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; 17f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 185e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 19c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 20c8587f34SStefano Zampini 21*684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 220fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 23*684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 24c8587f34SStefano Zampini 25c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 26b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 27c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 28c8587f34SStefano Zampini } 29c8587f34SStefano Zampini 308629588bSStefano Zampini /* 318629588bSStefano Zampini Setup local correction and local part of coarse basis. 328629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 338629588bSStefano Zampini */ 3447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 358629588bSStefano Zampini 368629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 378629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 388629588bSStefano Zampini 398629588bSStefano Zampini /* free */ 408629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 41c8587f34SStefano Zampini PetscFunctionReturn(0); 42c8587f34SStefano Zampini } 43c8587f34SStefano Zampini 44c8587f34SStefano Zampini #undef __FUNCT__ 45674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 46674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 47674ae819SStefano Zampini { 48674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 49674ae819SStefano Zampini PetscErrorCode ierr; 50674ae819SStefano Zampini 51674ae819SStefano Zampini PetscFunctionBegin; 52674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 53674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 54674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 55674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 56785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 57674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 58f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 59f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 60785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 6163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 6263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 63674ae819SStefano Zampini PetscFunctionReturn(0); 64674ae819SStefano Zampini } 65674ae819SStefano Zampini 66674ae819SStefano Zampini #undef __FUNCT__ 67674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 68674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 69674ae819SStefano Zampini { 70674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 71674ae819SStefano Zampini PetscErrorCode ierr; 72674ae819SStefano Zampini 73674ae819SStefano Zampini PetscFunctionBegin; 74b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 75674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 76674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 77674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 785ff63025SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 79674ae819SStefano Zampini PetscFunctionReturn(0); 80674ae819SStefano Zampini } 81674ae819SStefano Zampini 82674ae819SStefano Zampini #undef __FUNCT__ 83674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 84674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 85674ae819SStefano Zampini { 86674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 87674ae819SStefano Zampini PetscErrorCode ierr; 88674ae819SStefano Zampini 89674ae819SStefano Zampini PetscFunctionBegin; 90674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 91674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_rhs);CHKERRQ(ierr); 92674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 93674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 9415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 9515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 96674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 97674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 98674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 99674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 100674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 101674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 1028ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 103674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 104674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 105674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 106f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 107f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 108f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 109f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 110727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 111f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 11270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 1136e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 114674ae819SStefano Zampini PetscFunctionReturn(0); 115674ae819SStefano Zampini } 116674ae819SStefano Zampini 117674ae819SStefano Zampini #undef __FUNCT__ 118f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 119f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 1206bfb1811SStefano Zampini { 1216bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1226bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1236bfb1811SStefano Zampini VecType impVecType; 124e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 1256bfb1811SStefano Zampini PetscErrorCode ierr; 1266bfb1811SStefano Zampini 1276bfb1811SStefano Zampini PetscFunctionBegin; 128f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 129f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 130f4ddd8eeSStefano Zampini } 131e7b262bdSStefano Zampini /* get sizes */ 132e9189074SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_actual_vertices; 133e9189074SStefano Zampini n_R = pcis->n-pcbddc->n_actual_vertices; 1346bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 135e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 136e7b262bdSStefano Zampini /* R nodes */ 137e7b262bdSStefano Zampini old_size = -1; 138e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 139e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 140e7b262bdSStefano Zampini } 141e7b262bdSStefano Zampini if (n_R != old_size) { 142e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 143e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 1446bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 1456bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 1466bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 1476bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 148e7b262bdSStefano Zampini } 149e7b262bdSStefano Zampini /* local primal dofs */ 150e7b262bdSStefano Zampini old_size = -1; 151e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 152e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 153e7b262bdSStefano Zampini } 154e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 155e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 15683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 157e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 1586bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 159e7b262bdSStefano Zampini } 160e7b262bdSStefano Zampini /* local explicit constraints */ 161e7b262bdSStefano Zampini old_size = -1; 162e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 163e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 164e7b262bdSStefano Zampini } 165e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 166e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 16783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 16883b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 16983b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 17083b7ccabSStefano Zampini } 1716bfb1811SStefano Zampini PetscFunctionReturn(0); 1726bfb1811SStefano Zampini } 1736bfb1811SStefano Zampini 1746bfb1811SStefano Zampini #undef __FUNCT__ 17547f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 17647f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 17788ebb749SStefano Zampini { 17825084f0cSStefano Zampini PetscErrorCode ierr; 17925084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 18088ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 18188ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 18225084f0cSStefano Zampini /* submatrices of local problem */ 18388ebb749SStefano Zampini Mat A_RV,A_VR,A_VV; 18425084f0cSStefano Zampini /* working matrices */ 18525084f0cSStefano Zampini Mat M1,M2,M3,C_CR; 18625084f0cSStefano Zampini /* working vectors */ 18725084f0cSStefano Zampini Vec vec1_C,vec2_C,vec1_V,vec2_V; 18825084f0cSStefano Zampini /* additional working stuff */ 18925084f0cSStefano Zampini IS is_aux; 19025084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 19125084f0cSStefano Zampini const PetscScalar *array,*row_cmat_values; 19225084f0cSStefano Zampini const PetscInt *row_cmat_indices,*idx_R_local; 193e9189074SStefano Zampini PetscInt *idx_V_B,*auxindices; 19425084f0cSStefano Zampini PetscInt n_vertices,n_constraints,size_of_constraint; 19525084f0cSStefano Zampini PetscInt i,j,n_R,n_D,n_B; 196b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 19745a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 19888ebb749SStefano Zampini MatType impMatType; 19925084f0cSStefano Zampini /* some shortcuts to scalars */ 20025084f0cSStefano Zampini PetscScalar zero=0.0,one=1.0,m_one=-1.0; 20125084f0cSStefano Zampini /* for debugging purposes */ 20288ebb749SStefano Zampini PetscReal *coarsefunctions_errors,*constraints_errors; 20388ebb749SStefano Zampini 20488ebb749SStefano Zampini PetscFunctionBegin; 205e9189074SStefano Zampini /* get number of vertices (corners plus constraints with change of basis) 206e9189074SStefano Zampini pcbddc->n_actual_vertices stores the actual number of vertices, pcbddc->n_vertices the number of corners computed */ 207e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 20888ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 20988ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 21088ebb749SStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 21188ebb749SStefano Zampini n_R = pcis->n-n_vertices; 21288ebb749SStefano Zampini 21388ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 21488ebb749SStefano Zampini impMatType = MATSEQDENSE; 21588ebb749SStefano Zampini 21688ebb749SStefano Zampini /* Allocating some extra storage just to be safe */ 21788ebb749SStefano Zampini ierr = PetscMalloc (pcis->n*sizeof(PetscInt),&auxindices);CHKERRQ(ierr); 21888ebb749SStefano Zampini for (i=0;i<pcis->n;i++) auxindices[i]=i; 21988ebb749SStefano Zampini 22088ebb749SStefano Zampini /* vertices in boundary numbering */ 221785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 2225e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->primal_indices_local_idxs,&i,idx_V_B);CHKERRQ(ierr); 22388ebb749SStefano Zampini if (i != n_vertices) { 22422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 22588ebb749SStefano Zampini } 22688ebb749SStefano Zampini 22788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 22888ebb749SStefano Zampini if (n_constraints) { 229f4ddd8eeSStefano Zampini /* see if we can save some allocations */ 230f4ddd8eeSStefano Zampini if (pcbddc->local_auxmat2) { 231f4ddd8eeSStefano Zampini PetscInt on_R,on_constraints; 232f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->local_auxmat2,&on_R,&on_constraints);CHKERRQ(ierr); 233f4ddd8eeSStefano Zampini if (on_R != n_R || on_constraints != n_constraints) { 234f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 235f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 236f4ddd8eeSStefano Zampini } 237f4ddd8eeSStefano Zampini } 23845a1bb75SStefano Zampini /* work vectors */ 23945a1bb75SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_C,&vec1_C);CHKERRQ(ierr); 24045a1bb75SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_C,&vec2_C);CHKERRQ(ierr); 24145a1bb75SStefano Zampini /* auxiliary matrices */ 242f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat2) { 24388ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->local_auxmat2);CHKERRQ(ierr); 24425084f0cSStefano Zampini ierr = MatSetSizes(pcbddc->local_auxmat2,n_R,n_constraints,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 24588ebb749SStefano Zampini ierr = MatSetType(pcbddc->local_auxmat2,impMatType);CHKERRQ(ierr); 24625084f0cSStefano Zampini ierr = MatSetUp(pcbddc->local_auxmat2);CHKERRQ(ierr); 247f4ddd8eeSStefano Zampini } 24888ebb749SStefano Zampini 24925084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 25025084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 2518ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 25225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 25388ebb749SStefano Zampini 25488ebb749SStefano Zampini /* Assemble local_auxmat2 = - A_{RR}^{-1} C^T_{CR} needed by BDDC application */ 25588ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 25688ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 25788ebb749SStefano Zampini /* Get row of constraint matrix in R numbering */ 25825084f0cSStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 25925084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_R,size_of_constraint,row_cmat_indices,row_cmat_values,INSERT_VALUES);CHKERRQ(ierr); 26025084f0cSStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 26125084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_R);CHKERRQ(ierr); 26225084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_R);CHKERRQ(ierr); 26388ebb749SStefano Zampini /* Solve for row of constraint matrix in R numbering */ 26488ebb749SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 26525084f0cSStefano Zampini /* Set values in local_auxmat2 */ 26625084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec2_R,&array);CHKERRQ(ierr); 26788ebb749SStefano Zampini ierr = MatSetValues(pcbddc->local_auxmat2,n_R,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 26825084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec2_R,&array);CHKERRQ(ierr); 26988ebb749SStefano Zampini } 27088ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_auxmat2,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27188ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_auxmat2,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27225084f0cSStefano Zampini ierr = MatScale(pcbddc->local_auxmat2,m_one);CHKERRQ(ierr); 27388ebb749SStefano Zampini 27488ebb749SStefano Zampini /* Assemble explicitly M1 = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} needed in preproc */ 27525084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 27625084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 27788ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&M1);CHKERRQ(ierr); 27888ebb749SStefano Zampini ierr = MatSetSizes(M1,n_constraints,n_constraints,n_constraints,n_constraints);CHKERRQ(ierr); 27988ebb749SStefano Zampini ierr = MatSetType(M1,impMatType);CHKERRQ(ierr); 28025084f0cSStefano Zampini ierr = MatSetUp(M1);CHKERRQ(ierr); 28125084f0cSStefano Zampini ierr = MatDuplicate(M1,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 28225084f0cSStefano Zampini ierr = MatZeroEntries(M2);CHKERRQ(ierr); 28325084f0cSStefano Zampini ierr = VecSet(vec1_C,m_one);CHKERRQ(ierr); 28425084f0cSStefano Zampini ierr = MatDiagonalSet(M2,vec1_C,INSERT_VALUES);CHKERRQ(ierr); 28525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 28625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 28725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 28888ebb749SStefano Zampini /* Assemble local_auxmat1 = M1*C_{CR} needed by BDDC application in KSP and in preproc */ 289f4ddd8eeSStefano Zampini if (!pcbddc->local_auxmat1) { 29088ebb749SStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 291f4ddd8eeSStefano Zampini } else { 292f4ddd8eeSStefano Zampini ierr = MatMatMult(M1,C_CR,MAT_REUSE_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 293f4ddd8eeSStefano Zampini } 29488ebb749SStefano Zampini } 29588ebb749SStefano Zampini 29688ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 29788ebb749SStefano Zampini if (n_vertices) { 2983a50541eSStefano Zampini PetscInt ibs,mbs; 2993a50541eSStefano Zampini PetscBool issbaij; 300af732b37SStefano Zampini Mat newmat; 3013a50541eSStefano Zampini 3023a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 3039577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3049577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 3053a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 3069577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 3079577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3089577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3099577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 3109577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 3119577ea80SStefano Zampini } else { 3123a50541eSStefano Zampini /* this is safe */ 3139577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 3143a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3153a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 3163a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 31745a1bb75SStefano Zampini /* which of the two approaches is faster? */ 3183a50541eSStefano Zampini /* ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3193a50541eSStefano Zampini ierr = MatCreateTranspose(A_RV,&A_VR);CHKERRQ(ierr);*/ 3203a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3213a50541eSStefano Zampini ierr = MatCreateTranspose(A_VR,&A_RV);CHKERRQ(ierr); 3223a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 3233a50541eSStefano Zampini } else { 3249577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 3259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3269577ea80SStefano Zampini } 3273a50541eSStefano Zampini } 3282a7a6963SBarry Smith ierr = MatCreateVecs(A_RV,&vec1_V,NULL);CHKERRQ(ierr); 32945a1bb75SStefano Zampini ierr = VecDuplicate(vec1_V,&vec2_V);CHKERRQ(ierr); 33025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 33188ebb749SStefano Zampini } 33288ebb749SStefano Zampini 33388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 334f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 335f4ddd8eeSStefano Zampini PetscInt on_B,on_primal; 336f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 337f4ddd8eeSStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size) { 338f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 339f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 340f4ddd8eeSStefano Zampini } 341f4ddd8eeSStefano Zampini } 342f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_D) { 343f4ddd8eeSStefano Zampini PetscInt on_D,on_primal; 344f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,&on_primal);CHKERRQ(ierr); 345f4ddd8eeSStefano Zampini if (on_D != n_D || on_primal != pcbddc->local_primal_size) { 346f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 347f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 348f4ddd8eeSStefano Zampini } 349f4ddd8eeSStefano Zampini } 350f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 35188ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 35288ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_phi_B,n_B,pcbddc->local_primal_size,n_B,pcbddc->local_primal_size);CHKERRQ(ierr); 35388ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_phi_B,impMatType);CHKERRQ(ierr); 35425084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_phi_B);CHKERRQ(ierr); 355f4ddd8eeSStefano Zampini } 356f4ddd8eeSStefano Zampini if ( (pcbddc->switch_static || pcbddc->dbg_flag) && !pcbddc->coarse_phi_D ) { 35788ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 35888ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_phi_D,n_D,pcbddc->local_primal_size,n_D,pcbddc->local_primal_size);CHKERRQ(ierr); 35988ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_phi_D,impMatType);CHKERRQ(ierr); 36025084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_phi_D);CHKERRQ(ierr); 36188ebb749SStefano Zampini } 36288ebb749SStefano Zampini 36325084f0cSStefano Zampini if (pcbddc->dbg_flag) { 3648ce42a96SStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,&idx_R_local);CHKERRQ(ierr); 365785e854fSJed Brown ierr = PetscMalloc1(2*pcbddc->local_primal_size,&coarsefunctions_errors);CHKERRQ(ierr); 366785e854fSJed Brown ierr = PetscMalloc1(2*pcbddc->local_primal_size,&constraints_errors);CHKERRQ(ierr); 36788ebb749SStefano Zampini } 36888ebb749SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 369854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 37088ebb749SStefano Zampini 37188ebb749SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 37225084f0cSStefano Zampini 37325084f0cSStefano Zampini /* vertices */ 37488ebb749SStefano Zampini for (i=0;i<n_vertices;i++) { 37545a1bb75SStefano Zampini /* this should not be needed, but MatMult_BAIJ is broken when using compressed row routines */ 376f4ddd8eeSStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); /* TODO: REMOVE IT */ 37788ebb749SStefano Zampini ierr = VecSet(vec1_V,zero);CHKERRQ(ierr); 37888ebb749SStefano Zampini ierr = VecSetValue(vec1_V,i,one,INSERT_VALUES);CHKERRQ(ierr); 37988ebb749SStefano Zampini ierr = VecAssemblyBegin(vec1_V);CHKERRQ(ierr); 38088ebb749SStefano Zampini ierr = VecAssemblyEnd(vec1_V);CHKERRQ(ierr); 38125084f0cSStefano Zampini /* simplified solution of saddle point problem with null rhs on constraints multipliers */ 38288ebb749SStefano Zampini ierr = MatMult(A_RV,vec1_V,pcbddc->vec1_R);CHKERRQ(ierr); 38388ebb749SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 38488ebb749SStefano Zampini ierr = VecScale(pcbddc->vec1_R,m_one);CHKERRQ(ierr); 38588ebb749SStefano Zampini if (n_constraints) { 38688ebb749SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,pcbddc->vec1_R,vec1_C);CHKERRQ(ierr); 38788ebb749SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 38888ebb749SStefano Zampini ierr = VecScale(vec1_C,m_one);CHKERRQ(ierr); 38988ebb749SStefano Zampini } 39088ebb749SStefano Zampini ierr = MatMult(A_VR,pcbddc->vec1_R,vec2_V);CHKERRQ(ierr); 39188ebb749SStefano Zampini ierr = MatMultAdd(A_VV,vec1_V,vec2_V,vec2_V);CHKERRQ(ierr); 39288ebb749SStefano Zampini 39388ebb749SStefano Zampini /* Set values in coarse basis function and subdomain part of coarse_mat */ 39488ebb749SStefano Zampini /* coarse basis functions */ 39588ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 39688ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 39788ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 39825084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 39988ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_B,n_B,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 40025084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 40188ebb749SStefano Zampini ierr = MatSetValue(pcbddc->coarse_phi_B,idx_V_B[i],i,one,INSERT_VALUES);CHKERRQ(ierr); 4028eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 40388ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40488ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40525084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 40688ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_D,n_D,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 40725084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 40888ebb749SStefano Zampini } 40925084f0cSStefano Zampini /* subdomain contribution to coarse matrix. WARNING -> column major ordering */ 41025084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 41125084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[i*pcbddc->local_primal_size],array,n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 41225084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 41388ebb749SStefano Zampini if (n_constraints) { 41425084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 41525084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[i*pcbddc->local_primal_size+n_vertices],array,n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 41625084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 41788ebb749SStefano Zampini } 41888ebb749SStefano Zampini 41925084f0cSStefano Zampini /* check */ 42025084f0cSStefano Zampini if (pcbddc->dbg_flag) { 42125084f0cSStefano Zampini /* assemble subdomain vector on local nodes */ 42288ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 42325084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 4249f00e9b4SStefano Zampini if (n_R) { 42525084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 4269f00e9b4SStefano Zampini } 42725084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 428e9189074SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],one,INSERT_VALUES);CHKERRQ(ierr); 42925084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 43025084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 43188ebb749SStefano Zampini /* assemble subdomain vector of lagrange multipliers (i.e. primal nodes) */ 43288ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 43325084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 43425084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_vertices,auxindices,array,INSERT_VALUES);CHKERRQ(ierr); 43525084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 43688ebb749SStefano Zampini if (n_constraints) { 43725084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 43825084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_constraints,&auxindices[n_vertices],array,INSERT_VALUES);CHKERRQ(ierr); 43925084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 44088ebb749SStefano Zampini } 44125084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 44225084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 44388ebb749SStefano Zampini ierr = VecScale(pcbddc->vec1_P,m_one);CHKERRQ(ierr); 44488ebb749SStefano Zampini /* check saddle point solution */ 44588ebb749SStefano Zampini ierr = MatMult(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 44688ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 44788ebb749SStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[i]);CHKERRQ(ierr); 44888ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 44925084f0cSStefano Zampini /* shift by the identity matrix */ 45025084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,i,m_one,ADD_VALUES);CHKERRQ(ierr); 45125084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 45225084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 45388ebb749SStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[i]);CHKERRQ(ierr); 45488ebb749SStefano Zampini } 45588ebb749SStefano Zampini } 45688ebb749SStefano Zampini 45725084f0cSStefano Zampini /* constraints */ 45888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 45988ebb749SStefano Zampini ierr = VecSet(vec2_C,zero);CHKERRQ(ierr); 46088ebb749SStefano Zampini ierr = VecSetValue(vec2_C,i,m_one,INSERT_VALUES);CHKERRQ(ierr); 46188ebb749SStefano Zampini ierr = VecAssemblyBegin(vec2_C);CHKERRQ(ierr); 46288ebb749SStefano Zampini ierr = VecAssemblyEnd(vec2_C);CHKERRQ(ierr); 46325084f0cSStefano Zampini /* simplified solution of saddle point problem with null rhs on vertices multipliers */ 46488ebb749SStefano Zampini ierr = MatMult(M1,vec2_C,vec1_C);CHKERRQ(ierr); 46588ebb749SStefano Zampini ierr = MatMult(pcbddc->local_auxmat2,vec1_C,pcbddc->vec1_R);CHKERRQ(ierr); 46688ebb749SStefano Zampini ierr = VecScale(vec1_C,m_one);CHKERRQ(ierr); 46725084f0cSStefano Zampini if (n_vertices) { 46825084f0cSStefano Zampini ierr = MatMult(A_VR,pcbddc->vec1_R,vec2_V);CHKERRQ(ierr); 46925084f0cSStefano Zampini } 47088ebb749SStefano Zampini /* Set values in coarse basis function and subdomain part of coarse_mat */ 47188ebb749SStefano Zampini /* coarse basis functions */ 47225084f0cSStefano Zampini j = i+n_vertices; /* don't touch this! */ 47388ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 47488ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 47588ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 47625084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 47725084f0cSStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_B,n_B,auxindices,1,&j,array,INSERT_VALUES);CHKERRQ(ierr); 47825084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 4798eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 48088ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 48188ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 48225084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 48325084f0cSStefano Zampini ierr = MatSetValues(pcbddc->coarse_phi_D,n_D,auxindices,1,&j,array,INSERT_VALUES);CHKERRQ(ierr); 48425084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 48588ebb749SStefano Zampini } 48625084f0cSStefano Zampini /* subdomain contribution to coarse matrix. WARNING -> column major ordering */ 48788ebb749SStefano Zampini if (n_vertices) { 48825084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 48925084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[j*pcbddc->local_primal_size],array,n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 49025084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 49188ebb749SStefano Zampini } 49225084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 49325084f0cSStefano Zampini ierr = PetscMemcpy(&coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices],array,n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 49425084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 49588ebb749SStefano Zampini 49625084f0cSStefano Zampini if (pcbddc->dbg_flag) { 49788ebb749SStefano Zampini /* assemble subdomain vector on nodes */ 49888ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 49925084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 5009f00e9b4SStefano Zampini if (n_R) { 50125084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 5029f00e9b4SStefano Zampini } 50325084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 50425084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 50525084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 50688ebb749SStefano Zampini /* assemble subdomain vector of lagrange multipliers */ 50788ebb749SStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 50888ebb749SStefano Zampini if (n_vertices) { 50925084f0cSStefano Zampini ierr = VecGetArrayRead(vec2_V,&array);CHKERRQ(ierr); 51025084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_vertices,auxindices,array,INSERT_VALUES);CHKERRQ(ierr); 51125084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec2_V,&array);CHKERRQ(ierr); 51288ebb749SStefano Zampini } 51325084f0cSStefano Zampini ierr = VecGetArrayRead(vec1_C,&array);CHKERRQ(ierr); 51425084f0cSStefano Zampini ierr = VecSetValues(pcbddc->vec1_P,n_constraints,&auxindices[n_vertices],array,INSERT_VALUES);CHKERRQ(ierr); 51525084f0cSStefano Zampini ierr = VecRestoreArrayRead(vec1_C,&array);CHKERRQ(ierr); 51625084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 51725084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 51825084f0cSStefano Zampini ierr = VecScale(pcbddc->vec1_P,m_one);CHKERRQ(ierr); 51988ebb749SStefano Zampini /* check saddle point solution */ 52088ebb749SStefano Zampini ierr = MatMult(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 52188ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 52225084f0cSStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[j]);CHKERRQ(ierr); 52388ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 52425084f0cSStefano Zampini /* shift by the identity matrix */ 52525084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,j,m_one,ADD_VALUES);CHKERRQ(ierr); 52625084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 52725084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 52825084f0cSStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[j]);CHKERRQ(ierr); 52988ebb749SStefano Zampini } 53088ebb749SStefano Zampini } 53125084f0cSStefano Zampini /* call assembling routines for local coarse basis */ 53288ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_phi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53388ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_phi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5348eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 53588ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_phi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53688ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_phi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53788ebb749SStefano Zampini } 53825084f0cSStefano Zampini 53988ebb749SStefano Zampini /* compute other basis functions for non-symmetric problems */ 54051cb09feSStefano Zampini /*ierr = MatIsSymmetric(pc->pmat,0.,&pcbddc->issym);CHKERRQ(ierr);*/ 54151cb09feSStefano Zampini { /* this is a temporary workaround since seqbaij matrices does not have support for symmetry checking */ 54251cb09feSStefano Zampini PetscBool setsym; 54351cb09feSStefano Zampini ierr = MatIsSymmetricKnown(pc->pmat,&setsym,&pcbddc->issym);CHKERRQ(ierr); 54451cb09feSStefano Zampini if (!setsym) pcbddc->issym = PETSC_FALSE; 54551cb09feSStefano Zampini } 54651cb09feSStefano Zampini 547b9d89cd5SStefano Zampini if (!pcbddc->issym) { 548f4ddd8eeSStefano Zampini if (!pcbddc->coarse_psi_B) { 54988ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 55088ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_psi_B,n_B,pcbddc->local_primal_size,n_B,pcbddc->local_primal_size);CHKERRQ(ierr); 55188ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_psi_B,impMatType);CHKERRQ(ierr); 55225084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_psi_B);CHKERRQ(ierr); 553f4ddd8eeSStefano Zampini } 554f4ddd8eeSStefano Zampini if ( (pcbddc->switch_static || pcbddc->dbg_flag) && !pcbddc->coarse_psi_D) { 55588ebb749SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 55688ebb749SStefano Zampini ierr = MatSetSizes(pcbddc->coarse_psi_D,n_D,pcbddc->local_primal_size,n_D,pcbddc->local_primal_size);CHKERRQ(ierr); 55788ebb749SStefano Zampini ierr = MatSetType(pcbddc->coarse_psi_D,impMatType);CHKERRQ(ierr); 55825084f0cSStefano Zampini ierr = MatSetUp(pcbddc->coarse_psi_D);CHKERRQ(ierr); 55988ebb749SStefano Zampini } 56088ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 56188ebb749SStefano Zampini if (n_constraints) { 56288ebb749SStefano Zampini ierr = VecSet(vec1_C,zero);CHKERRQ(ierr); 56388ebb749SStefano Zampini for (j=0;j<n_constraints;j++) { 56425084f0cSStefano Zampini ierr = VecSetValue(vec1_C,j,coarse_submat_vals[(j+n_vertices)*pcbddc->local_primal_size+i],INSERT_VALUES);CHKERRQ(ierr); 56588ebb749SStefano Zampini } 56625084f0cSStefano Zampini ierr = VecAssemblyBegin(vec1_C);CHKERRQ(ierr); 56725084f0cSStefano Zampini ierr = VecAssemblyEnd(vec1_C);CHKERRQ(ierr); 56888ebb749SStefano Zampini } 56988ebb749SStefano Zampini if (i<n_vertices) { 57088ebb749SStefano Zampini ierr = VecSet(vec1_V,zero);CHKERRQ(ierr); 57188ebb749SStefano Zampini ierr = VecSetValue(vec1_V,i,m_one,INSERT_VALUES);CHKERRQ(ierr); 57288ebb749SStefano Zampini ierr = VecAssemblyBegin(vec1_V);CHKERRQ(ierr); 57388ebb749SStefano Zampini ierr = VecAssemblyEnd(vec1_V);CHKERRQ(ierr); 57488ebb749SStefano Zampini ierr = MatMultTranspose(A_VR,vec1_V,pcbddc->vec1_R);CHKERRQ(ierr); 57588ebb749SStefano Zampini if (n_constraints) { 57688ebb749SStefano Zampini ierr = MatMultTransposeAdd(C_CR,vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 57788ebb749SStefano Zampini } 57888ebb749SStefano Zampini } else { 57988ebb749SStefano Zampini ierr = MatMultTranspose(C_CR,vec1_C,pcbddc->vec1_R);CHKERRQ(ierr); 58088ebb749SStefano Zampini } 58188ebb749SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 58288ebb749SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 58388ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58488ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58525084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 58688ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_psi_B,n_B,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 58725084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_B,&array);CHKERRQ(ierr); 58888ebb749SStefano Zampini if (i<n_vertices) { 58988ebb749SStefano Zampini ierr = MatSetValue(pcbddc->coarse_psi_B,idx_V_B[i],i,one,INSERT_VALUES);CHKERRQ(ierr); 59088ebb749SStefano Zampini } 5918eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 59288ebb749SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 59388ebb749SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 59425084f0cSStefano Zampini ierr = VecGetArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 59588ebb749SStefano Zampini ierr = MatSetValues(pcbddc->coarse_psi_D,n_D,auxindices,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 59625084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcis->vec1_D,&array);CHKERRQ(ierr); 59788ebb749SStefano Zampini } 59888ebb749SStefano Zampini 59925084f0cSStefano Zampini if (pcbddc->dbg_flag) { 60088ebb749SStefano Zampini /* assemble subdomain vector on nodes */ 60188ebb749SStefano Zampini ierr = VecSet(pcis->vec1_N,zero);CHKERRQ(ierr); 60225084f0cSStefano Zampini ierr = VecGetArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 6039f00e9b4SStefano Zampini if (n_R) { 60425084f0cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,n_R,idx_R_local,array,INSERT_VALUES);CHKERRQ(ierr); 6059f00e9b4SStefano Zampini } 60625084f0cSStefano Zampini ierr = VecRestoreArrayRead(pcbddc->vec1_R,&array);CHKERRQ(ierr); 60725084f0cSStefano Zampini if (i<n_vertices) { 608e9189074SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],one,INSERT_VALUES);CHKERRQ(ierr); 60988ebb749SStefano Zampini } 61025084f0cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 61125084f0cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 61225084f0cSStefano Zampini /* assemble subdomain vector of lagrange multipliers */ 61325084f0cSStefano Zampini for (j=0;j<pcbddc->local_primal_size;j++) { 61425084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,j,-coarse_submat_vals[j*pcbddc->local_primal_size+i],INSERT_VALUES);CHKERRQ(ierr); 61525084f0cSStefano Zampini } 61625084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 61725084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 61888ebb749SStefano Zampini /* check saddle point solution */ 61988ebb749SStefano Zampini ierr = MatMultTranspose(pcbddc->local_mat,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 62088ebb749SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->ConstraintMatrix,pcbddc->vec1_P,pcis->vec2_N,pcis->vec2_N);CHKERRQ(ierr); 62188ebb749SStefano Zampini ierr = VecNorm(pcis->vec2_N,NORM_INFINITY,&coarsefunctions_errors[i+pcbddc->local_primal_size]);CHKERRQ(ierr); 62288ebb749SStefano Zampini ierr = MatMult(pcbddc->ConstraintMatrix,pcis->vec1_N,pcbddc->vec1_P);CHKERRQ(ierr); 62325084f0cSStefano Zampini /* shift by the identity matrix */ 62425084f0cSStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,i,m_one,ADD_VALUES);CHKERRQ(ierr); 62525084f0cSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 62625084f0cSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 62788ebb749SStefano Zampini ierr = VecNorm(pcbddc->vec1_P,NORM_INFINITY,&constraints_errors[i+pcbddc->local_primal_size]);CHKERRQ(ierr); 62888ebb749SStefano Zampini } 62988ebb749SStefano Zampini } 63088ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_psi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63188ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_psi_B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6328eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 63388ebb749SStefano Zampini ierr = MatAssemblyBegin(pcbddc->coarse_psi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63488ebb749SStefano Zampini ierr = MatAssemblyEnd(pcbddc->coarse_psi_D,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 63588ebb749SStefano Zampini } 636c0553b1fSStefano Zampini unsymmetric_check = PETSC_TRUE; 637c0553b1fSStefano Zampini } else { /* take references to already computed coarse basis */ 638c0553b1fSStefano Zampini unsymmetric_check = PETSC_FALSE; 639c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 640c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 641c0553b1fSStefano Zampini if (pcbddc->coarse_phi_D) { 642c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 643c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 644c0553b1fSStefano Zampini } 64588ebb749SStefano Zampini } 64688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 64788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 64888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 64988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 65025084f0cSStefano Zampini if (pcbddc->dbg_flag) { 65188ebb749SStefano Zampini Mat coarse_sub_mat; 65225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 65388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 65488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 65588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 65688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 65788ebb749SStefano Zampini PetscReal real_value; 65888ebb749SStefano Zampini 65988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 66088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 66188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 66288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 66388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 66488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 665c0553b1fSStefano Zampini if (unsymmetric_check) { 66688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 66788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 66888ebb749SStefano Zampini } 66988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 67088ebb749SStefano Zampini 67125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 67225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat and local basis functions\n");CHKERRQ(ierr); 67325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 674c0553b1fSStefano Zampini if (unsymmetric_check) { 67588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 67688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 67788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 67888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 67988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 68088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 68288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 68388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 68588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 68688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 68788ebb749SStefano Zampini } else { 68888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 68988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 69088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 69188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 69288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 69388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 69488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 69588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 69688ebb749SStefano Zampini } 69788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 69888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 69988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70088ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 70181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 70288ebb749SStefano Zampini ierr = MatNorm(TM1,NORM_INFINITY,&real_value);CHKERRQ(ierr); 7030fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 70425084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"----------------------------------\n");CHKERRQ(ierr); 70525084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d \n",PetscGlobalRank);CHKERRQ(ierr); 70625084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"matrix error = % 1.14e\n",real_value);CHKERRQ(ierr); 70725084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"coarse functions (phi) errors\n");CHKERRQ(ierr); 70888ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 70925084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i,coarsefunctions_errors[i]);CHKERRQ(ierr); 71088ebb749SStefano Zampini } 71125084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"constraints (phi) errors\n");CHKERRQ(ierr); 71288ebb749SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 71325084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i,constraints_errors[i]);CHKERRQ(ierr); 71488ebb749SStefano Zampini } 715c0553b1fSStefano Zampini if (unsymmetric_check) { 71625084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"coarse functions (psi) errors\n");CHKERRQ(ierr); 71788ebb749SStefano Zampini for (i=pcbddc->local_primal_size;i<2*pcbddc->local_primal_size;i++) { 71825084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i-pcbddc->local_primal_size,coarsefunctions_errors[i]);CHKERRQ(ierr); 71988ebb749SStefano Zampini } 72025084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"constraints (psi) errors\n");CHKERRQ(ierr); 72188ebb749SStefano Zampini for (i=pcbddc->local_primal_size;i<2*pcbddc->local_primal_size;i++) { 72225084f0cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local %02d-th function error = % 1.14e\n",i-pcbddc->local_primal_size,constraints_errors[i]);CHKERRQ(ierr); 72388ebb749SStefano Zampini } 72488ebb749SStefano Zampini } 72525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 72688ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 72788ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 72888ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 72988ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 73088ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 73188ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 73288ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 73388ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 73488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 73588ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 736c0553b1fSStefano Zampini if (unsymmetric_check) { 73788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 73888ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 73988ebb749SStefano Zampini } 74088ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 7418ce42a96SStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,&idx_R_local);CHKERRQ(ierr); 74288ebb749SStefano Zampini ierr = PetscFree(coarsefunctions_errors);CHKERRQ(ierr); 74388ebb749SStefano Zampini ierr = PetscFree(constraints_errors);CHKERRQ(ierr); 74488ebb749SStefano Zampini } 74588ebb749SStefano Zampini /* free memory */ 74688ebb749SStefano Zampini if (n_vertices) { 74788ebb749SStefano Zampini ierr = VecDestroy(&vec1_V);CHKERRQ(ierr); 74888ebb749SStefano Zampini ierr = VecDestroy(&vec2_V);CHKERRQ(ierr); 74988ebb749SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 75088ebb749SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 75188ebb749SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 75288ebb749SStefano Zampini } 75388ebb749SStefano Zampini if (n_constraints) { 75488ebb749SStefano Zampini ierr = VecDestroy(&vec1_C);CHKERRQ(ierr); 75588ebb749SStefano Zampini ierr = VecDestroy(&vec2_C);CHKERRQ(ierr); 75688ebb749SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 75788ebb749SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 75888ebb749SStefano Zampini } 75988ebb749SStefano Zampini ierr = PetscFree(auxindices);CHKERRQ(ierr); 7608629588bSStefano Zampini /* get back data */ 7618629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 76288ebb749SStefano Zampini PetscFunctionReturn(0); 76388ebb749SStefano Zampini } 76488ebb749SStefano Zampini 76588ebb749SStefano Zampini #undef __FUNCT__ 7665e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 7675e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 768aa0d41d4SStefano Zampini { 769aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7705e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7715e8657edSStefano Zampini Mat *change_mat_all; 7725e8657edSStefano Zampini IS is_local,is_global; 7735e8657edSStefano Zampini PetscBool sorted,isseqaij; 7745e8657edSStefano Zampini PetscInt local_size,*idxs_perm; 775aa0d41d4SStefano Zampini PetscErrorCode ierr; 776aa0d41d4SStefano Zampini 777aa0d41d4SStefano Zampini PetscFunctionBegin; 778aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7795e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 7805e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 781906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 782906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 783906d46d4SStefano Zampini ierr = ISSorted(is_global,&sorted);CHKERRQ(ierr); 784906d46d4SStefano Zampini if (!sorted) { 785906d46d4SStefano Zampini const PetscInt *idxs; 786906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 787aa0d41d4SStefano Zampini 7885e8657edSStefano Zampini ierr = PetscMalloc1(local_size,&idxs_perm);CHKERRQ(ierr); 7895e8657edSStefano Zampini ierr = PetscMalloc1(local_size,&idxs_sorted);CHKERRQ(ierr); 7905e8657edSStefano Zampini for (i=0;i<local_size;i++) { 791906d46d4SStefano Zampini idxs_perm[i] = i; 792aa0d41d4SStefano Zampini } 793906d46d4SStefano Zampini ierr = ISGetIndices(is_global,&idxs);CHKERRQ(ierr); 7945e8657edSStefano Zampini ierr = PetscSortIntWithPermutation(local_size,idxs,idxs_perm);CHKERRQ(ierr); 7955e8657edSStefano Zampini for (i=0;i<local_size;i++) { 796906d46d4SStefano Zampini idxs_sorted[i] = idxs[idxs_perm[i]]; 797aa0d41d4SStefano Zampini } 798906d46d4SStefano Zampini ierr = ISRestoreIndices(is_global,&idxs);CHKERRQ(ierr); 799906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 8005e8657edSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,idxs_sorted,PETSC_OWN_POINTER,&is_global);CHKERRQ(ierr); 801aa0d41d4SStefano Zampini } 802906d46d4SStefano Zampini 803906d46d4SStefano Zampini /* get change of basis on the whole set of local dofs */ 8045e8657edSStefano Zampini ierr = MatGetSubMatrices(ChangeOfBasisMatrix,1,&is_global,&is_global,MAT_INITIAL_MATRIX,&change_mat_all);CHKERRQ(ierr); 805906d46d4SStefano Zampini 806906d46d4SStefano Zampini if (!sorted) { 807906d46d4SStefano Zampini Mat new_mat; 808906d46d4SStefano Zampini IS is_perm; 809906d46d4SStefano Zampini PetscInt *idxs,i; 810906d46d4SStefano Zampini 8115e8657edSStefano Zampini ierr = PetscMalloc1(local_size,&idxs);CHKERRQ(ierr); 8125e8657edSStefano Zampini for (i=0;i<local_size;i++) { 813906d46d4SStefano Zampini idxs[idxs_perm[i]] = i; 814906d46d4SStefano Zampini } 815906d46d4SStefano Zampini ierr = PetscFree(idxs_perm);CHKERRQ(ierr); 8165e8657edSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,idxs,PETSC_OWN_POINTER,&is_perm);CHKERRQ(ierr); 817906d46d4SStefano Zampini ierr = ISSetPermutation(is_perm);CHKERRQ(ierr); 818906d46d4SStefano Zampini ierr = MatPermute(change_mat_all[0],is_perm,is_perm,&new_mat);CHKERRQ(ierr); 819906d46d4SStefano Zampini ierr = MatDestroy(&change_mat_all[0]);CHKERRQ(ierr); 820906d46d4SStefano Zampini change_mat_all[0] = new_mat; 821906d46d4SStefano Zampini ierr = ISDestroy(&is_perm);CHKERRQ(ierr); 822906d46d4SStefano Zampini } 823906d46d4SStefano Zampini 824906d46d4SStefano Zampini /* check */ 825906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 826906d46d4SStefano Zampini Vec x,x_change; 827906d46d4SStefano Zampini PetscReal error; 828906d46d4SStefano Zampini 8295e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 830906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 8315e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 8325e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8335e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8345e8657edSStefano Zampini ierr = MatMult(change_mat_all[0],matis->x,matis->y);CHKERRQ(ierr); 8355e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8365e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 837906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 838906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 839906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 840906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 841906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 842906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 843906d46d4SStefano Zampini } 844906d46d4SStefano Zampini 84522d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 8469b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 84722d5777bSStefano Zampini if (isseqaij) { 8485e8657edSStefano Zampini ierr = MatPtAP(matis->A,change_mat_all[0],MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 849aa0d41d4SStefano Zampini } else { 850aa0d41d4SStefano Zampini Mat work_mat; 851aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 8525e8657edSStefano Zampini ierr = MatPtAP(work_mat,change_mat_all[0],MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 853aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 854aa0d41d4SStefano Zampini } 85545a1bb75SStefano Zampini /* 85645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 85745a1bb75SStefano Zampini ierr = MatView(change_mat_all,(PetscViewer)0);CHKERRQ(ierr); 85845a1bb75SStefano Zampini */ 859906d46d4SStefano Zampini ierr = MatDestroyMatrices(1,&change_mat_all);CHKERRQ(ierr); 860906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 861aa0d41d4SStefano Zampini PetscFunctionReturn(0); 862aa0d41d4SStefano Zampini } 863aa0d41d4SStefano Zampini 864aa0d41d4SStefano Zampini #undef __FUNCT__ 865a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 8668ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 867a64d13efSStefano Zampini { 868a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 869a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8708ce42a96SStefano Zampini IS is_aux1,is_aux2; 871e9189074SStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 8723a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 8733a50541eSStefano Zampini PetscInt vbs,bs; 8744641a718SStefano Zampini PetscBT bitmask; 875a64d13efSStefano Zampini PetscErrorCode ierr; 876a64d13efSStefano Zampini 877a64d13efSStefano Zampini PetscFunctionBegin; 878b23d619eSStefano Zampini /* 879b23d619eSStefano Zampini No need to setup local scatters if 880b23d619eSStefano Zampini - primal space is unchanged 881b23d619eSStefano Zampini AND 882b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 883b23d619eSStefano Zampini AND 884b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 885b23d619eSStefano Zampini */ 886b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 887f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 888f4ddd8eeSStefano Zampini } 889f4ddd8eeSStefano Zampini /* destroy old objects */ 890f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 891f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 892f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 893a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 894a64d13efSStefano Zampini n_B = pcis->n_B; n_D = pcis->n - n_B; 895e9189074SStefano Zampini n_vertices = pcbddc->n_actual_vertices; 8964641a718SStefano Zampini /* create auxiliary bitmask */ 8974641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 8984641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 899e9189074SStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->primal_indices_local_idxs[i]);CHKERRQ(ierr); 9004641a718SStefano Zampini } 9013a50541eSStefano Zampini 902a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 903854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 904a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 9054641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 906a64d13efSStefano Zampini idx_R_local[n_R] = i; 907a64d13efSStefano Zampini n_R++; 908a64d13efSStefano Zampini } 909a64d13efSStefano Zampini } 9103a50541eSStefano Zampini 9113a50541eSStefano Zampini /* Block code */ 9123a50541eSStefano Zampini vbs = 1; 9133a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 9143a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 9153a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 9163a50541eSStefano Zampini PetscInt *vary; 9173a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 918785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 9193a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 920e9189074SStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->primal_indices_local_idxs[i]/bs]++; 9213a50541eSStefano Zampini for (i=0; i<n_vertices; i++) { 9223a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 9233a50541eSStefano Zampini is_blocked = PETSC_FALSE; 9243a50541eSStefano Zampini break; 9253a50541eSStefano Zampini } 9263a50541eSStefano Zampini } 9273a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 9283a50541eSStefano Zampini vbs = bs; 9293a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 9303a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 9313a50541eSStefano Zampini } 9323a50541eSStefano Zampini } 9333a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 9343a50541eSStefano Zampini } 9353a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 9363a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 937a64d13efSStefano Zampini 938a64d13efSStefano Zampini /* print some info if requested */ 939a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 940a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 941a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 9420fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 943a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 944a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 9453a50541eSStefano 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); 946a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"pcbddc->n_vertices = %d, pcbddc->n_constraints = %d\n",pcbddc->n_vertices,pcbddc->n_constraints);CHKERRQ(ierr); 947a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 948a64d13efSStefano Zampini } 949a64d13efSStefano Zampini 950a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 9513a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 952854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 953854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 954a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 9554641a718SStefano Zampini for (i=0; i<n_D; i++) { 9564641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 9574641a718SStefano Zampini } 958a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 959a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 9604641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 9614641a718SStefano Zampini aux_array1[j++] = i; 962a64d13efSStefano Zampini } 963a64d13efSStefano Zampini } 964a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 965a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 966a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 9674641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 9684641a718SStefano Zampini aux_array2[j++] = i; 969a64d13efSStefano Zampini } 970a64d13efSStefano Zampini } 971a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 972a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 973a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 974a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 975a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 976a64d13efSStefano Zampini 9778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 978785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 979a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 9804641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 9814641a718SStefano Zampini aux_array1[j++] = i; 982a64d13efSStefano Zampini } 983a64d13efSStefano Zampini } 984a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 985a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 986a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 987a64d13efSStefano Zampini } 9884641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 9893a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 990a64d13efSStefano Zampini PetscFunctionReturn(0); 991a64d13efSStefano Zampini } 992a64d13efSStefano Zampini 993304d26faSStefano Zampini 994304d26faSStefano Zampini #undef __FUNCT__ 995304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 996*684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 997304d26faSStefano Zampini { 998304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 999304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1000304d26faSStefano Zampini PC pc_temp; 1001304d26faSStefano Zampini Mat A_RR; 1002f4ddd8eeSStefano Zampini MatReuse reuse; 1003304d26faSStefano Zampini PetscScalar m_one = -1.0; 1004304d26faSStefano Zampini PetscReal value; 1005af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 10069577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1007304d26faSStefano Zampini PetscErrorCode ierr; 1008e604994aSStefano Zampini /* prefixes stuff */ 1009312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1010e604994aSStefano Zampini size_t len; 1011304d26faSStefano Zampini 1012304d26faSStefano Zampini PetscFunctionBegin; 1013304d26faSStefano Zampini 1014e604994aSStefano Zampini /* compute prefixes */ 1015e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1016e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1017e604994aSStefano Zampini if (!pcbddc->current_level) { 1018e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1019e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1020e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1021e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1022e604994aSStefano Zampini } else { 1023e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1024312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1025e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1026e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1027312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1028312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 102934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 103034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1031e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1032e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1033e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1034e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1035e604994aSStefano Zampini } 1036e604994aSStefano Zampini 1037304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1038*684f6988SStefano Zampini if (dirichlet) { 1039ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 10408ce42a96SStefano Zampini ierr = ISGetSize(pcis->is_I_local,&n_D);CHKERRQ(ierr); 1041304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1042304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1043304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1044304d26faSStefano Zampini /* default */ 1045304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1046e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 10479577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1048304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 10499577ea80SStefano Zampini if (issbaij) { 10509577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 10519577ea80SStefano Zampini } else { 1052304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 10539577ea80SStefano Zampini } 1054304d26faSStefano Zampini /* Allow user's customization */ 1055304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1056304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1057304d26faSStefano Zampini } 1058d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1059304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1060304d26faSStefano Zampini if (!n_D) { 1061304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1062304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1063304d26faSStefano Zampini } 1064304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1065304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1066304d26faSStefano Zampini /* set ksp_D into pcis data */ 1067304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1068304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1069304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1070*684f6988SStefano Zampini } 1071304d26faSStefano Zampini 1072304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1073*684f6988SStefano Zampini A_RR = 0; 1074*684f6988SStefano Zampini if (neumann) { 1075f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 10768ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1077f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1078f4ddd8eeSStefano Zampini PetscInt nn_R; 107981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1080f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1081f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1082f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1083f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1084f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1085f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1086f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1087727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1088f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1089f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1090f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1091f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1092f4ddd8eeSStefano Zampini } 1093f4ddd8eeSStefano Zampini } 1094f4ddd8eeSStefano Zampini /* last check */ 1095d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1096f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1097f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1098f4ddd8eeSStefano Zampini } 1099f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1100f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1101f4ddd8eeSStefano Zampini } 1102f4ddd8eeSStefano Zampini /* extract A_RR */ 1103af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1104af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 11053a50541eSStefano Zampini if (ibs != mbs) { 1106af732b37SStefano Zampini Mat newmat; 1107af732b37SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 1108f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1109af732b37SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 1110af732b37SStefano Zampini } else { 1111f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 1112af732b37SStefano Zampini } 1113f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1114304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1115304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1116304d26faSStefano Zampini /* default */ 1117304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1118e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1119304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 11209577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 11219577ea80SStefano Zampini if (issbaij) { 11229577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 11239577ea80SStefano Zampini } else { 1124304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 11259577ea80SStefano Zampini } 1126304d26faSStefano Zampini /* Allow user's customization */ 1127304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1128304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1129304d26faSStefano Zampini } 1130d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1131304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1132304d26faSStefano Zampini if (!n_R) { 1133304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1134304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1135304d26faSStefano Zampini } 1136304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1137304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1138*684f6988SStefano Zampini } 1139304d26faSStefano Zampini 1140304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 11410fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1142*684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1143*684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1144*684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1145*684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1146*684f6988SStefano Zampini } 1147*684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 11480fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 11490fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 11500fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 11510fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 11520fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1153304d26faSStefano Zampini /* need to be adapted? */ 1154b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1155b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1156b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1157304d26faSStefano Zampini /* print info */ 1158304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1159e604994aSStefano 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); 1160304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1161304d26faSStefano Zampini } 1162b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 11638ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1164304d26faSStefano Zampini } 1165*684f6988SStefano Zampini } 1166*684f6988SStefano Zampini if (neumann) { /* Neumann */ 11670fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 11680fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 11690fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 11700fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 11710fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1172304d26faSStefano Zampini /* need to be adapted? */ 1173b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1174b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1175304d26faSStefano Zampini /* print info */ 1176304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1177e604994aSStefano 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); 1178304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1179304d26faSStefano Zampini } 1180b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 11818ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1182304d26faSStefano Zampini } 11830fccc4e9SStefano Zampini } 1184*684f6988SStefano Zampini } 1185304d26faSStefano Zampini /* free Neumann problem's matrix */ 1186304d26faSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1187304d26faSStefano Zampini PetscFunctionReturn(0); 1188304d26faSStefano Zampini } 1189304d26faSStefano Zampini 1190304d26faSStefano Zampini #undef __FUNCT__ 1191ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 119220c7b377SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec rhs, Vec sol, Vec work, PetscBool applytranspose) 1193674ae819SStefano Zampini { 1194674ae819SStefano Zampini PetscErrorCode ierr; 1195674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1196674ae819SStefano Zampini 1197674ae819SStefano Zampini PetscFunctionBegin; 119820c7b377SStefano Zampini if (applytranspose) { 1199674ae819SStefano Zampini if (pcbddc->local_auxmat1) { 120020c7b377SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,rhs,work);CHKERRQ(ierr); 120120c7b377SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,work,rhs,rhs);CHKERRQ(ierr); 120220c7b377SStefano Zampini } 120320c7b377SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 120420c7b377SStefano Zampini } else { 120520c7b377SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr); 120620c7b377SStefano Zampini if (pcbddc->local_auxmat1) { 120720c7b377SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,sol,work);CHKERRQ(ierr); 120820c7b377SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,work,sol,sol);CHKERRQ(ierr); 120920c7b377SStefano Zampini } 1210674ae819SStefano Zampini } 1211674ae819SStefano Zampini PetscFunctionReturn(0); 1212674ae819SStefano Zampini } 1213674ae819SStefano Zampini 1214dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1215674ae819SStefano Zampini #undef __FUNCT__ 1216674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1217dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1218674ae819SStefano Zampini { 1219674ae819SStefano Zampini PetscErrorCode ierr; 1220674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1221674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1222674ae819SStefano Zampini const PetscScalar zero = 0.0; 1223674ae819SStefano Zampini 1224674ae819SStefano Zampini PetscFunctionBegin; 1225dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1226dc359a40SStefano Zampini if (applytranspose) { 1227674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 12288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1229dc359a40SStefano Zampini } else { 1230674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1231674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 123215aaf578SStefano Zampini } 123312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 123412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 123512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 123612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 123712edc857SStefano Zampini 12389f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 123912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 124012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 124112edc857SStefano Zampini if (applytranspose) { 124212edc857SStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,NULL,NULL);CHKERRQ(ierr); 124312edc857SStefano Zampini } else { 124412edc857SStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,NULL,NULL);CHKERRQ(ierr); 124512edc857SStefano Zampini } 124612edc857SStefano Zampini } 1247674ae819SStefano Zampini 1248674ae819SStefano Zampini /* Local solution on R nodes */ 12499f00e9b4SStefano Zampini if (pcis->n) { 1250674ae819SStefano Zampini ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr); 1251674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1252674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12538eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1254674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1255674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1256674ae819SStefano Zampini } 125720c7b377SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcbddc->vec1_R,pcbddc->vec2_R,pcbddc->vec1_C,applytranspose);CHKERRQ(ierr); 1258674ae819SStefano Zampini ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr); 1259674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1260674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12618eeda7d8SStefano Zampini if (pcbddc->switch_static) { 1262674ae819SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1263674ae819SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1264674ae819SStefano Zampini } 12659f00e9b4SStefano Zampini } 1266674ae819SStefano Zampini 12679f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 12689f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 126912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1270674ae819SStefano Zampini 1271674ae819SStefano Zampini /* Sum contributions from two levels */ 1272dc359a40SStefano Zampini if (applytranspose) { 1273dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1274dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1275dc359a40SStefano Zampini } else { 1276674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 12778eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1278dc359a40SStefano Zampini } 1279674ae819SStefano Zampini PetscFunctionReturn(0); 1280674ae819SStefano Zampini } 1281674ae819SStefano Zampini 128212edc857SStefano Zampini /* TODO: the following two function can be optimized using VecPlaceArray whenever possible and using overlap flag */ 1283674ae819SStefano Zampini #undef __FUNCT__ 1284674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 128512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1286674ae819SStefano Zampini { 1287674ae819SStefano Zampini PetscErrorCode ierr; 1288674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 128912edc857SStefano Zampini PetscScalar *array,*array2; 129012edc857SStefano Zampini Vec from,to; 1291674ae819SStefano Zampini 1292674ae819SStefano Zampini PetscFunctionBegin; 129312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 129412edc857SStefano Zampini from = pcbddc->coarse_vec; 129512edc857SStefano Zampini to = pcbddc->vec1_P; 129612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 129712edc857SStefano Zampini Vec tvec; 129812edc857SStefano Zampini PetscInt lsize; 129912edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 130012edc857SStefano Zampini ierr = VecGetLocalSize(tvec,&lsize);CHKERRQ(ierr); 130112edc857SStefano Zampini ierr = VecGetArrayRead(tvec,(const PetscScalar**)&array);CHKERRQ(ierr); 130212edc857SStefano Zampini ierr = VecGetArray(from,&array2);CHKERRQ(ierr); 130312edc857SStefano Zampini ierr = PetscMemcpy(array2,array,lsize*sizeof(PetscScalar));CHKERRQ(ierr); 130412edc857SStefano Zampini ierr = VecRestoreArrayRead(tvec,(const PetscScalar**)&array);CHKERRQ(ierr); 130512edc857SStefano Zampini ierr = VecRestoreArray(from,&array2);CHKERRQ(ierr); 130612edc857SStefano Zampini } 130712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 130812edc857SStefano Zampini from = pcbddc->vec1_P; 130912edc857SStefano Zampini to = pcbddc->coarse_vec; 131012edc857SStefano Zampini } 131112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1312674ae819SStefano Zampini PetscFunctionReturn(0); 1313674ae819SStefano Zampini } 1314674ae819SStefano Zampini 1315674ae819SStefano Zampini #undef __FUNCT__ 1316674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 131712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1318674ae819SStefano Zampini { 1319674ae819SStefano Zampini PetscErrorCode ierr; 1320674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 132112edc857SStefano Zampini PetscScalar *array,*array2; 132212edc857SStefano Zampini Vec from,to; 1323674ae819SStefano Zampini 1324674ae819SStefano Zampini PetscFunctionBegin; 132512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 132612edc857SStefano Zampini from = pcbddc->coarse_vec; 132712edc857SStefano Zampini to = pcbddc->vec1_P; 132812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 132912edc857SStefano Zampini from = pcbddc->vec1_P; 133012edc857SStefano Zampini to = pcbddc->coarse_vec; 133112edc857SStefano Zampini } 133212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 133312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 133412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 133512edc857SStefano Zampini Vec tvec; 133612edc857SStefano Zampini PetscInt lsize; 133712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 133812edc857SStefano Zampini ierr = VecGetLocalSize(tvec,&lsize);CHKERRQ(ierr); 133912edc857SStefano Zampini ierr = VecGetArrayRead(to,(const PetscScalar**)&array);CHKERRQ(ierr); 134012edc857SStefano Zampini ierr = VecGetArray(tvec,&array2);CHKERRQ(ierr); 134112edc857SStefano Zampini ierr = PetscMemcpy(array2,array,lsize*sizeof(PetscScalar));CHKERRQ(ierr); 134212edc857SStefano Zampini ierr = VecRestoreArrayRead(to,(const PetscScalar**)&array);CHKERRQ(ierr); 134312edc857SStefano Zampini ierr = VecRestoreArray(tvec,&array2);CHKERRQ(ierr); 134412edc857SStefano Zampini } 134512edc857SStefano Zampini } 1346674ae819SStefano Zampini PetscFunctionReturn(0); 1347674ae819SStefano Zampini } 1348674ae819SStefano Zampini 1349984c4197SStefano Zampini /* uncomment for testing purposes */ 1350984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1351674ae819SStefano Zampini #undef __FUNCT__ 1352674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1353674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1354674ae819SStefano Zampini { 1355674ae819SStefano Zampini PetscErrorCode ierr; 1356674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1357674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1358674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1359984c4197SStefano Zampini /* constraint and (optionally) change of basis matrix implemented as SeqAIJ */ 1360674ae819SStefano Zampini MatType impMatType=MATSEQAIJ; 1361984c4197SStefano Zampini /* one and zero */ 1362984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1363984c4197SStefano Zampini /* space to store constraints and their local indices */ 1364984c4197SStefano Zampini PetscScalar *temp_quadrature_constraint; 1365984c4197SStefano Zampini PetscInt *temp_indices,*temp_indices_to_constraint,*temp_indices_to_constraint_B; 1366984c4197SStefano Zampini /* iterators */ 1367984c4197SStefano Zampini PetscInt i,j,k,total_counts,temp_start_ptr; 1368984c4197SStefano Zampini /* stuff to store connected components stored in pcbddc->mat_graph */ 1369984c4197SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges,*used_IS; 1370984c4197SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges; 1371984c4197SStefano Zampini /* near null space stuff */ 1372674ae819SStefano Zampini MatNullSpace nearnullsp; 1373674ae819SStefano Zampini const Vec *nearnullvecs; 1374674ae819SStefano Zampini Vec *localnearnullsp; 1375984c4197SStefano Zampini PetscBool nnsp_has_cnst; 1376984c4197SStefano Zampini PetscInt nnsp_size; 1377984c4197SStefano Zampini PetscScalar *array; 1378984c4197SStefano Zampini /* BLAS integers */ 1379e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1380e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1381c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1382984c4197SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1383242a89d7SStefano Zampini PetscBool skip_lapack; 1384984c4197SStefano Zampini PetscScalar *work; 1385984c4197SStefano Zampini PetscReal *singular_vals; 1386984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1387984c4197SStefano Zampini PetscReal *rwork; 1388674ae819SStefano Zampini #endif 1389984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1390e310c8b4SStefano Zampini PetscBLASInt Blas_one_2=1; 1391984c4197SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1392b7d8b9f8SStefano Zampini #else 1393b7d8b9f8SStefano Zampini PetscBLASInt dummy_int_1=1,dummy_int_2=1; 1394b7d8b9f8SStefano Zampini PetscScalar dummy_scalar_1=0.0,dummy_scalar_2=0.0; 1395984c4197SStefano Zampini #endif 1396727cdba6SStefano Zampini /* reuse */ 1397727cdba6SStefano Zampini PetscInt olocal_primal_size; 1398727cdba6SStefano Zampini PetscInt *oprimal_indices_local_idxs; 1399984c4197SStefano Zampini /* change of basis */ 1400984c4197SStefano Zampini PetscInt *aux_primal_numbering,*aux_primal_minloc,*global_indices; 1401a717540cSStefano Zampini PetscBool boolforchange,qr_needed; 1402a717540cSStefano Zampini PetscBT touched,change_basis,qr_needed_idx; 1403984c4197SStefano Zampini /* auxiliary stuff */ 140439e2fb2aSStefano Zampini PetscInt *nnz,*is_indices,*aux_primal_numbering_B; 1405911cabfeSStefano Zampini PetscInt ncc,*gidxs,*permutation,*temp_indices_to_constraint_work; 1406911cabfeSStefano Zampini PetscScalar *temp_quadrature_constraint_work; 1407984c4197SStefano Zampini /* some quantities */ 140845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1409984c4197SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint,max_constraints,temp_constraints; 1410984c4197SStefano Zampini 1411674ae819SStefano Zampini 1412674ae819SStefano Zampini PetscFunctionBegin; 14138e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 14148e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 14158e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1416674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1417d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1418d06fc5fdSStefano Zampini /* free unneeded index sets */ 1419d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1420d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1421674ae819SStefano Zampini } 1422d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1423d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1424d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1425d06fc5fdSStefano Zampini } 1426d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1427d06fc5fdSStefano Zampini n_ISForEdges = 0; 1428d06fc5fdSStefano Zampini } 1429d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1430d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1431d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1432d06fc5fdSStefano Zampini } 1433d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1434d06fc5fdSStefano Zampini n_ISForFaces = 0; 1435d06fc5fdSStefano Zampini } 1436d06fc5fdSStefano Zampini /* HACKS (the following two blocks of code) */ 1437b9b85e73SStefano Zampini if (!ISForVertices && pcbddc->NullSpace && !pcbddc->user_ChangeOfBasisMatrix) { 1438b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 1439d06fc5fdSStefano Zampini if (!ISForEdges) { 1440d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 1441d06fc5fdSStefano Zampini } 1442b8ffe317SStefano Zampini } 144398a51de6SStefano Zampini if (pcbddc->NullSpace) { 144498a51de6SStefano Zampini /* use_change_of_basis should be consistent among processors */ 1445d06fc5fdSStefano Zampini PetscBool tbool[2],gbool[2]; 1446d06fc5fdSStefano Zampini tbool [0] = pcbddc->use_change_of_basis; 1447d06fc5fdSStefano Zampini tbool [1] = pcbddc->use_change_on_faces; 1448d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1449d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 1450d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 145198a51de6SStefano Zampini } 1452984c4197SStefano Zampini /* print some info */ 1453674ae819SStefano Zampini if (pcbddc->dbg_flag) { 14540fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1455674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1456674ae819SStefano Zampini i = 0; 1457674ae819SStefano Zampini if (ISForVertices) { 1458674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&i);CHKERRQ(ierr); 1459674ae819SStefano Zampini } 1460674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices\n",PetscGlobalRank,i);CHKERRQ(ierr); 1461674ae819SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges\n",PetscGlobalRank,n_ISForEdges);CHKERRQ(ierr); 146215aaf578SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces\n",PetscGlobalRank,n_ISForFaces);CHKERRQ(ierr); 1463674ae819SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1464674ae819SStefano Zampini } 1465674ae819SStefano Zampini /* check if near null space is attached to global mat */ 1466674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 1467674ae819SStefano Zampini if (nearnullsp) { 1468674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 1469f4ddd8eeSStefano Zampini /* remove any stored info */ 1470f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1471f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1472f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 1473f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 1474f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 1475473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1476f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 1477f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 1478f4ddd8eeSStefano Zampini } 1479984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 1480984c4197SStefano Zampini nnsp_size = 0; 1481674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 1482674ae819SStefano Zampini } 1483984c4197SStefano Zampini /* get max number of constraints on a single cc */ 1484984c4197SStefano Zampini max_constraints = nnsp_size; 1485984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 1486984c4197SStefano Zampini 1487674ae819SStefano Zampini /* 1488674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 1489674ae819SStefano Zampini - temp_indices will contain start index of each constraint stored as follows 1490911cabfeSStefano 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 1491911cabfeSStefano 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 1492911cabfeSStefano Zampini - temp_quadrature_constraint [temp_indices[i],...,temp_indices[i+1]-1] will contain the scalars representing the constraint itself 1493674ae819SStefano Zampini */ 1494674ae819SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 1495984c4197SStefano Zampini total_counts *= max_constraints; 1496674ae819SStefano Zampini n_vertices = 0; 1497674ae819SStefano Zampini if (ISForVertices) { 1498674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 1499674ae819SStefano Zampini } 1500674ae819SStefano Zampini total_counts += n_vertices; 1501854ce69bSBarry Smith ierr = PetscMalloc1(total_counts+1,&temp_indices);CHKERRQ(ierr); 15024641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 1503674ae819SStefano Zampini total_counts = 0; 1504674ae819SStefano Zampini max_size_of_constraint = 0; 1505674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 1506674ae819SStefano Zampini if (i<n_ISForEdges) { 1507674ae819SStefano Zampini used_IS = &ISForEdges[i]; 1508674ae819SStefano Zampini } else { 1509674ae819SStefano Zampini used_IS = &ISForFaces[i-n_ISForEdges]; 1510674ae819SStefano Zampini } 1511674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&j);CHKERRQ(ierr); 1512674ae819SStefano Zampini total_counts += j; 1513674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 1514674ae819SStefano Zampini } 1515984c4197SStefano Zampini total_counts *= max_constraints; 1516674ae819SStefano Zampini total_counts += n_vertices; 1517906d46d4SStefano Zampini ierr = PetscMalloc3(total_counts,&temp_quadrature_constraint,total_counts,&temp_indices_to_constraint,total_counts,&temp_indices_to_constraint_B);CHKERRQ(ierr); 1518984c4197SStefano Zampini /* get local part of global near null space vectors */ 1519785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 1520984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1521984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 1522984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1523984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1524984c4197SStefano Zampini } 1525674ae819SStefano Zampini 1526242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 1527242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 1528a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 1529242a89d7SStefano Zampini 15308f1c130eSStefano Zampini /* allocate some auxiliary stuff */ 15318f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 15328f1c130eSStefano 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); 15338f1c130eSStefano Zampini } 15348f1c130eSStefano Zampini 1535984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 1536a773dcb8SStefano Zampini if (!skip_lapack) { 1537674ae819SStefano Zampini PetscScalar temp_work; 1538911cabfeSStefano Zampini 1539674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1540984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 1541785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 1542785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 1543785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 1544674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1545785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 1546674ae819SStefano Zampini #endif 1547674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1548c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 1549c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 1550674ae819SStefano Zampini lwork = -1; 1551674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1552674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1553c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 1554674ae819SStefano Zampini #else 1555c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 1556674ae819SStefano Zampini #endif 1557674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1558984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 1559674ae819SStefano Zampini #else /* on missing GESVD */ 1560674ae819SStefano Zampini /* SVD */ 1561674ae819SStefano Zampini PetscInt max_n,min_n; 1562674ae819SStefano Zampini max_n = max_size_of_constraint; 1563984c4197SStefano Zampini min_n = max_constraints; 1564984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 1565674ae819SStefano Zampini min_n = max_size_of_constraint; 1566984c4197SStefano Zampini max_n = max_constraints; 1567674ae819SStefano Zampini } 1568785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 1569674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1570785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 1571674ae819SStefano Zampini #endif 1572674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 1573674ae819SStefano Zampini lwork = -1; 1574e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 1575e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 1576b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 1577674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1578674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1579e310c8b4SStefano 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)); 1580674ae819SStefano Zampini #else 1581e310c8b4SStefano 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)); 1582674ae819SStefano Zampini #endif 1583674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1584984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 1585984c4197SStefano Zampini #endif /* on missing GESVD */ 1586674ae819SStefano Zampini /* Allocate optimal workspace */ 1587674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 1588854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 1589674ae819SStefano Zampini } 1590674ae819SStefano Zampini /* Now we can loop on constraining sets */ 1591674ae819SStefano Zampini total_counts = 0; 1592674ae819SStefano Zampini temp_indices[0] = 0; 1593674ae819SStefano Zampini /* vertices */ 1594674ae819SStefano Zampini if (ISForVertices) { 1595674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1596674ae819SStefano Zampini if (nnsp_has_cnst) { /* consider all vertices */ 159739e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 1598674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1599674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1600674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1601674ae819SStefano Zampini total_counts++; 1602674ae819SStefano Zampini } 1603674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 1604984c4197SStefano Zampini PetscBool used_vertex; 1605674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 1606674ae819SStefano Zampini used_vertex = PETSC_FALSE; 1607674ae819SStefano Zampini k = 0; 1608674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 1609984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1610984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 1611674ae819SStefano Zampini temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i]; 1612674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]]=1.0; 1613674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+1; 1614674ae819SStefano Zampini total_counts++; 1615674ae819SStefano Zampini used_vertex = PETSC_TRUE; 1616674ae819SStefano Zampini } 1617984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1618674ae819SStefano Zampini k++; 1619674ae819SStefano Zampini } 1620674ae819SStefano Zampini } 1621674ae819SStefano Zampini } 1622674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1623674ae819SStefano Zampini n_vertices = total_counts; 1624674ae819SStefano Zampini } 1625984c4197SStefano Zampini 1626674ae819SStefano Zampini /* edges and faces */ 1627911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 1628911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 1629911cabfeSStefano Zampini used_IS = &ISForEdges[ncc]; 1630984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 1631674ae819SStefano Zampini } else { 1632911cabfeSStefano Zampini used_IS = &ISForFaces[ncc-n_ISForEdges]; 1633984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 1634674ae819SStefano Zampini } 1635674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 1636674ae819SStefano Zampini temp_start_ptr = total_counts; /* need to know the starting index of constraints stored */ 1637674ae819SStefano Zampini ierr = ISGetSize(*used_IS,&size_of_constraint);CHKERRQ(ierr); 1638674ae819SStefano Zampini ierr = ISGetIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1639984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 1640984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 1641674ae819SStefano Zampini if (nnsp_has_cnst) { 16425b08dc53SStefano Zampini PetscScalar quad_value; 1643674ae819SStefano Zampini temp_constraints++; 1644a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 1645674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 1646a773dcb8SStefano Zampini } else { 1647a773dcb8SStefano Zampini quad_value = 1.0; 1648a773dcb8SStefano Zampini } 164939e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1650674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1651674ae819SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=quad_value; 1652674ae819SStefano Zampini } 1653911cabfeSStefano Zampini /* sort by global ordering if using lapack subroutines */ 16548f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1655911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 1656911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1657911cabfeSStefano Zampini permutation[j]=j; 1658911cabfeSStefano Zampini } 1659911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 1660911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1661911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 1662911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 1663911cabfeSStefano Zampini } 1664911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1665911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 1666911cabfeSStefano Zampini } 1667674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 1668674ae819SStefano Zampini total_counts++; 1669674ae819SStefano Zampini } 1670674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 1671984c4197SStefano Zampini PetscReal real_value; 1672984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 167339e2fb2aSStefano Zampini ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1674674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1675984c4197SStefano Zampini temp_quadrature_constraint[temp_indices[total_counts]+j]=array[is_indices[j]]; 1676674ae819SStefano Zampini } 1677984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 1678984c4197SStefano Zampini /* check if array is null on the connected component */ 1679e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1680e310c8b4SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_one)); 16815b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 1682911cabfeSStefano Zampini /* sort by global ordering if using lapack subroutines */ 16838f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1684911cabfeSStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr); 1685911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1686911cabfeSStefano Zampini permutation[j]=j; 1687911cabfeSStefano Zampini } 1688911cabfeSStefano Zampini ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr); 1689911cabfeSStefano Zampini for (j=0;j<size_of_constraint;j++) { 1690911cabfeSStefano Zampini temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]]; 1691911cabfeSStefano Zampini temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]]; 1692911cabfeSStefano Zampini } 1693911cabfeSStefano Zampini ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 1694911cabfeSStefano Zampini ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr); 1695911cabfeSStefano Zampini } 1696674ae819SStefano Zampini temp_constraints++; 1697674ae819SStefano Zampini temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint; /* store new starting point */ 1698674ae819SStefano Zampini total_counts++; 1699674ae819SStefano Zampini } 1700674ae819SStefano Zampini } 1701674ae819SStefano Zampini ierr = ISRestoreIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr); 170245a1bb75SStefano Zampini valid_constraints = temp_constraints; 1703eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 1704a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 1705a773dcb8SStefano Zampini PetscScalar norm; 1706a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1707a773dcb8SStefano 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)); 1708a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 1709a773dcb8SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one)); 1710a773dcb8SStefano Zampini } else { /* perform SVD */ 1711984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 1712674ae819SStefano Zampini 1713674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1714984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 1715984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 1716984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 1717984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 1718984c4197SStefano Zampini from that computed using LAPACKgesvd 1719984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 1720984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 1721984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 1722674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 1723e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 1724984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1725674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 1726674ae819SStefano Zampini for (k=0;k<j+1;k++) { 1727e310c8b4SStefano 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)); 1728674ae819SStefano Zampini } 1729674ae819SStefano Zampini } 1730e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 1731e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1732e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 1733674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1734c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 1735674ae819SStefano Zampini #else 1736c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 1737674ae819SStefano Zampini #endif 1738674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1739984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 1740984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 1741674ae819SStefano Zampini j = 0; 1742984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 1743674ae819SStefano Zampini total_counts = total_counts-j; 174445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 1745e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 1746c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 1747c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1748c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 1749c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 1750c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 1751c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 1752674ae819SStefano Zampini if (j<temp_constraints) { 1753984c4197SStefano Zampini PetscInt ii; 1754984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k]=1.0/PetscSqrtReal(singular_vals[k]); 1755674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1756c4303822SStefano 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)); 1757674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1758984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 1759674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 1760984c4197SStefano 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]; 1761674ae819SStefano Zampini } 1762674ae819SStefano Zampini } 1763674ae819SStefano Zampini } 1764674ae819SStefano Zampini #else /* on missing GESVD */ 1765e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 1766e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 1767b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 1768674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1769674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1770e310c8b4SStefano 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)); 1771674ae819SStefano Zampini #else 1772e310c8b4SStefano 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)); 1773674ae819SStefano Zampini #endif 1774984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 1775674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1776984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 1777e310c8b4SStefano Zampini k = temp_constraints; 1778e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 1779674ae819SStefano Zampini j = 0; 1780e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 178145a1bb75SStefano Zampini valid_constraints = k-j; 1782911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 1783984c4197SStefano Zampini #endif /* on missing GESVD */ 1784674ae819SStefano Zampini } 1785a773dcb8SStefano Zampini } 178645a1bb75SStefano Zampini /* setting change_of_basis flag is safe now */ 178745a1bb75SStefano Zampini if (boolforchange) { 178845a1bb75SStefano Zampini for (j=0;j<valid_constraints;j++) { 178945a1bb75SStefano Zampini PetscBTSet(change_basis,total_counts-j-1); 179045a1bb75SStefano Zampini } 179145a1bb75SStefano Zampini } 1792674ae819SStefano Zampini } 1793674ae819SStefano Zampini /* free index sets of faces, edges and vertices */ 1794674ae819SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 1795674ae819SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 1796674ae819SStefano Zampini } 1797d06fc5fdSStefano Zampini if (n_ISForFaces) { 1798674ae819SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 1799d06fc5fdSStefano Zampini } 1800674ae819SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1801674ae819SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1802674ae819SStefano Zampini } 1803d06fc5fdSStefano Zampini if (n_ISForEdges) { 1804674ae819SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1805d06fc5fdSStefano Zampini } 1806674ae819SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 180739e2fb2aSStefano Zampini /* map temp_indices_to_constraint in boundary numbering */ 18085e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,temp_indices[total_counts],temp_indices_to_constraint,&i,temp_indices_to_constraint_B);CHKERRQ(ierr); 180939e2fb2aSStefano Zampini if (i != temp_indices[total_counts]) { 181039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",temp_indices[total_counts],i); 181139e2fb2aSStefano Zampini } 1812674ae819SStefano Zampini 1813984c4197SStefano Zampini /* free workspace */ 18148f1c130eSStefano Zampini if (!skip_lapack || pcbddc->use_qr_single) { 1815911cabfeSStefano Zampini ierr = PetscFree4(gidxs,permutation,temp_indices_to_constraint_work,temp_quadrature_constraint_work);CHKERRQ(ierr); 18168f1c130eSStefano Zampini } 18178f1c130eSStefano Zampini if (!skip_lapack) { 1818984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1819984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1820984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 1821984c4197SStefano Zampini #endif 1822984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 1823984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1824984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 1825984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 1826984c4197SStefano Zampini #endif 1827984c4197SStefano Zampini } 1828984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 1829984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 1830984c4197SStefano Zampini } 1831984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 1832984c4197SStefano Zampini 1833727cdba6SStefano Zampini /* set quantities in pcbddc data structure and store previous primal size */ 1834984c4197SStefano Zampini /* n_vertices defines the number of subdomain corners in the primal space */ 1835674ae819SStefano Zampini /* n_constraints defines the number of averages (they can be point primal dofs if change of basis is requested) */ 1836727cdba6SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 1837984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 1838674ae819SStefano Zampini pcbddc->n_vertices = n_vertices; 1839984c4197SStefano Zampini pcbddc->n_constraints = pcbddc->local_primal_size-pcbddc->n_vertices; 1840674ae819SStefano Zampini 1841674ae819SStefano Zampini /* Create constraint matrix */ 1842674ae819SStefano Zampini /* The constraint matrix is used to compute the l2g map of primal dofs */ 1843674ae819SStefano Zampini /* so we need to set it up properly either with or without change of basis */ 1844674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1845674ae819SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,impMatType);CHKERRQ(ierr); 1846984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 1847984c4197SStefano Zampini /* array to compute a local numbering of constraints : vertices first then constraints */ 1848785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_numbering);CHKERRQ(ierr); 1849984c4197SStefano Zampini /* array to select the proper local node (of minimum index with respect to global ordering) when changing the basis */ 1850984c4197SStefano 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 */ 1851785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_minloc);CHKERRQ(ierr); 1852984c4197SStefano Zampini /* auxiliary stuff for basis change */ 1853785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&global_indices);CHKERRQ(ierr); 18544641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&touched);CHKERRQ(ierr); 1855984c4197SStefano Zampini 1856984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 1857984c4197SStefano Zampini total_primal_vertices=0; 1858984c4197SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1859674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 1860984c4197SStefano Zampini if (size_of_constraint == 1) { 18614641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]]);CHKERRQ(ierr); 1862984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]]; 1863984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=0; 1864984c4197SStefano Zampini total_primal_vertices++; 18654641a718SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { /* Same procedure used in PCBDDCGetPrimalConstraintsLocalIdx */ 1866984c4197SStefano Zampini PetscInt min_loc,min_index; 1867984c4197SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],global_indices);CHKERRQ(ierr); 1868984c4197SStefano Zampini /* find first untouched local node */ 1869674ae819SStefano Zampini k = 0; 18704641a718SStefano Zampini while (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) k++; 1871984c4197SStefano Zampini min_index = global_indices[k]; 1872984c4197SStefano Zampini min_loc = k; 1873984c4197SStefano Zampini /* search the minimum among global nodes already untouched on the cc */ 1874984c4197SStefano Zampini for (k=1;k<size_of_constraint;k++) { 1875984c4197SStefano Zampini /* there can be more than one constraint on a single connected component */ 18764641a718SStefano Zampini if (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k]) && min_index > global_indices[k]) { 1877984c4197SStefano Zampini min_index = global_indices[k]; 1878984c4197SStefano Zampini min_loc = k; 1879674ae819SStefano Zampini } 1880674ae819SStefano Zampini } 18814641a718SStefano Zampini ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]+min_loc]);CHKERRQ(ierr); 1882984c4197SStefano Zampini aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]+min_loc]; 1883984c4197SStefano Zampini aux_primal_minloc[total_primal_vertices]=min_loc; 1884984c4197SStefano Zampini total_primal_vertices++; 1885984c4197SStefano Zampini } 1886984c4197SStefano Zampini } 1887a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 1888a717540cSStefano Zampini qr_needed = PETSC_FALSE; 1889a717540cSStefano Zampini ierr = PetscBTCreate(pcbddc->local_primal_size,&qr_needed_idx);CHKERRQ(ierr); 1890a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 1891a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 1892fa434743SStefano Zampini if (!pcbddc->use_qr_single) { 1893a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 1894a717540cSStefano Zampini j = 0; 1895a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 1896a717540cSStefano Zampini if (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) { 1897a717540cSStefano Zampini j++; 1898a717540cSStefano Zampini } 1899a717540cSStefano Zampini } 1900a717540cSStefano Zampini /* found more than one primal dof on the cc */ 1901a717540cSStefano Zampini if (j > 1) { 1902a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 1903a717540cSStefano Zampini qr_needed = PETSC_TRUE; 1904a717540cSStefano Zampini } 1905fa434743SStefano Zampini } else { 1906fa434743SStefano Zampini PetscBTSet(qr_needed_idx,i); 1907fa434743SStefano Zampini qr_needed = PETSC_TRUE; 1908fa434743SStefano Zampini } 1909a717540cSStefano Zampini } 1910a717540cSStefano Zampini } 1911984c4197SStefano Zampini /* free workspace */ 1912984c4197SStefano Zampini ierr = PetscFree(global_indices);CHKERRQ(ierr); 191345a1bb75SStefano Zampini 1914674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 1915a717540cSStefano Zampini ierr = PetscSortInt(total_primal_vertices,aux_primal_numbering);CHKERRQ(ierr); 1916984c4197SStefano Zampini 1917984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 1918785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 1919984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i]=1; 1920984c4197SStefano Zampini j=total_primal_vertices; 1921984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 19224641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 1923674ae819SStefano Zampini nnz[j]=temp_indices[i+1]-temp_indices[i]; 1924674ae819SStefano Zampini j++; 1925674ae819SStefano Zampini } 1926674ae819SStefano Zampini } 1927674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 1928674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1929674ae819SStefano Zampini /* set values in constraint matrix */ 1930984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 1931984c4197SStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,aux_primal_numbering[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 1932674ae819SStefano Zampini } 1933984c4197SStefano Zampini total_counts = total_primal_vertices; 1934984c4197SStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 19354641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 1936674ae819SStefano Zampini size_of_constraint=temp_indices[i+1]-temp_indices[i]; 1937674ae819SStefano 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); 1938674ae819SStefano Zampini total_counts++; 1939674ae819SStefano Zampini } 1940674ae819SStefano Zampini } 1941674ae819SStefano Zampini /* assembling */ 1942674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1943674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1944984c4197SStefano Zampini /* 194545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1946984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 1947984c4197SStefano Zampini */ 1948674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 1949674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 1950026de310SStefano Zampini /* dual and primal dofs on a single cc */ 1951984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 1952026de310SStefano Zampini /* iterator on aux_primal_minloc (ordered as read from nearnullspace: vertices, edges and then constraints) */ 1953026de310SStefano Zampini PetscInt primal_counter; 1954984c4197SStefano Zampini /* working stuff for GEQRF */ 195581d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 1956984c4197SStefano Zampini PetscBLASInt lqr_work; 1957984c4197SStefano Zampini /* working stuff for UNGQR */ 1958984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 1959984c4197SStefano Zampini PetscBLASInt lgqr_work; 1960984c4197SStefano Zampini /* working stuff for TRTRS */ 1961984c4197SStefano Zampini PetscScalar *trs_rhs; 19623f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 1963984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 1964984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 1965984c4197SStefano Zampini PetscScalar *start_vals; 1966984c4197SStefano Zampini /* working stuff for values insertion */ 19674641a718SStefano Zampini PetscBT is_primal; 1968906d46d4SStefano Zampini /* matrix sizes */ 1969906d46d4SStefano Zampini PetscInt global_size,local_size; 1970906d46d4SStefano Zampini /* work array for nonzeros */ 1971906d46d4SStefano Zampini PetscScalar *nnz_array; 1972906d46d4SStefano Zampini /* temporary change of basis */ 1973906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 1974906d46d4SStefano Zampini /* auxiliary work for global change of basis */ 1975906d46d4SStefano Zampini Vec nnz_vec; 197622bc73bbSStefano Zampini PetscInt *idxs_I,*idxs_B,*idxs_all,*d_nnz,*o_nnz; 1977906d46d4SStefano Zampini PetscInt nvtxs,*xadj,*adjncy,*idxs_mapped; 1978906d46d4SStefano Zampini PetscScalar *vals; 1979906d46d4SStefano Zampini PetscBool done; 1980984c4197SStefano Zampini 1981906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 1982906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 1983906d46d4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,impMatType);CHKERRQ(ierr); 1984906d46d4SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n_B,pcis->n_B,pcis->n_B,pcis->n_B);CHKERRQ(ierr); 1985906d46d4SStefano Zampini 1986906d46d4SStefano Zampini /* nonzeros for local mat */ 1987785e854fSJed Brown ierr = PetscMalloc1(pcis->n_B,&nnz);CHKERRQ(ierr); 1988a717540cSStefano Zampini for (i=0;i<pcis->n_B;i++) nnz[i]=1; 1989a717540cSStefano Zampini for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) { 1990a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 1991a717540cSStefano Zampini size_of_constraint = temp_indices[i+1]-temp_indices[i]; 1992a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 1993a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = size_of_constraint; 1994a717540cSStefano Zampini } else { 1995a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = 2; 1996a717540cSStefano Zampini /* get local primal index on the cc */ 1997a717540cSStefano Zampini j = 0; 1998a717540cSStefano Zampini while (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+j])) j++; 1999a717540cSStefano Zampini nnz[temp_indices_to_constraint_B[temp_indices[i]+j]] = size_of_constraint; 2000a717540cSStefano Zampini } 2001a717540cSStefano Zampini } 2002a717540cSStefano Zampini } 2003906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2004a717540cSStefano Zampini /* Set initial identity in the matrix */ 2005a717540cSStefano Zampini for (i=0;i<pcis->n_B;i++) { 2006906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2007a717540cSStefano Zampini } 2008a717540cSStefano Zampini 2009a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2010a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2011a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2012a717540cSStefano Zampini } 2013a717540cSStefano Zampini 2014a717540cSStefano Zampini 2015a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2016a717540cSStefano Zampini /* 2017a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2018a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2019a717540cSStefano Zampini 2020a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2021a717540cSStefano Zampini 2022a6b551f4SStefano 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) 2023a6b551f4SStefano Zampini 2024a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2025a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2026a717540cSStefano Zampini | ... | 2027a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2028a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2029a717540cSStefano Zampini 2030a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2031a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2032a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2033a6b551f4SStefano Zampini 2034a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2035a717540cSStefano Zampini */ 2036a717540cSStefano Zampini if (qr_needed) { 2037984c4197SStefano Zampini /* space to store Q */ 2038854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2039984c4197SStefano Zampini /* first we issue queries for optimal work */ 20403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 20413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 20423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2043984c4197SStefano Zampini lqr_work = -1; 20443f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2045984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2046984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2047785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2048984c4197SStefano Zampini lgqr_work = -1; 20493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 20503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 20513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 20523f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 20533f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 20543f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2055984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2056984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2057785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2058984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2059785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2060984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2061785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2062a717540cSStefano Zampini /* allocating workspace for check */ 2063a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2064785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&work);CHKERRQ(ierr); 2065a717540cSStefano Zampini } 2066a717540cSStefano Zampini } 2067984c4197SStefano Zampini /* array to store whether a node is primal or not */ 20684641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2069473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 20705e8657edSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,aux_primal_numbering,&i,aux_primal_numbering_B);CHKERRQ(ierr); 207139e2fb2aSStefano Zampini if (i != total_primal_vertices) { 207239e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 20734641a718SStefano Zampini } 207439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 207539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 207639e2fb2aSStefano Zampini } 207739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2078984c4197SStefano Zampini 2079a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 2080026de310SStefano Zampini /* -> using implicit ordering contained in temp_indices data */ 2081026de310SStefano Zampini total_counts = pcbddc->n_vertices; 2082026de310SStefano Zampini primal_counter = total_counts; 2083026de310SStefano Zampini while (total_counts<pcbddc->local_primal_size) { 2084026de310SStefano Zampini primal_dofs = 1; 20854641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2086026de310SStefano Zampini /* get all constraints with same support: if more then one constraint is present on the cc then surely indices are stored contiguosly */ 2087026de310SStefano 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]]) { 2088026de310SStefano Zampini primal_dofs++; 2089674ae819SStefano Zampini } 2090984c4197SStefano Zampini /* get constraint info */ 2091026de310SStefano Zampini size_of_constraint = temp_indices[total_counts+1]-temp_indices[total_counts]; 2092984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2093984c4197SStefano Zampini 2094984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2095a717540cSStefano 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); 2096674ae819SStefano Zampini } 2097984c4197SStefano Zampini 2098fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2099a717540cSStefano Zampini 2100a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2101a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2102a717540cSStefano Zampini ierr = PetscMemcpy(work,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2103a717540cSStefano Zampini } 2104984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 2105026de310SStefano Zampini ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2106984c4197SStefano Zampini 2107984c4197SStefano Zampini /* compute QR decomposition of constraints */ 21083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2111674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21123f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2113984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2114674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2115984c4197SStefano Zampini 2116984c4197SStefano Zampini /* explictly compute R^-T */ 2117984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2118984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 21193f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21203f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 21213f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 21223f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2123984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21243f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2125984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2126984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2127984c4197SStefano Zampini 2128a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 21293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 21313f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 21323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2133984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 21343f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2135984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2136984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2137984c4197SStefano Zampini 2138984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2139984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2140984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 21413f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 21423f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 21433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 21443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 21453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 21463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2147984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2148c4303822SStefano 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)); 2149984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 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 /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 2153026de310SStefano Zampini start_rows = &temp_indices_to_constraint_B[temp_indices[total_counts]]; 2154984c4197SStefano Zampini /* insert cols for primal dofs */ 2155984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2156984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 2157026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+aux_primal_minloc[primal_counter+j]]; 2158906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2159984c4197SStefano Zampini } 2160984c4197SStefano Zampini /* insert cols for dual dofs */ 2161984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 21624641a718SStefano Zampini if (!PetscBTLookup(is_primal,temp_indices_to_constraint_B[temp_indices[total_counts]+k])) { 2163984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 2164026de310SStefano Zampini start_cols = &temp_indices_to_constraint_B[temp_indices[total_counts]+k]; 2165906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2166984c4197SStefano Zampini j++; 2167674ae819SStefano Zampini } 2168674ae819SStefano Zampini } 2169984c4197SStefano Zampini 2170984c4197SStefano Zampini /* check change of basis */ 2171984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2172984c4197SStefano Zampini PetscInt ii,jj; 2173984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2174c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2175c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2176c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2177c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2178c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2179c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2180984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2181c4303822SStefano 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)); 2182984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2183984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2184984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2185984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2186984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE; 2187674ae819SStefano Zampini } 2188674ae819SStefano Zampini } 2189984c4197SStefano Zampini if (!valid_qr) { 219022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2191984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2192984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2193984c4197SStefano Zampini if (ii != jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2194984c4197SStefano 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])); 2195674ae819SStefano Zampini } 2196984c4197SStefano Zampini if (ii == jj && PetscAbsScalar(work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2197984c4197SStefano 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])); 2198984c4197SStefano Zampini } 2199984c4197SStefano Zampini } 2200984c4197SStefano Zampini } 2201674ae819SStefano Zampini } else { 220222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2203674ae819SStefano Zampini } 2204674ae819SStefano Zampini } 2205a717540cSStefano Zampini } else { /* simple transformation block */ 2206a717540cSStefano Zampini PetscInt row,col; 2207a6b551f4SStefano Zampini PetscScalar val,norm; 2208a6b551f4SStefano Zampini 2209a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2210a6b551f4SStefano 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)); 2211a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 2212a717540cSStefano Zampini row = temp_indices_to_constraint_B[temp_indices[total_counts]+j]; 2213a717540cSStefano Zampini if (!PetscBTLookup(is_primal,row)) { 2214a717540cSStefano Zampini col = temp_indices_to_constraint_B[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2215906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 2216a6b551f4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,temp_quadrature_constraint[temp_indices[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2217a717540cSStefano Zampini } else { 2218a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 2219a717540cSStefano Zampini col = temp_indices_to_constraint_B[temp_indices[total_counts]+k]; 2220a717540cSStefano Zampini if (row != col) { 2221a717540cSStefano Zampini val = -temp_quadrature_constraint[temp_indices[total_counts]+k]/temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]; 2222a717540cSStefano Zampini } else { 2223a6b551f4SStefano Zampini val = temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]/norm; 2224a717540cSStefano Zampini } 2225906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2226a717540cSStefano Zampini } 2227a717540cSStefano Zampini } 2228a717540cSStefano Zampini } 222998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 223022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2231a717540cSStefano Zampini } 2232674ae819SStefano Zampini } 2233026de310SStefano Zampini /* increment primal counter */ 2234026de310SStefano Zampini primal_counter += primal_dofs; 2235984c4197SStefano Zampini } else { 2236984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2237026de310SStefano 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); 2238674ae819SStefano Zampini } 2239674ae819SStefano Zampini } 2240026de310SStefano Zampini /* increment constraint counter total_counts */ 2241026de310SStefano Zampini total_counts += primal_dofs; 2242674ae819SStefano Zampini } 2243a717540cSStefano Zampini 2244a717540cSStefano Zampini /* free workspace */ 2245a717540cSStefano Zampini if (qr_needed) { 2246984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2247984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2248984c4197SStefano Zampini } 2249984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2250984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2251984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2252984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2253984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2254674ae819SStefano Zampini } 2255a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2256906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2257906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2258906d46d4SStefano Zampini 2259906d46d4SStefano Zampini /* assembling of global change of variable */ 2260906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2261906d46d4SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2262906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2263906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2264906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2265906d46d4SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->ChangeOfBasisMatrix,matis->mapping,matis->mapping);CHKERRQ(ierr); 2266906d46d4SStefano Zampini 2267906d46d4SStefano Zampini /* nonzeros (overestimated) */ 2268906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&nnz_vec);CHKERRQ(ierr); 2269906d46d4SStefano Zampini ierr = VecSetLocalToGlobalMapping(nnz_vec,matis->mapping);CHKERRQ(ierr); 2270906d46d4SStefano Zampini ierr = PetscMalloc2(pcis->n,&nnz_array,pcis->n,&idxs_all);CHKERRQ(ierr); 2271906d46d4SStefano Zampini for (i=0;i<pcis->n;i++) { 2272906d46d4SStefano Zampini nnz_array[i] = 1.0; 2273906d46d4SStefano Zampini idxs_all[i] = i; 2274906d46d4SStefano Zampini } 2275906d46d4SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&idxs_B);CHKERRQ(ierr); 2276906d46d4SStefano Zampini for (i=0;i<pcis->n_B;i++) { 2277906d46d4SStefano Zampini nnz_array[idxs_B[i]] = nnz[i]; 2278906d46d4SStefano Zampini } 2279a5187659SStefano Zampini if (pcis->n) { 2280906d46d4SStefano Zampini ierr = VecSetValuesLocal(nnz_vec,pcis->n,idxs_all,nnz_array,INSERT_VALUES);CHKERRQ(ierr); 2281a5187659SStefano Zampini } 2282906d46d4SStefano Zampini ierr = VecAssemblyBegin(nnz_vec);CHKERRQ(ierr); 2283906d46d4SStefano Zampini ierr = VecAssemblyEnd(nnz_vec);CHKERRQ(ierr); 2284906d46d4SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2285906d46d4SStefano Zampini ierr = PetscFree2(nnz_array,idxs_all);CHKERRQ(ierr); 228622bc73bbSStefano Zampini ierr = PetscMalloc2(local_size,&d_nnz,local_size,&o_nnz);CHKERRQ(ierr); 2287906d46d4SStefano Zampini ierr = VecGetArray(nnz_vec,&nnz_array);CHKERRQ(ierr); 2288906d46d4SStefano Zampini for (i=0;i<local_size;i++) { 228922bc73bbSStefano Zampini d_nnz[i] = PetscMin((PetscInt)(PetscRealPart(nnz_array[i])),local_size); 229022bc73bbSStefano Zampini o_nnz[i] = PetscMin((PetscInt)(PetscRealPart(nnz_array[i])),global_size-local_size); 2291906d46d4SStefano Zampini } 2292906d46d4SStefano Zampini ierr = VecRestoreArray(nnz_vec,&nnz_array);CHKERRQ(ierr); 2293906d46d4SStefano Zampini ierr = VecDestroy(&nnz_vec);CHKERRQ(ierr); 229422bc73bbSStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->ChangeOfBasisMatrix,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 229522bc73bbSStefano Zampini ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 2296906d46d4SStefano Zampini 2297906d46d4SStefano Zampini /* Set identity on dirichlet dofs */ 2298906d46d4SStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&idxs_I);CHKERRQ(ierr); 2299906d46d4SStefano Zampini for (i=0;i<pcis->n-pcis->n_B;i++) { 2300906d46d4SStefano Zampini PetscScalar one=1.0; 2301906d46d4SStefano Zampini ierr = MatSetValuesLocal(pcbddc->ChangeOfBasisMatrix,1,idxs_I+i,1,idxs_I+i,&one,INSERT_VALUES);CHKERRQ(ierr); 2302906d46d4SStefano Zampini } 2303906d46d4SStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&idxs_I);CHKERRQ(ierr); 2304906d46d4SStefano Zampini 2305906d46d4SStefano Zampini /* Set values at interface dofs */ 2306906d46d4SStefano Zampini done = PETSC_TRUE; 2307906d46d4SStefano Zampini ierr = MatGetRowIJ(localChangeOfBasisMatrix,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr); 2308906d46d4SStefano Zampini if (!done) { 2309906d46d4SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error in MatGetRowIJ called in %s\n",__FUNCT__); 2310906d46d4SStefano Zampini } 2311906d46d4SStefano Zampini ierr = MatSeqAIJGetArray(localChangeOfBasisMatrix,&vals);CHKERRQ(ierr); 2312906d46d4SStefano Zampini ierr = PetscMalloc1(xadj[nvtxs],&idxs_mapped);CHKERRQ(ierr); 23135e8657edSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->BtoNmap,xadj[nvtxs],adjncy,idxs_mapped);CHKERRQ(ierr); 2314906d46d4SStefano Zampini for (i=0;i<nvtxs;i++) { 2315906d46d4SStefano Zampini PetscInt row,*cols,ncols; 2316906d46d4SStefano Zampini PetscScalar *mat_vals; 2317906d46d4SStefano Zampini 2318906d46d4SStefano Zampini row = idxs_B[i]; 2319906d46d4SStefano Zampini ncols = xadj[i+1]-xadj[i]; 2320906d46d4SStefano Zampini cols = idxs_mapped+xadj[i]; 2321906d46d4SStefano Zampini mat_vals = vals+xadj[i]; 2322906d46d4SStefano Zampini ierr = MatSetValuesLocal(pcbddc->ChangeOfBasisMatrix,1,&row,ncols,cols,mat_vals,INSERT_VALUES);CHKERRQ(ierr); 2323906d46d4SStefano Zampini } 2324906d46d4SStefano Zampini ierr = MatRestoreRowIJ(localChangeOfBasisMatrix,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr); 2325906d46d4SStefano Zampini if (!done) { 2326906d46d4SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__); 2327906d46d4SStefano Zampini } 2328906d46d4SStefano Zampini ierr = MatSeqAIJRestoreArray(localChangeOfBasisMatrix,&vals);CHKERRQ(ierr); 2329906d46d4SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&idxs_B);CHKERRQ(ierr); 2330906d46d4SStefano Zampini ierr = PetscFree(idxs_mapped);CHKERRQ(ierr); 2331674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2332674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2333906d46d4SStefano Zampini 2334906d46d4SStefano Zampini /* check */ 2335906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2336906d46d4SStefano Zampini PetscReal error; 2337906d46d4SStefano Zampini Vec x,x_change; 2338906d46d4SStefano Zampini 2339906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2340906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2341906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2342906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2343906d46d4SStefano Zampini ierr = VecScatterBegin(pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2344906d46d4SStefano Zampini ierr = VecScatterEnd(pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2345906d46d4SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 2346906d46d4SStefano Zampini ierr = VecScatterBegin(pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2347906d46d4SStefano Zampini ierr = VecScatterEnd(pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2348906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2349906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2350906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2351906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2352906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on B: %1.6e\n",error);CHKERRQ(ierr); 2353906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2354906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2355906d46d4SStefano Zampini } 2356906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2357906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2358b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2359b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2360b9b85e73SStefano Zampini } 2361906d46d4SStefano Zampini 2362906d46d4SStefano Zampini /* set up change of basis context */ 2363906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2364906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2365906d46d4SStefano Zampini 2366906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2367906d46d4SStefano Zampini PetscInt global_size,local_size; 2368906d46d4SStefano Zampini 2369906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2370906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2371906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2372906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2373906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2374906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2375906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2376906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2377906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2378906d46d4SStefano Zampini } else { 2379906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2380906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2381906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2382906d46d4SStefano Zampini } 2383906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2384906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2385906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2386906d46d4SStefano Zampini } else { 2387906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2388906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2389906d46d4SStefano Zampini } 2390906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2391906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2392906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2393906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2394b9b85e73SStefano Zampini } 2395a717540cSStefano Zampini 2396727cdba6SStefano Zampini /* get indices in local ordering for vertices and constraints */ 2397727cdba6SStefano 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 */ 2398473ba861SJed Brown ierr = PetscMalloc1(olocal_primal_size,&oprimal_indices_local_idxs);CHKERRQ(ierr); 2399727cdba6SStefano Zampini ierr = PetscMemcpy(oprimal_indices_local_idxs,pcbddc->primal_indices_local_idxs,olocal_primal_size*sizeof(PetscInt));CHKERRQ(ierr); 2400727cdba6SStefano Zampini } 2401727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2402f347579bSStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2403473ba861SJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2404727cdba6SStefano Zampini ierr = PCBDDCGetPrimalVerticesLocalIdx(pc,&i,&aux_primal_numbering);CHKERRQ(ierr); 2405727cdba6SStefano Zampini ierr = PetscMemcpy(pcbddc->primal_indices_local_idxs,aux_primal_numbering,i*sizeof(PetscInt));CHKERRQ(ierr); 2406727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2407727cdba6SStefano Zampini ierr = PCBDDCGetPrimalConstraintsLocalIdx(pc,&j,&aux_primal_numbering);CHKERRQ(ierr); 2408727cdba6SStefano Zampini ierr = PetscMemcpy(&pcbddc->primal_indices_local_idxs[i],aux_primal_numbering,j*sizeof(PetscInt));CHKERRQ(ierr); 2409727cdba6SStefano Zampini ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr); 2410e9189074SStefano Zampini /* set quantities in PCBDDC data struct */ 2411e9189074SStefano Zampini pcbddc->n_actual_vertices = i; 2412727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2413727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2414727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2415727cdba6SStefano Zampini ierr = PetscMemcmp(pcbddc->primal_indices_local_idxs,oprimal_indices_local_idxs,olocal_primal_size,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2416c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2417727cdba6SStefano Zampini ierr = PetscFree(oprimal_indices_local_idxs);CHKERRQ(ierr); 2418727cdba6SStefano Zampini } 2419727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2420727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2421727cdba6SStefano Zampini 2422a717540cSStefano Zampini /* flush dbg viewer */ 2423b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2424b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2425b8ffe317SStefano Zampini } 2426a717540cSStefano Zampini 2427e310c8b4SStefano Zampini /* free workspace */ 2428a717540cSStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 2429a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 2430984c4197SStefano Zampini ierr = PetscFree(aux_primal_minloc);CHKERRQ(ierr); 24314641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 2432906d46d4SStefano Zampini ierr = PetscFree(temp_indices);CHKERRQ(ierr); 2433906d46d4SStefano Zampini ierr = PetscFree3(temp_quadrature_constraint,temp_indices_to_constraint,temp_indices_to_constraint_B);CHKERRQ(ierr); 2434674ae819SStefano Zampini PetscFunctionReturn(0); 2435674ae819SStefano Zampini } 2436674ae819SStefano Zampini 2437674ae819SStefano Zampini #undef __FUNCT__ 2438674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2439674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2440674ae819SStefano Zampini { 2441674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2442674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2443674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 244463602bcaSStefano Zampini PetscInt ierr,i,vertex_size; 2445674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2446674ae819SStefano Zampini 2447674ae819SStefano Zampini PetscFunctionBegin; 24488e61c736SStefano Zampini /* Reset previously computed graph */ 24498e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2450674ae819SStefano Zampini /* Init local Graph struct */ 2451674ae819SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping);CHKERRQ(ierr); 2452674ae819SStefano Zampini 2453575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2454575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2455575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2456575ad6abSStefano Zampini } 24579577ea80SStefano Zampini 2458674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 24594d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 2460674ae819SStefano Zampini Mat mat_adj; 24614d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 24624d379d7bSStefano Zampini PetscInt nvtxs; 2463674ae819SStefano Zampini PetscBool flg_row=PETSC_TRUE; 2464674ae819SStefano Zampini 2465674ae819SStefano Zampini ierr = MatConvert(matis->A,MATMPIADJ,MAT_INITIAL_MATRIX,&mat_adj);CHKERRQ(ierr); 24664d379d7bSStefano Zampini ierr = MatGetRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2467674ae819SStefano Zampini if (!flg_row) { 2468674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatGetRowIJ called in %s\n",__FUNCT__); 2469674ae819SStefano Zampini } 24704d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 24714d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 24724d379d7bSStefano Zampini pcbddc->deluxe_compute_rowadj = PETSC_FALSE; 24734d379d7bSStefano Zampini } else { /* just compute subdomain's connected components */ 24744d379d7bSStefano Zampini IS is_dummy; 24754d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 24764d379d7bSStefano Zampini PetscInt j,sum; 24774d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 24784d379d7bSStefano Zampini const PetscInt *idxs; 24794d379d7bSStefano Zampini PCBDDCGraph graph; 24804d379d7bSStefano Zampini PetscBT is_on_boundary; 24814d379d7bSStefano Zampini 24824d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 24834d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 24844d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 24854d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 24864d379d7bSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy);CHKERRQ(ierr); 24874d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 24884d379d7bSStefano Zampini graph->xadj = xadj; 24894d379d7bSStefano Zampini graph->adjncy = adjncy; 24904d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 24914d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 24924d379d7bSStefano Zampini 24934d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 24944d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains\n",PetscGlobalRank,graph->ncc);CHKERRQ(ierr); 24954d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 24964d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 24974d379d7bSStefano Zampini } 24984d379d7bSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 24994d379d7bSStefano Zampini } 25004d379d7bSStefano Zampini 25014d379d7bSStefano Zampini ierr = PetscBTCreate(nvtxs,&is_on_boundary);CHKERRQ(ierr); 25024d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 25034d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 25044d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 25054d379d7bSStefano Zampini } 25064d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 25074d379d7bSStefano Zampini 25084d379d7bSStefano Zampini ierr = PetscCalloc1(nvtxs+1,&cxadj);CHKERRQ(ierr); 25094d379d7bSStefano Zampini sum = 0; 25104d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 25114d379d7bSStefano Zampini PetscInt sizecc = 0; 25124d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25134d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25144d379d7bSStefano Zampini sizecc++; 25154d379d7bSStefano Zampini } 25164d379d7bSStefano Zampini } 25174d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25184d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25194d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 25204d379d7bSStefano Zampini } 25214d379d7bSStefano Zampini } 25224d379d7bSStefano Zampini sum += sizecc*sizecc; 25234d379d7bSStefano Zampini } 25244d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 25254d379d7bSStefano Zampini sum = 0; 25264d379d7bSStefano Zampini for (i=0;i<nvtxs;i++) { 25274d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 25284d379d7bSStefano Zampini cxadj[i] = sum; 25294d379d7bSStefano Zampini sum += temp; 25304d379d7bSStefano Zampini } 25314d379d7bSStefano Zampini cxadj[nvtxs] = sum; 25324d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 25334d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 25344d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 25354d379d7bSStefano Zampini PetscInt k,sizecc = 0; 25364d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 25374d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 25384d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 25394d379d7bSStefano Zampini sizecc++; 25404d379d7bSStefano Zampini } 25414d379d7bSStefano Zampini } 25424d379d7bSStefano Zampini } 25434d379d7bSStefano Zampini } 25444d379d7bSStefano Zampini } 25454d379d7bSStefano Zampini if (nvtxs) { 25464d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 25474d379d7bSStefano Zampini } else { 25484d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 25494d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 25504d379d7bSStefano Zampini } 25514d379d7bSStefano Zampini graph->xadj = 0; 25524d379d7bSStefano Zampini graph->adjncy = 0; 25534d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 25544d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 25554d379d7bSStefano Zampini } 25564d379d7bSStefano Zampini ierr = MatRestoreRowIJ(mat_adj,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 2557674ae819SStefano Zampini if (!flg_row) { 2558674ae819SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__); 2559674ae819SStefano Zampini } 2560674ae819SStefano Zampini ierr = MatDestroy(&mat_adj);CHKERRQ(ierr); 2561674ae819SStefano Zampini } 2562674ae819SStefano Zampini 256363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 2564674ae819SStefano Zampini vertex_size = 1; 256563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 256663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 256795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 256863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 2569a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 257063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 2571674ae819SStefano Zampini } 257263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 257363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 257463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 2575674ae819SStefano Zampini } 257663602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 2577674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 257863602bcaSStefano Zampini } else { 257963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 258063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 2581854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 258263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 258363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 258463602bcaSStefano Zampini } 258563602bcaSStefano Zampini } 2586674ae819SStefano Zampini } 2587674ae819SStefano Zampini 2588674ae819SStefano Zampini /* Setup of Graph */ 2589785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 2590a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 2591785d1243SStefano Zampini } 2592785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 2593a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2594785d1243SStefano Zampini } 259563602bcaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices); 2596674ae819SStefano Zampini 2597674ae819SStefano Zampini /* Graph's connected components analysis */ 2598674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 2599674ae819SStefano Zampini 2600674ae819SStefano Zampini /* print some info to stdout */ 2601674ae819SStefano Zampini if (pcbddc->dbg_flag) { 2602e49050b4SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer); 2603674ae819SStefano Zampini } 2604fb180af4SStefano Zampini 2605fb180af4SStefano Zampini /* mark topography has done */ 2606fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 2607674ae819SStefano Zampini PetscFunctionReturn(0); 2608674ae819SStefano Zampini } 2609674ae819SStefano Zampini 2610674ae819SStefano Zampini #undef __FUNCT__ 2611674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalVerticesLocalIdx" 2612f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalVerticesLocalIdx(PC pc, PetscInt *n_vertices, PetscInt **vertices_idx) 2613674ae819SStefano Zampini { 2614674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2615674ae819SStefano Zampini PetscInt *vertices,*row_cmat_indices,n,i,size_of_constraint,local_primal_size; 2616674ae819SStefano Zampini PetscErrorCode ierr; 2617674ae819SStefano Zampini 2618674ae819SStefano Zampini PetscFunctionBegin; 2619674ae819SStefano Zampini n = 0; 2620674ae819SStefano Zampini vertices = 0; 2621674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 2622674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&i);CHKERRQ(ierr); 2623b120a5c6SStefano Zampini for (i=0;i<local_primal_size;i++) { 2624b120a5c6SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2625b120a5c6SStefano Zampini if (size_of_constraint == 1) n++; 2626b120a5c6SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2627b120a5c6SStefano Zampini } 2628811e8ca2SStefano Zampini if (vertices_idx) { 2629785e854fSJed Brown ierr = PetscMalloc1(n,&vertices);CHKERRQ(ierr); 2630b120a5c6SStefano Zampini n = 0; 2631674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2632674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2633674ae819SStefano Zampini if (size_of_constraint == 1) { 2634674ae819SStefano Zampini vertices[n++]=row_cmat_indices[0]; 2635674ae819SStefano Zampini } 2636674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2637674ae819SStefano Zampini } 2638674ae819SStefano Zampini } 2639811e8ca2SStefano Zampini } 2640674ae819SStefano Zampini *n_vertices = n; 2641811e8ca2SStefano Zampini if (vertices_idx) *vertices_idx = vertices; 2642674ae819SStefano Zampini PetscFunctionReturn(0); 2643674ae819SStefano Zampini } 2644674ae819SStefano Zampini 2645674ae819SStefano Zampini #undef __FUNCT__ 2646674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalConstraintsLocalIdx" 2647f34684f1SStefano Zampini PetscErrorCode PCBDDCGetPrimalConstraintsLocalIdx(PC pc, PetscInt *n_constraints, PetscInt **constraints_idx) 2648674ae819SStefano Zampini { 2649674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2650674ae819SStefano Zampini PetscInt *constraints_index,*row_cmat_indices,*row_cmat_global_indices; 2651674ae819SStefano Zampini PetscInt n,i,j,size_of_constraint,local_primal_size,local_size,max_size_of_constraint,min_index,min_loc; 26524641a718SStefano Zampini PetscBT touched; 2653674ae819SStefano Zampini PetscErrorCode ierr; 2654674ae819SStefano Zampini 2655f34684f1SStefano Zampini /* This function assumes that the number of local constraints per connected component 2656f34684f1SStefano Zampini is not greater than the number of nodes defined for the connected component 2657f34684f1SStefano Zampini (otherwise we will surely have linear dependence between constraints and thus a singular coarse problem) */ 2658674ae819SStefano Zampini PetscFunctionBegin; 2659674ae819SStefano Zampini n = 0; 2660674ae819SStefano Zampini constraints_index = 0; 2661674ae819SStefano Zampini if (pcbddc->ConstraintMatrix) { 2662674ae819SStefano Zampini ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&local_size);CHKERRQ(ierr); 2663674ae819SStefano Zampini max_size_of_constraint = 0; 2664674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2665674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2666674ae819SStefano Zampini if (size_of_constraint > 1) { 2667674ae819SStefano Zampini n++; 2668674ae819SStefano Zampini } 2669674ae819SStefano Zampini max_size_of_constraint = PetscMax(size_of_constraint,max_size_of_constraint); 2670674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr); 2671674ae819SStefano Zampini } 2672811e8ca2SStefano Zampini if (constraints_idx) { 2673785e854fSJed Brown ierr = PetscMalloc1(n,&constraints_index);CHKERRQ(ierr); 2674785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint,&row_cmat_global_indices);CHKERRQ(ierr); 26754641a718SStefano Zampini ierr = PetscBTCreate(local_size,&touched);CHKERRQ(ierr); 2676674ae819SStefano Zampini n = 0; 2677674ae819SStefano Zampini for (i=0;i<local_primal_size;i++) { 2678674ae819SStefano Zampini ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2679674ae819SStefano Zampini if (size_of_constraint > 1) { 2680674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,row_cmat_indices,row_cmat_global_indices);CHKERRQ(ierr); 268182d3d8afSStefano Zampini /* find first untouched local node */ 268282d3d8afSStefano Zampini j = 0; 26834641a718SStefano Zampini while (PetscBTLookup(touched,row_cmat_indices[j])) j++; 268482d3d8afSStefano Zampini min_index = row_cmat_global_indices[j]; 268582d3d8afSStefano Zampini min_loc = j; 268682d3d8afSStefano Zampini /* search the minimum among nodes not yet touched on the connected component 268782d3d8afSStefano Zampini since there can be more than one constraint on a single cc */ 2688674ae819SStefano Zampini for (j=1;j<size_of_constraint;j++) { 26894641a718SStefano Zampini if (!PetscBTLookup(touched,row_cmat_indices[j]) && min_index > row_cmat_global_indices[j]) { 2690674ae819SStefano Zampini min_index = row_cmat_global_indices[j]; 2691674ae819SStefano Zampini min_loc = j; 2692674ae819SStefano Zampini } 2693674ae819SStefano Zampini } 26944641a718SStefano Zampini ierr = PetscBTSet(touched,row_cmat_indices[min_loc]);CHKERRQ(ierr); 2695674ae819SStefano Zampini constraints_index[n++] = row_cmat_indices[min_loc]; 2696674ae819SStefano Zampini } 2697674ae819SStefano Zampini ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr); 2698674ae819SStefano Zampini } 26994641a718SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 2700674ae819SStefano Zampini ierr = PetscFree(row_cmat_global_indices);CHKERRQ(ierr); 2701811e8ca2SStefano Zampini } 2702811e8ca2SStefano Zampini } 2703674ae819SStefano Zampini *n_constraints = n; 2704811e8ca2SStefano Zampini if (constraints_idx) *constraints_idx = constraints_index; 2705674ae819SStefano Zampini PetscFunctionReturn(0); 2706674ae819SStefano Zampini } 2707674ae819SStefano Zampini 2708674ae819SStefano Zampini #undef __FUNCT__ 2709674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 2710674ae819SStefano 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[]) 2711674ae819SStefano Zampini { 2712674ae819SStefano Zampini Vec local_vec,global_vec; 2713674ae819SStefano Zampini IS seqis,paris; 2714674ae819SStefano Zampini VecScatter scatter_ctx; 2715674ae819SStefano Zampini PetscScalar *array; 2716674ae819SStefano Zampini PetscInt *temp_global_dofs; 2717674ae819SStefano Zampini PetscScalar globalsum; 2718674ae819SStefano Zampini PetscInt i,j,s; 2719674ae819SStefano Zampini PetscInt nlocals,first_index,old_index,max_local; 2720674ae819SStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm,max_global; 2721674ae819SStefano Zampini PetscMPIInt *dof_sizes,*dof_displs; 2722674ae819SStefano Zampini PetscBool first_found; 2723674ae819SStefano Zampini PetscErrorCode ierr; 2724674ae819SStefano Zampini 2725674ae819SStefano Zampini PetscFunctionBegin; 2726674ae819SStefano Zampini /* mpi buffers */ 2727b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 2728b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 2729674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 2730785e854fSJed Brown ierr = PetscMalloc1(j,&dof_sizes);CHKERRQ(ierr); 2731785e854fSJed Brown ierr = PetscMalloc1(j,&dof_displs);CHKERRQ(ierr); 2732674ae819SStefano Zampini /* get maximum size of subset */ 2733785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 2734674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 2735674ae819SStefano Zampini max_local = 0; 2736b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2737674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 2738674ae819SStefano Zampini max_local = temp_global_dofs[i]; 2739674ae819SStefano Zampini } 2740674ae819SStefano Zampini } 2741b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 2742674ae819SStefano Zampini max_global++; 2743674ae819SStefano Zampini max_local = 0; 2744b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2745674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 2746674ae819SStefano Zampini max_local = local_dofs[i]; 2747674ae819SStefano Zampini } 2748674ae819SStefano Zampini } 2749674ae819SStefano Zampini max_local++; 2750674ae819SStefano Zampini /* allocate workspace */ 2751674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 2752674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 2753674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 2754674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 2755674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 2756674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 2757674ae819SStefano Zampini /* create scatter */ 2758674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 2759674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 2760674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 2761674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 2762674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 2763674ae819SStefano Zampini /* init array */ 2764674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 2765674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 2766674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 2767674ae819SStefano Zampini if (local_dofs_mult) { 2768674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2769674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 2770674ae819SStefano Zampini } 2771674ae819SStefano Zampini } else { 2772674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 2773674ae819SStefano Zampini array[local_dofs[i]]=1.0; 2774674ae819SStefano Zampini } 2775674ae819SStefano Zampini } 2776674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 2777674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 2778674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2779674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2780674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 27815b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 2782674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 2783674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 2784674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 2785674ae819SStefano Zampini nlocals = 0; 2786674ae819SStefano Zampini first_index = -1; 2787674ae819SStefano Zampini first_found = PETSC_FALSE; 2788674ae819SStefano Zampini for (i=0;i<s;i++) { 2789b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 2790674ae819SStefano Zampini first_found = PETSC_TRUE; 2791674ae819SStefano Zampini first_index = i; 2792674ae819SStefano Zampini } 27935b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 2794674ae819SStefano Zampini } 2795674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 2796674ae819SStefano Zampini if (!rank_prec_comm) { 2797674ae819SStefano Zampini dof_displs[0]=0; 2798674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 2799674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 2800674ae819SStefano Zampini } 2801674ae819SStefano Zampini } 2802674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 2803674ae819SStefano Zampini if (first_found) { 2804674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 2805674ae819SStefano Zampini old_index = first_index; 2806674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 2807b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 2808674ae819SStefano Zampini array[i] += array[old_index]; 2809674ae819SStefano Zampini old_index = i; 2810674ae819SStefano Zampini } 2811674ae819SStefano Zampini } 2812674ae819SStefano Zampini } 2813674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 2814674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 2815674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2816674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2817674ae819SStefano Zampini /* get global ordering of local dofs */ 2818674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 2819674ae819SStefano Zampini if (local_dofs_mult) { 2820674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 28215b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 2822674ae819SStefano Zampini } 2823674ae819SStefano Zampini } else { 2824674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 28255b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 2826674ae819SStefano Zampini } 2827674ae819SStefano Zampini } 2828674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 2829674ae819SStefano Zampini /* free workspace */ 2830674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 2831674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 2832674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 2833674ae819SStefano Zampini ierr = PetscFree(dof_sizes);CHKERRQ(ierr); 2834674ae819SStefano Zampini ierr = PetscFree(dof_displs);CHKERRQ(ierr); 2835674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 2836674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 2837674ae819SStefano Zampini PetscFunctionReturn(0); 2838674ae819SStefano Zampini } 28399a7d3425SStefano Zampini 28409a7d3425SStefano Zampini #undef __FUNCT__ 28419a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 28429a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 28439a7d3425SStefano Zampini { 28449a7d3425SStefano Zampini PetscInt i,j; 28459a7d3425SStefano Zampini PetscScalar *alphas; 28469a7d3425SStefano Zampini PetscErrorCode ierr; 28479a7d3425SStefano Zampini 28489a7d3425SStefano Zampini PetscFunctionBegin; 28499a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 2850785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 28519a7d3425SStefano Zampini for (i=0;i<n;i++) { 28529a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 28539a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 28549a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 28559a7d3425SStefano Zampini } 28569a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 28579a7d3425SStefano Zampini PetscFunctionReturn(0); 28589a7d3425SStefano Zampini } 28599a7d3425SStefano Zampini 2860e7931f94SStefano Zampini #undef __FUNCT__ 286170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 286228143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 2863e7931f94SStefano Zampini { 2864e7931f94SStefano Zampini Mat subdomain_adj; 2865e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 2866e7931f94SStefano Zampini MatPartitioning partitioner; 2867e7931f94SStefano Zampini Mat_IS *matis; 2868e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 2869e7931f94SStefano Zampini PetscInt prank; 2870e7931f94SStefano Zampini PetscMPIInt size,rank,color; 2871e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 2872e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 28733837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 2874e7931f94SStefano Zampini PetscErrorCode ierr; 28752b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 2876e7931f94SStefano Zampini PetscSubcomm subcomm; 2877a57a6d2fSStefano Zampini 2878e7931f94SStefano Zampini PetscFunctionBegin; 28792b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 28802b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 28812b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 2882e7931f94SStefano Zampini 2883e7931f94SStefano Zampini /* Get info on mapping */ 2884e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 2885e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 2886e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 2887e7931f94SStefano Zampini 2888e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 2889785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 2890e7931f94SStefano Zampini xadj[0] = 0; 2891e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 2892785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 2893785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 2894e7931f94SStefano Zampini 28952b510759SStefano Zampini if (threshold) { 2896d023bfaeSStefano Zampini PetscInt xadj_count = 0; 28972b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 2898d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 2899d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 2900d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 2901d023bfaeSStefano Zampini xadj_count++; 2902e7931f94SStefano Zampini } 2903e7931f94SStefano Zampini } 2904d023bfaeSStefano Zampini xadj[1] = xadj_count; 2905c8587f34SStefano Zampini } else { 2906e7931f94SStefano Zampini if (xadj[1]) { 2907e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 2908e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 2909c8587f34SStefano Zampini } 2910e7931f94SStefano Zampini } 2911e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 2912e7931f94SStefano Zampini if (use_square) { 2913e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 2914e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 2915e7931f94SStefano Zampini } 2916e7931f94SStefano Zampini } 2917e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 2918e7931f94SStefano Zampini 29193837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 2920e7931f94SStefano Zampini 2921e7931f94SStefano Zampini /* 2922e7931f94SStefano Zampini Restrict work on active processes only. 2923e7931f94SStefano Zampini */ 2924e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 2925e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 2926e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 29272b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 2928d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 2929e7931f94SStefano Zampini if (color) { 2930e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 2931e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 2932e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 2933c8587f34SStefano Zampini } else { 293428143c3dSStefano Zampini PetscInt coarsening_ratio; 2935e7931f94SStefano Zampini ierr = MPI_Comm_size(subcomm->comm,&size);CHKERRQ(ierr); 2936785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 2937e7931f94SStefano Zampini prank = rank; 2938e7931f94SStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm->comm);CHKERRQ(ierr); 29398002ef2cSStefano Zampini /* 2940e7931f94SStefano Zampini for (i=0;i<size;i++) { 2941e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 2942c8587f34SStefano Zampini } 29438002ef2cSStefano Zampini */ 2944e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 2945e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 2946c8587f34SStefano Zampini } 2947e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 2948e7931f94SStefano Zampini ierr = MatCreateMPIAdj(subcomm->comm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 294922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 2950e7931f94SStefano Zampini 2951e7931f94SStefano Zampini /* Partition */ 2952e7931f94SStefano Zampini ierr = MatPartitioningCreate(subcomm->comm,&partitioner);CHKERRQ(ierr); 2953e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 2954e7931f94SStefano Zampini if (use_vwgt) { 29553837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 2956e7931f94SStefano Zampini v_wgt[0] = local_size; 2957e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 2958c8587f34SStefano Zampini } 295928143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 296028143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 296128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 2962e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 2963e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 296422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 2965e7931f94SStefano Zampini 2966e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 296728143c3dSStefano Zampini if (contiguous) { 296828143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 296928143c3dSStefano Zampini } else { 297084ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 297128143c3dSStefano Zampini } 2972e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2973e7931f94SStefano Zampini /* clean up */ 2974e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 2975e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 2976e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 2977e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 2978e7931f94SStefano Zampini } 2979e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 2980e7931f94SStefano Zampini 2981e7931f94SStefano Zampini /* assemble parallel IS for sends */ 2982e7931f94SStefano Zampini i = 1; 2983e7931f94SStefano Zampini if (color) i=0; 2984e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 2985e7931f94SStefano Zampini 2986e7931f94SStefano Zampini /* get back IS */ 2987e7931f94SStefano Zampini *is_sends = ranks_send_to; 2988e7931f94SStefano Zampini PetscFunctionReturn(0); 2989e7931f94SStefano Zampini } 2990e7931f94SStefano Zampini 2991e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 2992e7931f94SStefano Zampini 2993e7931f94SStefano Zampini #undef __FUNCT__ 2994e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 299528143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 2996e7931f94SStefano Zampini { 299770cf5478SStefano Zampini Mat local_mat; 2998e7931f94SStefano Zampini Mat_IS *matis; 2999e7931f94SStefano Zampini IS is_sends_internal; 30009d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 300128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 30029d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3003e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3004e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3005e7931f94SStefano Zampini const PetscInt* is_indices; 3006e7931f94SStefano Zampini MatType new_local_type; 3007e7931f94SStefano Zampini /* buffers */ 3008e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 300928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 30109d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3011e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3012e7931f94SStefano Zampini /* MPI */ 301328143c3dSStefano Zampini MPI_Comm comm,comm_n; 301428143c3dSStefano Zampini PetscSubcomm subcomm; 3015e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 301628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 301728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 301828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 301928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 302028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3021e7931f94SStefano Zampini PetscErrorCode ierr; 3022e7931f94SStefano Zampini 3023e7931f94SStefano Zampini PetscFunctionBegin; 302428143c3dSStefano Zampini /* TODO: add missing checks */ 302528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 302628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 302728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 302828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3029e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 303028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3031e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3032e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3033e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3034e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3035e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 303628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 303770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 303870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 303928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 304070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 304170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 304270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 304370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 304470cf5478SStefano Zampini } 3045e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3046e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3047e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3048e7931f94SStefano Zampini if (!is_sends) { 304928143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 305028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 305128143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3052c8587f34SStefano Zampini } else { 3053e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3054e7931f94SStefano Zampini is_sends_internal = is_sends; 3055c8587f34SStefano Zampini } 3056e7931f94SStefano Zampini 3057e7931f94SStefano Zampini /* get pointer of MATIS data */ 3058e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3059e7931f94SStefano Zampini 3060e7931f94SStefano Zampini /* get comm */ 3061a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3062e7931f94SStefano Zampini 3063e7931f94SStefano Zampini /* compute number of sends */ 3064e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3065e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3066e7931f94SStefano Zampini 3067e7931f94SStefano Zampini /* compute number of receives */ 3068e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3069785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3070e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3071e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3072e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3073e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3074e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3075e7931f94SStefano Zampini 307628143c3dSStefano Zampini /* restrict comm if requested */ 307728143c3dSStefano Zampini subcomm = 0; 307828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 307928143c3dSStefano Zampini if (restrict_comm) { 308028143c3dSStefano Zampini PetscMPIInt color,rank,subcommsize; 308128143c3dSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 308228143c3dSStefano Zampini color = 0; 308328143c3dSStefano Zampini if (n_sends && !n_recvs) color = 1; /* sending only processes will not partecipate in new comm */ 308428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 308528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 308628143c3dSStefano Zampini /* check if reuse has been requested */ 308728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 308828143c3dSStefano Zampini if (*mat_n) { 308928143c3dSStefano Zampini PetscMPIInt subcommsize2; 309028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 309128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 309228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 309328143c3dSStefano Zampini } else { 309428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 309528143c3dSStefano Zampini } 309628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 309728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 309828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 309928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 310028143c3dSStefano Zampini comm_n = subcomm->comm; 310128143c3dSStefano Zampini } 310228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 310328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 310428143c3dSStefano Zampini } else { 310528143c3dSStefano Zampini comm_n = comm; 310628143c3dSStefano Zampini } 310728143c3dSStefano Zampini 3108e7931f94SStefano Zampini /* prepare send/receive buffers */ 3109785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3110e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3111785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3112e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 311328143c3dSStefano Zampini if (nis) { 3114854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 311528143c3dSStefano Zampini } 3116e7931f94SStefano Zampini 311728143c3dSStefano Zampini /* Get data from local matrices */ 3118e7931f94SStefano Zampini if (!isdense) { 3119a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3120e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3121e7931f94SStefano Zampini /* 3122e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3123e7931f94SStefano Zampini send_buffer_idxs should contain: 3124e7931f94SStefano Zampini - MatType_PRIVATE type 3125e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3126e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3127e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3128e7931f94SStefano Zampini */ 3129e7931f94SStefano Zampini } else { 3130e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3131e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3132854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3133e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3134e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3135e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3136e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3137e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3138e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3139e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3140e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3141e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3142c8587f34SStefano Zampini } 3143c8587f34SStefano Zampini } 3144e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 314528143c3dSStefano Zampini /* additional is (if any) */ 314628143c3dSStefano Zampini if (nis) { 314728143c3dSStefano Zampini PetscMPIInt psum; 314828143c3dSStefano Zampini PetscInt j; 314928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 315028143c3dSStefano Zampini PetscInt plen; 315128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 315228143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 315328143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 315428143c3dSStefano Zampini } 3155854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 315628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 315728143c3dSStefano Zampini PetscInt plen; 315828143c3dSStefano Zampini const PetscInt *is_array_idxs; 315928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 316028143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 316128143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 316228143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 316328143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 316428143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 316528143c3dSStefano Zampini } 316628143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 316728143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 316828143c3dSStefano Zampini } 316928143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 317028143c3dSStefano Zampini } 317128143c3dSStefano Zampini 3172e7931f94SStefano Zampini buf_size_idxs = 0; 3173e7931f94SStefano Zampini buf_size_vals = 0; 317428143c3dSStefano Zampini buf_size_idxs_is = 0; 3175e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3176e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3177e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 317828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3179e7931f94SStefano Zampini } 3180785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3181785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 318295ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3183e7931f94SStefano Zampini 3184e7931f94SStefano Zampini /* get new tags for clean communications */ 3185e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3186e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 318728143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3188e7931f94SStefano Zampini 3189e7931f94SStefano Zampini /* allocate for requests */ 3190785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3191785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 319295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3193785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3194785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 319595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3196e7931f94SStefano Zampini 3197e7931f94SStefano Zampini /* communications */ 3198e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3199e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 320028143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3201e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3202e7931f94SStefano Zampini source_dest = onodes[i]; 3203e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3204e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3205e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3206e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 320728143c3dSStefano Zampini if (nis) { 320828143c3dSStefano 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); 320928143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 321028143c3dSStefano Zampini } 3211e7931f94SStefano Zampini } 3212e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3213e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3214e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3215e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 321628143c3dSStefano Zampini if (nis) { 321728143c3dSStefano 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); 321828143c3dSStefano Zampini } 3219e7931f94SStefano Zampini } 3220e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3221e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3222e7931f94SStefano Zampini 3223e7931f94SStefano Zampini /* assemble new l2g map */ 3224e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3225e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 32269d30be91SStefano Zampini new_local_rows = 0; 3227e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 32289d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3229e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3230e7931f94SStefano Zampini } 32319d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3232e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 32339d30be91SStefano Zampini new_local_rows = 0; 3234e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 32359d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 32369d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3237e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3238e7931f94SStefano Zampini } 32399d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 32409d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3241e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3242e7931f94SStefano Zampini 3243e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3244e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3245e7931f94SStefano 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) */ 3246e7931f94SStefano Zampini if (n_recvs) { 324728143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3248e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3249e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3250e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3251e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3252e7931f94SStefano Zampini break; 3253e7931f94SStefano Zampini } 3254e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3255e7931f94SStefano Zampini } 3256e7931f94SStefano Zampini switch (new_local_type_private) { 325728143c3dSStefano Zampini case MATDENSE_PRIVATE: 325828143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3259e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3260e7931f94SStefano Zampini bs = 1; 326128143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 326228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 326328143c3dSStefano Zampini bs = 1; 326428143c3dSStefano Zampini } 3265e7931f94SStefano Zampini break; 3266e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3267e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3268e7931f94SStefano Zampini bs = 1; 3269e7931f94SStefano Zampini break; 3270e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3271e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3272e7931f94SStefano Zampini break; 3273e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3274e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3275e7931f94SStefano Zampini break; 3276e7931f94SStefano Zampini default: 32779d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3278e7931f94SStefano Zampini break; 3279e7931f94SStefano Zampini } 328028143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 328128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 328228143c3dSStefano Zampini bs = 1; 3283e7931f94SStefano Zampini } 3284e7931f94SStefano Zampini 328570cf5478SStefano Zampini /* create MATIS object if needed */ 328670cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3287e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 328828143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 328970cf5478SStefano Zampini } else { 329070cf5478SStefano Zampini /* it also destroys the local matrices */ 329170cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 329270cf5478SStefano Zampini } 329370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3294e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 32959d30be91SStefano Zampini 32969d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 32979d30be91SStefano Zampini 32989d30be91SStefano Zampini /* Global to local map of received indices */ 32999d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 33009d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 33019d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 33029d30be91SStefano Zampini 33039d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 33049d30be91SStefano Zampini buf_size_idxs = 0; 33059d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 33069d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 33079d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 33089d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 33099d30be91SStefano Zampini } 33109d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 33119d30be91SStefano Zampini 33129d30be91SStefano Zampini /* set preallocation */ 33139d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 33149d30be91SStefano Zampini if (!newisdense) { 33159d30be91SStefano Zampini PetscInt *new_local_nnz=0; 33169d30be91SStefano Zampini 33179d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 33189d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 33199d30be91SStefano Zampini if (n_recvs) { 33209d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 33219d30be91SStefano Zampini } 33229d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 33239d30be91SStefano Zampini PetscInt j; 33249d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 33259d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 33269d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 33279d30be91SStefano Zampini } 33289d30be91SStefano Zampini } else { 33299d30be91SStefano Zampini /* TODO */ 33309d30be91SStefano Zampini } 33319d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 33329d30be91SStefano Zampini } 33339d30be91SStefano Zampini if (new_local_nnz) { 33349d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 33359d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 33369d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 33379d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 33389d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 33399d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 33409d30be91SStefano Zampini } else { 33419d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 33429d30be91SStefano Zampini } 33439d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 33449d30be91SStefano Zampini } else { 33459d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 33469d30be91SStefano Zampini } 3347e7931f94SStefano Zampini 3348e7931f94SStefano Zampini /* set values */ 3349e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 33509d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3351e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3352e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3353e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 33549d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3355e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3356e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3357e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 335828143c3dSStefano Zampini } else { 335928143c3dSStefano Zampini /* TODO */ 3360e7931f94SStefano Zampini } 3361e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3362e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3363e7931f94SStefano Zampini } 3364e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3365e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 336670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 336770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 33689d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 33699d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3370e7931f94SStefano Zampini 3371dfd14d43SStefano Zampini #if 0 337228143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3373e7931f94SStefano Zampini Vec lvec,rvec; 3374e7931f94SStefano Zampini PetscReal infty_error; 3375e7931f94SStefano Zampini 33762a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3377e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3378e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3379e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 338070cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3381e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3382e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3383e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3384e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3385e7931f94SStefano Zampini } 338628143c3dSStefano Zampini #endif 3387e7931f94SStefano Zampini 338828143c3dSStefano Zampini /* assemble new additional is (if any) */ 338928143c3dSStefano Zampini if (nis) { 339028143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 339128143c3dSStefano Zampini 339228143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3393854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 339428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 339528143c3dSStefano Zampini psum = 0; 339628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 339728143c3dSStefano Zampini for (j=0;j<nis;j++) { 339828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 339928143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 340028143c3dSStefano Zampini psum += plen; 340128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 340228143c3dSStefano Zampini } 340328143c3dSStefano Zampini } 3404854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3405854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 340628143c3dSStefano Zampini for (i=1;i<nis;i++) { 340728143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 340828143c3dSStefano Zampini } 340928143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 341028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 341128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 341228143c3dSStefano Zampini for (j=0;j<nis;j++) { 341328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 341428143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 341528143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 341628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 341728143c3dSStefano Zampini } 341828143c3dSStefano Zampini } 341928143c3dSStefano Zampini for (i=0;i<nis;i++) { 342028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 342128143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 342228143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 342328143c3dSStefano Zampini } 342428143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 342528143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 342628143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 342728143c3dSStefano Zampini } 3428e7931f94SStefano Zampini /* free workspace */ 342928143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3431e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3432e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3433e7931f94SStefano Zampini if (isdense) { 3434e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3435e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3436e7931f94SStefano Zampini } else { 3437e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3438e7931f94SStefano Zampini } 343928143c3dSStefano Zampini if (nis) { 344028143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 344128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 344228143c3dSStefano Zampini } 3443e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3444e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 344528143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3446e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3447e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 344828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3449e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3450e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3451e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3452e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3453e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 345428143c3dSStefano Zampini if (nis) { 345528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 345628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 345728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 345828143c3dSStefano Zampini } 345928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 346028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 346128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 346228143c3dSStefano Zampini for (i=0;i<nis;i++) { 346328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 346428143c3dSStefano Zampini } 346528143c3dSStefano Zampini } 3466e7931f94SStefano Zampini PetscFunctionReturn(0); 3467e7931f94SStefano Zampini } 3468a57a6d2fSStefano Zampini 346912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 347012edc857SStefano Zampini #include <petsc-private/kspimpl.h> 347112edc857SStefano Zampini 3472c8587f34SStefano Zampini #undef __FUNCT__ 3473c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3474c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3475c8587f34SStefano Zampini { 3476c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3477c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 347820a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 34799881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 348020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 34816e683305SStefano Zampini IS coarse_is,*isarray; 34826e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 34836e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3484f9eb5b7dSStefano Zampini PC pc_temp; 3485c8587f34SStefano Zampini PCType coarse_pc_type; 3486c8587f34SStefano Zampini KSPType coarse_ksp_type; 3487f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 34884f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 34896e683305SStefano Zampini Mat t_coarse_mat_is; 34906e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 34916e683305SStefano Zampini PetscMPIInt all_procs; 349274e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 349368457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 349422bc73bbSStefano Zampini PetscScalar *array; 34959881197aSStefano Zampini PetscErrorCode ierr; 3496fdc09c96SStefano Zampini 3497c8587f34SStefano Zampini PetscFunctionBegin; 3498c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 349968457ee5SStefano 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 */ 350068457ee5SStefano Zampini compute_vecs = PETSC_TRUE; 3501fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 3502fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3503f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3504f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3505f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3506fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3507fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3508727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3509fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3510fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3511fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3512f4ddd8eeSStefano Zampini } 3513fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3514fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3515f4ddd8eeSStefano Zampini } 351670cf5478SStefano Zampini /* reset any subassembling information */ 351770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 35186e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 35196e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3520fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3521f4ddd8eeSStefano Zampini } 3522c8587f34SStefano Zampini 35236e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 35242b510759SStefano Zampini im_active = !!(pcis->n); 35252b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 35266e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 35276e683305SStefano Zampini void_procs = all_procs-active_procs; 35286e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 352974e2c79eSStefano Zampini redist = PETSC_FALSE; 353022bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 35316e683305SStefano Zampini csin_ml = PETSC_TRUE; 35326e683305SStefano Zampini ncoarse_ml = void_procs; 35336e683305SStefano Zampini csin_ds = PETSC_TRUE; 35346e683305SStefano Zampini ncoarse_ds = void_procs; 35356e683305SStefano Zampini } else { 35366e683305SStefano Zampini csin_ml = PETSC_FALSE; 35376e683305SStefano Zampini ncoarse_ml = all_procs; 35386e683305SStefano Zampini if (void_procs) { 35396e683305SStefano Zampini csin_ds = PETSC_TRUE; 35406e683305SStefano Zampini ncoarse_ds = void_procs; 35416e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 35426e683305SStefano Zampini } else { 354374e2c79eSStefano Zampini if (pcbddc->redistribute_coarse && pcbddc->redistribute_coarse < all_procs) { 354474e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 354574e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 354674e2c79eSStefano Zampini redist = PETSC_TRUE; 354774e2c79eSStefano Zampini } else { 35486e683305SStefano Zampini csin_ds = PETSC_FALSE; 35496e683305SStefano Zampini ncoarse_ds = all_procs; 35506e683305SStefano Zampini } 35516e683305SStefano Zampini } 355274e2c79eSStefano Zampini } 35536e683305SStefano Zampini 35546e683305SStefano Zampini /* 35556e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 35566e683305SStefano Zampini - we have not exceeded the number of levels requested 35576e683305SStefano Zampini - we can actually subassemble the active processes 35586e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 35596e683305SStefano Zampini */ 35606e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 35616e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 35626e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 35636e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 35646e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 3565f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 35662b510759SStefano Zampini } else { 3567f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 3568c8587f34SStefano Zampini } 3569c8587f34SStefano Zampini } 35706e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 35716e683305SStefano Zampini if (multilevel_allowed) { 35726e683305SStefano Zampini ncoarse = ncoarse_ml; 35736e683305SStefano Zampini csin = csin_ml; 35746e683305SStefano Zampini } else { 35756e683305SStefano Zampini ncoarse = ncoarse_ds; 35766e683305SStefano Zampini csin = csin_ds; 35776e683305SStefano Zampini } 3578e7931f94SStefano Zampini 3579abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 3580abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 3581abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 3582abbbba34SStefano Zampini 3583abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 358422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 358522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 358622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 358722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 3588b9b85e73SStefano Zampini #if 0 3589b9b85e73SStefano Zampini { 3590b9b85e73SStefano Zampini PetscViewer viewer; 3591b9b85e73SStefano Zampini char filename[256]; 3592b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 3593b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3594b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3595b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 3596b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3597b9b85e73SStefano Zampini } 3598b9b85e73SStefano Zampini #endif 35996e683305SStefano 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); 36006e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 36016e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 36026e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3603abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 3604abbbba34SStefano Zampini 36056e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 36066e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 36076e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 36086e683305SStefano Zampini const PetscInt *idxs; 36096e683305SStefano Zampini ISLocalToGlobalMapping tmap; 36106e683305SStefano Zampini 36116e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 36120be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 36136e683305SStefano Zampini /* allocate space for temporary storage */ 3614854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 3615854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 36166e683305SStefano Zampini /* allocate for IS array */ 36176e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 36186e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 36196e683305SStefano Zampini nis = nisdofs + nisneu; 3620854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 36216e683305SStefano Zampini /* dofs splitting */ 36226e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 36236e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 36246e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 36256e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 36266e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 36276e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 36286e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 36296e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 36306e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 36316e683305SStefano Zampini } 36326e683305SStefano Zampini /* neumann boundaries */ 36336e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 36346e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 36356e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 36366e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 36376e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 36386e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 36396e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 36406e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 36416e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 36426e683305SStefano Zampini } 36436e683305SStefano Zampini /* free memory */ 36446e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 36456e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 36466e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 36476e683305SStefano Zampini } else { 36486e683305SStefano Zampini nis = 0; 36496e683305SStefano Zampini nisdofs = 0; 36506e683305SStefano Zampini nisneu = 0; 36516e683305SStefano Zampini isarray = NULL; 36526e683305SStefano Zampini } 36536e683305SStefano Zampini /* destroy no longer needed map */ 36546e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 36556e683305SStefano Zampini 36566e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 36576e683305SStefano Zampini coarse_mat_is = NULL; 36586e683305SStefano Zampini if (csin) { 36596e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 366074e2c79eSStefano Zampini if (redist) { 366174e2c79eSStefano Zampini PetscMPIInt rank; 366274e2c79eSStefano Zampini PetscInt spc,n_spc_p1,dest[1]; 366374e2c79eSStefano Zampini 366474e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 366574e2c79eSStefano Zampini spc = all_procs/pcbddc->redistribute_coarse; 366674e2c79eSStefano Zampini n_spc_p1 = all_procs%pcbddc->redistribute_coarse; 366774e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 366874e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 366974e2c79eSStefano Zampini } else { 367074e2c79eSStefano Zampini dest[0] = rank/(spc+1); 367174e2c79eSStefano Zampini } 367274e2c79eSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),1,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 367374e2c79eSStefano Zampini } else { 36746e683305SStefano Zampini PetscInt j,tissize,*nisindices; 36756e683305SStefano Zampini PetscInt *coarse_candidates; 36766e683305SStefano Zampini const PetscInt* tisindices; 36776e683305SStefano Zampini /* get coarse candidates' ranks in pc communicator */ 3678854ce69bSBarry Smith ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 36796e683305SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 36806e683305SStefano Zampini for (i=0,j=0;i<all_procs;i++) { 36816e683305SStefano Zampini if (!coarse_candidates[i]) { 36826e683305SStefano Zampini coarse_candidates[j]=i; 36836e683305SStefano Zampini j++; 36846e683305SStefano Zampini } 36856e683305SStefano Zampini } 36866e683305SStefano Zampini if (j < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",j,ncoarse); 36876e683305SStefano Zampini /* get a suitable subassembling pattern */ 36886e683305SStefano Zampini if (csin_type_simple) { 36896e683305SStefano Zampini PetscMPIInt rank; 36906e683305SStefano Zampini PetscInt issize,isidx; 36916e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 36926e683305SStefano Zampini if (im_active) { 36936e683305SStefano Zampini issize = 1; 36946e683305SStefano Zampini isidx = (PetscInt)rank; 36956e683305SStefano Zampini } else { 36966e683305SStefano Zampini issize = 0; 36976e683305SStefano Zampini isidx = -1; 36986e683305SStefano Zampini } 36996e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 37006e683305SStefano Zampini } else { 37016e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 37026e683305SStefano Zampini } 37036e683305SStefano Zampini if (pcbddc->dbg_flag) { 37046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 37066e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 37076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 37086e683305SStefano Zampini for (i=0;i<j;i++) { 37096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 37106e683305SStefano Zampini } 37116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 37126e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37136e683305SStefano Zampini } 37146e683305SStefano Zampini /* shift the pattern on coarse candidates */ 37156e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 37166e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 3717854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 37186e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 37196e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 37206e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 37216e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 37226e683305SStefano Zampini } 372374e2c79eSStefano Zampini } 37246e683305SStefano Zampini if (pcbddc->dbg_flag) { 37256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 37276e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 37286e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37296e683305SStefano Zampini } 37306e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 37316e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 37326e683305SStefano Zampini } else { 37336e683305SStefano Zampini if (pcbddc->dbg_flag) { 37346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 37366e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37376e683305SStefano Zampini } 37386e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 37396e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 37406e683305SStefano Zampini } 37416e683305SStefano Zampini 37426e683305SStefano Zampini /* create local to global scatters for coarse problem */ 374368457ee5SStefano Zampini if (compute_vecs) { 37446e683305SStefano Zampini PetscInt lrows; 37456e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 37466e683305SStefano Zampini if (coarse_mat_is) { 37476e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 37486e683305SStefano Zampini } else { 37496e683305SStefano Zampini lrows = 0; 37506e683305SStefano Zampini } 37516e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 37526e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 37536e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 37546e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 37556e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 37566e683305SStefano Zampini } 37576e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 37586e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 3759c8587f34SStefano Zampini 3760f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 3761f9eb5b7dSStefano Zampini if (multilevel_allowed) { 3762f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 3763f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 3764f9eb5b7dSStefano Zampini } else { 3765f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 3766f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 3767c8587f34SStefano Zampini } 3768c8587f34SStefano Zampini 37696e683305SStefano Zampini /* print some info if requested */ 37706e683305SStefano Zampini if (pcbddc->dbg_flag) { 37716e683305SStefano Zampini if (!multilevel_allowed) { 37726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37736e683305SStefano Zampini if (multilevel_requested) { 37746e683305SStefano 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); 37756e683305SStefano Zampini } else if (pcbddc->max_levels) { 37766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 37776e683305SStefano Zampini } 37786e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 37796e683305SStefano Zampini } 37806e683305SStefano Zampini } 37816e683305SStefano Zampini 3782f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 37836e683305SStefano Zampini if (coarse_mat_is) { 37846e683305SStefano Zampini MatReuse coarse_mat_reuse; 37856a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 37866e683305SStefano Zampini if (pcbddc->dbg_flag) { 37876e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 37886e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 37896e683305SStefano Zampini } 3790f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 3791312be037SStefano Zampini char prefix[256],str_level[16]; 3792e604994aSStefano Zampini size_t len; 37936e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 3794c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 3795f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 37965f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 3797c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 37986e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 3799c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 3800c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 3801e604994aSStefano Zampini /* prefix */ 3802e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 3803e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3804e604994aSStefano Zampini if (!pcbddc->current_level) { 3805e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3806e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 3807c8587f34SStefano Zampini } else { 3808e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3809312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3810312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 381134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3812312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3813e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 3814e604994aSStefano Zampini } 3815e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 3816f9eb5b7dSStefano Zampini /* allow user customization */ 3817f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 38187e0def11SStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3819312be037SStefano Zampini } 3820f9eb5b7dSStefano Zampini 3821f9eb5b7dSStefano Zampini /* get some info after set from options */ 3822f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 3823f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 3824f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 38254f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 38266e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 3827f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 3828f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 3829f9eb5b7dSStefano Zampini } 38304f3a063dSStefano Zampini if (isredundant) { 38314f3a063dSStefano Zampini KSP inner_ksp; 38324f3a063dSStefano Zampini PC inner_pc; 38334f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 38344f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 38354f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 38364f3a063dSStefano Zampini } 3837f9eb5b7dSStefano Zampini 38386e683305SStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 3839f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 3840f9eb5b7dSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 3841f9eb5b7dSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 38426e683305SStefano Zampini if (nisdofs) { 38436e683305SStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 38446e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 38456e683305SStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 38461035eff8SStefano Zampini } 38471035eff8SStefano Zampini } 38486e683305SStefano Zampini if (nisneu) { 38496e683305SStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 38506e683305SStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 38511035eff8SStefano Zampini } 3852fdc09c96SStefano Zampini 3853f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 3854fa7f1dd8SStefano Zampini if (coarse_reuse) { 385581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 3856fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 38576e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 3858fa7f1dd8SStefano Zampini } else { 38596e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 3860fa7f1dd8SStefano Zampini } 3861c8587f34SStefano Zampini if (isbddc || isnn) { 386222bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 386370cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 386428143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 386522b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 38666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 38676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 38686e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 38696e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 387022b6e8a2SStefano Zampini } 387170cf5478SStefano Zampini } 38726e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 387370cf5478SStefano Zampini } else { 387422bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 387522bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 387622bc73bbSStefano Zampini } 387722bc73bbSStefano Zampini } else { 38782e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 3879c8587f34SStefano Zampini } 3880c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 3881c8587f34SStefano Zampini 3882c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 3883b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 38845a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 3885c8587f34SStefano Zampini 38866e683305SStefano Zampini /* set operators */ 38875f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 38886e683305SStefano Zampini if (pcbddc->dbg_flag) { 38896e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 38906e683305SStefano Zampini } 38916e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 38926e683305SStefano Zampini coarse_mat = 0; 38936e683305SStefano Zampini } 38946e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 3895b9b85e73SStefano Zampini #if 0 3896b9b85e73SStefano Zampini { 3897b9b85e73SStefano Zampini PetscViewer viewer; 3898b9b85e73SStefano Zampini char filename[256]; 3899b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 3900b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 3901b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3902b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 3903b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3904b9b85e73SStefano Zampini } 3905b9b85e73SStefano Zampini #endif 3906c8587f34SStefano Zampini 3907c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 3908c8587f34SStefano Zampini if (pcbddc->NullSpace) { 3909c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 391098a51de6SStefano Zampini } 391198a51de6SStefano Zampini 391298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 391398a51de6SStefano Zampini Vec crhs,csol; 391498a51de6SStefano Zampini PetscBool ispreonly; 391598a51de6SStefano Zampini if (CoarseNullSpace) { 3916c8587f34SStefano Zampini if (isbddc) { 3917c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 3918c8587f34SStefano Zampini } else { 3919c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 3920c8587f34SStefano Zampini } 3921c8587f34SStefano Zampini } 3922f9eb5b7dSStefano Zampini /* setup coarse ksp */ 3923f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 3924f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 3925f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 39266e683305SStefano Zampini /* hack */ 3927f347579bSStefano Zampini if (!csol) { 39282a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 3929f9eb5b7dSStefano Zampini } 3930f347579bSStefano Zampini if (!crhs) { 39312a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 3932f347579bSStefano Zampini } 3933cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 3934cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 39356e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 3936c8587f34SStefano Zampini KSP check_ksp; 39372b510759SStefano Zampini KSPType check_ksp_type; 3938c8587f34SStefano Zampini PC check_pc; 39396e683305SStefano Zampini Vec check_vec,coarse_vec; 39406a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 39412b510759SStefano Zampini PetscInt its; 39426e683305SStefano Zampini PetscBool compute_eigs; 39436e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 39446e683305SStefano Zampini PetscInt neigs; 39458e185a42SStefano Zampini const char *prefix; 3946c8587f34SStefano Zampini 39472b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 39486e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 394923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 3950f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 39512b510759SStefano Zampini if (ispreonly) { 39522b510759SStefano Zampini check_ksp_type = KSPPREONLY; 39536e683305SStefano Zampini compute_eigs = PETSC_FALSE; 39542b510759SStefano Zampini } else { 3955cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 39566e683305SStefano Zampini compute_eigs = PETSC_TRUE; 3957c8587f34SStefano Zampini } 3958c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 39596e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 39606e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 39616e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 3962a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 3963a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 3964a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 3965a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 3966c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 3967c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 3968c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 3969c8587f34SStefano Zampini /* create random vec */ 39706e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 39716e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 3972c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 3973c8587f34SStefano Zampini if (CoarseNullSpace) { 3974c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 3975c8587f34SStefano Zampini } 39766e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 3977c8587f34SStefano Zampini /* solve coarse problem */ 39786e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 3979c8587f34SStefano Zampini if (CoarseNullSpace) { 39806e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 3981c8587f34SStefano Zampini } 3982cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 39836e683305SStefano Zampini if (compute_eigs) { 3984854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 3985854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 39866e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 39876e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 39886e683305SStefano Zampini lambda_min = eigs_r[0]; 39896e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 39906e683305SStefano Zampini if (lambda_max>lambda_min) { 3991cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 3992cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 3993cbcc2c2aSStefano Zampini } 3994c8587f34SStefano Zampini } 3995c8587f34SStefano Zampini } 3996cbcc2c2aSStefano Zampini 3997c8587f34SStefano Zampini /* check coarse problem residual error */ 39986e683305SStefano Zampini if (pcbddc->dbg_flag) { 39996e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 40006e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 40016e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4002c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 40036e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 40046e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4005c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 40066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (%d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 40076e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 40086e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 40096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 40106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 40116e683305SStefano Zampini if (compute_eigs) { 40126e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4013deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4014c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 40156e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 40166e683305SStefano 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); 40176e683305SStefano Zampini for (i=0;i<neigs;i++) { 40186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4019c8587f34SStefano Zampini } 40206e683305SStefano Zampini } 40216e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 40226e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 40236e683305SStefano Zampini } 4024c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 40256e683305SStefano Zampini if (compute_eigs) { 40266e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 40276e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4028c8587f34SStefano Zampini } 40296e683305SStefano Zampini } 40306e683305SStefano Zampini } 4031cbcc2c2aSStefano Zampini /* print additional info */ 4032cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 40336e683305SStefano Zampini /* waits until all processes reaches this point */ 40346e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4035cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4036cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4037cbcc2c2aSStefano Zampini } 4038cbcc2c2aSStefano Zampini 40392b510759SStefano Zampini /* free memory */ 4040c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4041fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4042c8587f34SStefano Zampini PetscFunctionReturn(0); 4043c8587f34SStefano Zampini } 4044674ae819SStefano Zampini 4045f34684f1SStefano Zampini #undef __FUNCT__ 4046f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4047f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4048f34684f1SStefano Zampini { 4049f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4050f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4051f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4052727cdba6SStefano Zampini PetscInt i,coarse_size; 4053727cdba6SStefano Zampini PetscInt *local_primal_indices; 4054f34684f1SStefano Zampini PetscErrorCode ierr; 4055f34684f1SStefano Zampini 4056f34684f1SStefano Zampini PetscFunctionBegin; 4057f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 405889c96988SStefano Zampini if (!pcbddc->primal_indices_local_idxs && pcbddc->local_primal_size) { 405989c96988SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Local primal indices have not been created"); 4060727cdba6SStefano Zampini } 4061727cdba6SStefano 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); 4062f34684f1SStefano Zampini 4063f34684f1SStefano Zampini /* check numbering */ 4064f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4065f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4066b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4067f34684f1SStefano Zampini 4068f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4069f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4070f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 40710fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4072f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4073f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4074727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4075f34684f1SStefano Zampini } 4076f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4077f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4078f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4079f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4080f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4081f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4082f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4083f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4084f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4085f34684f1SStefano Zampini if (array[i] == 1.0) { 4086b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4087f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4088f34684f1SStefano Zampini } 4089f34684f1SStefano Zampini } 4090b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4091f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4092f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4093f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4094f34684f1SStefano Zampini } 4095f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4096f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4097f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4098f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4099f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4100f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4101b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4102f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4103f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4104f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4105f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4106727cdba6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%d]=%d (%d)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i]); 4107f34684f1SStefano Zampini } 4108f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4109f34684f1SStefano Zampini } 4110f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4111b9b85e73SStefano Zampini if (set_error_reduced) { 4112b9b85e73SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4113b9b85e73SStefano Zampini } 4114f34684f1SStefano Zampini } 4115f34684f1SStefano Zampini /* get back data */ 4116f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4117f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4118674ae819SStefano Zampini PetscFunctionReturn(0); 4119674ae819SStefano Zampini } 4120674ae819SStefano Zampini 4121e456f2a8SStefano Zampini #undef __FUNCT__ 4122e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4123a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4124e456f2a8SStefano Zampini { 4125e456f2a8SStefano Zampini IS localis_t; 4126a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4127e456f2a8SStefano Zampini PetscScalar *vals; 4128e456f2a8SStefano Zampini PetscErrorCode ierr; 4129e456f2a8SStefano Zampini 4130e456f2a8SStefano Zampini PetscFunctionBegin; 4131a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4132e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4133854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4134e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4135e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4136a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4137a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 41381035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4139a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 41401035eff8SStefano Zampini } 4141a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4142e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4143e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4144a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4145a7dc3881SStefano Zampini /* now compute set in local ordering */ 4146a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4147a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4148a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4149a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4150a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4151ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4152e456f2a8SStefano Zampini lsize++; 4153e456f2a8SStefano Zampini } 4154e456f2a8SStefano Zampini } 4155854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4156a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4157ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4158e456f2a8SStefano Zampini idxs[lsize++] = i; 4159e456f2a8SStefano Zampini } 4160e456f2a8SStefano Zampini } 4161a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4162a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4163e456f2a8SStefano Zampini *localis = localis_t; 4164e456f2a8SStefano Zampini PetscFunctionReturn(0); 4165e456f2a8SStefano Zampini } 4166906d46d4SStefano Zampini 4167906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4168906d46d4SStefano Zampini #undef __FUNCT__ 4169906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4170906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4171906d46d4SStefano Zampini { 4172906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4173906d46d4SStefano Zampini PetscErrorCode ierr; 4174906d46d4SStefano Zampini 4175906d46d4SStefano Zampini PetscFunctionBegin; 4176906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4177906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4178906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4179906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4180906d46d4SStefano Zampini PetscFunctionReturn(0); 4181906d46d4SStefano Zampini } 4182906d46d4SStefano Zampini 4183906d46d4SStefano Zampini #undef __FUNCT__ 4184906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4185906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4186906d46d4SStefano Zampini { 4187906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4188906d46d4SStefano Zampini PetscErrorCode ierr; 4189906d46d4SStefano Zampini 4190906d46d4SStefano Zampini PetscFunctionBegin; 4191906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4192906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4193906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4194906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4195906d46d4SStefano Zampini PetscFunctionReturn(0); 4196906d46d4SStefano Zampini } 4197