HYPRE_Int HYPRE_IJMatrixGetObject( HYPRE_IJMatrix matrix, void **object ) { hypre_IJMatrix *ijmatrix = (hypre_IJMatrix *) matrix; if (!ijmatrix) { hypre_error_in_arg(1); return hypre_error_flag; } *object = hypre_IJMatrixObject( ijmatrix ); return hypre_error_flag; }
/* Assume that we are given a fine and coarse topology and the coarse degrees of freedom (DOFs) have been chosen. Assume also, that the global interpolation matrix dof_DOF has a prescribed nonzero pattern. Then, the fine degrees of freedom can be split into 4 groups (here "i" stands for "interior"): NODEidof - dofs which are interpolated only from the DOF in one coarse vertex EDGEidof - dofs which are interpolated only from the DOFs in one coarse edge FACEidof - dofs which are interpolated only from the DOFs in one coarse face ELEMidof - dofs which are interpolated only from the DOFs in one coarse element The interpolation operator dof_DOF can be build in 4 steps, by consequently filling-in the rows corresponding to the above groups. The code below uses harmonic extension to extend the interpolation from one group to the next. */ HYPRE_Int hypre_ND1AMGeInterpolation (hypre_ParCSRMatrix * Aee, hypre_ParCSRMatrix * ELEM_idof, hypre_ParCSRMatrix * FACE_idof, hypre_ParCSRMatrix * EDGE_idof, hypre_ParCSRMatrix * ELEM_FACE, hypre_ParCSRMatrix * ELEM_EDGE, HYPRE_Int num_OffProcRows, hypre_MaxwellOffProcRow ** OffProcRows, hypre_IJMatrix * IJ_dof_DOF) { HYPRE_Int ierr = 0; HYPRE_Int i, j, k; HYPRE_Int *offproc_rnums, *swap; hypre_ParCSRMatrix * dof_DOF = hypre_IJMatrixObject(IJ_dof_DOF); hypre_ParCSRMatrix * ELEM_DOF = ELEM_EDGE; hypre_ParCSRMatrix * ELEM_FACEidof; hypre_ParCSRMatrix * ELEM_EDGEidof; hypre_CSRMatrix *A, *P; HYPRE_Int numELEM = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(ELEM_EDGE)); HYPRE_Int getrow_ierr; HYPRE_Int three_dimensional_problem; MPI_Comm comm= hypre_ParCSRMatrixComm(Aee); HYPRE_Int myproc; hypre_MPI_Comm_rank(comm, &myproc); #if 0 hypre_IJMatrix * ij_dof_DOF = hypre_CTAlloc(hypre_IJMatrix, 1); /* Convert dof_DOF to IJ matrix, so we can use AddToValues */ hypre_IJMatrixComm(ij_dof_DOF) = hypre_ParCSRMatrixComm(dof_DOF); hypre_IJMatrixRowPartitioning(ij_dof_DOF) = hypre_ParCSRMatrixRowStarts(dof_DOF); hypre_IJMatrixColPartitioning(ij_dof_DOF) = hypre_ParCSRMatrixColStarts(dof_DOF); hypre_IJMatrixObject(ij_dof_DOF) = dof_DOF; hypre_IJMatrixAssembleFlag(ij_dof_DOF) = 1; #endif /* sort the offproc rows to get quicker comparison for later */ if (num_OffProcRows) { offproc_rnums= hypre_TAlloc(HYPRE_Int, num_OffProcRows); swap = hypre_TAlloc(HYPRE_Int, num_OffProcRows); for (i= 0; i< num_OffProcRows; i++) { offproc_rnums[i]=(OffProcRows[i] -> row); swap[i] = i; } } if (num_OffProcRows > 1) { hypre_qsort2i(offproc_rnums, swap, 0, num_OffProcRows-1); } if (FACE_idof == EDGE_idof) three_dimensional_problem = 0; else three_dimensional_problem = 1; /* ELEM_FACEidof = ELEM_FACE x FACE_idof */ if (three_dimensional_problem) ELEM_FACEidof = hypre_ParMatmul(ELEM_FACE, FACE_idof); /* ELEM_EDGEidof = ELEM_EDGE x EDGE_idof */ ELEM_EDGEidof = hypre_ParMatmul(ELEM_EDGE, EDGE_idof); /* Loop over local coarse elements */ k = hypre_ParCSRMatrixFirstRowIndex(ELEM_EDGE); for (i = 0; i < numELEM; i++, k++) { HYPRE_Int size1, size2; HYPRE_Int *col_ind0, *col_ind1, *col_ind2; HYPRE_Int num_DOF, *DOF0, *DOF; HYPRE_Int num_idof, *idof0, *idof; HYPRE_Int num_bdof, *bdof; double *boolean_data; /* Determine the coarse DOFs */ hypre_ParCSRMatrixGetRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data); DOF= hypre_TAlloc(HYPRE_Int, num_DOF); for (j= 0; j< num_DOF; j++) { DOF[j]= DOF0[j]; } hypre_ParCSRMatrixRestoreRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data); qsort0(DOF,0,num_DOF-1); /* Find the fine dofs interior for the current coarse element */ hypre_ParCSRMatrixGetRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data); idof= hypre_TAlloc(HYPRE_Int, num_idof); for (j= 0; j< num_idof; j++) { idof[j]= idof0[j]; } hypre_ParCSRMatrixRestoreRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data); /* Sort the interior dofs according to their global number */ qsort0(idof,0,num_idof-1); /* Find the fine dofs on the boundary of the current coarse element */ if (three_dimensional_problem) { hypre_ParCSRMatrixGetRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data); col_ind1= hypre_TAlloc(HYPRE_Int, size1); for (j= 0; j< size1; j++) { col_ind1[j]= col_ind0[j]; } hypre_ParCSRMatrixRestoreRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data); } else size1 = 0; hypre_ParCSRMatrixGetRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data); col_ind2= hypre_TAlloc(HYPRE_Int, size2); for (j= 0; j< size2; j++) { col_ind2[j]= col_ind0[j]; } hypre_ParCSRMatrixRestoreRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data); /* Merge and sort the boundary dofs according to their global number */ num_bdof = size1 + size2; bdof = hypre_CTAlloc(HYPRE_Int, num_bdof); if (three_dimensional_problem) memcpy(bdof, col_ind1, size1*sizeof(HYPRE_Int)); memcpy(bdof+size1, col_ind2, size2*sizeof(HYPRE_Int)); qsort0(bdof,0,num_bdof-1); /* A = extract_rows(Aee, idof) */ A = hypre_CSRMatrixCreate (num_idof, num_idof + num_bdof, num_idof * (num_idof + num_bdof)); hypre_CSRMatrixInitialize(A); { HYPRE_Int *I = hypre_CSRMatrixI(A); HYPRE_Int *J = hypre_CSRMatrixJ(A); double *data = hypre_CSRMatrixData(A); HYPRE_Int *tmp_J; double *tmp_data; I[0] = 0; for (j = 0; j < num_idof; j++) { getrow_ierr= hypre_ParCSRMatrixGetRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data); if (getrow_ierr <0) hypre_printf("getrow Aee off proc[%d] = \n",myproc); memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int)); memcpy(data, tmp_data, I[j+1]*sizeof(double)); J+= I[j+1]; data+= I[j+1]; hypre_ParCSRMatrixRestoreRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data); I[j+1] += I[j]; } } /* P = extract_rows(dof_DOF, idof+bdof) */ P = hypre_CSRMatrixCreate (num_idof + num_bdof, num_DOF, (num_idof + num_bdof) * num_DOF); hypre_CSRMatrixInitialize(P); { HYPRE_Int *I = hypre_CSRMatrixI(P); HYPRE_Int *J = hypre_CSRMatrixJ(P); double *data = hypre_CSRMatrixData(P); HYPRE_Int m; HYPRE_Int *tmp_J; double *tmp_data; I[0] = 0; for (j = 0; j < num_idof; j++) { getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data); if (getrow_ierr >= 0) { memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int)); memcpy(data, tmp_data, I[j+1]*sizeof(double)); J+= I[j+1]; data+= I[j+1]; hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data); I[j+1] += I[j]; } else /* row offproc */ { hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data); /* search for OffProcRows */ m= 0; while (m < num_OffProcRows) { if (offproc_rnums[m] == idof[j]) { break; } else { m++; } } I[j+1]= (OffProcRows[swap[m]] -> ncols); tmp_J = (OffProcRows[swap[m]] -> cols); tmp_data= (OffProcRows[swap[m]] -> data); memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int)); memcpy(data, tmp_data, I[j+1]*sizeof(double)); J+= I[j+1]; data+= I[j+1]; I[j+1] += I[j]; } } for ( ; j < num_idof + num_bdof; j++) { getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data); if (getrow_ierr >= 0) { memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int)); memcpy(data, tmp_data, I[j+1]*sizeof(double)); J+= I[j+1]; data+= I[j+1]; hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data); I[j+1] += I[j]; } else /* row offproc */ { hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data); /* search for OffProcRows */ m= 0; while (m < num_OffProcRows) { if (offproc_rnums[m] == bdof[j-num_idof]) { break; } else { m++; } } if (m>= num_OffProcRows)hypre_printf("here the mistake\n"); I[j+1]= (OffProcRows[swap[m]] -> ncols); tmp_J = (OffProcRows[swap[m]] -> cols); tmp_data= (OffProcRows[swap[m]] -> data); memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int)); memcpy(data, tmp_data, I[j+1]*sizeof(double)); J+= I[j+1]; data+= I[j+1]; I[j+1] += I[j]; } } } /* Pi = Aii^{-1} Aib Pb */ hypre_HarmonicExtension (A, P, num_DOF, DOF, num_idof, idof, num_bdof, bdof); /* Insert Pi in dof_DOF */ { HYPRE_Int * ncols = hypre_CTAlloc(HYPRE_Int, num_idof); for (j = 0; j < num_idof; j++) ncols[j] = num_DOF; hypre_IJMatrixAddToValuesParCSR (IJ_dof_DOF, num_idof, ncols, idof, hypre_CSRMatrixJ(P), hypre_CSRMatrixData(P)); hypre_TFree(ncols); } hypre_TFree(DOF); hypre_TFree(idof); if (three_dimensional_problem) { hypre_TFree(col_ind1); } hypre_TFree(col_ind2); hypre_TFree(bdof); hypre_CSRMatrixDestroy(A); hypre_CSRMatrixDestroy(P); } #if 0 hypre_TFree(ij_dof_DOF); #endif if (three_dimensional_problem) hypre_ParCSRMatrixDestroy(ELEM_FACEidof); hypre_ParCSRMatrixDestroy(ELEM_EDGEidof); if (num_OffProcRows) { hypre_TFree(offproc_rnums); hypre_TFree(swap); } return ierr; }
HYPRE_Int hypre_MaxwellSolve2( void * maxwell_vdata, hypre_SStructMatrix * A_in, hypre_SStructVector * f, hypre_SStructVector * u ) { hypre_MaxwellData *maxwell_data = maxwell_vdata; hypre_ParVector *f_edge; hypre_ParVector *u_edge; HYPRE_Int max_iter = maxwell_data-> max_iter; double tol = maxwell_data-> tol; HYPRE_Int rel_change = maxwell_data-> rel_change; HYPRE_Int zero_guess = maxwell_data-> zero_guess; HYPRE_Int npre_relax = maxwell_data-> num_pre_relax; HYPRE_Int npost_relax = maxwell_data-> num_post_relax; hypre_ParCSRMatrix **Ann_l = maxwell_data-> Ann_l; hypre_ParCSRMatrix **Pn_l = maxwell_data-> Pn_l; hypre_ParCSRMatrix **RnT_l = maxwell_data-> RnT_l; hypre_ParVector **bn_l = maxwell_data-> bn_l; hypre_ParVector **xn_l = maxwell_data-> xn_l; hypre_ParVector **resn_l = maxwell_data-> resn_l; hypre_ParVector **en_l = maxwell_data-> en_l; hypre_ParVector **nVtemp2_l = maxwell_data-> nVtemp2_l; HYPRE_Int **nCF_marker_l = maxwell_data-> nCF_marker_l; double *nrelax_weight= maxwell_data-> nrelax_weight; double *nomega = maxwell_data-> nomega; HYPRE_Int nrelax_type = maxwell_data-> nrelax_type; HYPRE_Int node_numlevs = maxwell_data-> node_numlevels; hypre_ParCSRMatrix *Tgrad = maxwell_data-> Tgrad; hypre_ParCSRMatrix *T_transpose = maxwell_data-> T_transpose; hypre_ParCSRMatrix **Aee_l = maxwell_data-> Aee_l; hypre_IJMatrix **Pe_l = maxwell_data-> Pe_l; hypre_IJMatrix **ReT_l = maxwell_data-> ReT_l; hypre_ParVector **be_l = maxwell_data-> be_l; hypre_ParVector **xe_l = maxwell_data-> xe_l; hypre_ParVector **rese_l = maxwell_data-> rese_l; hypre_ParVector **ee_l = maxwell_data-> ee_l; hypre_ParVector **eVtemp2_l = maxwell_data-> eVtemp2_l; HYPRE_Int **eCF_marker_l = maxwell_data-> eCF_marker_l; double *erelax_weight= maxwell_data-> erelax_weight; double *eomega = maxwell_data-> eomega; HYPRE_Int erelax_type = maxwell_data-> erelax_type; HYPRE_Int edge_numlevs = maxwell_data-> edge_numlevels; HYPRE_Int **BdryRanks_l = maxwell_data-> BdryRanks_l; HYPRE_Int *BdryRanksCnts_l= maxwell_data-> BdryRanksCnts_l; HYPRE_Int logging = maxwell_data-> logging; double *norms = maxwell_data-> norms; double *rel_norms = maxwell_data-> rel_norms; HYPRE_Int Solve_err_flag; HYPRE_Int relax_local, cycle_param; double b_dot_b = 0, r_dot_r, eps = 0; double e_dot_e, x_dot_x; HYPRE_Int i, j; HYPRE_Int level; HYPRE_Int ierr= 0; /* added for the relaxation routines */ hypre_ParVector *ze = NULL; if (hypre_NumThreads() > 1) { /* Aee is always bigger than Ann */ ze = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(Aee_l[0]), hypre_ParCSRMatrixGlobalNumRows(Aee_l[0]), hypre_ParCSRMatrixRowStarts(Aee_l[0])); hypre_ParVectorInitialize(ze); hypre_ParVectorSetPartitioningOwner(ze,0); } hypre_BeginTiming(maxwell_data-> time_index); hypre_SStructVectorConvert(f, &f_edge); hypre_SStructVectorConvert(u, &u_edge); hypre_ParVectorZeroBCValues(f_edge, BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParVectorZeroBCValues(u_edge, BdryRanks_l[0], BdryRanksCnts_l[0]); be_l[0]= f_edge; xe_l[0]= u_edge; /* the nodal fine vectors: xn= 0. bn= T'*(be- Aee*xe) is updated in the cycle. */ hypre_ParVectorSetConstantValues(xn_l[0], 0.0); relax_local= 0; cycle_param= 0; (maxwell_data-> num_iterations) = 0; /* if max_iter is zero, return */ if (max_iter == 0) { /* if using a zero initial guess, return zero */ if (zero_guess) { hypre_ParVectorSetConstantValues(xe_l[0], 0.0); } hypre_EndTiming(maxwell_data -> time_index); return ierr; } /* part of convergence check */ if (tol > 0.0) { /* eps = (tol^2) */ b_dot_b= hypre_ParVectorInnerProd(be_l[0], be_l[0]); eps = tol*tol; /* if rhs is zero, return a zero solution */ if (b_dot_b == 0.0) { hypre_ParVectorSetConstantValues(xe_l[0], 0.0); if (logging > 0) { norms[0] = 0.0; rel_norms[0] = 0.0; } hypre_EndTiming(maxwell_data -> time_index); return ierr; } } /*----------------------------------------------------- * Do V-cycles: * For each index l, "fine" = l, "coarse" = (l-1) * * solution update: * edge_sol= edge_sol + T*node_sol *-----------------------------------------------------*/ for (i = 0; i < max_iter; i++) { /* compute fine grid residual & nodal rhs. */ hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); hypre_ParVectorZeroBCValues(rese_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParCSRMatrixMatvec(1.0, T_transpose, rese_l[0], 0.0, bn_l[0]); /* convergence check */ if (tol > 0.0) { r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]); if (logging > 0) { norms[i] = sqrt(r_dot_r); if (b_dot_b > 0) rel_norms[i] = sqrt(r_dot_r/b_dot_b); else rel_norms[i] = 0.0; } /* always do at least 1 V-cycle */ if ((r_dot_r/b_dot_b < eps) && (i > 0)) { if (rel_change) { if ((e_dot_e/x_dot_x) < eps) break; } else { break; } } } hypre_ParVectorCopy(bn_l[0], resn_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]); r_dot_r= hypre_ParVectorInnerProd(resn_l[0], resn_l[0]); for (level= 0; level<= node_numlevs-2; level++) { /*----------------------------------------------- * Down cycle *-----------------------------------------------*/ for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); } /*for (j= 0; j< npre_relax; j++) */ /* compute residuals */ hypre_ParVectorCopy(bn_l[level], resn_l[level]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[level], xn_l[level], 1.0, resn_l[level]); /* restrict residuals */ hypre_ParCSRMatrixMatvecT(1.0, RnT_l[level], resn_l[level], 0.0, bn_l[level+1]); /* zero off initial guess for the next level */ hypre_ParVectorSetConstantValues(xn_l[level+1], 0.0); } /* for (level= 0; level<= node_numlevs-2; level++) */ /* coarsest node solve */ level= node_numlevs-1; Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); /*--------------------------------------------------------------------- * Cycle up the levels. *---------------------------------------------------------------------*/ for (level= (node_numlevs - 2); level>= 1; level--) { hypre_ParCSRMatrixMatvec(1.0, Pn_l[level], xn_l[level+1], 0.0, en_l[level]); hypre_ParVectorAxpy(1.0, en_l[level], xn_l[level]); /* post smooth */ for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); } } /* for (level= (en_numlevs - 2); level>= 1; level--) */ /* interpolate error and correct on finest grids */ hypre_ParCSRMatrixMatvec(1.0, Pn_l[0], xn_l[1], 0.0, en_l[0]); hypre_ParVectorAxpy(1.0, en_l[0], xn_l[0]); for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[0], bn_l[0], nCF_marker_l[0], nrelax_type, relax_local, cycle_param, nrelax_weight[0], nomega[0], NULL, xn_l[0], nVtemp2_l[0], ze); } /* for (j= 0; j< npost_relax; j++) */ hypre_ParVectorCopy(bn_l[0], resn_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]); /* add the gradient solution component to xe_l[0] */ hypre_ParCSRMatrixMatvec(1.0, Tgrad, xn_l[0], 1.0, xe_l[0]); hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]); for (level= 0; level<= edge_numlevs-2; level++) { /*----------------------------------------------- * Down cycle *-----------------------------------------------*/ for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } /*for (j= 0; j< npre_relax; j++) */ /* compute residuals */ hypre_ParVectorCopy(be_l[level], rese_l[level]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[level], xe_l[level], 1.0, rese_l[level]); /* restrict residuals */ hypre_ParCSRMatrixMatvecT(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(ReT_l[level]), rese_l[level], 0.0, be_l[level+1]); hypre_ParVectorZeroBCValues(be_l[level+1], BdryRanks_l[level+1], BdryRanksCnts_l[level+1]); /* zero off initial guess for the next level */ hypre_ParVectorSetConstantValues(xe_l[level+1], 0.0); } /* for (level= 1; level<= edge_numlevels-2; level++) */ /* coarsest edge solve */ level= edge_numlevs-1; for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } /*--------------------------------------------------------------------- * Up cycle. *---------------------------------------------------------------------*/ for (level= (edge_numlevs - 2); level>= 1; level--) { hypre_ParCSRMatrixMatvec(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[level]), xe_l[level+1], 0.0, ee_l[level]); hypre_ParVectorZeroBCValues(ee_l[level], BdryRanks_l[level], BdryRanksCnts_l[level]); hypre_ParVectorAxpy(1.0, ee_l[level], xe_l[level]); /* post smooth */ for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } } /* for (level= (edge_numlevs - 2); level>= 1; level--) */ /* interpolate error and correct on finest grids */ hypre_ParCSRMatrixMatvec(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[0]), xe_l[1], 0.0, ee_l[0]); hypre_ParVectorZeroBCValues(ee_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParVectorAxpy(1.0, ee_l[0], xe_l[0]); for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[0], be_l[0], eCF_marker_l[0], erelax_type, relax_local, cycle_param, erelax_weight[0], eomega[0], NULL, xe_l[0], eVtemp2_l[0], ze); } /* for (j= 0; j< npost_relax; j++) */ e_dot_e= hypre_ParVectorInnerProd(ee_l[0], ee_l[0]); x_dot_x= hypre_ParVectorInnerProd(xe_l[0], xe_l[0]); hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); (maxwell_data -> num_iterations) = (i + 1); } hypre_EndTiming(maxwell_data -> time_index); if (ze) hypre_ParVectorDestroy(ze); return ierr; }
HYPRE_Int HYPRE_IJMatrixCreate( MPI_Comm comm, HYPRE_Int ilower, HYPRE_Int iupper, HYPRE_Int jlower, HYPRE_Int jupper, HYPRE_IJMatrix *matrix ) { HYPRE_Int *row_partitioning; HYPRE_Int *col_partitioning; HYPRE_Int *info; HYPRE_Int num_procs; HYPRE_Int myid; hypre_IJMatrix *ijmatrix; #ifdef HYPRE_NO_GLOBAL_PARTITION HYPRE_Int row0, col0, rowN, colN; #else HYPRE_Int *recv_buf; HYPRE_Int i, i4; HYPRE_Int square; #endif ijmatrix = hypre_CTAlloc(hypre_IJMatrix, 1); hypre_IJMatrixComm(ijmatrix) = comm; hypre_IJMatrixObject(ijmatrix) = NULL; hypre_IJMatrixTranslator(ijmatrix) = NULL; hypre_IJMatrixObjectType(ijmatrix) = HYPRE_UNITIALIZED; hypre_IJMatrixAssembleFlag(ijmatrix) = 0; hypre_IJMatrixPrintLevel(ijmatrix) = 0; hypre_MPI_Comm_size(comm,&num_procs); hypre_MPI_Comm_rank(comm, &myid); if (ilower > iupper+1 || ilower < 0) { hypre_error_in_arg(2); hypre_TFree(ijmatrix); return hypre_error_flag; } if (iupper < -1) { hypre_error_in_arg(3); hypre_TFree(ijmatrix); return hypre_error_flag; } if (jlower > jupper+1 || jlower < 0) { hypre_error_in_arg(4); hypre_TFree(ijmatrix); return hypre_error_flag; } if (jupper < -1) { hypre_error_in_arg(5); hypre_TFree(ijmatrix); return hypre_error_flag; } #ifdef HYPRE_NO_GLOBAL_PARTITION info = hypre_CTAlloc(HYPRE_Int,2); row_partitioning = hypre_CTAlloc(HYPRE_Int, 2); col_partitioning = hypre_CTAlloc(HYPRE_Int, 2); row_partitioning[0] = ilower; row_partitioning[1] = iupper+1; col_partitioning[0] = jlower; col_partitioning[1] = jupper+1; /* now we need the global number of rows and columns as well as the global first row and column index */ /* proc 0 has the first row and col */ if (myid==0) { info[0] = ilower; info[1] = jlower; } hypre_MPI_Bcast(info, 2, HYPRE_MPI_INT, 0, comm); row0 = info[0]; col0 = info[1]; /* proc (num_procs-1) has the last row and col */ if (myid == (num_procs-1)) { info[0] = iupper; info[1] = jupper; } hypre_MPI_Bcast(info, 2, HYPRE_MPI_INT, num_procs-1, comm); rowN = info[0]; colN = info[1]; hypre_IJMatrixGlobalFirstRow(ijmatrix) = row0; hypre_IJMatrixGlobalFirstCol(ijmatrix) = col0; hypre_IJMatrixGlobalNumRows(ijmatrix) = rowN - row0 + 1; hypre_IJMatrixGlobalNumCols(ijmatrix) = colN - col0 + 1; hypre_TFree(info); #else info = hypre_CTAlloc(HYPRE_Int,4); recv_buf = hypre_CTAlloc(HYPRE_Int,4*num_procs); row_partitioning = hypre_CTAlloc(HYPRE_Int, num_procs+1); info[0] = ilower; info[1] = iupper; info[2] = jlower; info[3] = jupper; /* Generate row- and column-partitioning through information exchange across all processors, check whether the matrix is square, and if the partitionings match. i.e. no overlaps or gaps, if there are overlaps or gaps in the row partitioning or column partitioning , ierr will be set to -9 or -10, respectively */ hypre_MPI_Allgather(info,4,HYPRE_MPI_INT,recv_buf,4,HYPRE_MPI_INT,comm); row_partitioning[0] = recv_buf[0]; square = 1; for (i=0; i < num_procs-1; i++) { i4 = 4*i; if ( recv_buf[i4+1] != (recv_buf[i4+4]-1) ) { hypre_error(HYPRE_ERROR_GENERIC); hypre_TFree(ijmatrix); hypre_TFree(info); hypre_TFree(recv_buf); hypre_TFree(row_partitioning); return hypre_error_flag; } else row_partitioning[i+1] = recv_buf[i4+4]; if ((square && (recv_buf[i4] != recv_buf[i4+2])) || (recv_buf[i4+1] != recv_buf[i4+3]) ) { square = 0; } } i4 = (num_procs-1)*4; row_partitioning[num_procs] = recv_buf[i4+1]+1; if ((recv_buf[i4] != recv_buf[i4+2]) || (recv_buf[i4+1] != recv_buf[i4+3])) square = 0; if (square) col_partitioning = row_partitioning; else { col_partitioning = hypre_CTAlloc(HYPRE_Int,num_procs+1); col_partitioning[0] = recv_buf[2]; for (i=0; i < num_procs-1; i++) { i4 = 4*i; if (recv_buf[i4+3] != recv_buf[i4+6]-1) { hypre_error(HYPRE_ERROR_GENERIC); hypre_TFree(ijmatrix); hypre_TFree(info); hypre_TFree(recv_buf); hypre_TFree(row_partitioning); hypre_TFree(col_partitioning); return hypre_error_flag; } else col_partitioning[i+1] = recv_buf[i4+6]; } col_partitioning[num_procs] = recv_buf[num_procs*4-1]+1; } hypre_IJMatrixGlobalFirstRow(ijmatrix) = row_partitioning[0]; hypre_IJMatrixGlobalFirstCol(ijmatrix) = col_partitioning[0]; hypre_IJMatrixGlobalNumRows(ijmatrix) = row_partitioning[num_procs] - row_partitioning[0]; hypre_IJMatrixGlobalNumCols(ijmatrix) = col_partitioning[num_procs] - col_partitioning[0]; hypre_TFree(info); hypre_TFree(recv_buf); #endif hypre_IJMatrixRowPartitioning(ijmatrix) = row_partitioning; hypre_IJMatrixColPartitioning(ijmatrix) = col_partitioning; *matrix = (HYPRE_IJMatrix) ijmatrix; return hypre_error_flag; }