int zdominate(int n, NCformat *Astore) /* make the matrix diagonally dominant */ { doublecomplex *nzval = (doublecomplex *)Astore->nzval; int *rowind = Astore->rowind; int *colptr = Astore->colptr; int nnz = colptr[n]; int fill = 0; doublecomplex *nzval_new; int *rowind_new; int i, j, diag; double s; for (i = 0; i < n; i++) { diag = -1; for (j = colptr[i]; j < colptr[i + 1]; j++) if (rowind[j] == i) diag = j; if (diag < 0) fill++; } if (fill) { nzval_new = doublecomplexMalloc(nnz + fill); rowind_new = intMalloc(nnz+ fill); fill = 0; for (i = 0; i < n; i++) { s = 1e-6; diag = -1; for (j = colptr[i] - fill; j < colptr[i + 1]; j++) { if ((rowind_new[j + fill] = rowind[j]) == i) diag = j; nzval_new[j + fill] = nzval[j]; s += z_abs1(&nzval_new[j + fill]); } if (diag >= 0) { nzval_new[diag+fill].r = s * 3.0; nzval_new[diag+fill].i = 0.0; } else { rowind_new[colptr[i + 1] + fill] = i; nzval_new[colptr[i + 1] + fill].r = s * 3.0; nzval_new[colptr[i + 1] + fill].i = 0.0; fill++; } colptr[i + 1] += fill; } Astore->nzval = nzval_new; Astore->rowind = rowind_new; SUPERLU_FREE(nzval); SUPERLU_FREE(rowind); } else { for (i = 0; i < n; i++) { s = 1e-6; diag = -1; for (j = colptr[i]; j < colptr[i + 1]; j++) { if (rowind[j] == i) diag = j; s += z_abs1(&nzval[j]); } nzval[diag].r = s * 3.0; nzval[diag].i = 0.0; } } Astore->nnz += fill; return fill; }
main(int argc, char *argv[]) { SuperMatrix A; NCformat *Astore; doublecomplex *a; int *asub, *xa; int *perm_c; /* column permutation vector */ int *perm_r; /* row permutations from partial pivoting */ SuperMatrix L; /* factor L */ SCformat *Lstore; SuperMatrix U; /* factor U */ NCformat *Ustore; SuperMatrix B; int nrhs, ldx, info, m, n, nnz; doublecomplex *xact, *rhs; mem_usage_t mem_usage; superlu_options_t options; SuperLUStat_t stat; #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Enter main()"); #endif /* Set the default input options: options.Fact = DOFACT; options.Equil = YES; options.ColPerm = COLAMD; options.DiagPivotThresh = 1.0; options.Trans = NOTRANS; options.IterRefine = NOREFINE; options.SymmetricMode = NO; options.PivotGrowth = NO; options.ConditionNumber = NO; options.PrintStat = YES; */ set_default_options(&options); /* Read the matrix in Harwell-Boeing format. */ zreadhb(&m, &n, &nnz, &a, &asub, &xa); zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); nrhs = 1; if ( !(rhs = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[]."); zCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); ldx = n; zGenXtrue(n, nrhs, xact, ldx); zFillRHS(options.Trans, nrhs, xact, ldx, &A, &B); if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[]."); if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[]."); /* Initialize the statistics variables. */ StatInit(&stat); zgssv(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info); if ( info == 0 ) { /* This is how you could access the solution matrix. */ doublecomplex *sol = (doublecomplex*) ((DNformat*) B.Store)->nzval; /* Compute the infinity norm of the error. */ zinf_norm_error(nrhs, &B, xact); Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz); zQuerySpace(&L, &U, &mem_usage); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); } else { printf("zgssv() error returns INFO= %d\n", info); if ( info <= n ) { /* factorization completes */ zQuerySpace(&L, &U, &mem_usage); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); } } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); SUPERLU_FREE (rhs); SUPERLU_FREE (xact); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); Destroy_CompCol_Matrix(&A); Destroy_SuperMatrix_Store(&B); Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Exit main()"); #endif }
void zgstrs (trans_t trans, SuperMatrix *L, SuperMatrix *U, int *perm_c, int *perm_r, SuperMatrix *B, SuperLUStat_t *stat, int *info) { /* * Purpose * ======= * * ZGSTRS solves a system of linear equations A*X=B or A'*X=B * with A sparse and B dense, using the LU factorization computed by * ZGSTRF. * * See supermatrix.h for the definition of 'SuperMatrix' structure. * * Arguments * ========= * * trans (input) trans_t * Specifies the form of the system of equations: * = NOTRANS: A * X = B (No transpose) * = TRANS: A'* X = B (Transpose) * = CONJ: A**H * X = B (Conjugate transpose) * * L (input) SuperMatrix* * The factor L from the factorization Pr*A*Pc=L*U as computed by * zgstrf(). Use compressed row subscripts storage for supernodes, * i.e., L has types: Stype = SLU_SC, Dtype = SLU_Z, Mtype = SLU_TRLU. * * U (input) SuperMatrix* * The factor U from the factorization Pr*A*Pc=L*U as computed by * zgstrf(). Use column-wise storage scheme, i.e., U has types: * Stype = SLU_NC, Dtype = SLU_Z, Mtype = SLU_TRU. * * perm_c (input) int*, dimension (L->ncol) * Column permutation vector, which defines the * permutation matrix Pc; perm_c[i] = j means column i of A is * in position j in A*Pc. * * perm_r (input) int*, dimension (L->nrow) * Row permutation vector, which defines the permutation matrix Pr; * perm_r[i] = j means row i of A is in position j in Pr*A. * * B (input/output) SuperMatrix* * B has types: Stype = SLU_DN, Dtype = SLU_Z, Mtype = SLU_GE. * On entry, the right hand side matrix. * On exit, the solution matrix if info = 0; * * stat (output) SuperLUStat_t* * Record the statistics on runtime and floating-point operation count. * See util.h for the definition of 'SuperLUStat_t'. * * info (output) int* * = 0: successful exit * < 0: if info = -i, the i-th argument had an illegal value * */ #ifdef _CRAY _fcd ftcs1, ftcs2, ftcs3, ftcs4; #endif int incx = 1, incy = 1; #ifdef USE_VENDOR_BLAS doublecomplex alpha = {1.0, 0.0}, beta = {1.0, 0.0}; doublecomplex *work_col; #endif doublecomplex temp_comp; DNformat *Bstore; doublecomplex *Bmat; SCformat *Lstore; NCformat *Ustore; doublecomplex *Lval, *Uval; int fsupc, nrow, nsupr, nsupc, luptr, istart, irow; int i, j, k, iptr, jcol, n, ldb, nrhs; doublecomplex *work, *rhs_work, *soln; flops_t solve_ops; void zprint_soln(); /* Test input parameters ... */ *info = 0; Bstore = B->Store; ldb = Bstore->lda; nrhs = B->ncol; if ( trans != NOTRANS && trans != TRANS && trans != CONJ ) *info = -1; else if ( L->nrow != L->ncol || L->nrow < 0 || L->Stype != SLU_SC || L->Dtype != SLU_Z || L->Mtype != SLU_TRLU ) *info = -2; else if ( U->nrow != U->ncol || U->nrow < 0 || U->Stype != SLU_NC || U->Dtype != SLU_Z || U->Mtype != SLU_TRU ) *info = -3; else if ( ldb < SUPERLU_MAX(0, L->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_Z || B->Mtype != SLU_GE ) *info = -6; if ( *info ) { i = -(*info); xerbla_("zgstrs", &i); return; } n = L->nrow; work = doublecomplexCalloc(n * nrhs); if ( !work ) ABORT("Malloc fails for local work[]."); soln = doublecomplexMalloc(n); if ( !soln ) ABORT("Malloc fails for local soln[]."); Bmat = Bstore->nzval; Lstore = L->Store; Lval = Lstore->nzval; Ustore = U->Store; Uval = Ustore->nzval; solve_ops = 0; if ( trans == NOTRANS ) { /* Permute right hand sides to form Pr*B */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } /* Forward solve PLy=Pb. */ for (k = 0; k <= Lstore->nsuper; k++) { fsupc = L_FST_SUPC(k); istart = L_SUB_START(fsupc); nsupr = L_SUB_START(fsupc+1) - istart; nsupc = L_FST_SUPC(k+1) - fsupc; nrow = nsupr - nsupc; solve_ops += 4 * nsupc * (nsupc - 1) * nrhs; solve_ops += 8 * nrow * nsupc * nrhs; if ( nsupc == 1 ) { for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; luptr = L_NZ_START(fsupc); for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){ irow = L_SUB(iptr); ++luptr; zz_mult(&temp_comp, &rhs_work[fsupc], &Lval[luptr]); z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp); } } } else { luptr = L_NZ_START(fsupc); #ifdef USE_VENDOR_BLAS #ifdef _CRAY ftcs1 = _cptofcd("L", strlen("L")); ftcs2 = _cptofcd("N", strlen("N")); ftcs3 = _cptofcd("U", strlen("U")); CTRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); CGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha, &Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb, &beta, &work[0], &n ); #else ztrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); zgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha, &Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb, &beta, &work[0], &n ); #endif for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; work_col = &work[j*n]; iptr = istart + nsupc; for (i = 0; i < nrow; i++) { irow = L_SUB(iptr); z_sub(&rhs_work[irow], &rhs_work[irow], &work_col[i]); work_col[i].r = 0.0; work_col[i].i = 0.0; iptr++; } } #else for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; zlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]); zmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc], &rhs_work[fsupc], &work[0] ); iptr = istart + nsupc; for (i = 0; i < nrow; i++) { irow = L_SUB(iptr); z_sub(&rhs_work[irow], &rhs_work[irow], &work[i]); work[i].r = 0.; work[i].i = 0.; iptr++; } } #endif } /* else ... */ } /* for L-solve */ #ifdef DEBUG printf("After L-solve: y=\n"); zprint_soln(n, nrhs, Bmat); #endif /* * Back solve Ux=y. */ for (k = Lstore->nsuper; k >= 0; k--) { fsupc = L_FST_SUPC(k); istart = L_SUB_START(fsupc); nsupr = L_SUB_START(fsupc+1) - istart; nsupc = L_FST_SUPC(k+1) - fsupc; luptr = L_NZ_START(fsupc); solve_ops += 4 * nsupc * (nsupc + 1) * nrhs; if ( nsupc == 1 ) { rhs_work = &Bmat[0]; for (j = 0; j < nrhs; j++) { z_div(&rhs_work[fsupc], &rhs_work[fsupc], &Lval[luptr]); rhs_work += ldb; } } else { #ifdef USE_VENDOR_BLAS #ifdef _CRAY ftcs1 = _cptofcd("L", strlen("L")); ftcs2 = _cptofcd("U", strlen("U")); ftcs3 = _cptofcd("N", strlen("N")); CTRSM( ftcs1, ftcs2, ftcs3, ftcs3, &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); #else ztrsm_("L", "U", "N", "N", &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); #endif #else for (j = 0; j < nrhs; j++) zusolve ( nsupr, nsupc, &Lval[luptr], &Bmat[fsupc+j*ldb] ); #endif } for (j = 0; j < nrhs; ++j) { rhs_work = &Bmat[j*ldb]; for (jcol = fsupc; jcol < fsupc + nsupc; jcol++) { solve_ops += 8*(U_NZ_START(jcol+1) - U_NZ_START(jcol)); for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1); i++ ){ irow = U_SUB(i); zz_mult(&temp_comp, &rhs_work[jcol], &Uval[i]); z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp); } } } } /* for U-solve */ #ifdef DEBUG printf("After U-solve: x=\n"); zprint_soln(n, nrhs, Bmat); #endif /* Compute the final solution X := Pc*X. */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[k] = rhs_work[perm_c[k]]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } stat->ops[SOLVE] = solve_ops; } else { /* Solve A'*X=B or CONJ(A)*X=B */ /* Permute right hand sides to form Pc'*B. */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[perm_c[k]] = rhs_work[k]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } stat->ops[SOLVE] = 0; if (trans == TRANS) { for (k = 0; k < nrhs; ++k) { /* Multiply by inv(U'). */ sp_ztrsv("U", "T", "N", L, U, &Bmat[k*ldb], stat, info); /* Multiply by inv(L'). */ sp_ztrsv("L", "T", "U", L, U, &Bmat[k*ldb], stat, info); } } else { /* trans == CONJ */ for (k = 0; k < nrhs; ++k) { /* Multiply by conj(inv(U')). */ sp_ztrsv("U", "C", "N", L, U, &Bmat[k*ldb], stat, info); /* Multiply by conj(inv(L')). */ sp_ztrsv("L", "C", "U", L, U, &Bmat[k*ldb], stat, info); } } /* Compute the final solution X := Pr'*X (=inv(Pr)*X) */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[k] = rhs_work[perm_r[k]]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } } SUPERLU_FREE(work); SUPERLU_FREE(soln); }
main(int argc, char *argv[]) { char equed[1]; yes_no_t equil; trans_t trans; SuperMatrix A, L, U; SuperMatrix B, X; NCformat *Astore; NCformat *Ustore; SCformat *Lstore; doublecomplex *a; int *asub, *xa; int *perm_r; /* row permutations from partial pivoting */ int *perm_c; /* column permutation vector */ int *etree; void *work; int info, lwork, nrhs, ldx; int i, m, n, nnz; doublecomplex *rhsb, *rhsx, *xact; double *R, *C; double *ferr, *berr; double u, rpg, rcond; mem_usage_t mem_usage; superlu_options_t options; SuperLUStat_t stat; extern void parse_command_line(); #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Enter main()"); #endif /* Defaults */ lwork = 0; nrhs = 1; equil = YES; u = 1.0; trans = NOTRANS; /* Set the default input options: options.Fact = DOFACT; options.Equil = YES; options.ColPerm = COLAMD; options.DiagPivotThresh = 1.0; options.Trans = NOTRANS; options.IterRefine = NOREFINE; options.SymmetricMode = NO; options.PivotGrowth = NO; options.ConditionNumber = NO; options.PrintStat = YES; */ set_default_options(&options); /* Can use command line input to modify the defaults. */ parse_command_line(argc, argv, &lwork, &u, &equil, &trans); options.Equil = equil; options.DiagPivotThresh = u; options.Trans = trans; /* Add more functionalities that the defaults. */ options.PivotGrowth = YES; /* Compute reciprocal pivot growth */ options.ConditionNumber = YES;/* Compute reciprocal condition number */ options.IterRefine = DOUBLE; /* Perform double-precision refinement */ if ( lwork > 0 ) { work = SUPERLU_MALLOC(lwork); if ( !work ) { ABORT("ZLINSOLX: cannot allocate work[]"); } } /* Read matrix A from a file in Harwell-Boeing format.*/ zreadhb(&m, &n, &nnz, &a, &asub, &xa); zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); if ( !(rhsb = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[]."); if ( !(rhsx = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[]."); zCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_Z, SLU_GE); zCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); ldx = n; zGenXtrue(n, nrhs, xact, ldx); zFillRHS(trans, nrhs, xact, ldx, &A, &B); if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[]."); if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[]."); if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[]."); if ( !(R = (double *) SUPERLU_MALLOC(A.nrow * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for R[]."); if ( !(C = (double *) SUPERLU_MALLOC(A.ncol * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for C[]."); if ( !(ferr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for ferr[]."); if ( !(berr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for berr[]."); /* Initialize the statistics variables. */ StatInit(&stat); /* Solve the system and compute the condition number and error bounds using dgssvx. */ zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); printf("zgssvx(): info %d\n", info); if ( info == 0 || info == n+1 ) { /* This is how you could access the solution matrix. */ doublecomplex *sol = (doublecomplex*) ((DNformat*) X.Store)->nzval; if ( options.PivotGrowth == YES ) printf("Recip. pivot growth = %e\n", rpg); if ( options.ConditionNumber == YES ) printf("Recip. condition number = %e\n", rcond); if ( options.IterRefine != NOREFINE ) { printf("Iterative Refinement:\n"); printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR"); for (i = 0; i < nrhs; ++i) printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]); } Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); fflush(stdout); } else if ( info > 0 && lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); SUPERLU_FREE (rhsb); SUPERLU_FREE (rhsx); SUPERLU_FREE (xact); SUPERLU_FREE (etree); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); SUPERLU_FREE (R); SUPERLU_FREE (C); SUPERLU_FREE (ferr); SUPERLU_FREE (berr); Destroy_CompCol_Matrix(&A); Destroy_SuperMatrix_Store(&B); Destroy_SuperMatrix_Store(&X); if ( lwork >= 0 ) { Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Exit main()"); #endif }
/*! \brief * * <pre> * Purpose * ======= * * ZGSRFS improves the computed solution to a system of linear * equations and provides error bounds and backward error estimates for * the solution. * * If equilibration was performed, the system becomes: * (diag(R)*A_original*diag(C)) * X = diag(R)*B_original. * * See supermatrix.h for the definition of 'SuperMatrix' structure. * * Arguments * ========= * * trans (input) trans_t * Specifies the form of the system of equations: * = NOTRANS: A * X = B (No transpose) * = TRANS: A'* X = B (Transpose) * = CONJ: A**H * X = B (Conjugate transpose) * * A (input) SuperMatrix* * The original matrix A in the system, or the scaled A if * equilibration was done. The type of A can be: * Stype = SLU_NC, Dtype = SLU_Z, Mtype = SLU_GE. * * L (input) SuperMatrix* * The factor L from the factorization Pr*A*Pc=L*U. Use * compressed row subscripts storage for supernodes, * i.e., L has types: Stype = SLU_SC, Dtype = SLU_Z, Mtype = SLU_TRLU. * * U (input) SuperMatrix* * The factor U from the factorization Pr*A*Pc=L*U as computed by * zgstrf(). Use column-wise storage scheme, * i.e., U has types: Stype = SLU_NC, Dtype = SLU_Z, Mtype = SLU_TRU. * * perm_c (input) int*, dimension (A->ncol) * Column permutation vector, which defines the * permutation matrix Pc; perm_c[i] = j means column i of A is * in position j in A*Pc. * * perm_r (input) int*, dimension (A->nrow) * Row permutation vector, which defines the permutation matrix Pr; * perm_r[i] = j means row i of A is in position j in Pr*A. * * equed (input) Specifies the form of equilibration that was done. * = 'N': No equilibration. * = 'R': Row equilibration, i.e., A was premultiplied by diag(R). * = 'C': Column equilibration, i.e., A was postmultiplied by * diag(C). * = 'B': Both row and column equilibration, i.e., A was replaced * by diag(R)*A*diag(C). * * R (input) double*, dimension (A->nrow) * The row scale factors for A. * If equed = 'R' or 'B', A is premultiplied by diag(R). * If equed = 'N' or 'C', R is not accessed. * * C (input) double*, dimension (A->ncol) * The column scale factors for A. * If equed = 'C' or 'B', A is postmultiplied by diag(C). * If equed = 'N' or 'R', C is not accessed. * * B (input) SuperMatrix* * B has types: Stype = SLU_DN, Dtype = SLU_Z, Mtype = SLU_GE. * The right hand side matrix B. * if equed = 'R' or 'B', B is premultiplied by diag(R). * * X (input/output) SuperMatrix* * X has types: Stype = SLU_DN, Dtype = SLU_Z, Mtype = SLU_GE. * On entry, the solution matrix X, as computed by zgstrs(). * On exit, the improved solution matrix X. * if *equed = 'C' or 'B', X should be premultiplied by diag(C) * in order to obtain the solution to the original system. * * FERR (output) double*, dimension (B->ncol) * The estimated forward error bound for each solution vector * X(j) (the j-th column of the solution matrix X). * If XTRUE is the true solution corresponding to X(j), FERR(j) * is an estimated upper bound for the magnitude of the largest * element in (X(j) - XTRUE) divided by the magnitude of the * largest element in X(j). The estimate is as reliable as * the estimate for RCOND, and is almost always a slight * overestimate of the true error. * * BERR (output) double*, dimension (B->ncol) * The componentwise relative backward error of each solution * vector X(j) (i.e., the smallest relative change in * any element of A or B that makes X(j) an exact solution). * * stat (output) SuperLUStat_t* * Record the statistics on runtime and floating-point operation count. * See util.h for the definition of 'SuperLUStat_t'. * * info (output) int* * = 0: successful exit * < 0: if INFO = -i, the i-th argument had an illegal value * * Internal Parameters * =================== * * ITMAX is the maximum number of steps of iterative refinement. * * </pre> */ void zgsrfs(trans_t trans, SuperMatrix *A, SuperMatrix *L, SuperMatrix *U, int *perm_c, int *perm_r, char *equed, double *R, double *C, SuperMatrix *B, SuperMatrix *X, double *ferr, double *berr, SuperLUStat_t *stat, int *info) { #define ITMAX 5 /* Table of constant values */ int ione = 1; doublecomplex ndone = {-1., 0.}; doublecomplex done = {1., 0.}; /* Local variables */ NCformat *Astore; doublecomplex *Aval; SuperMatrix Bjcol; DNformat *Bstore, *Xstore, *Bjcol_store; doublecomplex *Bmat, *Xmat, *Bptr, *Xptr; int kase; double safe1, safe2; int i, j, k, irow, nz, count, notran, rowequ, colequ; int ldb, ldx, nrhs; double s, xk, lstres, eps, safmin; char transc[1]; trans_t transt; doublecomplex *work; double *rwork; int *iwork; extern double dlamch_(char *); extern int zlacon_(int *, doublecomplex *, doublecomplex *, double *, int *); #ifdef _CRAY extern int CCOPY(int *, doublecomplex *, int *, doublecomplex *, int *); extern int CSAXPY(int *, doublecomplex *, doublecomplex *, int *, doublecomplex *, int *); #else extern int zcopy_(int *, doublecomplex *, int *, doublecomplex *, int *); extern int zaxpy_(int *, doublecomplex *, doublecomplex *, int *, doublecomplex *, int *); #endif Astore = A->Store; Aval = Astore->nzval; Bstore = B->Store; Xstore = X->Store; Bmat = Bstore->nzval; Xmat = Xstore->nzval; ldb = Bstore->lda; ldx = Xstore->lda; nrhs = B->ncol; /* Test the input parameters */ *info = 0; notran = (trans == NOTRANS); if ( !notran && trans != TRANS && trans != CONJ ) *info = -1; else if ( A->nrow != A->ncol || A->nrow < 0 || A->Stype != SLU_NC || A->Dtype != SLU_Z || A->Mtype != SLU_GE ) *info = -2; else if ( L->nrow != L->ncol || L->nrow < 0 || L->Stype != SLU_SC || L->Dtype != SLU_Z || L->Mtype != SLU_TRLU ) *info = -3; else if ( U->nrow != U->ncol || U->nrow < 0 || U->Stype != SLU_NC || U->Dtype != SLU_Z || U->Mtype != SLU_TRU ) *info = -4; else if ( ldb < SUPERLU_MAX(0, A->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_Z || B->Mtype != SLU_GE ) *info = -10; else if ( ldx < SUPERLU_MAX(0, A->nrow) || X->Stype != SLU_DN || X->Dtype != SLU_Z || X->Mtype != SLU_GE ) *info = -11; if (*info != 0) { i = -(*info); xerbla_("zgsrfs", &i); return; } /* Quick return if possible */ if ( A->nrow == 0 || nrhs == 0) { for (j = 0; j < nrhs; ++j) { ferr[j] = 0.; berr[j] = 0.; } return; } rowequ = lsame_(equed, "R") || lsame_(equed, "B"); colequ = lsame_(equed, "C") || lsame_(equed, "B"); /* Allocate working space */ work = doublecomplexMalloc(2*A->nrow); rwork = (double *) SUPERLU_MALLOC( A->nrow * sizeof(double) ); iwork = intMalloc(A->nrow); if ( !work || !rwork || !iwork ) ABORT("Malloc fails for work/rwork/iwork."); if ( notran ) { *(unsigned char *)transc = 'N'; transt = TRANS; } else { *(unsigned char *)transc = 'T'; transt = NOTRANS; } /* NZ = maximum number of nonzero elements in each row of A, plus 1 */ nz = A->ncol + 1; eps = dlamch_("Epsilon"); safmin = dlamch_("Safe minimum"); /* Set SAFE1 essentially to be the underflow threshold times the number of additions in each row. */ safe1 = nz * safmin; safe2 = safe1 / eps; /* Compute the number of nonzeros in each row (or column) of A */ for (i = 0; i < A->nrow; ++i) iwork[i] = 0; if ( notran ) { for (k = 0; k < A->ncol; ++k) for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) ++iwork[Astore->rowind[i]]; } else { for (k = 0; k < A->ncol; ++k) iwork[k] = Astore->colptr[k+1] - Astore->colptr[k]; } /* Copy one column of RHS B into Bjcol. */ Bjcol.Stype = B->Stype; Bjcol.Dtype = B->Dtype; Bjcol.Mtype = B->Mtype; Bjcol.nrow = B->nrow; Bjcol.ncol = 1; Bjcol.Store = (void *) SUPERLU_MALLOC( sizeof(DNformat) ); if ( !Bjcol.Store ) ABORT("SUPERLU_MALLOC fails for Bjcol.Store"); Bjcol_store = Bjcol.Store; Bjcol_store->lda = ldb; Bjcol_store->nzval = work; /* address aliasing */ /* Do for each right hand side ... */ for (j = 0; j < nrhs; ++j) { count = 0; lstres = 3.; Bptr = &Bmat[j*ldb]; Xptr = &Xmat[j*ldx]; while (1) { /* Loop until stopping criterion is satisfied. */ /* Compute residual R = B - op(A) * X, where op(A) = A, A**T, or A**H, depending on TRANS. */ #ifdef _CRAY CCOPY(&A->nrow, Bptr, &ione, work, &ione); #else zcopy_(&A->nrow, Bptr, &ione, work, &ione); #endif sp_zgemv(transc, ndone, A, Xptr, ione, done, work, ione); /* Compute componentwise relative backward error from formula max(i) ( abs(R(i)) / ( abs(op(A))*abs(X) + abs(B) )(i) ) where abs(Z) is the componentwise absolute value of the matrix or vector Z. If the i-th component of the denominator is less than SAFE2, then SAFE1 is added to the i-th component of the numerator before dividing. */ for (i = 0; i < A->nrow; ++i) rwork[i] = z_abs1( &Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if (notran) { for (k = 0; k < A->ncol; ++k) { xk = z_abs1( &Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += z_abs1(&Aval[i]) * xk; } } else { for (k = 0; k < A->ncol; ++k) { s = 0.; for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) { irow = Astore->rowind[i]; s += z_abs1(&Aval[i]) * z_abs1(&Xptr[irow]); } rwork[k] += s; } } s = 0.; for (i = 0; i < A->nrow; ++i) { if (rwork[i] > safe2) { s = SUPERLU_MAX( s, z_abs1(&work[i]) / rwork[i] ); } else if ( rwork[i] != 0.0 ) { s = SUPERLU_MAX( s, (z_abs1(&work[i]) + safe1) / rwork[i] ); } /* If rwork[i] is exactly 0.0, then we know the true residual also must be exactly 0.0. */ } berr[j] = s; /* Test stopping criterion. Continue iterating if 1) The residual BERR(J) is larger than machine epsilon, and 2) BERR(J) decreased by at least a factor of 2 during the last iteration, and 3) At most ITMAX iterations tried. */ if (berr[j] > eps && berr[j] * 2. <= lstres && count < ITMAX) { /* Update solution and try again. */ zgstrs (trans, L, U, perm_c, perm_r, &Bjcol, stat, info); #ifdef _CRAY CAXPY(&A->nrow, &done, work, &ione, &Xmat[j*ldx], &ione); #else zaxpy_(&A->nrow, &done, work, &ione, &Xmat[j*ldx], &ione); #endif lstres = berr[j]; ++count; } else { break; } } /* end while */ stat->RefineSteps = count; /* Bound error from formula: norm(X - XTRUE) / norm(X) .le. FERR = norm( abs(inv(op(A)))* ( abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) ))) / norm(X) where norm(Z) is the magnitude of the largest component of Z inv(op(A)) is the inverse of op(A) abs(Z) is the componentwise absolute value of the matrix or vector Z NZ is the maximum number of nonzeros in any row of A, plus 1 EPS is machine epsilon The i-th component of abs(R)+NZ*EPS*(abs(op(A))*abs(X)+abs(B)) is incremented by SAFE1 if the i-th component of abs(op(A))*abs(X) + abs(B) is less than SAFE2. Use ZLACON to estimate the infinity-norm of the matrix inv(op(A)) * diag(W), where W = abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) ))) */ for (i = 0; i < A->nrow; ++i) rwork[i] = z_abs1( &Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if ( notran ) { for (k = 0; k < A->ncol; ++k) { xk = z_abs1( &Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += z_abs1(&Aval[i]) * xk; } } else { for (k = 0; k < A->ncol; ++k) { s = 0.; for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) { irow = Astore->rowind[i]; xk = z_abs1( &Xptr[irow] ); s += z_abs1(&Aval[i]) * xk; } rwork[k] += s; } } for (i = 0; i < A->nrow; ++i) if (rwork[i] > safe2) rwork[i] = z_abs(&work[i]) + (iwork[i]+1)*eps*rwork[i]; else rwork[i] = z_abs(&work[i])+(iwork[i]+1)*eps*rwork[i]+safe1; kase = 0; do { zlacon_(&A->nrow, &work[A->nrow], work, &ferr[j], &kase); if (kase == 0) break; if (kase == 1) { /* Multiply by diag(W)*inv(op(A)**T)*(diag(C) or diag(R)). */ if ( notran && colequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], C[i]); } else if ( !notran && rowequ ) for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], R[i]); } zgstrs (transt, L, U, perm_c, perm_r, &Bjcol, stat, info); for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], rwork[i]); } } else { /* Multiply by (diag(C) or diag(R))*inv(op(A))*diag(W). */ for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], rwork[i]); } zgstrs (trans, L, U, perm_c, perm_r, &Bjcol, stat, info); if ( notran && colequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], C[i]); } else if ( !notran && rowequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], R[i]); } } } while ( kase != 0 ); /* Normalize error. */ lstres = 0.; if ( notran && colequ ) { for (i = 0; i < A->nrow; ++i) lstres = SUPERLU_MAX( lstres, C[i] * z_abs1( &Xptr[i]) ); } else if ( !notran && rowequ ) { for (i = 0; i < A->nrow; ++i) lstres = SUPERLU_MAX( lstres, R[i] * z_abs1( &Xptr[i]) ); } else { for (i = 0; i < A->nrow; ++i) lstres = SUPERLU_MAX( lstres, z_abs1( &Xptr[i]) ); } if ( lstres != 0. ) ferr[j] /= lstres; } /* for each RHS j ... */ SUPERLU_FREE(work); SUPERLU_FREE(rwork); SUPERLU_FREE(iwork); SUPERLU_FREE(Bjcol.Store); return; } /* zgsrfs */
int main(int argc, char *argv[]) { /* * Purpose * ======= * * The driver program ZLINSOLX1. * * This example illustrates how to use ZGSSVX to solve systems with the same * A but different right-hand side. * In this case, we factorize A only once in the first call to DGSSVX, * and reuse the following data structures in the subsequent call to ZGSSVX: * perm_c, perm_r, R, C, L, U. * */ char equed[1]; yes_no_t equil; trans_t trans; SuperMatrix A, L, U; SuperMatrix B, X; NCformat *Astore; NCformat *Ustore; SCformat *Lstore; doublecomplex *a; int *asub, *xa; int *perm_c; /* column permutation vector */ int *perm_r; /* row permutations from partial pivoting */ int *etree; void *work; int info, lwork, nrhs, ldx; int i, m, n, nnz; doublecomplex *rhsb, *rhsx, *xact; double *R, *C; double *ferr, *berr; double u, rpg, rcond; mem_usage_t mem_usage; superlu_options_t options; SuperLUStat_t stat; extern void parse_command_line(); #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Enter main()"); #endif /* Defaults */ lwork = 0; nrhs = 1; equil = YES; u = 1.0; trans = NOTRANS; /* Set the default values for options argument: options.Fact = DOFACT; options.Equil = YES; options.ColPerm = COLAMD; options.DiagPivotThresh = 1.0; options.Trans = NOTRANS; options.IterRefine = NOREFINE; options.SymmetricMode = NO; options.PivotGrowth = NO; options.ConditionNumber = NO; options.PrintStat = YES; */ set_default_options(&options); /* Can use command line input to modify the defaults. */ parse_command_line(argc, argv, &lwork, &u, &equil, &trans); options.Equil = equil; options.DiagPivotThresh = u; options.Trans = trans; if ( lwork > 0 ) { work = SUPERLU_MALLOC(lwork); if ( !work ) { ABORT("ZLINSOLX: cannot allocate work[]"); } } /* Read matrix A from a file in Harwell-Boeing format.*/ zreadhb(&m, &n, &nnz, &a, &asub, &xa); zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); if ( !(rhsb = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[]."); if ( !(rhsx = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[]."); zCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_Z, SLU_GE); zCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); ldx = n; zGenXtrue(n, nrhs, xact, ldx); zFillRHS(trans, nrhs, xact, ldx, &A, &B); if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[]."); if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[]."); if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[]."); if ( !(R = (double *) SUPERLU_MALLOC(A.nrow * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for R[]."); if ( !(C = (double *) SUPERLU_MALLOC(A.ncol * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for C[]."); if ( !(ferr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for ferr[]."); if ( !(berr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for berr[]."); /* Initialize the statistics variables. */ StatInit(&stat); /* ONLY PERFORM THE LU DECOMPOSITION */ B.ncol = 0; /* Indicate not to solve the system */ zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); printf("LU factorization: zgssvx() returns info %d\n", info); if ( info == 0 || info == n+1 ) { if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg); if ( options.ConditionNumber ) printf("Recip. condition number = %e\n", rcond); Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); fflush(stdout); } else if ( info > 0 && lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); /* ------------------------------------------------------------ NOW WE SOLVE THE LINEAR SYSTEM USING THE FACTORED FORM OF A. ------------------------------------------------------------*/ options.Fact = FACTORED; /* Indicate the factored form of A is supplied. */ B.ncol = nrhs; /* Set the number of right-hand side */ /* Initialize the statistics variables. */ StatInit(&stat); zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); printf("Triangular solve: zgssvx() returns info %d\n", info); if ( info == 0 || info == n+1 ) { /* This is how you could access the solution matrix. */ doublecomplex *sol = (doublecomplex*) ((DNformat*) X.Store)->nzval; if ( options.IterRefine ) { printf("Iterative Refinement:\n"); printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR"); for (i = 0; i < nrhs; ++i) printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]); } fflush(stdout); } else if ( info > 0 && lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); SUPERLU_FREE (rhsb); SUPERLU_FREE (rhsx); SUPERLU_FREE (xact); SUPERLU_FREE (etree); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); SUPERLU_FREE (R); SUPERLU_FREE (C); SUPERLU_FREE (ferr); SUPERLU_FREE (berr); Destroy_CompCol_Matrix(&A); Destroy_SuperMatrix_Store(&B); Destroy_SuperMatrix_Store(&X); if ( lwork == 0 ) { Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } else if ( lwork > 0 ) { SUPERLU_FREE(work); } #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Exit main()"); #endif }
main(int argc, char *argv[]) { SuperMatrix A; NCformat *Astore; doublecomplex *a; int *asub, *xa; int *perm_r; /* row permutations from partial pivoting */ int *perm_c; /* column permutation vector */ SuperMatrix L; /* factor L */ SCformat *Lstore; SuperMatrix U; /* factor U */ NCformat *Ustore; SuperMatrix B; int nrhs, ldx, info, panel_size, m, n, nnz, permc_spec; char trans[1]; doublecomplex *xact, *rhs; mem_usage_t mem_usage; nrhs = 1; *trans = 'N'; zreadhb(&m, &n, &nnz, &a, &asub, &xa); zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); if ( !(rhs = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[]."); zCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); ldx = n; zGenXtrue(n, nrhs, xact, ldx); zFillRHS(trans, nrhs, xact, ldx, &A, &B); if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[]."); if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[]."); /* * Get column permutation vector perm_c[], according to permc_spec: * permc_spec = 0: natural ordering * permc_spec = 1: minimum degree on structure of A'*A * permc_spec = 2: minimum degree on structure of A'+A * permc_spec = 3: approximate minimum degree for unsymmetric matrices */ permc_spec = 1; get_perm_c(permc_spec, &A, perm_c); panel_size = sp_ienv(1); zgssv(&A, perm_c, perm_r, &L, &U, &B, &info); if ( info == 0 ) { zinf_norm_error(nrhs, &B, xact); /* Inf. norm of the error */ Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); zQuerySpace(&L, &U, panel_size, &mem_usage); printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6, mem_usage.expansions); } else { printf("zgssv() error returns INFO= %d\n", info); if ( info <= n ) { /* factorization completes */ zQuerySpace(&L, &U, panel_size, &mem_usage); printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6, mem_usage.expansions); } } SUPERLU_FREE (rhs); SUPERLU_FREE (xact); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); Destroy_CompCol_Matrix(&A); Destroy_SuperMatrix_Store(&B); Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); }
void zgstrs (trans_t trans, SuperMatrix *L, SuperMatrix *U, int *perm_c, int *perm_r, SuperMatrix *B, SuperLUStat_t *stat, int *info) { #ifdef _CRAY _fcd ftcs1, ftcs2, ftcs3, ftcs4; #endif int incx = 1, incy = 1; #ifdef USE_VENDOR_BLAS doublecomplex alpha = {1.0, 0.0}, beta = {1.0, 0.0}; doublecomplex *work_col; #endif doublecomplex temp_comp; DNformat *Bstore; doublecomplex *Bmat; SCformat *Lstore; NCformat *Ustore; doublecomplex *Lval, *Uval; int fsupc, nrow, nsupr, nsupc, luptr, istart, irow; int i, j, k, iptr, jcol, n, ldb, nrhs; doublecomplex *work, *rhs_work, *soln; flops_t solve_ops; void zprint_soln(); /* Test input parameters ... */ *info = 0; Bstore = B->Store; ldb = Bstore->lda; nrhs = B->ncol; if ( trans != NOTRANS && trans != TRANS && trans != CONJ ) *info = -1; else if ( L->nrow != L->ncol || L->nrow < 0 || L->Stype != SLU_SC || L->Dtype != SLU_Z || L->Mtype != SLU_TRLU ) *info = -2; else if ( U->nrow != U->ncol || U->nrow < 0 || U->Stype != SLU_NC || U->Dtype != SLU_Z || U->Mtype != SLU_TRU ) *info = -3; else if ( ldb < SUPERLU_MAX(0, L->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_Z || B->Mtype != SLU_GE ) *info = -6; if ( *info ) { i = -(*info); input_error("zgstrs", &i); return; } n = L->nrow; work = doublecomplexCalloc(n * nrhs); if ( !work ) ABORT("Malloc fails for local work[]."); soln = doublecomplexMalloc(n); if ( !soln ) ABORT("Malloc fails for local soln[]."); Bmat = Bstore->nzval; Lstore = L->Store; Lval = Lstore->nzval; Ustore = U->Store; Uval = Ustore->nzval; solve_ops = 0; if ( trans == NOTRANS ) { /* Permute right hand sides to form Pr*B */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } /* Forward solve PLy=Pb. */ for (k = 0; k <= Lstore->nsuper; k++) { fsupc = L_FST_SUPC(k); istart = L_SUB_START(fsupc); nsupr = L_SUB_START(fsupc+1) - istart; nsupc = L_FST_SUPC(k+1) - fsupc; nrow = nsupr - nsupc; solve_ops += 4 * nsupc * (nsupc - 1) * nrhs; solve_ops += 8 * nrow * nsupc * nrhs; if ( nsupc == 1 ) { for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; luptr = L_NZ_START(fsupc); for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){ irow = L_SUB(iptr); ++luptr; zz_mult(&temp_comp, &rhs_work[fsupc], &Lval[luptr]); z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp); } } } else { luptr = L_NZ_START(fsupc); #ifdef USE_VENDOR_BLAS #ifdef _CRAY ftcs1 = _cptofcd("L", strlen("L")); ftcs2 = _cptofcd("N", strlen("N")); ftcs3 = _cptofcd("U", strlen("U")); CTRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); CGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha, &Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb, &beta, &work[0], &n ); #else ztrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); zgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha, &Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb, &beta, &work[0], &n ); #endif for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; work_col = &work[j*n]; iptr = istart + nsupc; for (i = 0; i < nrow; i++) { irow = L_SUB(iptr); z_sub(&rhs_work[irow], &rhs_work[irow], &work_col[i]); work_col[i].r = 0.0; work_col[i].i = 0.0; iptr++; } } #else for (j = 0; j < nrhs; j++) { rhs_work = &Bmat[j*ldb]; zlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]); zmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc], &rhs_work[fsupc], &work[0] ); iptr = istart + nsupc; for (i = 0; i < nrow; i++) { irow = L_SUB(iptr); z_sub(&rhs_work[irow], &rhs_work[irow], &work[i]); work[i].r = 0.; work[i].i = 0.; iptr++; } } #endif } /* else ... */ } /* for L-solve */ #ifdef DEBUG printf("After L-solve: y=\n"); zprint_soln(n, nrhs, Bmat); #endif /* * Back solve Ux=y. */ for (k = Lstore->nsuper; k >= 0; k--) { fsupc = L_FST_SUPC(k); istart = L_SUB_START(fsupc); nsupr = L_SUB_START(fsupc+1) - istart; nsupc = L_FST_SUPC(k+1) - fsupc; luptr = L_NZ_START(fsupc); solve_ops += 4 * nsupc * (nsupc + 1) * nrhs; if ( nsupc == 1 ) { rhs_work = &Bmat[0]; for (j = 0; j < nrhs; j++) { z_div(&rhs_work[fsupc], &rhs_work[fsupc], &Lval[luptr]); rhs_work += ldb; } } else { #ifdef USE_VENDOR_BLAS #ifdef _CRAY ftcs1 = _cptofcd("L", strlen("L")); ftcs2 = _cptofcd("U", strlen("U")); ftcs3 = _cptofcd("N", strlen("N")); CTRSM( ftcs1, ftcs2, ftcs3, ftcs3, &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); #else ztrsm_("L", "U", "N", "N", &nsupc, &nrhs, &alpha, &Lval[luptr], &nsupr, &Bmat[fsupc], &ldb); #endif #else for (j = 0; j < nrhs; j++) zusolve ( nsupr, nsupc, &Lval[luptr], &Bmat[fsupc+j*ldb] ); #endif } for (j = 0; j < nrhs; ++j) { rhs_work = &Bmat[j*ldb]; for (jcol = fsupc; jcol < fsupc + nsupc; jcol++) { solve_ops += 8*(U_NZ_START(jcol+1) - U_NZ_START(jcol)); for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1); i++ ){ irow = U_SUB(i); zz_mult(&temp_comp, &rhs_work[jcol], &Uval[i]); z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp); } } } } /* for U-solve */ #ifdef DEBUG printf("After U-solve: x=\n"); zprint_soln(n, nrhs, Bmat); #endif /* Compute the final solution X := Pc*X. */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[k] = rhs_work[perm_c[k]]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } stat->ops[SOLVE] = solve_ops; } else { /* Solve A'*X=B or CONJ(A)*X=B */ /* Permute right hand sides to form Pc'*B. */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[perm_c[k]] = rhs_work[k]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } stat->ops[SOLVE] = 0; if (trans == TRANS) { for (k = 0; k < nrhs; ++k) { /* Multiply by inv(U'). */ sp_ztrsv("U", "T", "N", L, U, &Bmat[k*ldb], stat, info); /* Multiply by inv(L'). */ sp_ztrsv("L", "T", "U", L, U, &Bmat[k*ldb], stat, info); } } else { /* trans == CONJ */ for (k = 0; k < nrhs; ++k) { /* Multiply by conj(inv(U')). */ sp_ztrsv("U", "C", "N", L, U, &Bmat[k*ldb], stat, info); /* Multiply by conj(inv(L')). */ sp_ztrsv("L", "C", "U", L, U, &Bmat[k*ldb], stat, info); } } /* Compute the final solution X := Pr'*X (=inv(Pr)*X) */ for (i = 0; i < nrhs; i++) { rhs_work = &Bmat[i*ldb]; for (k = 0; k < n; k++) soln[k] = rhs_work[perm_r[k]]; for (k = 0; k < n; k++) rhs_work[k] = soln[k]; } } SUPERLU_FREE(work); SUPERLU_FREE(soln); }
void zgsrfs(char *trans, SuperMatrix *A, SuperMatrix *L, SuperMatrix *U, int *perm_r, int *perm_c, char *equed, double *R, double *C, SuperMatrix *B, SuperMatrix *X, double *ferr, double *berr, int *info) { /* * Purpose * ======= * * ZGSRFS improves the computed solution to a system of linear * equations and provides error bounds and backward error estimates for * the solution. * * If equilibration was performed, the system becomes: * (diag(R)*A_original*diag(C)) * X = diag(R)*B_original. * * See supermatrix.h for the definition of 'SuperMatrix' structure. * * Arguments * ========= * * trans (input) char* * Specifies the form of the system of equations: * = 'N': A * X = B (No transpose) * = 'T': A**T * X = B (Transpose) * = 'C': A**H * X = B (Conjugate transpose = Transpose) * * A (input) SuperMatrix* * The original matrix A in the system, or the scaled A if * equilibration was done. The type of A can be: * Stype = NC, Dtype = _Z, Mtype = GE. * * L (input) SuperMatrix* * The factor L from the factorization Pr*A*Pc=L*U. Use * compressed row subscripts storage for supernodes, * i.e., L has types: Stype = SC, Dtype = _Z, Mtype = TRLU. * * U (input) SuperMatrix* * The factor U from the factorization Pr*A*Pc=L*U as computed by * zgstrf(). Use column-wise storage scheme, * i.e., U has types: Stype = NC, Dtype = _Z, Mtype = TRU. * * perm_r (input) int*, dimension (A->nrow) * Row permutation vector, which defines the permutation matrix Pr; * perm_r[i] = j means row i of A is in position j in Pr*A. * * perm_c (input) int*, dimension (A->ncol) * Column permutation vector, which defines the * permutation matrix Pc; perm_c[i] = j means column i of A is * in position j in A*Pc. * * equed (input) Specifies the form of equilibration that was done. * = 'N': No equilibration. * = 'R': Row equilibration, i.e., A was premultiplied by diag(R). * = 'C': Column equilibration, i.e., A was postmultiplied by * diag(C). * = 'B': Both row and column equilibration, i.e., A was replaced * by diag(R)*A*diag(C). * * R (input) double*, dimension (A->nrow) * The row scale factors for A. * If equed = 'R' or 'B', A is premultiplied by diag(R). * If equed = 'N' or 'C', R is not accessed. * * C (input) double*, dimension (A->ncol) * The column scale factors for A. * If equed = 'C' or 'B', A is postmultiplied by diag(C). * If equed = 'N' or 'R', C is not accessed. * * B (input) SuperMatrix* * B has types: Stype = DN, Dtype = _Z, Mtype = GE. * The right hand side matrix B. * if equed = 'R' or 'B', B is premultiplied by diag(R). * * X (input/output) SuperMatrix* * X has types: Stype = DN, Dtype = _Z, Mtype = GE. * On entry, the solution matrix X, as computed by zgstrs(). * On exit, the improved solution matrix X. * if *equed = 'C' or 'B', X should be premultiplied by diag(C) * in order to obtain the solution to the original system. * * FERR (output) double*, dimension (B->ncol) * The estimated forward error bound for each solution vector * X(j) (the j-th column of the solution matrix X). * If XTRUE is the true solution corresponding to X(j), FERR(j) * is an estimated upper bound for the magnitude of the largest * element in (X(j) - XTRUE) divided by the magnitude of the * largest element in X(j). The estimate is as reliable as * the estimate for RCOND, and is almost always a slight * overestimate of the true error. * * BERR (output) double*, dimension (B->ncol) * The componentwise relative backward error of each solution * vector X(j) (i.e., the smallest relative change in * any element of A or B that makes X(j) an exact solution). * * info (output) int* * = 0: successful exit * < 0: if INFO = -i, the i-th argument had an illegal value * * Internal Parameters * =================== * * ITMAX is the maximum number of steps of iterative refinement. * */ #define ITMAX 5 /* Table of constant values */ int ione = 1; doublecomplex ndone = {-1., 0.}; doublecomplex done = {1., 0.}; /* Local variables */ NCformat *Astore; doublecomplex *Aval; SuperMatrix Bjcol; DNformat *Bstore, *Xstore, *Bjcol_store; doublecomplex *Bmat, *Xmat, *Bptr, *Xptr; int kase; double safe1, safe2; int i, j, k, irow, nz, count, notran, rowequ, colequ; int ldb, ldx, nrhs; double s, xk, lstres, eps, safmin; char transt[1]; doublecomplex *work; double *rwork; int *iwork; extern double dlamch_(char *); extern int zlacon_(int *, doublecomplex *, doublecomplex *, double *, int *); #ifdef _CRAY extern int CCOPY(int *, doublecomplex *, int *, doublecomplex *, int *); extern int CSAXPY(int *, doublecomplex *, doublecomplex *, int *, doublecomplex *, int *); #else extern int zcopy_(int *, doublecomplex *, int *, doublecomplex *, int *); extern int zaxpy_(int *, doublecomplex *, doublecomplex *, int *, doublecomplex *, int *); #endif Astore = A->Store; Aval = Astore->nzval; Bstore = B->Store; Xstore = X->Store; Bmat = Bstore->nzval; Xmat = Xstore->nzval; ldb = Bstore->lda; ldx = Xstore->lda; nrhs = B->ncol; /* Test the input parameters */ *info = 0; notran = lsame_(trans, "N"); if ( !notran && !lsame_(trans, "T") && !lsame_(trans, "C")) *info = -1; else if ( A->nrow != A->ncol || A->nrow < 0 || A->Stype != NC || A->Dtype != _Z || A->Mtype != GE ) *info = -2; else if ( L->nrow != L->ncol || L->nrow < 0 || L->Stype != SC || L->Dtype != _Z || L->Mtype != TRLU ) *info = -3; else if ( U->nrow != U->ncol || U->nrow < 0 || U->Stype != NC || U->Dtype != _Z || U->Mtype != TRU ) *info = -4; else if ( ldb < MAX(0, A->nrow) || B->Stype != DN || B->Dtype != _Z || B->Mtype != GE ) *info = -10; else if ( ldx < MAX(0, A->nrow) || X->Stype != DN || X->Dtype != _Z || X->Mtype != GE ) *info = -11; if (*info != 0) { i = -(*info); xerbla_("zgsrfs", &i); return; } /* Quick return if possible */ if ( A->nrow == 0 || nrhs == 0) { for (j = 0; j < nrhs; ++j) { ferr[j] = 0.; berr[j] = 0.; } return; } rowequ = lsame_(equed, "R") || lsame_(equed, "B"); colequ = lsame_(equed, "C") || lsame_(equed, "B"); /* Allocate working space */ work = doublecomplexMalloc(2*A->nrow); rwork = (double *) SUPERLU_MALLOC( A->nrow * sizeof(double) ); iwork = intMalloc(A->nrow); if ( !work || !rwork || !iwork ) ABORT("Malloc fails for work/rwork/iwork."); if ( notran ) { *(unsigned char *)transt = 'T'; } else { *(unsigned char *)transt = 'N'; } /* NZ = maximum number of nonzero elements in each row of A, plus 1 */ nz = A->ncol + 1; eps = dlamch_("Epsilon"); safmin = dlamch_("Safe minimum"); safe1 = nz * safmin; safe2 = safe1 / eps; /* Compute the number of nonzeros in each row (or column) of A */ for (i = 0; i < A->nrow; ++i) iwork[i] = 0; if ( notran ) { for (k = 0; k < A->ncol; ++k) for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) ++iwork[Astore->rowind[i]]; } else { for (k = 0; k < A->ncol; ++k) iwork[k] = Astore->colptr[k+1] - Astore->colptr[k]; } /* Copy one column of RHS B into Bjcol. */ Bjcol.Stype = B->Stype; Bjcol.Dtype = B->Dtype; Bjcol.Mtype = B->Mtype; Bjcol.nrow = B->nrow; Bjcol.ncol = 1; Bjcol.Store = (void *) SUPERLU_MALLOC( sizeof(DNformat) ); if ( !Bjcol.Store ) ABORT("SUPERLU_MALLOC fails for Bjcol.Store"); Bjcol_store = Bjcol.Store; Bjcol_store->lda = ldb; Bjcol_store->nzval = work; /* address aliasing */ /* Do for each right hand side ... */ for (j = 0; j < nrhs; ++j) { count = 0; lstres = 3.; Bptr = &Bmat[j*ldb]; Xptr = &Xmat[j*ldx]; while (1) { /* Loop until stopping criterion is satisfied. */ /* Compute residual R = B - op(A) * X, where op(A) = A, A**T, or A**H, depending on TRANS. */ #ifdef _CRAY CCOPY(&A->nrow, Bptr, &ione, work, &ione); #else zcopy_(&A->nrow, Bptr, &ione, work, &ione); #endif sp_zgemv(trans, ndone, A, Xptr, ione, done, work, ione); /* Compute componentwise relative backward error from formula max(i) ( abs(R(i)) / ( abs(op(A))*abs(X) + abs(B) )(i) ) where abs(Z) is the componentwise absolute value of the matrix or vector Z. If the i-th component of the denominator is less than SAFE2, then SAFE1 is added to the i-th component of the numerator and denominator before dividing. */ for (i = 0; i < A->nrow; ++i) rwork[i] = z_abs1( &Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if (notran) { for (k = 0; k < A->ncol; ++k) { xk = z_abs1( &Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += z_abs1(&Aval[i]) * xk; } } else { for (k = 0; k < A->ncol; ++k) { s = 0.; for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) { irow = Astore->rowind[i]; s += z_abs1(&Aval[i]) * z_abs1(&Xptr[irow]); } rwork[k] += s; } } s = 0.; for (i = 0; i < A->nrow; ++i) { if (rwork[i] > safe2) s = MAX( s, z_abs1(&work[i]) / rwork[i] ); else s = MAX( s, (z_abs1(&work[i]) + safe1) / (rwork[i] + safe1) ); } berr[j] = s; /* Test stopping criterion. Continue iterating if 1) The residual BERR(J) is larger than machine epsilon, and 2) BERR(J) decreased by at least a factor of 2 during the last iteration, and 3) At most ITMAX iterations tried. */ if (berr[j] > eps && berr[j] * 2. <= lstres && count < ITMAX) { /* Update solution and try again. */ zgstrs (trans, L, U, perm_r, perm_c, &Bjcol, info); #ifdef _CRAY CAXPY(&A->nrow, &done, work, &ione, &Xmat[j*ldx], &ione); #else zaxpy_(&A->nrow, &done, work, &ione, &Xmat[j*ldx], &ione); #endif lstres = berr[j]; ++count; } else { break; } } /* end while */ /* Bound error from formula: norm(X - XTRUE) / norm(X) .le. FERR = norm( abs(inv(op(A)))* ( abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) ))) / norm(X) where norm(Z) is the magnitude of the largest component of Z inv(op(A)) is the inverse of op(A) abs(Z) is the componentwise absolute value of the matrix or vector Z NZ is the maximum number of nonzeros in any row of A, plus 1 EPS is machine epsilon The i-th component of abs(R)+NZ*EPS*(abs(op(A))*abs(X)+abs(B)) is incremented by SAFE1 if the i-th component of abs(op(A))*abs(X) + abs(B) is less than SAFE2. Use ZLACON to estimate the infinity-norm of the matrix inv(op(A)) * diag(W), where W = abs(R) + NZ*EPS*( abs(op(A))*abs(X)+abs(B) ))) */ for (i = 0; i < A->nrow; ++i) rwork[i] = z_abs1( &Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if ( notran ) { for (k = 0; k < A->ncol; ++k) { xk = z_abs1( &Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += z_abs1(&Aval[i]) * xk; } } else { for (k = 0; k < A->ncol; ++k) { s = 0.; for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) { irow = Astore->rowind[i]; xk = z_abs1( &Xptr[irow] ); s += z_abs1(&Aval[i]) * xk; } rwork[k] += s; } } for (i = 0; i < A->nrow; ++i) if (rwork[i] > safe2) rwork[i] = z_abs(&work[i]) + (iwork[i]+1)*eps*rwork[i]; else rwork[i] = z_abs(&work[i])+(iwork[i]+1)*eps*rwork[i]+safe1; kase = 0; do { zlacon_(&A->nrow, &work[A->nrow], work, &ferr[j], &kase); if (kase == 0) break; if (kase == 1) { /* Multiply by diag(W)*inv(op(A)**T)*(diag(C) or diag(R)). */ if ( notran && colequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], C[i]); } else if ( !notran && rowequ ) for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], R[i]); } zgstrs (transt, L, U, perm_r, perm_c, &Bjcol, info); for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], rwork[i]); } } else { /* Multiply by (diag(C) or diag(R))*inv(op(A))*diag(W). */ for (i = 0; i < A->nrow; ++i) { zd_mult(&work[i], &work[i], rwork[i]); } zgstrs (trans, L, U, perm_r, perm_c, &Bjcol, info); if ( notran && colequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], C[i]); } else if ( !notran && rowequ ) for (i = 0; i < A->ncol; ++i) { zd_mult(&work[i], &work[i], R[i]); } } } while ( kase != 0 ); /* Normalize error. */ lstres = 0.; if ( notran && colequ ) { for (i = 0; i < A->nrow; ++i) lstres = MAX( lstres, C[i] * z_abs1( &Xptr[i]) ); } else if ( !notran && rowequ ) { for (i = 0; i < A->nrow; ++i) lstres = MAX( lstres, R[i] * z_abs1( &Xptr[i]) ); } else { for (i = 0; i < A->nrow; ++i) lstres = MAX( lstres, z_abs1( &Xptr[i]) ); } if ( lstres != 0. ) ferr[j] /= lstres; } /* for each RHS j ... */ SUPERLU_FREE(work); SUPERLU_FREE(rwork); SUPERLU_FREE(iwork); SUPERLU_FREE(Bjcol.Store); return; } /* zgsrfs */
main(int argc, char *argv[]) { /* * Purpose * ======= * * ZDRIVE is the main test program for the DOUBLE COMPLEX linear * equation driver routines ZGSSV and ZGSSVX. * * The program is invoked by a shell script file -- ztest.csh. * The output from the tests are written into a file -- ztest.out. * * ===================================================================== */ doublecomplex *a, *a_save; int *asub, *asub_save; int *xa, *xa_save; SuperMatrix A, B, X, L, U; SuperMatrix ASAV, AC; mem_usage_t mem_usage; int *perm_r; /* row permutation from partial pivoting */ int *perm_c, *pc_save; /* column permutation */ int *etree; doublecomplex zero = {0.0, 0.0}; double *R, *C; double *ferr, *berr; double *rwork; doublecomplex *wwork; void *work; int info, lwork, nrhs, panel_size, relax; int m, n, nnz; doublecomplex *xact; doublecomplex *rhsb, *solx, *bsav; int ldb, ldx; double rpg, rcond; int i, j, k1; double rowcnd, colcnd, amax; int maxsuper, rowblk, colblk; int prefact, nofact, equil, iequed; int nt, nrun, nfail, nerrs, imat, fimat, nimat; int nfact, ifact, itran; int kl, ku, mode, lda; int zerot, izero, ioff; double u; double anorm, cndnum; doublecomplex *Afull; double result[NTESTS]; superlu_options_t options; fact_t fact; trans_t trans; SuperLUStat_t stat; static char matrix_type[8]; static char equed[1], path[4], sym[1], dist[1]; /* Fixed set of parameters */ int iseed[] = {1988, 1989, 1990, 1991}; static char equeds[] = {'N', 'R', 'C', 'B'}; static fact_t facts[] = {FACTORED, DOFACT, SamePattern, SamePattern_SameRowPerm}; static trans_t transs[] = {NOTRANS, TRANS, CONJ}; /* Some function prototypes */ extern int zgst01(int, int, SuperMatrix *, SuperMatrix *, SuperMatrix *, int *, int *, double *); extern int zgst02(trans_t, int, int, int, SuperMatrix *, doublecomplex *, int, doublecomplex *, int, double *resid); extern int zgst04(int, int, doublecomplex *, int, doublecomplex *, int, double rcond, double *resid); extern int zgst07(trans_t, int, int, SuperMatrix *, doublecomplex *, int, doublecomplex *, int, doublecomplex *, int, double *, double *, double *); extern int zlatb4_(char *, int *, int *, int *, char *, int *, int *, double *, int *, double *, char *); extern int zlatms_(int *, int *, char *, int *, char *, double *d, int *, double *, double *, int *, int *, char *, doublecomplex *, int *, doublecomplex *, int *); extern int sp_zconvert(int, int, doublecomplex *, int, int, int, doublecomplex *a, int *, int *, int *); /* Executable statements */ strcpy(path, "ZGE"); nrun = 0; nfail = 0; nerrs = 0; /* Defaults */ lwork = 0; n = 1; nrhs = 1; panel_size = sp_ienv(1); relax = sp_ienv(2); u = 1.0; strcpy(matrix_type, "LA"); parse_command_line(argc, argv, matrix_type, &n, &panel_size, &relax, &nrhs, &maxsuper, &rowblk, &colblk, &lwork, &u); if ( lwork > 0 ) { work = SUPERLU_MALLOC(lwork); if ( !work ) { fprintf(stderr, "expert: cannot allocate %d bytes\n", lwork); exit (-1); } } /* Set the default input options. */ set_default_options(&options); options.DiagPivotThresh = u; options.PrintStat = NO; options.PivotGrowth = YES; options.ConditionNumber = YES; options.IterRefine = DOUBLE; if ( strcmp(matrix_type, "LA") == 0 ) { /* Test LAPACK matrix suite. */ m = n; lda = SUPERLU_MAX(n, 1); nnz = n * n; /* upper bound */ fimat = 1; nimat = NTYPES; Afull = doublecomplexCalloc(lda * n); zallocateA(n, nnz, &a, &asub, &xa); } else { /* Read a sparse matrix */ fimat = nimat = 0; zreadhb(&m, &n, &nnz, &a, &asub, &xa); } zallocateA(n, nnz, &a_save, &asub_save, &xa_save); rhsb = doublecomplexMalloc(m * nrhs); bsav = doublecomplexMalloc(m * nrhs); solx = doublecomplexMalloc(n * nrhs); ldb = m; ldx = n; zCreate_Dense_Matrix(&B, m, nrhs, rhsb, ldb, SLU_DN, SLU_Z, SLU_GE); zCreate_Dense_Matrix(&X, n, nrhs, solx, ldx, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); etree = intMalloc(n); perm_r = intMalloc(n); perm_c = intMalloc(n); pc_save = intMalloc(n); R = (double *) SUPERLU_MALLOC(m*sizeof(double)); C = (double *) SUPERLU_MALLOC(n*sizeof(double)); ferr = (double *) SUPERLU_MALLOC(nrhs*sizeof(double)); berr = (double *) SUPERLU_MALLOC(nrhs*sizeof(double)); j = SUPERLU_MAX(m,n) * SUPERLU_MAX(4,nrhs); rwork = (double *) SUPERLU_MALLOC(j*sizeof(double)); for (i = 0; i < j; ++i) rwork[i] = 0.; if ( !R ) ABORT("SUPERLU_MALLOC fails for R"); if ( !C ) ABORT("SUPERLU_MALLOC fails for C"); if ( !ferr ) ABORT("SUPERLU_MALLOC fails for ferr"); if ( !berr ) ABORT("SUPERLU_MALLOC fails for berr"); if ( !rwork ) ABORT("SUPERLU_MALLOC fails for rwork"); wwork = doublecomplexCalloc( SUPERLU_MAX(m,n) * SUPERLU_MAX(4,nrhs) ); for (i = 0; i < n; ++i) perm_c[i] = pc_save[i] = i; options.ColPerm = MY_PERMC; for (imat = fimat; imat <= nimat; ++imat) { /* All matrix types */ if ( imat ) { /* Skip types 5, 6, or 7 if the matrix size is too small. */ zerot = (imat >= 5 && imat <= 7); if ( zerot && n < imat-4 ) continue; /* Set up parameters with ZLATB4 and generate a test matrix with ZLATMS. */ zlatb4_(path, &imat, &n, &n, sym, &kl, &ku, &anorm, &mode, &cndnum, dist); zlatms_(&n, &n, dist, iseed, sym, &rwork[0], &mode, &cndnum, &anorm, &kl, &ku, "No packing", Afull, &lda, &wwork[0], &info); if ( info ) { printf(FMT3, "ZLATMS", info, izero, n, nrhs, imat, nfail); continue; } /* For types 5-7, zero one or more columns of the matrix to test that INFO is returned correctly. */ if ( zerot ) { if ( imat == 5 ) izero = 1; else if ( imat == 6 ) izero = n; else izero = n / 2 + 1; ioff = (izero - 1) * lda; if ( imat < 7 ) { for (i = 0; i < n; ++i) Afull[ioff + i] = zero; } else { for (j = 0; j < n - izero + 1; ++j) for (i = 0; i < n; ++i) Afull[ioff + i + j*lda] = zero; } } else { izero = 0; } /* Convert to sparse representation. */ sp_zconvert(n, n, Afull, lda, kl, ku, a, asub, xa, &nnz); } else { izero = 0; zerot = 0; } zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); /* Save a copy of matrix A in ASAV */ zCreate_CompCol_Matrix(&ASAV, m, n, nnz, a_save, asub_save, xa_save, SLU_NC, SLU_Z, SLU_GE); zCopy_CompCol_Matrix(&A, &ASAV); /* Form exact solution. */ zGenXtrue(n, nrhs, xact, ldx); StatInit(&stat); for (iequed = 0; iequed < 4; ++iequed) { *equed = equeds[iequed]; if (iequed == 0) nfact = 4; else nfact = 1; /* Only test factored, pre-equilibrated matrix */ for (ifact = 0; ifact < nfact; ++ifact) { fact = facts[ifact]; options.Fact = fact; for (equil = 0; equil < 2; ++equil) { options.Equil = equil; prefact = ( options.Fact == FACTORED || options.Fact == SamePattern_SameRowPerm ); /* Need a first factor */ nofact = (options.Fact != FACTORED); /* Not factored */ /* Restore the matrix A. */ zCopy_CompCol_Matrix(&ASAV, &A); if ( zerot ) { if ( prefact ) continue; } else if ( options.Fact == FACTORED ) { if ( equil || iequed ) { /* Compute row and column scale factors to equilibrate matrix A. */ zgsequ(&A, R, C, &rowcnd, &colcnd, &amax, &info); /* Force equilibration. */ if ( !info && n > 0 ) { if ( lsame_(equed, "R") ) { rowcnd = 0.; colcnd = 1.; } else if ( lsame_(equed, "C") ) { rowcnd = 1.; colcnd = 0.; } else if ( lsame_(equed, "B") ) { rowcnd = 0.; colcnd = 0.; } } /* Equilibrate the matrix. */ zlaqgs(&A, R, C, rowcnd, colcnd, amax, equed); } } if ( prefact ) { /* Need a factor for the first time */ /* Save Fact option. */ fact = options.Fact; options.Fact = DOFACT; /* Preorder the matrix, obtain the column etree. */ sp_preorder(&options, &A, perm_c, etree, &AC); /* Factor the matrix AC. */ zgstrf(&options, &AC, relax, panel_size, etree, work, lwork, perm_c, perm_r, &L, &U, &stat, &info); if ( info ) { printf("** First factor: info %d, equed %c\n", info, *equed); if ( lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); exit(0); } } Destroy_CompCol_Permuted(&AC); /* Restore Fact option. */ options.Fact = fact; } /* if .. first time factor */ for (itran = 0; itran < NTRAN; ++itran) { trans = transs[itran]; options.Trans = trans; /* Restore the matrix A. */ zCopy_CompCol_Matrix(&ASAV, &A); /* Set the right hand side. */ zFillRHS(trans, nrhs, xact, ldx, &A, &B); zCopy_Dense_Matrix(m, nrhs, rhsb, ldb, bsav, ldb); /*---------------- * Test zgssv *----------------*/ if ( options.Fact == DOFACT && itran == 0) { /* Not yet factored, and untransposed */ zCopy_Dense_Matrix(m, nrhs, rhsb, ldb, solx, ldx); zgssv(&options, &A, perm_c, perm_r, &L, &U, &X, &stat, &info); if ( info && info != izero ) { printf(FMT3, "zgssv", info, izero, n, nrhs, imat, nfail); } else { /* Reconstruct matrix from factors and compute residual. */ zgst01(m, n, &A, &L, &U, perm_c, perm_r, &result[0]); nt = 1; if ( izero == 0 ) { /* Compute residual of the computed solution. */ zCopy_Dense_Matrix(m, nrhs, rhsb, ldb, wwork, ldb); zgst02(trans, m, n, nrhs, &A, solx, ldx, wwork,ldb, &result[1]); nt = 2; } /* Print information about the tests that did not pass the threshold. */ for (i = 0; i < nt; ++i) { if ( result[i] >= THRESH ) { printf(FMT1, "zgssv", n, i, result[i]); ++nfail; } } nrun += nt; } /* else .. info == 0 */ /* Restore perm_c. */ for (i = 0; i < n; ++i) perm_c[i] = pc_save[i]; if (lwork == 0) { Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } } /* if .. end of testing zgssv */ /*---------------- * Test zgssvx *----------------*/ /* Equilibrate the matrix if fact = FACTORED and equed = 'R', 'C', or 'B'. */ if ( options.Fact == FACTORED && (equil || iequed) && n > 0 ) { zlaqgs(&A, R, C, rowcnd, colcnd, amax, equed); } /* Solve the system and compute the condition number and error bounds using zgssvx. */ zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); if ( info && info != izero ) { printf(FMT3, "zgssvx", info, izero, n, nrhs, imat, nfail); if ( lwork == -1 ) { printf("** Estimated memory: %.0f bytes\n", mem_usage.total_needed); exit(0); } } else { if ( !prefact ) { /* Reconstruct matrix from factors and compute residual. */ zgst01(m, n, &A, &L, &U, perm_c, perm_r, &result[0]); k1 = 0; } else { k1 = 1; } if ( !info ) { /* Compute residual of the computed solution.*/ zCopy_Dense_Matrix(m, nrhs, bsav, ldb, wwork, ldb); zgst02(trans, m, n, nrhs, &ASAV, solx, ldx, wwork, ldb, &result[1]); /* Check solution from generated exact solution. */ zgst04(n, nrhs, solx, ldx, xact, ldx, rcond, &result[2]); /* Check the error bounds from iterative refinement. */ zgst07(trans, n, nrhs, &ASAV, bsav, ldb, solx, ldx, xact, ldx, ferr, berr, &result[3]); /* Print information about the tests that did not pass the threshold. */ for (i = k1; i < NTESTS; ++i) { if ( result[i] >= THRESH ) { printf(FMT2, "zgssvx", options.Fact, trans, *equed, n, imat, i, result[i]); ++nfail; } } nrun += NTESTS; } /* if .. info == 0 */ } /* else .. end of testing zgssvx */ } /* for itran ... */ if ( lwork == 0 ) { Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } } /* for equil ... */ } /* for ifact ... */ } /* for iequed ... */ #if 0 if ( !info ) { PrintPerf(&L, &U, &mem_usage, rpg, rcond, ferr, berr, equed); } #endif } /* for imat ... */ /* Print a summary of the results. */ PrintSumm("ZGE", nfail, nrun, nerrs); SUPERLU_FREE (rhsb); SUPERLU_FREE (bsav); SUPERLU_FREE (solx); SUPERLU_FREE (xact); SUPERLU_FREE (etree); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); SUPERLU_FREE (pc_save); SUPERLU_FREE (R); SUPERLU_FREE (C); SUPERLU_FREE (ferr); SUPERLU_FREE (berr); SUPERLU_FREE (rwork); SUPERLU_FREE (wwork); Destroy_SuperMatrix_Store(&B); Destroy_SuperMatrix_Store(&X); Destroy_CompCol_Matrix(&A); Destroy_CompCol_Matrix(&ASAV); if ( lwork > 0 ) { SUPERLU_FREE (work); Destroy_SuperMatrix_Store(&L); Destroy_SuperMatrix_Store(&U); } StatFree(&stat); return 0; }
int main(int argc, char *argv[]) { /* * Purpose * ======= * * The driver program ZLINSOLX2. * * This example illustrates how to use ZGSSVX to solve systems repeatedly * with the same sparsity pattern of matrix A. * In this case, the column permutation vector perm_c is computed once. * The following data structures will be reused in the subsequent call to * ZGSSVX: perm_c, etree * */ char equed[1]; yes_no_t equil; trans_t trans; SuperMatrix A, A1, L, U; SuperMatrix B, B1, X; NCformat *Astore; NCformat *Ustore; SCformat *Lstore; doublecomplex *a, *a1; int *asub, *xa, *asub1, *xa1; int *perm_r; /* row permutations from partial pivoting */ int *perm_c; /* column permutation vector */ int *etree; void *work; int info, lwork, nrhs, ldx; int i, j, m, n, nnz; doublecomplex *rhsb, *rhsb1, *rhsx, *xact; double *R, *C; double *ferr, *berr; double u, rpg, rcond; mem_usage_t mem_usage; superlu_options_t options; SuperLUStat_t stat; extern void parse_command_line(); #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Enter main()"); #endif /* Defaults */ lwork = 0; nrhs = 1; equil = YES; u = 1.0; trans = NOTRANS; /* Set the default input options: options.Fact = DOFACT; options.Equil = YES; options.ColPerm = COLAMD; options.DiagPivotThresh = 1.0; options.Trans = NOTRANS; options.IterRefine = NOREFINE; options.SymmetricMode = NO; options.PivotGrowth = NO; options.ConditionNumber = NO; options.PrintStat = YES; */ set_default_options(&options); /* Can use command line input to modify the defaults. */ parse_command_line(argc, argv, &lwork, &u, &equil, &trans); options.Equil = equil; options.DiagPivotThresh = u; options.Trans = trans; if ( lwork > 0 ) { work = SUPERLU_MALLOC(lwork); if ( !work ) { ABORT("DLINSOLX: cannot allocate work[]"); } } /* Read matrix A from a file in Harwell-Boeing format.*/ zreadhb(&m, &n, &nnz, &a, &asub, &xa); if ( !(a1 = doublecomplexMalloc(nnz)) ) ABORT("Malloc fails for a1[]."); if ( !(asub1 = intMalloc(nnz)) ) ABORT("Malloc fails for asub1[]."); if ( !(xa1 = intMalloc(n+1)) ) ABORT("Malloc fails for xa1[]."); for (i = 0; i < nnz; ++i) { a1[i] = a[i]; asub1[i] = asub[i]; } for (i = 0; i < n+1; ++i) xa1[i] = xa[i]; zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); if ( !(rhsb = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[]."); if ( !(rhsb1 = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb1[]."); if ( !(rhsx = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[]."); zCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_Z, SLU_GE); zCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_Z, SLU_GE); xact = doublecomplexMalloc(n * nrhs); ldx = n; zGenXtrue(n, nrhs, xact, ldx); zFillRHS(trans, nrhs, xact, ldx, &A, &B); for (j = 0; j < nrhs; ++j) for (i = 0; i < m; ++i) rhsb1[i+j*m] = rhsb[i+j*m]; if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[]."); if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[]."); if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[]."); if ( !(R = (double *) SUPERLU_MALLOC(A.nrow * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for R[]."); if ( !(C = (double *) SUPERLU_MALLOC(A.ncol * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for C[]."); if ( !(ferr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for ferr[]."); if ( !(berr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) ) ABORT("SUPERLU_MALLOC fails for berr[]."); /* Initialize the statistics variables. */ StatInit(&stat); /* ------------------------------------------------------------ WE SOLVE THE LINEAR SYSTEM FOR THE FIRST TIME: AX = B ------------------------------------------------------------*/ zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); printf("First system: zgssvx() returns info %d\n", info); if ( info == 0 || info == n+1 ) { /* This is how you could access the solution matrix. */ doublecomplex *sol = (doublecomplex*) ((DNformat*) X.Store)->nzval; if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg); if ( options.ConditionNumber ) printf("Recip. condition number = %e\n", rcond); Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); if ( options.IterRefine ) { printf("Iterative Refinement:\n"); printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR"); for (i = 0; i < nrhs; ++i) printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]); } fflush(stdout); } else if ( info > 0 && lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); Destroy_CompCol_Matrix(&A); Destroy_Dense_Matrix(&B); if ( lwork >= 0 ) { /* Deallocate storage associated with L and U. */ Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } /* ------------------------------------------------------------ NOW WE SOLVE ANOTHER LINEAR SYSTEM: A1*X = B1 ONLY THE SPARSITY PATTERN OF A1 IS THE SAME AS THAT OF A. ------------------------------------------------------------*/ options.Fact = SamePattern; StatInit(&stat); /* Initialize the statistics variables. */ zCreate_CompCol_Matrix(&A1, m, n, nnz, a1, asub1, xa1, SLU_NC, SLU_Z, SLU_GE); zCreate_Dense_Matrix(&B1, m, nrhs, rhsb1, m, SLU_DN, SLU_Z, SLU_GE); zgssvx(&options, &A1, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B1, &X, &rpg, &rcond, ferr, berr, &mem_usage, &stat, &info); printf("\nSecond system: zgssvx() returns info %d\n", info); if ( info == 0 || info == n+1 ) { /* This is how you could access the solution matrix. */ doublecomplex *sol = (doublecomplex*) ((DNformat*) X.Store)->nzval; if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg); if ( options.ConditionNumber ) printf("Recip. condition number = %e\n", rcond); Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf("No of nonzeros in factor L = %d\n", Lstore->nnz); printf("No of nonzeros in factor U = %d\n", Ustore->nnz); printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n); printf("L\\U MB %.3f\ttotal MB needed %.3f\n", mem_usage.for_lu/1e6, mem_usage.total_needed/1e6); if ( options.IterRefine ) { printf("Iterative Refinement:\n"); printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR"); for (i = 0; i < nrhs; ++i) printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]); } fflush(stdout); } else if ( info > 0 && lwork == -1 ) { printf("** Estimated memory: %d bytes\n", info - n); } if ( options.PrintStat ) StatPrint(&stat); StatFree(&stat); SUPERLU_FREE (xact); SUPERLU_FREE (etree); SUPERLU_FREE (perm_r); SUPERLU_FREE (perm_c); SUPERLU_FREE (R); SUPERLU_FREE (C); SUPERLU_FREE (ferr); SUPERLU_FREE (berr); Destroy_CompCol_Matrix(&A1); Destroy_Dense_Matrix(&B1); Destroy_Dense_Matrix(&X); if ( lwork == 0 ) { Destroy_SuperNode_Matrix(&L); Destroy_CompCol_Matrix(&U); } else if ( lwork > 0 ) { SUPERLU_FREE(work); } #if ( DEBUGlevel>=1 ) CHECK_MALLOC("Exit main()"); #endif }