKERRQ(ierr); 9524641a718SStefano Zampini } 9533a50541eSStefano Zampini 954a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 955854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 956a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 9574641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 958a64d13efSStefano Zampini idx_R_local[n_R] = i; 959a64d13efSStefano Zampini n_R++; 960a64d13efSStefano Zampini } 961a64d13efSStefano Zampini } 9623a50541eSStefano Zampini 9633a50541eSStefano Zampini /* Block code */ 9643a50541eSStefano Zampini vbs = 1; 9653a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 9663a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 9673a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 9683a50541eSStefano Zampini PetscInt *vary; 9693a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 970785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 9713a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 972e9189074SStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->primal_indices_local_idxs[i]/bs]++; 9733a50541eSStefano Zampini for (i=0; i<n_vertices; i++) { 9743a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 9753a50541eSStefano Zampini is_blocked = PETSC_FALSE; 9763a50541eSStefano Zampini break; 9773a50541eSStefano Zampini } 9783a50541eSStefano Zampini } 9793a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 9803a50541eSStefano Zampini vbs = bs; 9813a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 9823a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 9833a50541eSStefano Zampini } 9843a50541eSStefano Zampini } 9853a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 9863a50541eSStefano Zampini } 9873a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 9883a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 989a64d13efSStefano Zampini 990a64d13efSStefano Zampini /* print some info if requested */ 991a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 992a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 993a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 9940fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 995a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 996a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 9973a50541eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %d, v_size = %d, constraints = %d, local_primal_size = %d\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices,pcbddc->local_primal_size);CHKERRQ(ierr); 998a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"pcbddc->n_vertices = %d, pcbddc->n_constraints = %d\n",pcbddc->n_vertices,pcbddc->n_constraints);CHKERRQ(ierr); 999a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1000a64d13efSStefano Zampini } 1001a64d13efSStefano Zampini 1002a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 10033a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1004854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1005854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1006a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 10074641a718SStefano Zampini for (i=0; i<n_D; i++) { 10084641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 10094641a718SStefano Zampini } 1010a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1011a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 10124641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 10134641a718SStefano Zampini aux_array1[j++] = i; 1014a64d13efSStefano Zampini } 1015a64d13efSStefano Zampini } 1016a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1017a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1018a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 10194641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 10204641a718SStefano Zampini aux_array2[j++] = i; 1021a64d13efSStefano Zampini } 1022a64d13efSStefano Zampini } 1023a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1024a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1025a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1026a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1027a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1028a64d13efSStefano Zampini 10298eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1030785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1031a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 10324641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 10334641a718SStefano Zampini aux_array1[j++] = i; 1034a64d13efSStefano Zampini } 1035a64d13efSStefano Zampini } 1036a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1037a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1038a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1039a64d13efSStefano Zampini } 10404641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 10413a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1042a64d13efSStefano Zampini PetscFunctionReturn(0); 1043a64d13efSStefano Zampini } 1044a64d13efSStefano Zampini 1045304d26faSStefano Zampini 1046304d26faSStefano Zampini #undef __FUNCT__ 1047304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 10488ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc) 1049304d26faSStefano Zampini { 1050304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1051304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1052304d26faSStefano Zampini PC pc_temp; 1053304d26faSStefano Zampini Mat A_RR; 1054f4ddd8eeSStefano Zampini MatReuse reuse; 1055304d26faSStefano Zampini PetscScalar m_one = -1.0; 1056304d26faSStefano Zampini PetscReal value; 1057af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 10589577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1059304d26faSStefano Zampini PetscErrorCode ierr; 1060e604994aSStefano Zampini /* prefixes stuff */ 1061312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1062e604994aSStefano Zampini size_t len; 1063304d26faSStefano Zampini 1064304d26faSStefano Zampini PetscFunctionBegin; 1065304d26faSStefano Zampini 1066e604994aSStefano Zampini /* compute prefixes */ 1067e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1068e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1069e604994aSStefano Zampini if (!pcbddc->current_level) { 1070e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1071e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1072e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1073e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1074e604994aSStefano Zampini } else { 1075e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1076312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1077e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1078e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1079312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1080312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 108134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 108234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1083e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1084e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1085e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1086e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1087e604994aSStefano Zampini } 1088e604994aSStefano Zampini 1089304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1090ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 10918ce42a96SStefano Zampini ierr = ISGetSize(pcis->is_I_local,&n_D);CHKERRQ(ierr); 1092304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1093304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1094304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1095304d26faSStefano Zampini /* default */ 1096304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1097e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 10989577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1099304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 11009577ea80SStefano Zampini if (issbaij) { 11019577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 11029577ea80SStefano Zampini } else { 1103304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 11049577ea80SStefano Zampini } 1105304d26faSStefano Zampini /* Allow user's customization */ 1106304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1107304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1108304d26faSStefano Zampini } 1109d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1110304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1111304d26faSStefano Zampini if (!n_D) { 1112304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1113304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1114304d26faSStefano Zampini } 1115304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1116304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1117304d26faSStefano Zampini /* set ksp_D into pcis data */ 1118304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1119304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1120304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1121304d26faSStefano Zampini 1122304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1123f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 11248ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1125f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1126f4ddd8eeSStefano Zampini PetscInt nn_R; 112781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1128f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1129f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1130f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1131f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1132f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1133f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1134f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1135727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1136f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1137f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1138f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1139f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1140f4ddd8eeSStefano Zampini } 1141f4ddd8eeSStefano Zampini } 1142f4ddd8eeSStefano Zampini /* last check */ 1143d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1144f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1145f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1146f4ddd8eeSStefano Zampini } 1147f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1148f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1149f4ddd8eeSStefano Zampini } 1150f4ddd8eeSStefano Zampini /* extract A_RR */ 1151af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1152af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 11533a50541eSStefano Zampini if (ibs != mbs) { 1154af732b37SStefano Zampini Mat newmat; 1155af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1156f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1157af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1158af732b37SStefano Zampini } else { 1159f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1160af732b37SStefano Zampini } 1161f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1162304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1163304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1164304d26faSStefano Zampini /* default */ 1165304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1166e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1167304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 11689577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 11699577ea80SStefano Zampini if (issbaij) { 11709577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 11719577ea80SStefano Zampini } else { 1172304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 11739577ea80SStefano Zampini } 1174304d26faSStefano Zampini /* Allow user's customization */ 1175304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1176304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1177304d26faSStefano Zampini } 1178d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1179304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1180304d26faSStefano Zampini if (!n_R) { 1181304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1182304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1183304d26faSStefano Zampini } 1184304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1185304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1186304d26faSStefano Zampini 1187304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 11880fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1189304d26faSStefano Zampini /* Dirichlet */ 11900fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 11910fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 11920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 11930fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 11940fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1195304d26faSStefano Zampini /* need to be adapted? */ 1196b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1197b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1198b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1199304d26faSStefano Zampini /* print info */ 1200304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1201304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 12020fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1203304d26faSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1204304d26faSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Checking solution of Dirichlet and Neumann problems\n");CHKERRQ(ierr); 1205e604994aSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Dirichlet solve (%s) = % 1.14e \n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_D))->prefix,value);CHKERRQ(ierr); 1206304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1207304d26faSStefano Zampini } 1208b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 12098ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1210304d26faSStefano Zampini } 1211304d26faSStefano Zampini 1212304d26faSStefano Zampini /* Neumann */ 12130fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 12140fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 12150fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 12160fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 12170fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1218304d26faSStefano Zampini /* need to be adapted? */ 1219b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1220b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1221304d26faSStefano Zampini /* print info */ 1222304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1223e604994aSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Neumann solve (%s) = % 1.14e \n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_R))->prefix,value);CHKERRQ(ierr); 1224304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1225304d26faSStefano Zampini } 1226b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 12278ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1228304d26faSStefano Zampini } 12290fccc4e9SStefano Zampini } 1230304d26faSStefano Zampini /* free Neumann problem's matrix */ 1231304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1232304d26faSStefano Zampini PetscFunctionReturn(0); 1233304d26faSStefano Zampini } 1234304d26faSStefano Zampini 1235304d26faSStefano Zampini #undef __FUNCT__ 1236ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 123720c7b377SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec rhs, Vec sol, Vec work, PetscBool applytranspose) 1238674ae819SStefano Zampini { 1239674ae819SStefano Zampini PetscErrorCode ierr; 1240674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1241674ae819SStefano Zampini 1242674ae819SStefano Zampini PetscFunctionBegin; 124320c7b377SStefano Zampini if (applytranspose) { 1244674ae819SStefano Zampini if (pcbddc->local_auxmat1) { 124520c7b377SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,rhs,work);CHKERRQ(ierr); 124620c7b377SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,work,rhs,rhs);CHKERRQ(ierr); 124720c7b377SStefano Zampini } 124820c7b377SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 124920c7b377SStefano Zampini } else { 125020c7b377SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 125120c7b377SStefano Zampini if (pcbddc->local_auxmat1) { 125220c7b377SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,sol,work);CHKERRQ(ierr); 125320c7b377SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,work,sol,sol);CHKERRQ(ierr); 125420c7b377SStefano Zampini } 1255674ae819SStefano Zampini } 1256674ae819SStefano Zampini PetscFunctionReturn(0); 1257674ae819SStefano Zampini } 1258674ae819SStefano Zampini 1259dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1260674ae819SStefano Zampini #undef __FUNCT__ 1261674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1262dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1263674ae819SStefano Zampini { 1264674ae819SStefano Zampini PetscErrorCode ierr; 1265674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1266674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1267674ae819SStefano Zampini const PetscScalar zero = 0.0; 1268674ae819SStefano Zampini 1269674ae819SStefano Zampini PetscFunctionBegin; 1270dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1271dc359a40SStefano Zampini if (applytranspose) { 1272674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 12738eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1274dc359a40SStefano Zampini } else { 1275674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1276674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 127715aaf578SStefano Zampini } 127812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 127912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 128012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 128112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 128212edc857SStefano Zampini 12839f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 128412edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 128512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 128612edc857SStefano Zampini if (applytranspose) { 128712edc857SStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,NULL,NULL);CHKERRQ(ierr); 128812edc857SStefano Zampini } else { 128912edc857SStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,NULL,NULL);CHKERRQ(ierr); 129012edc857SStefano Zampini } 129112edc857SStefano Zampini } 1292674ae819SStefano Zampini 1293674ae819SStefano Zampini /* Local solution on R nodes */ 12949f00e9b4SStefano Zampini if (pcis->n) { 1295674ae819SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 1296674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1297674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12988eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1299674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1300674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1301674ae819SStefano Zampini } 130220c7b377SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcbddc->vec1_R,pcbddc->vec2_R,pcbddc->vec1_C,applytranspose);CHKERRQ(ierr); 1303674ae819SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 1304674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1305674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 13068eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1307674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1308674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1309674ae819SStefano Zampini } 13109f00e9b4SStefano Zampini } 1311674ae819SStefano Zampini 13129f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 13139f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 131412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1315674ae819SStefano Zampini 1316674ae819SStefano Zampini /* Sum contributions from two levels */ 1317dc359a40SStefano Zampini if (applytranspose) { 1318dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1319dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1320dc359a40SStefano Zampini } else { 1321674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 13228eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1323dc359a40SStefano Zampini } 1324674ae819SStefano Zampini PetscFunctionReturn(0); 1325674ae819SStefano Zampini } 1326674ae819SStefano Zampini 132712edc857SStefano Zampini /* TODO: the following two function can be optimized using VecPlaceArray whenever possible and using overlap flag */ 1328674ae819SStefano Zampini #undef __FUNCT__ 1329674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 133012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1331674ae819SStefano Zampini { 1332674ae819SStefano Zampini PetscErrorCode ierr; 1333674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 133412edc857SStefano Zampini PetscScalar *array,*array2; 133512edc857SStefano Zampini Vec from,to; 1336674ae819SStefano Zampini 1337674ae819SStefano Zampini PetscFunctionBegin; 133812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 133912edc857SStefano Zampini from = pcbddc->coarse_vec; 134012edc857SStefano Zampini to = pcbddc->vec1_P; 134112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 134212edc857SStefano Zampini Vec tvec; 134312edc857SStefano Zampini PetscInt lsize; 134412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 134512edc857SStefano Zampini ierr = VecGetLocalSize(tvec,&lsize);CHKERRQ(ierr); 134612edc857SStefano Zampini ierr = VecGetArrayRead(tvec,(const PetscScalar**)&array);CHKERRQ(ierr); 134712edc857SStefano Zampini ierr = VecGetArray(from,&array2);CHKERRQ(ierr); 134812edc857SStefano Zampini ierr = PetscMemcpy(array2,array,lsize*sizeof(PetscScalar));CHKERRQ(ierr); 134912edc857SStefano Zampini ierr = VecRestoreArrayRead(tvec,(const PetscScalar**)&array);CHKERRQ(ierr); 135012edc857SStefano Zampini ierr = VecRestoreArray(from,&array2);CHKERRQ(ierr); 135112edc857SStefano Zampini } 135212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 135312edc857SStefano Zampini from = pcbddc->vec1_P; 135412edc857SStefano Zampini to = pcbddc->coarse_vec; 135512edc857SStefano Zampini } 135612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1357674ae819SStefano Zampini PetscFunctionReturn(0); 1358674ae819SStefano Zampini } 1359674ae819SStefano Zampini 1360674ae819SStefano Zampini #undef __FUNCT__ 1361674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 136212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1363674ae819SStefano Zampini { 1364674ae819SStefano Zampini PetscErrorCode ierr; 1365674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 136612edc857SStefano Zampini PetscScalar *array,*array2; 136712edc857SStefano Zampini Vec from,to; 1368674ae819SStefano Zampini 1369674ae819SStefano Zampini PetscFunctionBegin; 137012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 137112edc857SStefano Zampini from = pcbddc->coarse_vec; 137212edc857SStefano Zampini to = pcbddc->vec1_P; 137312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 137412edc857SStefano Zampini from = pcbddc->vec1_P; 137512edc857SStefano Zampini to = pcbddc->coarse_vec; 137612edc857SStefano Zampini } 137712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 137812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 137912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 138012edc857SStefano Zampini Vec tvec; 138112edc857SStefano Zampini PetscInt lsize; 138212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 138312edc857SStefano Zampini ierr = VecGetLocalSize(tvec,&lsize);CHKERRQ(ierr); 138412edc857SStefano Zampini ierr = VecGetArrayRead(to,(const PetscScalar**)&array);CHKERRQ(ierr); 138512edc857SStefano Zampini ierr = VecGetArray(tvec,&array2);CHKERRQ(ierr); 138612edc857SStefano Zampini ierr = PetscMemcpy(array2,array,lsize*sizeof(PetscScalar));CHKERRQ(ierr); 138712edc857SStefano Zampini ierr = VecRestoreArrayRead(to,(const PetscScalar**)&array);CHKERRQ(ierr); 138812edc857SStefano Zampini ierr = VecRestoreArray(tvec,&array2);CHKERRQ(ierr); 138912edc857SStefano Zampini } 139012edc857SStefano Zampini } 1391674ae819SStefano Zampini PetscFunctionReturn(0); 1392674ae819SStefano Zampini } 1393674ae819SStefano Zampini 1394984c4197SStefano Zampini /* uncomment for testing purposes */ 1395984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1396674ae819SStefano Zampini #undef __FUNCT__ 1397674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1398674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1399674ae819SStefano Zampini { 1400674ae819SStefano Zampini PetscErrorCode ierr; 1401674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1402674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1403674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1404984c4197SStefano Zampini /* constraint and (optionally) change of basis matrix implemented as SeqAIJ */ 1405674ae819SStefano Zampini MatType impMatType=MATSEQAIJ; 1406984c4197SStefano Zampini /* one and zero */ 1407984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1408984c4197SStefano Zampini /* space to store constraints and their local indices */ 1409984c4197SStefano Zampini PetscScalar *temp_quadrature_constraint; 1410984c4197SStefano Zampini PetscInt *temp_indices,*temp_indices_to_constraint,*temp_indices_to_constraint_B; 1411984c4197SStefano Zampini /* iterators */ 1412984c4197SStefano Zampini PetscInt i,j,k,total_counts,temp_start_ptr; 1413984c4197SStefano Zampini /* stuff to store connected components stored in pcbddc->mat_graph */ 1414984c4197SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges,*used_IS; 1415984c4197SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges; 1416984c4197SStefano Zampini /* near null space stuff */ 1417674ae819SStefano Zampini MatNullSpace nearnullsp; 1418674ae819SStefano Zampini const Vec *nearnullvecs; 1419674ae819SStefano Zampini Vec *localnearnullsp; 1420984c4197SStefano Zampini PetscBool nnsp_has_cnst; 1421984c4197SStefano Zampini PetscInt nnsp_size; 1422984c4197SStefano Zampini PetscScalar *array; 1423984c4197SStefano Zampini /* BLAS integers */ 1424e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1425e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1426c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1427984c4197SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1428242a89d7SStefano Zampini PetscBool skip_lapack; 1429984c4197SStefano Zampini PetscScalar *work; 1430984c4197SStefano Zampini PetscReal *singular_vals; 1431984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1432984c4197SStefano Zampini PetscReal *rwork; 1433674ae819SStefano Zampini #endif 1434984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1435e310c8b4SStefano Zampini PetscBLASInt Blas_one_2=1; 1436984c4197SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1437b7d8b9f8SStefano Zampini #else 1438b7d8b9f8SStefano Zampini PetscBLASInt dummy_int_1=1,dummy_int_2=1; 1439b7d8b9f8SStefano Zampini PetscScalar dummy_scalar_1=0.0,dummy_scalar_2=0.0; 1440984c4197SStefano Zampini #endif 1441727cdba6SStefano Zampini /* reuse */ 1442727cdba6SStefano Zampini PetscInt olocal_primal_size; 1443727cdba6SStefano Zampini PetscInt *oprimal_indices_local_idxs; 1444984c4197SStefano Zampini /* change of basis */ 1445984c4197SStefano Zampini PetscInt *aux_primal_numbering,*aux_primal_minloc,*global_indices; 1446a717540cSStefano Zampini PetscBool boolforchange,qr_needed; 1447a717540cSStefano Zampini PetscBT touched,change_basis,qr_needed_idx; 1448984c4197SStefano Zampini /* auxiliary stuff */ 144939e2fb2aSStefano Zampini PetscInt *nnz,*is_indices,*aux_primal_numbering_B; 1450911cabfeSStefano Zampini PetscInt ncc,*gidxs,*permutation,*temp_indices_to_constraint_work; 1451911cabfeSStefano Zampini PetscScalar *temp_quadrature_constraint_work; 1452984c4197SStefano Zampini /* some quantities */ 145345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1454984c4197SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint,max_constraints,temp_constraints; 1455984c4197SStefano Zampini 1456674ae819SStefano Zampini 1457674ae819SStefano Zampini PetscFunctionBegin; 14588e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 14598e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 14608e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1461674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1462d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1463d06fc5fdSStefano Zampini /* free unneeded index sets */ 1464d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1465d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1466674ae819SStefano Zampini } 1467d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1468d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1469d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1470d06fc5fdSStefano Zampini } 1471d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1472d06fc5fdSStefano Zampini n_ISForEdges = 0; 1473d06fc5fdSStefano Zampini } 1474d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1475d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1476d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1477d06fc5fdSStefano Zampini } 1478d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1479d06fc5fdSStefano Zampini n_ISForFaces = 0; 1480d06fc5fdSStefano Zampini } 1481d06fc5fdSStefano Zampini /* HACKS (the following two blocks of code) */ 1482b9b85e73SStefano Zampini if (!ISForVertices && pcbddc->NullSpace && !pcbddc->user_ChangeOfBasisMatrix) { 1483b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1484d06fc5fdSStefano Zampini if (!ISForEdges) { 1485d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1486d06fc5fdSStefano Zampini } 1487b8ffe317SStefano Zampini } 148898a51de6SStefano Zampini if (pcbddc->NullSpace) { 148998a51de6SStefano Zampini /* use_change_of_basis should be consistent among processors */ 1490d06fc5fdSStefano Zampini PetscBool tbool[2],gbool[2]; 1491d06fc5fdSStefano Zampini tbool [0] = pcbddc->use_change_of_basis; 1492d06fc5fdSStefano Zampini tbool [1] = pcbddc->use_change_on_faces; 1493d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1494d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1495d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 149698a51de6SStefano Zampini } 1497984c4197SStefano Zampini /* print some info */ 1498674ae819SStefano Zampini if (pcbddc->dbg_flag) { 14990fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1500674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1501674ae819SStefano Zampini i = 0; 1502674ae819SStefano Zampini if (ISForVertices) { 1503674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&i);CHKERRQ(ierr); 1504674ae819SStefano Zampini } 1505674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices\n",PetscGlobalRank,i);CHKERRQ(ierr); 1506674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges\n",PetscGlobalRank,n_ISForEdges);CHKERRQ(ierr); 150715aaf578SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces\n",PetscGlobalRank,n_ISForFaces);CHKERRQ(ierr); 1508674ae819SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1509674ae819SStefano Zampini } 1510674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1511674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1512674ae819SStefano Zampini if (nearnullsp) { 1513674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1514f4ddd8eeSStefano Zampini /* remove any stored info */ 1515f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1516f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1517f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1518f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1519f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1520473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1521f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1522f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1523f4ddd8eeSStefano Zampini } 1524984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1525984c4197SStefano Zampini nnsp_size = 0; 1526674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1527674ae819SStefano Zampini } 1528984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1529984c4197SStefano Zampini max_constraints = nnsp_size; 1530984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1531984c4197SStefano Zampini 1532674ae819SStefano Zampini /* 1533674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 1534674ae819SStefano Zampini - temp_indices will contain start index of each constraint stored as follows 1535911cabfeSStefano 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 1536911cabfeSStefano 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 1537911cabfeSStefano Zampini - temp_quadrature_constraint [temp_indices[i],...,temp_indices[i+1]-1] will contain the scalars representing the constraint itself 1538674ae819SStefano Zampini */ 1539674ae819SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 1540984c4197SStefano Zampini total_counts *= max_constraints; 1541674ae819SStefano Zampini n_vertices = 0; 1542674ae819SStefano Zampini if (ISForVertices) { 1543674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1544674ae819SStefano Zampini } 1545674ae819SStefano Zampini total_counts += n_vertices; 1546854ce69bSBarry Smith ierr = PetscMalloc1(total_counts+1,&temp_indices);CHKERRQ(ierr); 15474641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 1548674ae819SStefano Zampini total_counts = 0; 1549674ae819SStefano Zampini max_size_of_constraint = 0; 1550674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 1551674ae819SStefano Zampini if (i<n_ISForEdges) { 1552674ae819SStefano Zampini used_IS = &ISForEdges[i]; 1553674ae819SStefano Zampini } else { 1554674ae819SStefano Zampini used_IS = &ISForFaces[i-n_ISForEdges]; 1555674ae819SStefano Zampini } 1556674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&j);CHKERRQ(ierr); 1557674ae819SStefano Zampini total_counts += j; 1558674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1559674ae819SStefano Zampini } 1560984c4197SStefano Zampini total_counts *= max_constraints; 1561674ae819SStefano Zampini total_counts += n_vertices; 1562906d46d4SStefano Zampini ierr = PetscMalloc3(total_counts,&temp_quadrature_constraint,total_counts,&temp_indices_to_constraint,total_counts,&temp_indices_to_constraint_B);CHKERRQ(ierr); 1563984c4197SStefano Zampini /* get local part of global near null space vectors */ 1564785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1565984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1566984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1567984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1568984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1569984c4197SStefano Zampini } 1570674ae819SStefano Zampini 1571242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1572242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1573a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1574242a89d7SStefano Zampini 15758f1c130eSStefano Zampini /* allocate some auxiliary stuff */ 15768f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 15778f1c130eSStefano Zampini ierr = PetscMalloc4(max_size_of_constraint,&gidxs,max_size_of_constraint,&permutation,max_size_of_constraint,&temp_indices_to_constraint_work,max_size_of_constraint,&temp_quadrature_constraint_work);CHKERRQ(ierr); 15788f1c130eSStefano Zampini } 15798f1c130eSStefano Zampini 1580984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1581a773dcb8SStefano Zampini if (!skip_lapack) { 1582674ae819SStefano Zampini PetscScalar temp_work; 1583911cabfeSStefano Zampini 1584674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1585984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1586785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 1587785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 1588785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 1589674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1590785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 1591674ae819SStefano Zampini #endif 1592674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1593c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 1594c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 1595674ae819SStefano Zampini lwork = -1; 1596674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1597674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1598c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 1599674ae819SStefano Zampini #else 1600c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 1601674ae819SStefano Zampini #endif 1602674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1603984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 1604674ae819SStefano Zampini #else /* on missing GESVD */ 1605674ae819SStefano Zampini /* SVD */ 1606674ae819SStefano Zampini PetscInt max_n,min_n; 1607674ae819SStefano Zampini max_n = max_size_of_constraint; 1608984c4197SStefano Zampini min_n = max_constraints; 1609984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 1610674ae819SStefano Zampini min_n = max_size_of_constraint; 1611984c4197SStefano Zampini max_n = max_constraints; 1612674ae819SStefano Zampini } 1613785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 1614674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1615785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 1616674ae819SStefano Zampini #endif 1617674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1618674ae819SStefano Zampini lwork = -1; 1619e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 1620e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 1621b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 1622674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1623674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1624e310c8b4SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar_1,&dummy_int_1,&dummy_scalar_2,&dummy_int_2,&temp_work,&lwork,&lierr)); 1625674ae819SStefano Zampini #else 1626e310c8b4SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&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)); 1627674ae819SStefano Zampini #endif 1628674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1629984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 1630984c4197SStefano Zampini #endif /* on missing GESVD */ 1631674ae819SStefano Zampini /* Allocate optimal workspace */ 1632674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 1633854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 1634674ae819SStefano Zampini } 1635674ae819SStefano Zampini /* Now we can loop on constraining sets */ 1636674ae819SStefano Zampini total_counts = 0; 1637674ae819SStefano Zampini temp_indices[0] = 0; 1638674ae819SStefano Zampini /* vertices */ 1639674ae819SStefano Zampini if (ISForVertices) { 1640674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1641674ae819SStefano Zampini if (nnsp_has_cnst) { /* consider all vertices */ 164239e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 1643674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1644674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1645674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1646674ae819SStefano Zampini total_counts++; 1647674ae819SStefano Zampini } 1648674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 1649984c4197SStefano Zampini PetscBool used_vertex; 1650674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1651674ae819SStefano Zampini used_vertex = PETSC_FALSE; 1652674ae819SStefano Zampini k = 0; 1653674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 1654984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1655984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 1656674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i]; 1657674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1658674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1659674ae819SStefano Zampini total_counts++; 1660674ae819SStefano Zampini used_vertex = PETSC_TRUE; 1661674ae819SStefano Zampini } 1662984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1663674ae819SStefano Zampini k++; 1664674ae819SStefano Zampini } 1665674ae819SStefano Zampini } 1666674ae819SStefano Zampini } 1667674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1668674ae819SStefano Zampini n_vertices = total_counts; 1669674ae819SStefano Zampini } 1670984c4197SStefano Zampini 1671674ae819SStefano Zampini /* edges and faces */ 1672911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 1673911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 1674911cabfeSStefano Zampini used_IS = &ISForEdges[ncc]; 1675984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 1676674ae819SStefano Zampini } else { 1677911cabfeSStefano Zampini used_IS = &ISForFaces[ncc-n_ISForEdges]; 1678984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 1679674ae819SStefano Zampini } 1680674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 1681674ae819SStefano Zampini temp_start_ptr = total_counts; /* need to know the starting index of constraints stored */ 1682674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&size_of_constraint);CHKERRQ(ierr); 1683674ae819SStefano Zampini ierr = ISGetIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1684984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 1685984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 1686674ae819SStefano Zampini if (nnsp_has_cnst) { 16875b08dc53SStefano Zampini PetscScalar quad_value; 1688674ae819SStefano Zampini temp_constraints++; 1689a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 1690674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 1691a773dcb8SStefano Zampini } else { 1692a773dcb8SStefano Zampini quad_value = 1.0; 1693a773dcb8SStefano Zampini } 169439e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1695674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1696674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=quad_value; 1697674ae819SStefano Zampini } 1698911cabfeSStefano Zampini /* sort by global ordering if using lapack subroutines */ 16998f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1700911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 1701911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1702911cabfeSStefano Zampini permutation[j]=j; 1703911cabfeSStefano Zampini } 1704911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 1705911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1706911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 1707911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 1708911cabfeSStefano Zampini } 1709911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1710911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 1711911cabfeSStefano Zampini } 1712674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 1713674ae819SStefano Zampini total_counts++; 1714674ae819SStefano Zampini } 1715674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 1716984c4197SStefano Zampini PetscReal real_value; 1717984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 171839e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1719674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1720984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=array[is_indices[j]]; 1721674ae819SStefano Zampini } 1722984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1723984c4197SStefano Zampini /* check if array is null on the connected component */ 1724e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1725e310c8b4SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_one)); 17265b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 1727911cabfeSStefano Zampini /* sort by global ordering if using lapack subroutines */ 17288f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1729911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 1730911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1731911cabfeSStefano Zampini permutation[j]=j; 1732911cabfeSStefano Zampini } 1733911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 1734911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1735911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 1736911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 1737911cabfeSStefano Zampini } 1738911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1739911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 1740911cabfeSStefano Zampini } 1741674ae819SStefano Zampini temp_constraints++; 1742674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 1743674ae819SStefano Zampini total_counts++; 1744674ae819SStefano Zampini } 1745674ae819SStefano Zampini } 1746674ae819SStefano Zampini ierr = ISRestoreIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 174745a1bb75SStefano Zampini valid_constraints = temp_constraints; 1748eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 1749a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 1750a773dcb8SStefano Zampini PetscScalar norm; 1751a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1752a773dcb8SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one)); 1753a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 1754a773dcb8SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one)); 1755a773dcb8SStefano Zampini } else { /* perform SVD */ 1756984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 1757674ae819SStefano Zampini 1758674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1759984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 1760984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 1761984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 1762984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 1763984c4197SStefano Zampini from that computed using LAPACKgesvd 1764984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 1765984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 1766984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 1767674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 1768e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1769984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1770674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 1771674ae819SStefano Zampini for (k=0;k<j+1;k++) { 1772e310c8b4SStefano Zampini PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k]=BLASdot_(&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr+k]],&Blas_one,&temp_quadrature_constraint[temp_indices[temp_start_ptr+j]],&Blas_one_2)); 1773674ae819SStefano Zampini } 1774674ae819SStefano Zampini } 1775e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 1776e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1777e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 1778674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1779c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 1780674ae819SStefano Zampini #else 1781c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 1782674ae819SStefano Zampini #endif 1783674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1784984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 1785984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 1786674ae819SStefano Zampini j = 0; 1787984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 1788674ae819SStefano Zampini total_counts = total_counts-j; 178945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 1790e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 1791c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 1792c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1793c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 1794c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 1795c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 1796c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 1797674ae819SStefano Zampini if (j<temp_constraints) { 1798984c4197SStefano Zampini PetscInt ii; 1799984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k]=1.0/PetscSqrtReal(singular_vals[k]); 1800674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1801c4303822SStefano 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)); 1802674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1803984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 1804674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 1805984c4197SStefano 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]; 1806674ae819SStefano Zampini } 1807674ae819SStefano Zampini } 1808674ae819SStefano Zampini } 1809674ae819SStefano Zampini #else /* on missing GESVD */ 1810e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 1811e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1812b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 1813674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1814674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1815e310c8b4SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&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)); 1816674ae819SStefano Zampini #else 1817e310c8b4SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&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)); 1818674ae819SStefano Zampini #endif 1819984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 1820674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1821984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 1822e310c8b4SStefano Zampini k = temp_constraints; 1823e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 1824674ae819SStefano Zampini j = 0; 1825e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 182645a1bb75SStefano Zampini valid_constraints = k-j; 1827911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 1828984c4197SStefano Zampini #endif /* on missing GESVD */ 1829674ae819SStefano Zampini } 1830a773dcb8SStefano Zampini } 183145a1bb75SStefano Zampini /* setting change_of_basis flag is safe now */ 183245a1bb75SStefano Zampini if (boolforchange) { 183345a1bb75SStefano Zampini for (j=0;j<valid_constraints;j++) { 183445a1bb75SStefano Zampini PetscBTSet(change_basis,total_counts-j-1); 183545a1bb75SStefano Zampini } 183645a1bb75SStefano Zampini } 1837674ae819SStefano Zampini } 1838674ae819SStefano Zampini /* free index sets of faces, edges and vertices */ 1839674ae819SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1840674ae819SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1841674ae819SStefano Zampini } 1842d06fc5fdSStefano Zampini if (n_ISForFaces) { 1843674ae819SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1844d06fc5fdSStefano Zampini } 1845674ae819SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1846674ae819SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1847674ae819SStefano Zampini } 1848d06fc5fdSStefano Zampini if (n_ISForEdges) { 1849674ae819SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1850d06fc5fdSStefano Zampini } 1851674ae819SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 185239e2fb2aSStefano Zampini /* map temp_indices_to_constraint in boundary numbering */ 185339e2fb2aSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcbddc->BtoNmap,IS_GTOLM_DROP,temp_indices[total_counts],temp_indices_to_constraint,&i,temp_indices_to_constraint_B);CHKERRQ(ierr); 185439e2fb2aSStefano Zampini if (i != temp_indices[total_counts]) { 185539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",temp_indices[total_counts],i); 185639e2fb2aSStefano Zampini } 1857674ae819SStefano Zampini 1858984c4197SStefano Zampini /* free workspace */ 18598f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1860911cabfeSStefano Zampini ierr = PetscFree4(gidxs,permutation,temp_indices_to_constraint_work,temp_quadrature_constraint_work);CHKERRQ(ierr); 18618f1c130eSStefano Zampini } 18628f1c130eSStefano Zampini if (!skip_lapack) { 1863984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1864984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1865984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 1866984c4197SStefano Zampini #endif 1867984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 1868984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1869984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 1870984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 1871984c4197SStefano Zampini #endif 1872984c4197SStefano Zampini } 1873984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1874984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 1875984c4197SStefano Zampini } 1876984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 1877984c4197SStefano Zampini 1878727cdba6SStefano Zampini /* set quantities in pcbddc data structure and store previous primal size */ 1879984c4197SStefano Zampini /* n_vertices defines the number of subdomain corners in the primal space */ 1880674ae819SStefano Zampini /* n_constraints defines the number of averages (they can be point primal dofs if change of basis is requested) */ 1881727cdba6SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 1882984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 1883674ae819SStefano Zampini pcbddc->n_vertices = n_vertices; 1884984c4197SStefano Zampini pcbddc->n_constraints = pcbddc->local_primal_size-pcbddc->n_vertices; 1885674ae819SStefano Zampini 1886674ae819SStefano Zampini /* Create constraint matrix */ 1887674ae819SStefano Zampini /* The constraint matrix is used to compute the l2g map of primal dofs */ 1888674ae819SStefano Zampini /* so we need to set it up properly either with or without change of basis */ 1889674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1890674ae819SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,impMatType);CHKERRQ(ierr); 1891984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 1892984c4197SStefano Zampini /* array to compute a local numbering of constraints : vertices first then constraints */ 1893785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_numbering);CHKERRQ(ierr); 1894984c4197SStefano Zampini /* array to select the proper local node (of minimum index with respect to global ordering) when changing the basis */ 1895984c4197SStefano 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 */ 1896785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_minloc);CHKERRQ(ierr); 1897984c4197SStefano Zampini /* auxiliary stuff for basis change */ 1898785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&global_indices);CHKERRQ(ierr); 18994641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&touched);CHKERRQ(ierr); 1900984c4197SStefano Zampini 1901984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 1902984c4197SStefano Zampini total_primal_vertices=0; 1903984c4197SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1904674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 1905984c4197SStefano Zampini if (size_of_constraint == 1) { 19064641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]]);CHKERRQ(ierr); 1907984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]]; 1908984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=0; 1909984c4197SStefano Zampini total_primal_vertices++; 19104641a718SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { /* Same procedure used in PCBDDCGetPrimalConstraintsLocalIdx */ 1911984c4197SStefano Zampini PetscInt min_loc,min_index; 1912984c4197SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],global_indices);CHKERRQ(ierr); 1913984c4197SStefano Zampini /* find first untouched local node */ 1914674ae819SStefano Zampini k = 0; 19154641a718SStefano Zampini while (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) k++; 1916984c4197SStefano Zampini min_index = global_indices[k]; 1917984c4197SStefano Zampini min_loc = k; 1918984c4197SStefano Zampini /* search the minimum among global nodes already untouched on the cc */ 1919984c4197SStefano Zampini for (k=1;k<size_of_constraint;k++) { 1920984c4197SStefano Zampini /* there can be more than one constraint on a single connected component */ 19214641a718SStefano Zampini if (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k]) && min_index > global_indices[k]) { 1922984c4197SStefano Zampini min_index = global_indices[k]; 1923984c4197SStefano Zampini min_loc = k; 1924674ae819SStefano Zampini } 1925674ae819SStefano Zampini } 19264641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]+min_loc]);CHKERRQ(ierr); 1927984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]+min_loc]; 1928984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=min_loc; 1929984c4197SStefano Zampini total_primal_vertices++; 1930984c4197SStefano Zampini } 1931984c4197SStefano Zampini } 1932a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 1933a717540cSStefano Zampini qr_needed = PETSC_FALSE; 1934a717540cSStefano Zampini ierr = PetscBTCreate(pcbddc->local_primal_size,&qr_needed_idx);CHKERRQ(ierr); 1935a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 1936a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 1937fa434743SStefano Zampini if (!pcbddc->use_qr_single) { 1938a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 1939a717540cSStefano Zampini j = 0; 1940a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 1941a717540cSStefano Zampini if (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) { 1942a717540cSStefano Zampini j++; 1943a717540cSStefano Zampini } 1944a717540cSStefano Zampini } 1945a717540cSStefano Zampini /* found more than one primal dof on the cc */ 1946a717540cSStefano Zampini if (j > 1) { 1947a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 1948a717540cSStefano Zampini qr_needed = PETSC_TRUE; 1949a717540cSStefano Zampini } 1950fa434743SStefano Zampini } else { 1951fa434743SStefano Zampini PetscBTSet(qr_needed_idx,i); 1952fa434743SStefano Zampini qr_needed = PETSC_TRUE; 1953fa434743SStefano Zampini } 1954a717540cSStefano Zampini } 1955a717540cSStefano Zampini } 1956984c4197SStefano Zampini /* free workspace */ 1957984c4197SStefano Zampini ierr = PetscFree(global_indices);CHKERRQ(ierr); 195845a1bb75SStefano Zampini 1959674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 1960a717540cSStefano Zampini ierr = PetscSortInt(total_primal_vertices,aux_primal_numbering);CHKERRQ(ierr); 1961984c4197SStefano Zampini 1962984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 1963785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 1964984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i]=1; 1965984c4197SStefano Zampini j=total_primal_vertices; 1966984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 19674641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 1968674ae819SStefano Zampini nnz[j]=temp_indices[i+1]-temp_indices[i]; 1969674ae819SStefano Zampini j++; 1970674ae819SStefano Zampini } 1971674ae819SStefano Zampini } 1972674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 1973674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1974674ae819SStefano Zampini /* set values in constraint matrix */ 1975984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 1976984c4197SStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,aux_primal_numbering[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 1977674ae819SStefano Zampini } 1978984c4197SStefano Zampini total_counts = total_primal_vertices; 1979984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 19804641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 1981674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 1982674ae819SStefano 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); 1983674ae819SStefano Zampini total_counts++; 1984674ae819SStefano Zampini } 1985674ae819SStefano Zampini } 1986674ae819SStefano Zampini /* assembling */ 1987674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1988674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1989984c4197SStefano Zampini /* 199045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1991984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 1992984c4197SStefano Zampini */ 1993674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 1994674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 1995026de310SStefano Zampini /* dual and primal dofs on a single cc */ 1996984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 1997026de310SStefano Zampini /* iterator on aux_primal_minloc (ordered as read from nearnullspace: vertices, edges and then constraints) */ 1998026de310SStefano Zampini PetscInt primal_counter; 1999984c4197SStefano Zampini /* working stuff for GEQRF */ 200081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2001984c4197SStefano Zampini PetscBLASInt lqr_work; 2002984c4197SStefano Zampini /* working stuff for UNGQR */ 2003984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2004984c4197SStefano Zampini PetscBLASInt lgqr_work; 2005984c4197SStefano Zampini /* working stuff for TRTRS */ 2006984c4197SStefano Zampini PetscScalar *trs_rhs; 20073f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2008984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2009984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2010984c4197SStefano Zampini PetscScalar *start_vals; 2011984c4197SStefano Zampini /* working stuff for values insertion */ 20124641a718SStefano Zampini PetscBT is_primal; 2013906d46d4SStefano Zampini /* matrix sizes */ 2014906d46d4SStefano Zampini PetscInt global_size,local_size; 2015906d46d4SStefano Zampini /* work array for nonzeros */ 2016906d46d4SStefano Zampini PetscScalar *nnz_array; 2017906d46d4SStefano Zampini /* temporary change of basis */ 2018906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2019906d46d4SStefano Zampini /* auxiliary work for global change of basis */ 2020906d46d4SStefano Zampini Vec nnz_vec; 202122bc73bbSStefano Zampini PetscInt *idxs_I,*idxs_B,*idxs_all,*d_nnz,*o_nnz; 2022906d46d4SStefano Zampini PetscInt nvtxs,*xadj,*adjncy,*idxs_mapped; 2023906d46d4SStefano Zampini PetscScalar *vals; 2024906d46d4SStefano Zampini PetscBool done; 2025984c4197SStefano Zampini 2026906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2027906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 2028906d46d4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,impMatType);CHKERRQ(ierr); 2029906d46d4SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n_B,pcis->n_B,pcis->n_B,pcis->n_B);CHKERRQ(ierr); 2030906d46d4SStefano Zampini 2031906d46d4SStefano Zampini /* nonzeros for local mat */ 2032785e854fSJed Brown ierr = PetscMalloc1(pcis->n_B,&nnz);CHKERRQ(ierr); 2033a717540cSStefano Zampini for (i=0;i<pcis->n_B;i++) nnz[i]=1; 2034a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 2035a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 2036a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 2037a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 2038a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = size_of_constraint; 2039a717540cSStefano Zampini } else { 2040a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = 2; 2041a717540cSStefano Zampini /* get local primal index on the cc */ 2042a717540cSStefano Zampini j = 0; 2043a717540cSStefano Zampini while (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+j])) j++; 2044a717540cSStefano Zampini nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = size_of_constraint; 2045a717540cSStefano Zampini } 2046a717540cSStefano Zampini } 2047a717540cSStefano Zampini } 2048906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2049a717540cSStefano Zampini /* Set initial identity in the matrix */ 2050a717540cSStefano Zampini for (i=0;i<pcis->n_B;i++) { 2051906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2052a717540cSStefano Zampini } 2053a717540cSStefano Zampini 2054a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2055a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2056a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2057a717540cSStefano Zampini } 2058a717540cSStefano Zampini 2059a717540cSStefano Zampini 2060a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2061a717540cSStefano Zampini /* 2062a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2063a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2064a717540cSStefano Zampini 2065a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2066a717540cSStefano Zampini 2067a6b551f4SStefano Zampini - Using the following block transformation if there is only a primal dof on the cc (and -pc_bddc_use_qr_single is not specified) 2068a6b551f4SStefano Zampini 2069a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2070a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2071a717540cSStefano Zampini | ... | 2072a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2073a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2074a717540cSStefano Zampini 2075a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2076a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2077a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2078a6b551f4SStefano Zampini 2079a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2080a717540cSStefano Zampini */ 2081a717540cSStefano Zampini if (qr_needed) { 2082984c4197SStefano Zampini /* space to store Q */ 2083854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2084984c4197SStefano Zampini /* first we issue queries for optimal work */ 20853f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 20863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 20873f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2088984c4197SStefano Zampini lqr_work = -1; 20893f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2090984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2091984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2092785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2093984c4197SStefano Zampini lgqr_work = -1; 20943f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 20953f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 20963f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 20973f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 20983f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 20993f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2100984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2101984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2102785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2103984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2104785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2105984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2106785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2107a717540cSStefano Zampini /* allocating workspace for check */ 2108a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2109785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&work);CHKERRQ(ierr); 2110a717540cSStefano Zampini } 2111a717540cSStefano Zampini } 2112984c4197SStefano Zampini /* array to store whether a node is primal or not */ 21134641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2114473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 211539e2fb2aSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcbddc->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,aux_primal_numbering,&i,aux_primal_numbering_B);CHKERRQ(ierr); 211639e2fb2aSStefano Zampini if (i != total_primal_vertices) { 211739e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 21184641a718SStefano Zampini } 211939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 212039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 212139e2fb2aSStefano Zampini } 212239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2123984c4197SStefano Zampini 2124a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 2125026de310SStefano Zampini /* -> using implicit ordering contained in temp_indices data */ 2126026de310SStefano Zampini total_counts = pcbddc->n_vertices; 2127026de310SStefano Zampini primal_counter = total_counts; 2128026de310SStefano Zampini while (total_counts<pcbddc->local_primal_size) { 2129026de310SStefano Zampini primal_dofs = 1; 21304641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2131026de310SStefano Zampini /* get all constraints with same support: if more then one constraint is present on the cc then surely indices are stored contiguosly */ 2132026de310SStefano 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]]) { 2133026de310SStefano Zampini primal_dofs++; 2134674ae819SStefano Zampini } 2135984c4197SStefano Zampini /* get constraint info */ 2136026de310SStefano Zampini size_of_constraint = temp_indices[total_counts+1]-temp_indices[total_counts]; 2137984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2138984c4197SStefano Zampini 2139984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2140a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d to %d (incl) need a change of basis (size %d)\n",total_counts,total_counts+primal_dofs-1,size_of_constraint);CHKERRQ(ierr); 2141674ae819SStefano Zampini } 2142984c4197SStefano Zampini 2143fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2144a717540cSStefano Zampini 2145a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2146a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2147a717540cSStefano Zampini ierr = PetscMemcpy(work,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2148a717540cSStefano Zampini } 2149984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 2150026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2151984c4197SStefano Zampini 2152984c4197SStefano Zampini /* compute QR decomposition of constraints */ 21533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21543f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21553f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2156674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21573f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2158984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2159674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2160984c4197SStefano Zampini 2161984c4197SStefano Zampini /* explictly compute R^-T */ 2162984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2163984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 21643f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21653f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 21663f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 21673f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2168984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21693f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2170984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2171984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2172984c4197SStefano Zampini 2173a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 21743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21753f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21763f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 21773f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2178984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21793f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2180984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2181984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2182984c4197SStefano Zampini 2183984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2184984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2185984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 21863f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21873f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21883f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 21893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 21903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 21913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2192984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2193c4303822SStefano 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)); 2194984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2195026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2196984c4197SStefano Zampini 2197984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 2198026de310SStefano Zampini start_rows = &temp_indices_to_constraint_B[temp_indices[total_counts]]; 2199984c4197SStefano Zampini /* insert cols for primal dofs */ 2200984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2201984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 2202026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+aux_primal_minloc[primal_counter+j]]; 2203906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2204984c4197SStefano Zampini } 2205984c4197SStefano Zampini /* insert cols for dual dofs */ 2206984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 22074641a718SStefano Zampini if (!PetscBTLookup(is_primal,temp_indices_to_constraint_B[temp_indices[total_counts]+k])) { 2208984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 2209026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+k]; 2210906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2211984c4197SStefano Zampini j++; 2212674ae819SStefano Zampini } 2213674ae819SStefano Zampini } 2214984c4197SStefano Zampini 2215984c4197SStefano Zampini /* check change of basis */ 2216984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2217984c4197SStefano Zampini PetscInt ii,jj; 2218984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2219c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2220c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2221c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2222c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2223c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2224c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2225984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2226c4303822SStefano 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)); 2227984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2228984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2229984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2230984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2231984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE; 2232674ae819SStefano Zampini } 2233674ae819SStefano Zampini } 2234984c4197SStefano Zampini if (!valid_qr) { 223522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2236984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2237984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2238984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2239984c4197SStefano 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])); 2240674ae819SStefano Zampini } 2241984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2242984c4197SStefano 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])); 2243984c4197SStefano Zampini } 2244984c4197SStefano Zampini } 2245984c4197SStefano Zampini } 2246674ae819SStefano Zampini } else { 224722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2248674ae819SStefano Zampini } 2249674ae819SStefano Zampini } 2250a717540cSStefano Zampini } else { /* simple transformation block */ 2251a717540cSStefano Zampini PetscInt row,col; 2252a6b551f4SStefano Zampini PetscScalar val,norm; 2253a6b551f4SStefano Zampini 2254a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2255a6b551f4SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one)); 2256a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2257a717540cSStefano Zampini row = temp_indices_to_constraint_B[temp_indices[total_counts]+j]; 2258a717540cSStefano Zampini if (!PetscBTLookup(is_primal,row)) { 2259a717540cSStefano Zampini col = temp_indices_to_constraint_B[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2260906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 2261a6b551f4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,temp_quadrature_constraint[temp_indices[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2262a717540cSStefano Zampini } else { 2263a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 2264a717540cSStefano Zampini col = temp_indices_to_constraint_B[temp_indices[total_counts]+k]; 2265a717540cSStefano Zampini if (row != col) { 2266a717540cSStefano Zampini val = -temp_quadrature_constraint[temp_indices[total_counts]+k]/temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2267a717540cSStefano Zampini } else { 2268a6b551f4SStefano Zampini val = temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]/norm; 2269a717540cSStefano Zampini } 2270906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2271a717540cSStefano Zampini } 2272a717540cSStefano Zampini } 2273a717540cSStefano Zampini } 227498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 227522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2276a717540cSStefano Zampini } 2277674ae819SStefano Zampini } 2278026de310SStefano Zampini /* increment primal counter */ 2279026de310SStefano Zampini primal_counter += primal_dofs; 2280984c4197SStefano Zampini } else { 2281984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2282026de310SStefano 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); 2283674ae819SStefano Zampini } 2284674ae819SStefano Zampini } 2285026de310SStefano Zampini /* increment constraint counter total_counts */ 2286026de310SStefano Zampini total_counts += primal_dofs; 2287674ae819SStefano Zampini } 2288a717540cSStefano Zampini 2289a717540cSStefano Zampini /* free workspace */ 2290a717540cSStefano Zampini if (qr_needed) { 2291984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2292984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2293984c4197SStefano Zampini } 2294984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2295984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2296984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2297984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2298984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2299674ae819SStefano Zampini } 2300a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2301906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2302906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2303906d46d4SStefano Zampini 2304906d46d4SStefano Zampini /* assembling of global change of variable */ 2305906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2306906d46d4SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2307906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2308906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2309906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2310906d46d4SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->ChangeOfBasisMatrix,matis->mapping,matis->mapping);CHKERRQ(ierr); 2311906d46d4SStefano Zampini 2312906d46d4SStefano Zampini /* nonzeros (overestimated) */ 2313906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&nnz_vec);CHKERRQ(ierr); 2314906d46d4SStefano Zampini ierr = VecSetLocalToGlobalMapping(nnz_vec,matis->mapping);CHKERRQ(ierr); 2315906d46d4SStefano Zampini ierr = PetscMalloc2(pcis->n,&nnz_array,pcis->n,&idxs_all);CHKERRQ(ierr); 2316906d46d4SStefano Zampini for (i=0;i<pcis->n;i++) { 2317906d46d4SStefano Zampini nnz_array[i] = 1.0; 2318906d46d4SStefano Zampini idxs_all[i] = i; 2319906d46d4SStefano Zampini } 2320906d46d4SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&idxs_B);CHKERRQ(ierr); 2321906d46d4SStefano Zampini for (i=0;i<pcis->n_B;i++) { 2322906d46d4SStefano Zampini nnz_array[idxs_B[i]] = nnz[i]; 2323906d46d4SStefano Zampini } 2324a5187659SStefano Zampini if (pcis->n) { 2325906d46d4SStefano Zampini ierr = VecSetValuesLocal(nnz_vec,pcis->n,idxs_all,nnz_array,INSERT_VALUES);CHKERRQ(ierr); 2326a5187659SStefano Zampini } 2327906d46d4SStefano Zampini ierr = VecAssemblyBegin(nnz_vec);CHKERRQ(ierr); 2328906d46d4SStefano Zampini ierr = VecAssemblyEnd(nnz_vec);CHKERRQ(ierr); 2329906d46d4SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2330906d46d4SStefano Zampini ierr = PetscFree2(nnz_array,idxs_all);CHKERRQ(ierr); 233122bc73bbSStefano Zampini ierr = PetscMalloc2(local_size,&d_nnz,local_size,&o_nnz);CHKERRQ(ierr); 2332906d46d4SStefano Zampini ierr = VecGetArray(nnz_vec,&nnz_array);CHKERRQ(ierr); 2333906d46d4SStefano Zampini for (i=0;i<local_size;i++) { 233422bc73bbSStefano Zampini d_nnz[i] = PetscMin((PetscInt)(PetscRealPart(nnz_array[i])),local_size); 233522bc73bbSStefano Zampini o_nnz[i] = PetscMin((PetscInt)(PetscRealPart(nnz_array[i])),global_size-local_size); 2336906d46d4SStefano Zampini } 2337906d46d4SStefano Zampini ierr = VecRestoreArray(nnz_vec,&nnz_array);CHKERRQ(ierr); 2338906d46d4SStefano Zampini ierr = VecDestroy(&nnz_vec);CHKERRQ(ierr); 233922bc73bbSStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->ChangeOfBasisMatrix,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 234022bc73bbSStefano Zampini ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 2341906d46d4SStefano Zampini 2342906d46d4SStefano Zampini /* Set identity on dirichlet dofs */ 2343906d46d4SStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&idxs_I);CHKERRQ(ierr); 2344906d46d4SStefano Zampini for (i=0;i<pcis->n-pcis->n_B;i++) { 2345906d46d4SStefano Zampini PetscScalar one=1.0; 2346906d46d4SStefano Zampini ierr = MatSetValuesLocal(pcbddc->ChangeOfBasisMatrix,1,idxs_I+i,1,idxs_I+i,&one,INSERT_VALUES);CHKERRQ(ierr); 2347906d46d4SStefano Zampini } 2348906d46d4SStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&idxs_I);CHKERRQ(ierr); 2349906d46d4SStefano Zampini 2350906d46d4SStefano Zampini /* Set values at interface dofs */ 2351906d46d4SStefano Zampini done = PETSC_TRUE; 2352906d46d4SStefano Zampini ierr = MatGetRowIJ(localChangeOfBasisMatrix,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr); 2353906d46d4SStefano Zampini if (!done) { 2354906d46d4SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error in MatGetRowIJ called in %s\n",__FUNCT__); 2355906d46d4SStefano Zampini } 2356906d46d4SStefano Zampini ierr = MatSeqAIJGetArray(localChangeOfBasisMatrix,&vals);CHKERRQ(ierr); 2357906d46d4SStefano Zampini ierr = PetscMalloc1(xadj[nvtxs],&idxs_mapped);CHKERRQ(ierr); 2358906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->BtoNmap,xadj[nvtxs],adjncy,idxs_mapped);CHKERRQ(ierr); 2359906d46d4SStefano Zampini for (i=0;i<nvtxs;i++) { 2360906d46d4SStefano Zampini PetscInt row,*cols,ncols; 2361906d46d4SStefano Zampini PetscScalar *mat_vals; 2362906d46d4SStefano Zampini 2363906d46d4SStefano Zampini row = idxs_B[i]; 2364906d46d4SStefano Zampini ncols = xadj[i+1]-xadj[i]; 2365906d46d4SStefano Zampini cols = idxs_mapped+xadj[i]; 2366906d46d4SStefano Zampini mat_vals = vals+xadj[i]; 2367906d46d4SStefano Zampini ierr = MatSetValuesLocal(pcbddc->ChangeOfBasisMatrix,1,&row,ncols,cols,mat_vals,INSERT_VALUES);CHKERRQ(ierr); 2368906d46d4SStefano Zampini } 2369906d46d4SStefano Zampini ierr = MatRestoreRowIJ(localChangeOfBasisMatrix,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr); 2370906d46d4SStefano Zampini if (!done) { 2371906d46d4SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__); 2372906d46d4SStefano Zampini } 2373906d46d4SStefano Zampini ierr = MatSeqAIJRestoreArray(localChangeOfBasisMatrix,&vals);CHKERRQ(ierr); 2374906d46d4SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&idxs_B);CHKERRQ(ierr); 2375906d46d4SStefano Zampini ierr = PetscFree(idxs_mapped);CHKERRQ(ierr); 2376674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2377674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2378906d46d4SStefano Zampini 2379906d46d4SStefano Zampini /* check */ 2380906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2381906d46d4SStefano Zampini PetscReal error; 2382906d46d4SStefano Zampini Vec x,x_change; 2383906d46d4SStefano Zampini 2384906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2385906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2386906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2387906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2388906d46d4SStefano Zampini ierr = VecScatterBegin(pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2389906d46d4SStefano Zampini ierr = VecScatterEnd(pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2390906d46d4SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 2391906d46d4SStefano Zampini ierr = VecScatterBegin(pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2392906d46d4SStefano Zampini ierr = VecScatterEnd(pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2393906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2394906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2395906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2396906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2397906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on B: %1.6e\n",error);CHKERRQ(ierr); 2398906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2399906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2400906d46d4SStefano Zampini } 2401906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2402906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2403b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2404b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2405b9b85e73SStefano Zampini } 2406906d46d4SStefano Zampini 2407906d46d4SStefano Zampini /* set up change of basis context */ 2408906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2409906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2410906d46d4SStefano Zampini 2411906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2412906d46d4SStefano Zampini PetscInt global_size,local_size; 2413906d46d4SStefano Zampini 2414906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2415906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2416906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2417906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2418906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2419906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2420906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2421906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2422906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2423906d46d4SStefano Zampini } else { 2424906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2425906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2426906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2427906d46d4SStefano Zampini } 2428906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2429906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2430906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2431906d46d4SStefano Zampini } else { 2432906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2433906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2434906d46d4SStefano Zampini } 2435906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2436906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2437906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2438906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2439b9b85e73SStefano Zampini } 2440a717540cSStefano Zampini 2441727cdba6SStefano Zampini /* get indices in local ordering for vertices and constraints */ 2442727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { /* if this is true, I need to check if a new primal space has been introduced */ 2443473ba861SJed Brown ierr = PetscMalloc1(olocal_primal_size,&oprimal_indices_local_idxs);CHKERRQ(ierr); 2444727cdba6SStefano Zampini ierr = PetscMemcpy(oprimal_indices_local_idxs,pcbddc->primal_indices_local_idxs,olocal_primal_size*sizeof(PetscInt));CHKERRQ(ierr); 2445727cdba6SStefano Zampini } 2446727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2447f347579bSStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2448473ba861SJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2449727cdba6SStefano Zampini ierr = PCBDDCGetPrimalVerticesLocalIdx(pc,&i,&aux_primal_numbering);CHKERRQ(ierr); 2450727cdba6SStefano Zampini ierr = PetscMemcpy(pcbddc->primal_indices_local_idxs,aux_primal_numbering,i*sizeof(PetscInt));CHKERRQ(ierr); 2451727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2452727cdba6SStefano Zampini ierr = PCBDDCGetPrimalConstraintsLocalIdx(pc,&j,&aux_primal_numbering);CHKERRQ(ierr); 2453727cdba6SStefano Zampini ierr = PetscMemcpy(&pcbddc->primal_indices_local_idxs[i],aux_primal_numbering,j*sizeof(PetscInt));CHKERRQ(ierr); 2454727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2455e9189074SStefano Zampini /* set quantities in PCBDDC data struct */ 2456e9189074SStefano Zampini pcbddc->n_actual_vertices = i; 2457727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2458727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2459727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2460727cdba6SStefano Zampini ierr = PetscMemcmp(pcbddc->primal_indices_local_idxs,oprimal_indices_local_idxs,olocal_primal_size,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2461c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2462727cdba6SStefano Zampini ierr = PetscFree(oprimal_indices_local_idxs);CHKERRQ(ierr); 2463727cdba6SStefano Zampini } 2464727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2465727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2466727cdba6SStefano Zampini 2467a717540cSStefano Zampini /* flush dbg viewer */ 2468b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2469b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2470b8ffe317SStefano Zampini } 2471a717540cSStefano Zampini 2472e310c8b4SStefano Zampini /* free workspace */ 2473a717540cSStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 2474a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 2475984c4197SStefano Zampini ierr = PetscFree(aux_primal_minloc);CHKERRQ(ierr); 24764641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 2477906d46d4SStefano Zampini ierr = PetscFree(temp_indices);CHKERRQ(ierr); 2478906d46d4SStefano Zampini ierr = PetscFree3(temp_quadrature_constraint,temp_indices_to_constraint,temp_indices_to_constraint_B);CHKERRQ(ierr); 2479674ae819SStefano Zampini PetscFunctionReturn(0); 2480674ae819SStefano Zampini } 2481674ae819SStefano Zampini 2482674ae819SStefano Zampini #undef __FUNCT__ 2483674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2484674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2485674ae819SStefano Zampini { 2486674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2487674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2488674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 248963602bcaSStefano Zampini PetscInt ierr,i,vertex_size; 2490674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2491674ae819SStefano Zampini 2492674ae819SStefano Zampini PetscFunctionBegin; 24938e61c736SStefano Zampini /* Reset previously computed graph */ 24948e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2495674ae819SStefano Zampini /* Init local Graph struct */ 2496674ae819SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping);CHKERRQ(ierr); 2497674ae819SStefano Zampini 2498575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2499575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2500575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2501575ad6abSStefano Zampini } 25029577ea80SStefano Zampini 2503674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 25044d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 2505674ae819SStefano Zampini Mat mat_adj; 25064d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 25074d379d7bSStefano Zampini PetscInt nvtxs; 2508674ae819SStefano Zampini PetscBool flg_row=PETSC_TRUE; 2509674ae819SStefano Zampini 2510674ae819SStefano Zampini ierr = MatConvert(matis->A,MATMPIADJ,MAT_INITIAL_MATRIX,&mat_adj);CHKERRQ(ierr); 25114d379d7bSStefano Zampini ierr = MatGetRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2512674ae819SStefano Zampini if (!flg_row) { 2513674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatGetRowIJ called in %s\n",__FUNCT__); 2514674ae819SStefano Zampini } 25154d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 25164d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 25174d379d7bSStefano Zampini pcbddc->deluxe_compute_rowadj = PETSC_FALSE; 25184d379d7bSStefano Zampini } else { /* just compute subdomain's connected components */ 25194d379d7bSStefano Zampini IS is_dummy; 25204d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 25214d379d7bSStefano Zampini PetscInt j,sum; 25224d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 25234d379d7bSStefano Zampini const PetscInt *idxs; 25244d379d7bSStefano Zampini PCBDDCGraph graph; 25254d379d7bSStefano Zampini PetscBT is_on_boundary; 25264d379d7bSStefano Zampini 25274d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 25284d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 25294d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 25304d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 25314d379d7bSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy);CHKERRQ(ierr); 25324d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 25334d379d7bSStefano Zampini graph->xadj = xadj; 25344d379d7bSStefano Zampini graph->adjncy = adjncy; 25354d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 25364d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 25374d379d7bSStefano Zampini 25384d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 25394d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains\n",PetscGlobalRank,graph->ncc);CHKERRQ(ierr); 25404d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 25414d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 25424d379d7bSStefano Zampini } 25434d379d7bSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 25444d379d7bSStefano Zampini } 25454d379d7bSStefano Zampini 25464d379d7bSStefano Zampini ierr = PetscBTCreate(nvtxs,&is_on_boundary);CHKERRQ(ierr); 25474d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 25484d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 25494d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 25504d379d7bSStefano Zampini } 25514d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 25524d379d7bSStefano Zampini 25534d379d7bSStefano Zampini ierr = PetscCalloc1(nvtxs+1,&cxadj);CHKERRQ(ierr); 25544d379d7bSStefano Zampini sum = 0; 25554d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 25564d379d7bSStefano Zampini PetscInt sizecc = 0; 25574d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25584d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25594d379d7bSStefano Zampini sizecc++; 25604d379d7bSStefano Zampini } 25614d379d7bSStefano Zampini } 25624d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25634d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25644d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 25654d379d7bSStefano Zampini } 25664d379d7bSStefano Zampini } 25674d379d7bSStefano Zampini sum += sizecc*sizecc; 25684d379d7bSStefano Zampini } 25694d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 25704d379d7bSStefano Zampini sum = 0; 25714d379d7bSStefano Zampini for (i=0;i<nvtxs;i++) { 25724d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 25734d379d7bSStefano Zampini cxadj[i] = sum; 25744d379d7bSStefano Zampini sum += temp; 25754d379d7bSStefano Zampini } 25764d379d7bSStefano Zampini cxadj[nvtxs] = sum; 25774d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 25784d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25794d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25804d379d7bSStefano Zampini PetscInt k,sizecc = 0; 25814d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 25824d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 25834d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 25844d379d7bSStefano Zampini sizecc++; 25854d379d7bSStefano Zampini } 25864d379d7bSStefano Zampini } 25874d379d7bSStefano Zampini } 25884d379d7bSStefano Zampini } 25894d379d7bSStefano Zampini } 25904d379d7bSStefano Zampini if (nvtxs) { 25914d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 25924d379d7bSStefano Zampini } else { 25934d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 25944d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 25954d379d7bSStefano Zampini } 25964d379d7bSStefano Zampini graph->xadj = 0; 25974d379d7bSStefano Zampini graph->adjncy = 0; 25984d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 25994d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 26004d379d7bSStefano Zampini } 26014d379d7bSStefano Zampini ierr = MatRestoreRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2602674ae819SStefano Zampini if (!flg_row) { 2603674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__); 2604674ae819SStefano Zampini } 2605674ae819SStefano Zampini ierr = MatDestroy(&mat_adj);CHKERRQ(ierr); 2606674ae819SStefano Zampini } 2607674ae819SStefano Zampini 260863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 2609674ae819SStefano Zampini vertex_size = 1; 261063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 261163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 261295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 261363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 2614a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 261563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 2616674ae819SStefano Zampini } 261763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 261863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 261963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 2620674ae819SStefano Zampini } 262163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 2622674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 262363602bcaSStefano Zampini } else { 262463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 262563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 2626854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 262763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 262863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 262963602bcaSStefano Zampini } 263063602bcaSStefano Zampini } 2631674ae819SStefano Zampini } 2632674ae819SStefano Zampini 2633674ae819SStefano Zampini /* Setup of Graph */ 2634785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 2635a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 2636785d1243SStefano Zampini } 2637785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 2638a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2639785d1243SStefano Zampini } 264063602bcaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices); 2641674ae819SStefano Zampini 2642674ae819SStefano Zampini /* Graph's connected components analysis */ 2643674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 2644674ae819SStefano Zampini 2645674ae819SStefano Zampini /* print some info to stdout */ 2646674ae819SStefano Zampini if (pcbddc->dbg_flag) { 2647e49050b4SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer); 2648674ae819SStefano Zampini } 2649fb180af4SStefano Zampini 2650fb180af4SStefano Zampini /* mark topography has done */ 2651fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 2652674ae819SStefano Zampini PetscFunctionReturn(0); 2653674ae819SStefano Zampini } 2654674ae819SStefano Zampini 2655674ae819SStefano Zampini #undef __FUNCT__ 2656674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalVerticesLocalIdx" 2657f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalVerticesLocalIdx(PC pc, PetscInt *n_vertices, PetscInt **vertices_idx) 2658674ae819SStefano Zampini { 2659674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2660674ae819SStefano Zampini PetscInt *vertices,*row_cmat_indices,n,i,size_of_constraint,local_primal_size; 2661674ae819SStefano Zampini PetscErrorCode ierr; 2662674ae819SStefano Zampini 2663674ae819SStefano Zampini PetscFunctionBegin; 2664674ae819SStefano Zampini n = 0; 2665674ae819SStefano Zampini vertices = 0; 2666674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 2667674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&i);CHKERRQ(ierr); 2668b120a5c6SStefano Zampini for (i=0;i<local_primal_size;i++) { 2669b120a5c6SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2670b120a5c6SStefano Zampini if (size_of_constraint == 1) n++; 2671b120a5c6SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2672b120a5c6SStefano Zampini } 2673811e8ca2SStefano Zampini if (vertices_idx) { 2674785e854fSJed Brown ierr = PetscMalloc1(n,&vertices);CHKERRQ(ierr); 2675b120a5c6SStefano Zampini n = 0; 2676674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2677674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2678674ae819SStefano Zampini if (size_of_constraint == 1) { 2679674ae819SStefano Zampini vertices[n++]=row_cmat_indices[0]; 2680674ae819SStefano Zampini } 2681674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2682674ae819SStefano Zampini } 2683674ae819SStefano Zampini } 2684811e8ca2SStefano Zampini } 2685674ae819SStefano Zampini *n_vertices = n; 2686811e8ca2SStefano Zampini if (vertices_idx) *vertices_idx = vertices; 2687674ae819SStefano Zampini PetscFunctionReturn(0); 2688674ae819SStefano Zampini } 2689674ae819SStefano Zampini 2690674ae819SStefano Zampini #undef __FUNCT__ 2691674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalConstraintsLocalIdx" 2692f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalConstraintsLocalIdx(PC pc, PetscInt *n_constraints, PetscInt **constraints_idx) 2693674ae819SStefano Zampini { 2694674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2695674ae819SStefano Zampini PetscInt *constraints_index,*row_cmat_indices,*row_cmat_global_indices; 2696674ae819SStefano Zampini PetscInt n,i,j,size_of_constraint,local_primal_size,local_size,max_size_of_constraint,min_index,min_loc; 26974641a718SStefano Zampini PetscBT touched; 2698674ae819SStefano Zampini PetscErrorCode ierr; 2699674ae819SStefano Zampini 2700f34684f1SStefano Zampini /* This function assumes that the number of local constraints per connected component 2701f34684f1SStefano Zampini is not greater than the number of nodes defined for the connected component 2702f34684f1SStefano Zampini (otherwise we will surely have linear dependence between constraints and thus a singular coarse problem) */ 2703674ae819SStefano Zampini PetscFunctionBegin; 2704674ae819SStefano Zampini n = 0; 2705674ae819SStefano Zampini constraints_index = 0; 2706674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 2707674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&local_size);CHKERRQ(ierr); 2708674ae819SStefano Zampini max_size_of_constraint = 0; 2709674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2710674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2711674ae819SStefano Zampini if (size_of_constraint > 1) { 2712674ae819SStefano Zampini n++; 2713674ae819SStefano Zampini } 2714674ae819SStefano Zampini max_size_of_constraint = PetscMax(size_of_constraint,max_size_of_constraint); 2715674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2716674ae819SStefano Zampini } 2717811e8ca2SStefano Zampini if (constraints_idx) { 2718785e854fSJed Brown ierr = PetscMalloc1(n,&constraints_index);CHKERRQ(ierr); 2719785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&row_cmat_global_indices);CHKERRQ(ierr); 27204641a718SStefano Zampini ierr = PetscBTCreate(local_size,&touched);CHKERRQ(ierr); 2721674ae819SStefano Zampini n = 0; 2722674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2723674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2724674ae819SStefano Zampini if (size_of_constraint > 1) { 2725674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,row_cmat_indices,row_cmat_global_indices);CHKERRQ(ierr); 272682d3d8afSStefano Zampini /* find first untouched local node */ 272782d3d8afSStefano Zampini j = 0; 27284641a718SStefano Zampini while (PetscBTLookup(touched,row_cmat_indices[j])) j++; 272982d3d8afSStefano Zampini min_index = row_cmat_global_indices[j]; 273082d3d8afSStefano Zampini min_loc = j; 273182d3d8afSStefano Zampini /* search the minimum among nodes not yet touched on the connected component 273282d3d8afSStefano Zampini since there can be more than one constraint on a single cc */ 2733674ae819SStefano Zampini for (j=1;j<size_of_constraint;j++) { 27344641a718SStefano Zampini if (!PetscBTLookup(touched,row_cmat_indices[j]) && min_index > row_cmat_global_indices[j]) { 2735674ae819SStefano Zampini min_index = row_cmat_global_indices[j]; 2736674ae819SStefano Zampini min_loc = j; 2737674ae819SStefano Zampini } 2738674ae819SStefano Zampini } 27394641a718SStefano Zampini ierr = PetscBTSet(touched,row_cmat_indices[min_loc]);CHKERRQ(ierr); 2740674ae819SStefano Zampini constraints_index[n++] = row_cmat_indices[min_loc]; 2741674ae819SStefano Zampini } 2742674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2743674ae819SStefano Zampini } 27444641a718SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 2745674ae819SStefano Zampini ierr = PetscFree(row_cmat_global_indices);CHKERRQ(ierr); 2746811e8ca2SStefano Zampini } 2747811e8ca2SStefano Zampini } 2748674ae819SStefano Zampini *n_constraints = n; 2749811e8ca2SStefano Zampini if (constraints_idx) *constraints_idx = constraints_index; 2750674ae819SStefano Zampini PetscFunctionReturn(0); 2751674ae819SStefano Zampini } 2752674ae819SStefano Zampini 2753674ae819SStefano Zampini #undef __FUNCT__ 2754674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 2755674ae819SStefano 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[]) 2756674ae819SStefano Zampini { 2757674ae819SStefano Zampini Vec local_vec,global_vec; 2758674ae819SStefano Zampini IS seqis,paris; 2759674ae819SStefano Zampini VecScatter scatter_ctx; 2760674ae819SStefano Zampini PetscScalar *array; 2761674ae819SStefano Zampini PetscInt *temp_global_dofs; 2762674ae819SStefano Zampini PetscScalar globalsum; 2763674ae819SStefano Zampini PetscInt i,j,s; 2764674ae819SStefano Zampini PetscInt nlocals,first_index,old_index,max_local; 2765674ae819SStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm,max_global; 2766674ae819SStefano Zampini PetscMPIInt *dof_sizes,*dof_displs; 2767674ae819SStefano Zampini PetscBool first_found; 2768674ae819SStefano Zampini PetscErrorCode ierr; 2769674ae819SStefano Zampini 2770674ae819SStefano Zampini PetscFunctionBegin; 2771674ae819SStefano Zampini /* mpi buffers */ 2772b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 2773b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 2774674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 2775785e854fSJed Brown ierr = PetscMalloc1(j,&dof_sizes);CHKERRQ(ierr); 2776785e854fSJed Brown ierr = PetscMalloc1(j,&dof_displs);CHKERRQ(ierr); 2777674ae819SStefano Zampini /* get maximum size of subset */ 2778785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 2779674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 2780674ae819SStefano Zampini max_local = 0; 2781b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2782674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 2783674ae819SStefano Zampini max_local = temp_global_dofs[i]; 2784674ae819SStefano Zampini } 2785674ae819SStefano Zampini } 2786b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 2787674ae819SStefano Zampini max_global++; 2788674ae819SStefano Zampini max_local = 0; 2789b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2790674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 2791674ae819SStefano Zampini max_local = local_dofs[i]; 2792674ae819SStefano Zampini } 2793674ae819SStefano Zampini } 2794674ae819SStefano Zampini max_local++; 2795674ae819SStefano Zampini /* allocate workspace */ 2796674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 2797674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 2798674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 2799674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 2800674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 2801674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 2802674ae819SStefano Zampini /* create scatter */ 2803674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 2804674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 2805674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 2806674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 2807674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 2808674ae819SStefano Zampini /* init array */ 2809674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 2810674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 2811674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 2812674ae819SStefano Zampini if (local_dofs_mult) { 2813674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2814674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 2815674ae819SStefano Zampini } 2816674ae819SStefano Zampini } else { 2817674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2818674ae819SStefano Zampini array[local_dofs[i]]=1.0; 2819674ae819SStefano Zampini } 2820674ae819SStefano Zampini } 2821674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 2822674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 2823674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2824674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2825674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 28265b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 2827674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 2828674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 2829674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 2830674ae819SStefano Zampini nlocals = 0; 2831674ae819SStefano Zampini first_index = -1; 2832674ae819SStefano Zampini first_found = PETSC_FALSE; 2833674ae819SStefano Zampini for (i=0;i<s;i++) { 2834b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 2835674ae819SStefano Zampini first_found = PETSC_TRUE; 2836674ae819SStefano Zampini first_index = i; 2837674ae819SStefano Zampini } 28385b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 2839674ae819SStefano Zampini } 2840674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 2841674ae819SStefano Zampini if (!rank_prec_comm) { 2842674ae819SStefano Zampini dof_displs[0]=0; 2843674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 2844674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 2845674ae819SStefano Zampini } 2846674ae819SStefano Zampini } 2847674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 2848674ae819SStefano Zampini if (first_found) { 2849674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 2850674ae819SStefano Zampini old_index = first_index; 2851674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 2852b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 2853674ae819SStefano Zampini array[i] += array[old_index]; 2854674ae819SStefano Zampini old_index = i; 2855674ae819SStefano Zampini } 2856674ae819SStefano Zampini } 2857674ae819SStefano Zampini } 2858674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 2859674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 2860674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2861674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2862674ae819SStefano Zampini /* get global ordering of local dofs */ 2863674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 2864674ae819SStefano Zampini if (local_dofs_mult) { 2865674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 28665b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 2867674ae819SStefano Zampini } 2868674ae819SStefano Zampini } else { 2869674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 28705b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 2871674ae819SStefano Zampini } 2872674ae819SStefano Zampini } 2873674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 2874674ae819SStefano Zampini /* free workspace */ 2875674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 2876674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 2877674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 2878674ae819SStefano Zampini ierr = PetscFree(dof_sizes);CHKERRQ(ierr); 2879674ae819SStefano Zampini ierr = PetscFree(dof_displs);CHKERRQ(ierr); 2880674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 2881674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 2882674ae819SStefano Zampini PetscFunctionReturn(0); 2883674ae819SStefano Zampini } 28849a7d3425SStefano Zampini 28859a7d3425SStefano Zampini #undef __FUNCT__ 28869a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 28879a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 28889a7d3425SStefano Zampini { 28899a7d3425SStefano Zampini PetscInt i,j; 28909a7d3425SStefano Zampini PetscScalar *alphas; 28919a7d3425SStefano Zampini PetscErrorCode ierr; 28929a7d3425SStefano Zampini 28939a7d3425SStefano Zampini PetscFunctionBegin; 28949a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 2895785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 28969a7d3425SStefano Zampini for (i=0;i<n;i++) { 28979a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 28989a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 28999a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 29009a7d3425SStefano Zampini } 29019a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 29029a7d3425SStefano Zampini PetscFunctionReturn(0); 29039a7d3425SStefano Zampini } 29049a7d3425SStefano Zampini 2905e7931f94SStefano Zampini #undef __FUNCT__ 290670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 290728143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 2908e7931f94SStefano Zampini { 2909e7931f94SStefano Zampini Mat subdomain_adj; 2910e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 2911e7931f94SStefano Zampini MatPartitioning partitioner; 2912e7931f94SStefano Zampini Mat_IS *matis; 2913e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 2914e7931f94SStefano Zampini PetscInt prank; 2915e7931f94SStefano Zampini PetscMPIInt size,rank,color; 2916e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 2917e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 29183837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 2919e7931f94SStefano Zampini PetscErrorCode ierr; 29202b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 2921e7931f94SStefano Zampini PetscSubcomm subcomm; 2922a57a6d2fSStefano Zampini 2923e7931f94SStefano Zampini PetscFunctionBegin; 29242b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 29252b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 29262b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 2927e7931f94SStefano Zampini 2928e7931f94SStefano Zampini /* Get info on mapping */ 2929e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 2930e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 2931e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 2932e7931f94SStefano Zampini 2933e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 2934785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 2935e7931f94SStefano Zampini xadj[0] = 0; 2936e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 2937785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 2938785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 2939e7931f94SStefano Zampini 29402b510759SStefano Zampini if (threshold) { 2941d023bfaeSStefano Zampini PetscInt xadj_count = 0; 29422b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 2943d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 2944d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 2945d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 2946d023bfaeSStefano Zampini xadj_count++; 2947e7931f94SStefano Zampini } 2948e7931f94SStefano Zampini } 2949d023bfaeSStefano Zampini xadj[1] = xadj_count; 2950c8587f34SStefano Zampini } else { 2951e7931f94SStefano Zampini if (xadj[1]) { 2952e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 2953e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 2954c8587f34SStefano Zampini } 2955e7931f94SStefano Zampini } 2956e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 2957e7931f94SStefano Zampini if (use_square) { 2958e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 2959e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 2960e7931f94SStefano Zampini } 2961e7931f94SStefano Zampini } 2962e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 2963e7931f94SStefano Zampini 29643837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 2965e7931f94SStefano Zampini 2966e7931f94SStefano Zampini /* 2967e7931f94SStefano Zampini Restrict work on active processes only. 2968e7931f94SStefano Zampini */ 2969e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 2970e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 2971e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 29722b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 2973d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 2974e7931f94SStefano Zampini if (color) { 2975e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 2976e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 2977e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 2978c8587f34SStefano Zampini } else { 297928143c3dSStefano Zampini PetscInt coarsening_ratio; 2980e7931f94SStefano Zampini ierr = MPI_Comm_size(subcomm->comm,&size);CHKERRQ(ierr); 2981785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 2982e7931f94SStefano Zampini prank = rank; 2983e7931f94SStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm->comm);CHKERRQ(ierr); 29848002ef2cSStefano Zampini /* 2985e7931f94SStefano Zampini for (i=0;i<size;i++) { 2986e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 2987c8587f34SStefano Zampini } 29888002ef2cSStefano Zampini */ 2989e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 2990e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 2991c8587f34SStefano Zampini } 2992e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 2993e7931f94SStefano Zampini ierr = MatCreateMPIAdj(subcomm->comm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 299422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 2995e7931f94SStefano Zampini 2996e7931f94SStefano Zampini /* Partition */ 2997e7931f94SStefano Zampini ierr = MatPartitioningCreate(subcomm->comm,&partitioner);CHKERRQ(ierr); 2998e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 2999e7931f94SStefano Zampini if (use_vwgt) { 30003837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3001e7931f94SStefano Zampini v_wgt[0] = local_size; 3002e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3003c8587f34SStefano Zampini } 300428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 300528143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 300628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3007e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3008e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 300922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3010e7931f94SStefano Zampini 3011e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 301228143c3dSStefano Zampini if (contiguous) { 301328143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 301428143c3dSStefano Zampini } else { 301584ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 301628143c3dSStefano Zampini } 3017e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3018e7931f94SStefano Zampini /* clean up */ 3019e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3020e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3021e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3022e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3023e7931f94SStefano Zampini } 3024e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3025e7931f94SStefano Zampini 3026e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3027e7931f94SStefano Zampini i = 1; 3028e7931f94SStefano Zampini if (color) i=0; 3029e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3030e7931f94SStefano Zampini 3031e7931f94SStefano Zampini /* get back IS */ 3032e7931f94SStefano Zampini *is_sends = ranks_send_to; 3033e7931f94SStefano Zampini PetscFunctionReturn(0); 3034e7931f94SStefano Zampini } 3035e7931f94SStefano Zampini 3036e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3037e7931f94SStefano Zampini 3038e7931f94SStefano Zampini #undef __FUNCT__ 3039e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 304028143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3041e7931f94SStefano Zampini { 304270cf5478SStefano Zampini Mat local_mat; 3043e7931f94SStefano Zampini Mat_IS *matis; 3044e7931f94SStefano Zampini IS is_sends_internal; 30459d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 304628143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 30479d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3048e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3049e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3050e7931f94SStefano Zampini const PetscInt* is_indices; 3051e7931f94SStefano Zampini MatType new_local_type; 3052e7931f94SStefano Zampini /* buffers */ 3053e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 305428143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 30559d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3056e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3057e7931f94SStefano Zampini /* MPI */ 305828143c3dSStefano Zampini MPI_Comm comm,comm_n; 305928143c3dSStefano Zampini PetscSubcomm subcomm; 3060e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 306128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 306228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 306328143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 306428143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 306528143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3066e7931f94SStefano Zampini PetscErrorCode ierr; 3067e7931f94SStefano Zampini 3068e7931f94SStefano Zampini PetscFunctionBegin; 306928143c3dSStefano Zampini /* TODO: add missing checks */ 307028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 307128143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 307228143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 307328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3074e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 307528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3076e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3077e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3078e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3079e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3080e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 308128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 308270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 308370cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 308428143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 308570cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 308670cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 308770cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 308870cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 308970cf5478SStefano Zampini } 3090e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3091e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3092e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3093e7931f94SStefano Zampini if (!is_sends) { 309428143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 309528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 309628143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3097c8587f34SStefano Zampini } else { 3098e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3099e7931f94SStefano Zampini is_sends_internal = is_sends; 3100c8587f34SStefano Zampini } 3101e7931f94SStefano Zampini 3102e7931f94SStefano Zampini /* get pointer of MATIS data */ 3103e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3104e7931f94SStefano Zampini 3105e7931f94SStefano Zampini /* get comm */ 3106a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3107e7931f94SStefano Zampini 3108e7931f94SStefano Zampini /* compute number of sends */ 3109e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3110e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3111e7931f94SStefano Zampini 3112e7931f94SStefano Zampini /* compute number of receives */ 3113e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3114785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3115e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3116e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3117e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3118e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3119e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3120e7931f94SStefano Zampini 312128143c3dSStefano Zampini /* restrict comm if requested */ 312228143c3dSStefano Zampini subcomm = 0; 312328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 312428143c3dSStefano Zampini if (restrict_comm) { 312528143c3dSStefano Zampini PetscMPIInt color,rank,subcommsize; 312628143c3dSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 312728143c3dSStefano Zampini color = 0; 312828143c3dSStefano Zampini if (n_sends && !n_recvs) color = 1; /* sending only processes will not partecipate in new comm */ 312928143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 313028143c3dSStefano Zampini subcommsize = commsize - subcommsize; 313128143c3dSStefano Zampini /* check if reuse has been requested */ 313228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 313328143c3dSStefano Zampini if (*mat_n) { 313428143c3dSStefano Zampini PetscMPIInt subcommsize2; 313528143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 313628143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 313728143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 313828143c3dSStefano Zampini } else { 313928143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 314028143c3dSStefano Zampini } 314128143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 314228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 314328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 314428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 314528143c3dSStefano Zampini comm_n = subcomm->comm; 314628143c3dSStefano Zampini } 314728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 314828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 314928143c3dSStefano Zampini } else { 315028143c3dSStefano Zampini comm_n = comm; 315128143c3dSStefano Zampini } 315228143c3dSStefano Zampini 3153e7931f94SStefano Zampini /* prepare send/receive buffers */ 3154785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3155e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3156785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3157e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 315828143c3dSStefano Zampini if (nis) { 3159854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 316028143c3dSStefano Zampini } 3161e7931f94SStefano Zampini 316228143c3dSStefano Zampini /* Get data from local matrices */ 3163e7931f94SStefano Zampini if (!isdense) { 3164a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3165e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3166e7931f94SStefano Zampini /* 3167e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3168e7931f94SStefano Zampini send_buffer_idxs should contain: 3169e7931f94SStefano Zampini - MatType_PRIVATE type 3170e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3171e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3172e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3173e7931f94SStefano Zampini */ 3174e7931f94SStefano Zampini } else { 3175e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3176e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3177854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3178e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3179e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3180e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3181e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3182e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3183e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3184e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3185e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3186e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3187c8587f34SStefano Zampini } 3188c8587f34SStefano Zampini } 3189e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 319028143c3dSStefano Zampini /* additional is (if any) */ 319128143c3dSStefano Zampini if (nis) { 319228143c3dSStefano Zampini PetscMPIInt psum; 319328143c3dSStefano Zampini PetscInt j; 319428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 319528143c3dSStefano Zampini PetscInt plen; 319628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 319728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 319828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 319928143c3dSStefano Zampini } 3200854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 320128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 320228143c3dSStefano Zampini PetscInt plen; 320328143c3dSStefano Zampini const PetscInt *is_array_idxs; 320428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 320528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 320628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 320728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 320828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 320928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 321028143c3dSStefano Zampini } 321128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 321228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 321328143c3dSStefano Zampini } 321428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 321528143c3dSStefano Zampini } 321628143c3dSStefano Zampini 3217e7931f94SStefano Zampini buf_size_idxs = 0; 3218e7931f94SStefano Zampini buf_size_vals = 0; 321928143c3dSStefano Zampini buf_size_idxs_is = 0; 3220e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3221e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3222e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 322328143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3224e7931f94SStefano Zampini } 3225785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3226785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 322795ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3228e7931f94SStefano Zampini 3229e7931f94SStefano Zampini /* get new tags for clean communications */ 3230e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3231e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 323228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3233e7931f94SStefano Zampini 3234e7931f94SStefano Zampini /* allocate for requests */ 3235785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3236785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 323795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3238785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3239785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 324095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3241e7931f94SStefano Zampini 3242e7931f94SStefano Zampini /* communications */ 3243e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3244e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 324528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3246e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3247e7931f94SStefano Zampini source_dest = onodes[i]; 3248e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3249e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3250e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3251e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 325228143c3dSStefano Zampini if (nis) { 325328143c3dSStefano Zampini ierr = MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);CHKERRQ(ierr); 325428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 325528143c3dSStefano Zampini } 3256e7931f94SStefano Zampini } 3257e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3258e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3259e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3260e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 326128143c3dSStefano Zampini if (nis) { 326228143c3dSStefano Zampini ierr = MPI_Isend(send_buffer_idxs_is,ilengths_idxs_is[source_dest],MPIU_INT,source_dest,tag_idxs_is,comm,&send_req_idxs_is[i]);CHKERRQ(ierr); 326328143c3dSStefano Zampini } 3264e7931f94SStefano Zampini } 3265e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3266e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3267e7931f94SStefano Zampini 3268e7931f94SStefano Zampini /* assemble new l2g map */ 3269e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3270e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 32719d30be91SStefano Zampini new_local_rows = 0; 3272e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 32739d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3274e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3275e7931f94SStefano Zampini } 32769d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3277e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 32789d30be91SStefano Zampini new_local_rows = 0; 3279e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 32809d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 32819d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3282e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3283e7931f94SStefano Zampini } 32849d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 32859d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3286e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3287e7931f94SStefano Zampini 3288e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3289e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3290e7931f94SStefano Zampini /* it also assumes that if the block size is set, than it is the same among all local matrices (see checks at the beginning of the function) */ 3291e7931f94SStefano Zampini if (n_recvs) { 329228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3293e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3294e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3295e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3296e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3297e7931f94SStefano Zampini break; 3298e7931f94SStefano Zampini } 3299e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3300e7931f94SStefano Zampini } 3301e7931f94SStefano Zampini switch (new_local_type_private) { 330228143c3dSStefano Zampini case MATDENSE_PRIVATE: 330328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3304e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3305e7931f94SStefano Zampini bs = 1; 330628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 330728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 330828143c3dSStefano Zampini bs = 1; 330928143c3dSStefano Zampini } 3310e7931f94SStefano Zampini break; 3311e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3312e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3313e7931f94SStefano Zampini bs = 1; 3314e7931f94SStefano Zampini break; 3315e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3316e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3317e7931f94SStefano Zampini break; 3318e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3319e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3320e7931f94SStefano Zampini break; 3321e7931f94SStefano Zampini default: 33229d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3323e7931f94SStefano Zampini break; 3324e7931f94SStefano Zampini } 332528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 332628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 332728143c3dSStefano Zampini bs = 1; 3328e7931f94SStefano Zampini } 3329e7931f94SStefano Zampini 333070cf5478SStefano Zampini /* create MATIS object if needed */ 333170cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3332e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 333328143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 333470cf5478SStefano Zampini } else { 333570cf5478SStefano Zampini /* it also destroys the local matrices */ 333670cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 333770cf5478SStefano Zampini } 333870cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3339e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 33409d30be91SStefano Zampini 33419d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 33429d30be91SStefano Zampini 33439d30be91SStefano Zampini /* Global to local map of received indices */ 33449d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 33459d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 33469d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 33479d30be91SStefano Zampini 33489d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 33499d30be91SStefano Zampini buf_size_idxs = 0; 33509d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 33519d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 33529d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 33539d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 33549d30be91SStefano Zampini } 33559d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 33569d30be91SStefano Zampini 33579d30be91SStefano Zampini /* set preallocation */ 33589d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 33599d30be91SStefano Zampini if (!newisdense) { 33609d30be91SStefano Zampini PetscInt *new_local_nnz=0; 33619d30be91SStefano Zampini 33629d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 33639d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 33649d30be91SStefano Zampini if (n_recvs) { 33659d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 33669d30be91SStefano Zampini } 33679d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 33689d30be91SStefano Zampini PetscInt j; 33699d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 33709d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 33719d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 33729d30be91SStefano Zampini } 33739d30be91SStefano Zampini } else { 33749d30be91SStefano Zampini /* TODO */ 33759d30be91SStefano Zampini } 33769d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 33779d30be91SStefano Zampini } 33789d30be91SStefano Zampini if (new_local_nnz) { 33799d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 33809d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 33819d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 33829d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 33839d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 33849d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 33859d30be91SStefano Zampini } else { 33869d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 33879d30be91SStefano Zampini } 33889d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 33899d30be91SStefano Zampini } else { 33909d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 33919d30be91SStefano Zampini } 3392e7931f94SStefano Zampini 3393e7931f94SStefano Zampini /* set values */ 3394e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 33959d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3396e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3397e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3398e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 33999d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3400e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3401e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3402e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 340328143c3dSStefano Zampini } else { 340428143c3dSStefano Zampini /* TODO */ 3405e7931f94SStefano Zampini } 3406e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3407e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3408e7931f94SStefano Zampini } 3409e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3410e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 341170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 341270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 34139d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 34149d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3415e7931f94SStefano Zampini 3416dfd14d43SStefano Zampini #if 0 341728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3418e7931f94SStefano Zampini Vec lvec,rvec; 3419e7931f94SStefano Zampini PetscReal infty_error; 3420e7931f94SStefano Zampini 34212a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3422e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3423e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3424e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 342570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3426e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3427e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3428e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3429e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3430e7931f94SStefano Zampini } 343128143c3dSStefano Zampini #endif 3432e7931f94SStefano Zampini 343328143c3dSStefano Zampini /* assemble new additional is (if any) */ 343428143c3dSStefano Zampini if (nis) { 343528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 343628143c3dSStefano Zampini 343728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3438854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 343928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 344028143c3dSStefano Zampini psum = 0; 344128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 344228143c3dSStefano Zampini for (j=0;j<nis;j++) { 344328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 344428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 344528143c3dSStefano Zampini psum += plen; 344628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 344728143c3dSStefano Zampini } 344828143c3dSStefano Zampini } 3449854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3450854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 345128143c3dSStefano Zampini for (i=1;i<nis;i++) { 345228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 345328143c3dSStefano Zampini } 345428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 345528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 345628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 345728143c3dSStefano Zampini for (j=0;j<nis;j++) { 345828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 345928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 346028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 346128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 346228143c3dSStefano Zampini } 346328143c3dSStefano Zampini } 346428143c3dSStefano Zampini for (i=0;i<nis;i++) { 346528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 346628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 346728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 346828143c3dSStefano Zampini } 346928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 347028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 347128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 347228143c3dSStefano Zampini } 3473e7931f94SStefano Zampini /* free workspace */ 347428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3475e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3476e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3477e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3478e7931f94SStefano Zampini if (isdense) { 3479e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3480e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3481e7931f94SStefano Zampini } else { 3482e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3483e7931f94SStefano Zampini } 348428143c3dSStefano Zampini if (nis) { 348528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 348628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 348728143c3dSStefano Zampini } 3488e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3489e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 349028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3491e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3492e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 349328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3494e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3495e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3496e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3497e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3498e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 349928143c3dSStefano Zampini if (nis) { 350028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 350128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 350228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 350328143c3dSStefano Zampini } 350428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 350528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 350628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 350728143c3dSStefano Zampini for (i=0;i<nis;i++) { 350828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 350928143c3dSStefano Zampini } 351028143c3dSStefano Zampini } 3511e7931f94SStefano Zampini PetscFunctionReturn(0); 3512e7931f94SStefano Zampini } 3513a57a6d2fSStefano Zampini 351412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 351512edc857SStefano Zampini #include <petsc-private/kspimpl.h> 351612edc857SStefano Zampini 3517c8587f34SStefano Zampini #undef __FUNCT__ 3518c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3519c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3520c8587f34SStefano Zampini { 3521c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3522c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 352320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 35249881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 352520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 35266e683305SStefano Zampini IS coarse_is,*isarray; 35276e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 35286e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3529f9eb5b7dSStefano Zampini PC pc_temp; 3530c8587f34SStefano Zampini PCType coarse_pc_type; 3531c8587f34SStefano Zampini KSPType coarse_ksp_type; 3532f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 35334f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 35346e683305SStefano Zampini Mat t_coarse_mat_is; 35356e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 35366e683305SStefano Zampini PetscMPIInt all_procs; 353774e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 353868457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 353922bc73bbSStefano Zampini PetscScalar *array; 35409881197aSStefano Zampini PetscErrorCode ierr; 3541fdc09c96SStefano Zampini 3542c8587f34SStefano Zampini PetscFunctionBegin; 3543c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 354468457ee5SStefano Zampini if (pcbddc->new_primal_space || pcbddc->coarse_size == -1) { /* a new primal space is present or it is the first initialization, so recompute global numbering */ 354568457ee5SStefano Zampini compute_vecs = PETSC_TRUE; 3546fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 3547fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3548f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3549f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3550f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3551fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3552fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3553727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3554fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3555fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3556fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3557f4ddd8eeSStefano Zampini } 3558fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3559fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3560f4ddd8eeSStefano Zampini } 356170cf5478SStefano Zampini /* reset any subassembling information */ 356270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 35636e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 35646e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3565fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3566f4ddd8eeSStefano Zampini } 3567c8587f34SStefano Zampini 35686e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 35692b510759SStefano Zampini im_active = !!(pcis->n); 35702b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 35716e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 35726e683305SStefano Zampini void_procs = all_procs-active_procs; 35736e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 357474e2c79eSStefano Zampini redist = PETSC_FALSE; 357522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 35766e683305SStefano Zampini csin_ml = PETSC_TRUE; 35776e683305SStefano Zampini ncoarse_ml = void_procs; 35786e683305SStefano Zampini csin_ds = PETSC_TRUE; 35796e683305SStefano Zampini ncoarse_ds = void_procs; 35806e683305SStefano Zampini } else { 35816e683305SStefano Zampini csin_ml = PETSC_FALSE; 35826e683305SStefano Zampini ncoarse_ml = all_procs; 35836e683305SStefano Zampini if (void_procs) { 35846e683305SStefano Zampini csin_ds = PETSC_TRUE; 35856e683305SStefano Zampini ncoarse_ds = void_procs; 35866e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 35876e683305SStefano Zampini } else { 358874e2c79eSStefano Zampini if (pcbddc->redistribute_coarse && pcbddc->redistribute_coarse < all_procs) { 358974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 359074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 359174e2c79eSStefano Zampini redist = PETSC_TRUE; 359274e2c79eSStefano Zampini } else { 35936e683305SStefano Zampini csin_ds = PETSC_FALSE; 35946e683305SStefano Zampini ncoarse_ds = all_procs; 35956e683305SStefano Zampini } 35966e683305SStefano Zampini } 359774e2c79eSStefano Zampini } 35986e683305SStefano Zampini 35996e683305SStefano Zampini /* 36006e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 36016e683305SStefano Zampini - we have not exceeded the number of levels requested 36026e683305SStefano Zampini - we can actually subassemble the active processes 36036e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 36046e683305SStefano Zampini */ 36056e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 36066e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 36076e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 36086e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 36096e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3610f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 36112b510759SStefano Zampini } else { 3612f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3613c8587f34SStefano Zampini } 3614c8587f34SStefano Zampini } 36156e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 36166e683305SStefano Zampini if (multilevel_allowed) { 36176e683305SStefano Zampini ncoarse = ncoarse_ml; 36186e683305SStefano Zampini csin = csin_ml; 36196e683305SStefano Zampini } else { 36206e683305SStefano Zampini ncoarse = ncoarse_ds; 36216e683305SStefano Zampini csin = csin_ds; 36226e683305SStefano Zampini } 3623e7931f94SStefano Zampini 3624abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3625abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3626abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3627abbbba34SStefano Zampini 3628abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 362922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 363022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 363122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 363222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3633b9b85e73SStefano Zampini #if 0 3634b9b85e73SStefano Zampini { 3635b9b85e73SStefano Zampini PetscViewer viewer; 3636b9b85e73SStefano Zampini char filename[256]; 3637b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 3638b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3639b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3640b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 3641b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3642b9b85e73SStefano Zampini } 3643b9b85e73SStefano Zampini #endif 36446e683305SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,&t_coarse_mat_is);CHKERRQ(ierr); 36456e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 36466e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 36476e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3648abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 3649abbbba34SStefano Zampini 36506e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 36516e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 36526e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 36536e683305SStefano Zampini const PetscInt *idxs; 36546e683305SStefano Zampini ISLocalToGlobalMapping tmap; 36556e683305SStefano Zampini 36566e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 36570be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 36586e683305SStefano Zampini /* allocate space for temporary storage */ 3659854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 3660854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 36616e683305SStefano Zampini /* allocate for IS array */ 36626e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 36636e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 36646e683305SStefano Zampini nis = nisdofs + nisneu; 3665854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 36666e683305SStefano Zampini /* dofs splitting */ 36676e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 36686e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 36696e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 36706e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 36716e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 36726e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 36736e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 36746e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 36756e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 36766e683305SStefano Zampini } 36776e683305SStefano Zampini /* neumann boundaries */ 36786e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 36796e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 36806e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 36816e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 36826e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 36836e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 36846e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 36856e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 36866e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 36876e683305SStefano Zampini } 36886e683305SStefano Zampini /* free memory */ 36896e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 36906e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 36916e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 36926e683305SStefano Zampini } else { 36936e683305SStefano Zampini nis = 0; 36946e683305SStefano Zampini nisdofs = 0; 36956e683305SStefano Zampini nisneu = 0; 36966e683305SStefano Zampini isarray = NULL; 36976e683305SStefano Zampini } 36986e683305SStefano Zampini /* destroy no longer needed map */ 36996e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 37006e683305SStefano Zampini 37016e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 37026e683305SStefano Zampini coarse_mat_is = NULL; 37036e683305SStefano Zampini if (csin) { 37046e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 370574e2c79eSStefano Zampini if (redist) { 370674e2c79eSStefano Zampini PetscMPIInt rank; 370774e2c79eSStefano Zampini PetscInt spc,n_spc_p1,dest[1]; 370874e2c79eSStefano Zampini 370974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 371074e2c79eSStefano Zampini spc = all_procs/pcbddc->redistribute_coarse; 371174e2c79eSStefano Zampini n_spc_p1 = all_procs%pcbddc->redistribute_coarse; 371274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 371374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 371474e2c79eSStefano Zampini } else { 371574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 371674e2c79eSStefano Zampini } 371774e2c79eSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),1,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 371874e2c79eSStefano Zampini } else { 37196e683305SStefano Zampini PetscInt j,tissize,*nisindices; 37206e683305SStefano Zampini PetscInt *coarse_candidates; 37216e683305SStefano Zampini const PetscInt* tisindices; 37226e683305SStefano Zampini /* get coarse candidates' ranks in pc communicator */ 3723854ce69bSBarry Smith ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 37246e683305SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 37256e683305SStefano Zampini for (i=0,j=0;i<all_procs;i++) { 37266e683305SStefano Zampini if (!coarse_candidates[i]) { 37276e683305SStefano Zampini coarse_candidates[j]=i; 37286e683305SStefano Zampini j++; 37296e683305SStefano Zampini } 37306e683305SStefano Zampini } 37316e683305SStefano Zampini if (j < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",j,ncoarse); 37326e683305SStefano Zampini /* get a suitable subassembling pattern */ 37336e683305SStefano Zampini if (csin_type_simple) { 37346e683305SStefano Zampini PetscMPIInt rank; 37356e683305SStefano Zampini PetscInt issize,isidx; 37366e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 37376e683305SStefano Zampini if (im_active) { 37386e683305SStefano Zampini issize = 1; 37396e683305SStefano Zampini isidx = (PetscInt)rank; 37406e683305SStefano Zampini } else { 37416e683305SStefano Zampini issize = 0; 37426e683305SStefano Zampini isidx = -1; 37436e683305SStefano Zampini } 37446e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 37456e683305SStefano Zampini } else { 37466e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 37476e683305SStefano Zampini } 37486e683305SStefano Zampini if (pcbddc->dbg_flag) { 37496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 37516e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 37526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 37536e683305SStefano Zampini for (i=0;i<j;i++) { 37546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 37556e683305SStefano Zampini } 37566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 37576e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37586e683305SStefano Zampini } 37596e683305SStefano Zampini /* shift the pattern on coarse candidates */ 37606e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 37616e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 3762854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 37636e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 37646e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 37656e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 37666e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 37676e683305SStefano Zampini } 376874e2c79eSStefano Zampini } 37696e683305SStefano Zampini if (pcbddc->dbg_flag) { 37706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 37726e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 37736e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37746e683305SStefano Zampini } 37756e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 37766e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 37776e683305SStefano Zampini } else { 37786e683305SStefano Zampini if (pcbddc->dbg_flag) { 37796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 37816e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37826e683305SStefano Zampini } 37836e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 37846e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 37856e683305SStefano Zampini } 37866e683305SStefano Zampini 37876e683305SStefano Zampini /* create local to global scatters for coarse problem */ 378868457ee5SStefano Zampini if (compute_vecs) { 37896e683305SStefano Zampini PetscInt lrows; 37906e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 37916e683305SStefano Zampini if (coarse_mat_is) { 37926e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 37936e683305SStefano Zampini } else { 37946e683305SStefano Zampini lrows = 0; 37956e683305SStefano Zampini } 37966e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 37976e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 37986e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 37996e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 38006e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 38016e683305SStefano Zampini } 38026e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 38036e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 3804c8587f34SStefano Zampini 3805f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 3806f9eb5b7dSStefano Zampini if (multilevel_allowed) { 3807f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 3808f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 3809f9eb5b7dSStefano Zampini } else { 3810f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 3811f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 3812c8587f34SStefano Zampini } 3813c8587f34SStefano Zampini 38146e683305SStefano Zampini /* print some info if requested */ 38156e683305SStefano Zampini if (pcbddc->dbg_flag) { 38166e683305SStefano Zampini if (!multilevel_allowed) { 38176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 38186e683305SStefano Zampini if (multilevel_requested) { 38196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Not enough active processes on level %d (active processes %d, coarsening ratio %d)\n",pcbddc->current_level,active_procs,pcbddc->coarsening_ratio);CHKERRQ(ierr); 38206e683305SStefano Zampini } else if (pcbddc->max_levels) { 38216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 38226e683305SStefano Zampini } 38236e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 38246e683305SStefano Zampini } 38256e683305SStefano Zampini } 38266e683305SStefano Zampini 3827f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 38286e683305SStefano Zampini if (coarse_mat_is) { 38296e683305SStefano Zampini MatReuse coarse_mat_reuse; 38306a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 38316e683305SStefano Zampini if (pcbddc->dbg_flag) { 38326e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 38336e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 38346e683305SStefano Zampini } 3835f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 3836312be037SStefano Zampini char prefix[256],str_level[16]; 3837e604994aSStefano Zampini size_t len; 38386e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 3839c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 3840f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 38415f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 3842c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 38436e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 3844c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 3845c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 3846e604994aSStefano Zampini /* prefix */ 3847e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 3848e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3849e604994aSStefano Zampini if (!pcbddc->current_level) { 3850e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3851e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 3852c8587f34SStefano Zampini } else { 3853e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3854312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3855312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 385634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3857312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3858e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 3859e604994aSStefano Zampini } 3860e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 3861f9eb5b7dSStefano Zampini /* allow user customization */ 3862f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 38637e0def11SStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3864312be037SStefano Zampini } 3865f9eb5b7dSStefano Zampini 3866f9eb5b7dSStefano Zampini /* get some info after set from options */ 3867f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 3868f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 3869f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 38704f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 38716e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 3872f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 3873f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 3874f9eb5b7dSStefano Zampini } 38754f3a063dSStefano Zampini if (isredundant) { 38764f3a063dSStefano Zampini KSP inner_ksp; 38774f3a063dSStefano Zampini PC inner_pc; 38784f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 38794f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 38804f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 38814f3a063dSStefano Zampini } 3882f9eb5b7dSStefano Zampini 38836e683305SStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 3884f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 3885f9eb5b7dSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 3886f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 38876e683305SStefano Zampini if (nisdofs) { 38886e683305SStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 38896e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 38906e683305SStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 38911035eff8SStefano Zampini } 38921035eff8SStefano Zampini } 38936e683305SStefano Zampini if (nisneu) { 38946e683305SStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 38956e683305SStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 38961035eff8SStefano Zampini } 3897fdc09c96SStefano Zampini 3898f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 3899fa7f1dd8SStefano Zampini if (coarse_reuse) { 390081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 3901fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 39026e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 3903fa7f1dd8SStefano Zampini } else { 39046e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 3905fa7f1dd8SStefano Zampini } 3906c8587f34SStefano Zampini if (isbddc || isnn) { 390722bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 390870cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 390928143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 391022b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 39116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 39126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 39136e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 39146e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 391522b6e8a2SStefano Zampini } 391670cf5478SStefano Zampini } 39176e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 391870cf5478SStefano Zampini } else { 391922bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 392022bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 392122bc73bbSStefano Zampini } 392222bc73bbSStefano Zampini } else { 39232e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 3924c8587f34SStefano Zampini } 3925c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 3926c8587f34SStefano Zampini 3927c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 3928b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 39295a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 3930c8587f34SStefano Zampini 39316e683305SStefano Zampini /* set operators */ 39325f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 39336e683305SStefano Zampini if (pcbddc->dbg_flag) { 39346e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 39356e683305SStefano Zampini } 39366e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 39376e683305SStefano Zampini coarse_mat = 0; 39386e683305SStefano Zampini } 39396e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 3940b9b85e73SStefano Zampini #if 0 3941b9b85e73SStefano Zampini { 3942b9b85e73SStefano Zampini PetscViewer viewer; 3943b9b85e73SStefano Zampini char filename[256]; 3944b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 3945b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 3946b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3947b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 3948b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3949b9b85e73SStefano Zampini } 3950b9b85e73SStefano Zampini #endif 3951c8587f34SStefano Zampini 3952c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 3953c8587f34SStefano Zampini if (pcbddc->NullSpace) { 3954c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 395598a51de6SStefano Zampini } 395698a51de6SStefano Zampini 395798a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 395898a51de6SStefano Zampini Vec crhs,csol; 395998a51de6SStefano Zampini PetscBool ispreonly; 396098a51de6SStefano Zampini if (CoarseNullSpace) { 3961c8587f34SStefano Zampini if (isbddc) { 3962c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 3963c8587f34SStefano Zampini } else { 3964c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 3965c8587f34SStefano Zampini } 3966c8587f34SStefano Zampini } 3967f9eb5b7dSStefano Zampini /* setup coarse ksp */ 3968f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 3969f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 3970f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 39716e683305SStefano Zampini /* hack */ 3972f347579bSStefano Zampini if (!csol) { 39732a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 3974f9eb5b7dSStefano Zampini } 3975f347579bSStefano Zampini if (!crhs) { 39762a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 3977f347579bSStefano Zampini } 3978cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 3979cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 39806e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 3981c8587f34SStefano Zampini KSP check_ksp; 39822b510759SStefano Zampini KSPType check_ksp_type; 3983c8587f34SStefano Zampini PC check_pc; 39846e683305SStefano Zampini Vec check_vec,coarse_vec; 39856a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 39862b510759SStefano Zampini PetscInt its; 39876e683305SStefano Zampini PetscBool compute_eigs; 39886e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 39896e683305SStefano Zampini PetscInt neigs; 39908e185a42SStefano Zampini const char *prefix; 3991c8587f34SStefano Zampini 39922b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 39936e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 399423ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 3995f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 39962b510759SStefano Zampini if (ispreonly) { 39972b510759SStefano Zampini check_ksp_type = KSPPREONLY; 39986e683305SStefano Zampini compute_eigs = PETSC_FALSE; 39992b510759SStefano Zampini } else { 4000cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 40016e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4002c8587f34SStefano Zampini } 4003c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 40046e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 40056e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 40066e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4007a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4008a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4009a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4010a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4011c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4012c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4013c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4014c8587f34SStefano Zampini /* create random vec */ 40156e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 40166e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4017c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4018c8587f34SStefano Zampini if (CoarseNullSpace) { 4019c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4020c8587f34SStefano Zampini } 40216e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4022c8587f34SStefano Zampini /* solve coarse problem */ 40236e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4024c8587f34SStefano Zampini if (CoarseNullSpace) { 40256e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4026c8587f34SStefano Zampini } 4027cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 40286e683305SStefano Zampini if (compute_eigs) { 4029854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4030854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 40316e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 40326e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 40336e683305SStefano Zampini lambda_min = eigs_r[0]; 40346e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 40356e683305SStefano Zampini if (lambda_max>lambda_min) { 4036cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4037cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4038cbcc2c2aSStefano Zampini } 4039c8587f34SStefano Zampini } 4040c8587f34SStefano Zampini } 4041cbcc2c2aSStefano Zampini 4042c8587f34SStefano Zampini /* check coarse problem residual error */ 40436e683305SStefano Zampini if (pcbddc->dbg_flag) { 40446e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 40456e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 40466e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4047c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 40486e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 40496e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4050c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 40516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (%d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 40526e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 40536e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 40546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 40556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 40566e683305SStefano Zampini if (compute_eigs) { 40576e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4058deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4059c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 40606e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 40616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,lambda_min,lambda_max,lambda_min_s,lambda_max_s);CHKERRQ(ierr); 40626e683305SStefano Zampini for (i=0;i<neigs;i++) { 40636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4064c8587f34SStefano Zampini } 40656e683305SStefano Zampini } 40666e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 40676e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 40686e683305SStefano Zampini } 4069c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 40706e683305SStefano Zampini if (compute_eigs) { 40716e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 40726e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4073c8587f34SStefano Zampini } 40746e683305SStefano Zampini } 40756e683305SStefano Zampini } 4076cbcc2c2aSStefano Zampini /* print additional info */ 4077cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 40786e683305SStefano Zampini /* waits until all processes reaches this point */ 40796e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4080cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4081cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4082cbcc2c2aSStefano Zampini } 4083cbcc2c2aSStefano Zampini 40842b510759SStefano Zampini /* free memory */ 4085c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4086fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4087c8587f34SStefano Zampini PetscFunctionReturn(0); 4088c8587f34SStefano Zampini } 4089674ae819SStefano Zampini 4090f34684f1SStefano Zampini #undef __FUNCT__ 4091f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4092f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4093f34684f1SStefano Zampini { 4094f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4095f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4096f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4097727cdba6SStefano Zampini PetscInt i,coarse_size; 4098727cdba6SStefano Zampini PetscInt *local_primal_indices; 4099f34684f1SStefano Zampini PetscErrorCode ierr; 4100f34684f1SStefano Zampini 4101f34684f1SStefano Zampini PetscFunctionBegin; 4102f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 410389c96988SStefano Zampini if (!pcbddc->primal_indices_local_idxs && pcbddc->local_primal_size) { 410489c96988SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Local primal indices have not been created"); 4105727cdba6SStefano Zampini } 4106727cdba6SStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,NULL,&coarse_size,&local_primal_indices);CHKERRQ(ierr); 4107f34684f1SStefano Zampini 4108f34684f1SStefano Zampini /* check numbering */ 4109f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4110f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4111b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4112f34684f1SStefano Zampini 4113f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4114f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4115f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 41160fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4117f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4118f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4119727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4120f34684f1SStefano Zampini } 4121f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4122f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4123f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4124f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4125f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4126f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4127f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4128f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4129f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4130f34684f1SStefano Zampini if (array[i] == 1.0) { 4131b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4132f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4133f34684f1SStefano Zampini } 4134f34684f1SStefano Zampini } 4135b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4136f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4137f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4138f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4139f34684f1SStefano Zampini } 4140f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4141f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4142f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4143f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4144f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4145f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4146b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4147f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4148f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4149f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4150f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4151727cdba6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%d]=%d (%d)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i]); 4152f34684f1SStefano Zampini } 4153f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4154f34684f1SStefano Zampini } 4155f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4156b9b85e73SStefano Zampini if (set_error_reduced) { 4157b9b85e73SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4158b9b85e73SStefano Zampini } 4159f34684f1SStefano Zampini } 4160f34684f1SStefano Zampini /* get back data */ 4161f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4162f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4163674ae819SStefano Zampini PetscFunctionReturn(0); 4164674ae819SStefano Zampini } 4165674ae819SStefano Zampini 4166e456f2a8SStefano Zampini #undef __FUNCT__ 4167e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4168a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4169e456f2a8SStefano Zampini { 4170e456f2a8SStefano Zampini IS localis_t; 4171a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4172e456f2a8SStefano Zampini PetscScalar *vals; 4173e456f2a8SStefano Zampini PetscErrorCode ierr; 4174e456f2a8SStefano Zampini 4175e456f2a8SStefano Zampini PetscFunctionBegin; 4176a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4177e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4178854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4179e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4180e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4181a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4182a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 41831035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4184a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 41851035eff8SStefano Zampini } 4186a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4187e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4188e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4189a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4190a7dc3881SStefano Zampini /* now compute set in local ordering */ 4191a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4192a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4193a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4194a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4195a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4196ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4197e456f2a8SStefano Zampini lsize++; 4198e456f2a8SStefano Zampini } 4199e456f2a8SStefano Zampini } 4200854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4201a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4202ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4203e456f2a8SStefano Zampini idxs[lsize++] = i; 4204e456f2a8SStefano Zampini } 4205e456f2a8SStefano Zampini } 4206a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4207a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4208e456f2a8SStefano Zampini *localis = localis_t; 4209e456f2a8SStefano Zampini PetscFunctionReturn(0); 4210e456f2a8SStefano Zampini } 4211906d46d4SStefano Zampini 4212906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4213906d46d4SStefano Zampini #undef __FUNCT__ 4214906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4215906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4216906d46d4SStefano Zampini { 4217906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4218906d46d4SStefano Zampini PetscErrorCode ierr; 4219906d46d4SStefano Zampini 4220906d46d4SStefano Zampini PetscFunctionBegin; 4221906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4222906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4223906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4224906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4225906d46d4SStefano Zampini PetscFunctionReturn(0); 4226906d46d4SStefano Zampini } 4227906d46d4SStefano Zampini 4228906d46d4SStefano Zampini #undef __FUNCT__ 4229906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4230906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4231906d46d4SStefano Zampini { 4232906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4233906d46d4SStefano Zampini PetscErrorCode ierr; 4234906d46d4SStefano Zampini 4235906d46d4SStefano Zampini PetscFunctionBegin; 4236906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4237906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4238906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4239906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4240906d46d4SStefano Zampini PetscFunctionReturn(0); 4241906d46d4SStefano Zampini } 4242