int main(int argc, char *argv[]) { SuperMatrix A; NCformat *Astore; float *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; float *xact, *rhs; mem_usage_t mem_usage; superlu_options_t options; SuperLUStat_t stat; FILE *fp = stdin; #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. */ sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa); sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE); Astore = A.Store; printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz); nrhs = 1; if ( !(rhs = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[]."); sCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE); xact = floatMalloc(n * nrhs); ldx = n; sGenXtrue(n, nrhs, xact, ldx); sFillRHS(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); sgssv(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info); if ( info == 0 ) { /* This is how you could access the solution matrix. */ float *sol = (float*) ((DNformat*) B.Store)->nzval; /* Compute the infinity norm of the error. */ sinf_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); sQuerySpace(&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("sgssv() error returns INFO= %d\n", info); if ( info <= n ) { /* factorization completes */ sQuerySpace(&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 sgssvx(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r, int *etree, char *equed, float *R, float *C, SuperMatrix *L, SuperMatrix *U, void *work, int lwork, SuperMatrix *B, SuperMatrix *X, float *recip_pivot_growth, float *rcond, float *ferr, float *berr, mem_usage_t *mem_usage, SuperLUStat_t *stat, int *info ) { DNformat *Bstore, *Xstore; float *Bmat, *Xmat; int ldb, ldx, nrhs; SuperMatrix *AA;/* A in SLU_NC format used by the factorization routine.*/ SuperMatrix AC; /* Matrix postmultiplied by Pc */ int colequ, equil, nofact, notran, rowequ, permc_spec; trans_t trant; char norm[1]; int i, j, info1; float amax, anorm, bignum, smlnum, colcnd, rowcnd, rcmax, rcmin; int relax, panel_size; float diag_pivot_thresh; double t0; /* temporary time */ double *utime; /* External functions */ extern float slangs(char *, SuperMatrix *); Bstore = B->Store; Xstore = X->Store; Bmat = Bstore->nzval; Xmat = Xstore->nzval; ldb = Bstore->lda; ldx = Xstore->lda; nrhs = B->ncol; *info = 0; nofact = (options->Fact != FACTORED); equil = (options->Equil == YES); notran = (options->Trans == NOTRANS); if ( nofact ) { *(unsigned char *)equed = 'N'; rowequ = FALSE; colequ = FALSE; } else { rowequ = lsame_(equed, "R") || lsame_(equed, "B"); colequ = lsame_(equed, "C") || lsame_(equed, "B"); smlnum = slamch_("Safe minimum"); bignum = 1. / smlnum; } #if 0 printf("dgssvx: Fact=%4d, Trans=%4d, equed=%c\n", options->Fact, options->Trans, *equed); #endif /* Test the input parameters */ if (options->Fact != DOFACT && options->Fact != SamePattern && options->Fact != SamePattern_SameRowPerm && options->Fact != FACTORED && options->Trans != NOTRANS && options->Trans != TRANS && options->Trans != CONJ && options->Equil != NO && options->Equil != YES) *info = -1; else if ( A->nrow != A->ncol || A->nrow < 0 || (A->Stype != SLU_NC && A->Stype != SLU_NR) || A->Dtype != SLU_S || A->Mtype != SLU_GE ) *info = -2; else if (options->Fact == FACTORED && !(rowequ || colequ || lsame_(equed, "N"))) *info = -6; else { if (rowequ) { rcmin = bignum; rcmax = 0.; for (j = 0; j < A->nrow; ++j) { rcmin = SUPERLU_MIN(rcmin, R[j]); rcmax = SUPERLU_MAX(rcmax, R[j]); } if (rcmin <= 0.) *info = -7; else if ( A->nrow > 0) rowcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum); else rowcnd = 1.; } if (colequ && *info == 0) { rcmin = bignum; rcmax = 0.; for (j = 0; j < A->nrow; ++j) { rcmin = SUPERLU_MIN(rcmin, C[j]); rcmax = SUPERLU_MAX(rcmax, C[j]); } if (rcmin <= 0.) *info = -8; else if (A->nrow > 0) colcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum); else colcnd = 1.; } if (*info == 0) { if ( lwork < -1 ) *info = -12; else if ( B->ncol < 0 ) *info = -13; else if ( B->ncol > 0 ) { /* no checking if B->ncol=0 */ if ( Bstore->lda < SUPERLU_MAX(0, A->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_S || B->Mtype != SLU_GE ) *info = -13; } if ( X->ncol < 0 ) *info = -14; else if ( X->ncol > 0 ) { /* no checking if X->ncol=0 */ if ( Xstore->lda < SUPERLU_MAX(0, A->nrow) || (B->ncol != 0 && B->ncol != X->ncol) || X->Stype != SLU_DN || X->Dtype != SLU_S || X->Mtype != SLU_GE ) *info = -14; } } } if (*info != 0) { i = -(*info); xerbla_("sgssvx", &i); return; } /* Initialization for factor parameters */ panel_size = sp_ienv(1); relax = sp_ienv(2); diag_pivot_thresh = options->DiagPivotThresh; utime = stat->utime; /* Convert A to SLU_NC format when necessary. */ if ( A->Stype == SLU_NR ) { NRformat *Astore = A->Store; AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) ); sCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz, Astore->nzval, Astore->colind, Astore->rowptr, SLU_NC, A->Dtype, A->Mtype); if ( notran ) { /* Reverse the transpose argument. */ trant = TRANS; notran = 0; } else { trant = NOTRANS; notran = 1; } } else { /* A->Stype == SLU_NC */ trant = options->Trans; AA = A; } if ( nofact && equil ) { t0 = SuperLU_timer_(); /* Compute row and column scalings to equilibrate the matrix A. */ sgsequ(AA, R, C, &rowcnd, &colcnd, &amax, &info1); if ( info1 == 0 ) { /* Equilibrate matrix A. */ slaqgs(AA, R, C, rowcnd, colcnd, amax, equed); rowequ = lsame_(equed, "R") || lsame_(equed, "B"); colequ = lsame_(equed, "C") || lsame_(equed, "B"); } utime[EQUIL] = SuperLU_timer_() - t0; } if ( nofact ) { t0 = SuperLU_timer_(); /* * Gnet column permutation vector perm_c[], according to permc_spec: * permc_spec = NATURAL: natural ordering * permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A * permc_spec = MMD_ATA: minimum degree on structure of A'*A * permc_spec = COLAMD: approximate minimum degree column ordering * permc_spec = MY_PERMC: the ordering already supplied in perm_c[] */ permc_spec = options->ColPerm; if ( permc_spec != MY_PERMC && options->Fact == DOFACT ) get_perm_c(permc_spec, AA, perm_c); utime[COLPERM] = SuperLU_timer_() - t0; t0 = SuperLU_timer_(); sp_preorder(options, AA, perm_c, etree, &AC); utime[ETREE] = SuperLU_timer_() - t0; /* printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n", relax, panel_size, sp_ienv(3), sp_ienv(4)); fflush(stdout); */ /* Compute the LU factorization of A*Pc. */ t0 = SuperLU_timer_(); sgstrf(options, &AC, relax, panel_size, etree, work, lwork, perm_c, perm_r, L, U, stat, info); utime[FACT] = SuperLU_timer_() - t0; if ( lwork == -1 ) { mem_usage->total_needed = *info - A->ncol; return; } } if ( options->PivotGrowth ) { if ( *info > 0 ) { if ( *info <= A->ncol ) { /* Compute the reciprocal pivot growth factor of the leading rank-deficient *info columns of A. */ *recip_pivot_growth = sPivotGrowth(*info, AA, perm_c, L, U); } return; } /* Compute the reciprocal pivot growth factor *recip_pivot_growth. */ *recip_pivot_growth = sPivotGrowth(A->ncol, AA, perm_c, L, U); } if ( options->ConditionNumber ) { /* Estimate the reciprocal of the condition number of A. */ t0 = SuperLU_timer_(); if ( notran ) { *(unsigned char *)norm = '1'; } else { *(unsigned char *)norm = 'I'; } anorm = slangs(norm, AA); sgscon(norm, L, U, anorm, rcond, stat, info); utime[RCOND] = SuperLU_timer_() - t0; } if ( nrhs > 0 ) { /* Scale the right hand side if equilibration was performed. */ if ( notran ) { if ( rowequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < A->nrow; ++i) Bmat[i + j*ldb] *= R[i]; } } else if ( colequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < A->nrow; ++i) Bmat[i + j*ldb] *= C[i]; } /* Compute the solution matrix X. */ for (j = 0; j < nrhs; j++) /* Save a copy of the right hand sides */ for (i = 0; i < B->nrow; i++) Xmat[i + j*ldx] = Bmat[i + j*ldb]; t0 = SuperLU_timer_(); sgstrs (trant, L, U, perm_c, perm_r, X, stat, info); utime[SOLVE] = SuperLU_timer_() - t0; /* Use iterative refinement to improve the computed solution and compute error bounds and backward error estimates for it. */ t0 = SuperLU_timer_(); if ( options->IterRefine != NOREFINE ) { sgsrfs(trant, AA, L, U, perm_c, perm_r, equed, R, C, B, X, ferr, berr, stat, info); } else { for (j = 0; j < nrhs; ++j) ferr[j] = berr[j] = 1.0; } utime[REFINE] = SuperLU_timer_() - t0; /* Transform the solution matrix X to a solution of the original system. */ if ( notran ) { if ( colequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < A->nrow; ++i) Xmat[i + j*ldx] *= C[i]; } } else if ( rowequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < A->nrow; ++i) Xmat[i + j*ldx] *= R[i]; } } /* end if nrhs > 0 */ if ( options->ConditionNumber ) { /* Set INFO = A->ncol+1 if the matrix is singular to working precision. */ if ( *rcond < slamch_("E") ) *info = A->ncol + 1; } if ( nofact ) { sQuerySpace(L, U, mem_usage); Destroy_CompCol_Permuted(&AC); } if ( A->Stype == SLU_NR ) { Destroy_SuperMatrix_Store(AA); SUPERLU_FREE(AA); } }
int main ( int argc, char *argv[] ) /**********************************************************************/ /* Purpose: SUPER_LU_S2 solves a symmetric sparse system read from a file. Discussion: The sparse matrix is stored in a file using the Harwell-Boeing sparse matrix format. The file should be assigned to the standard input of this program. For instance, if the matrix is stored in the file "g10_rua.txt", the execution command might be: super_lu_s2 < g10_rua.txt Modified: 25 April 2004 Reference: James Demmel, John Gilbert, Xiaoye Li, SuperLU Users's Guide, Sections 1 and 2. Local parameters: SuperMatrix L, the computed L factor. int *perm_c, the column permutation vector. int *perm_r, the row permutations from partial pivoting. SuperMatrix U, the computed U factor. */ { SuperMatrix A; NCformat *Astore; float *a; int *asub; SuperMatrix B; int info; SuperMatrix L; int ldx; SCformat *Lstore; int m; mem_usage_t mem_usage; int n; int nnz; int nrhs; superlu_options_t options; int *perm_c; int *perm_r; float *rhs; float *sol; SuperLUStat_t stat; SuperMatrix U; NCformat *Ustore; int *xa; float *xact; /* Say hello. */ printf ( "\n" ); printf ( "SUPER_LU_S2:\n" ); printf ( " Read a symmetric sparse matrix A from standard input,\n"); printf ( " stored in Harwell-Boeing Sparse Matrix format.\n" ); printf ( "\n" ); printf ( " Solve a linear system A * X = B.\n" ); /* 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 ); /* Now we modify the default options to use the symmetric mode. */ options.SymmetricMode = YES; options.ColPerm = MMD_AT_PLUS_A; options.DiagPivotThresh = 0.001; /* Read the matrix in Harwell-Boeing format. */ sreadhb ( &m, &n, &nnz, &a, &asub, &xa ); /* Create storage for a compressed column matrix. */ sCreate_CompCol_Matrix ( &A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE ); Astore = A.Store; printf ( "\n" ); printf ( " Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz ); /* Set up the right hand side. */ nrhs = 1; rhs = floatMalloc ( m * nrhs ); if ( !rhs ) { ABORT ( " Malloc fails for rhs[]." ); } sCreate_Dense_Matrix ( &B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE ); xact = floatMalloc ( n * nrhs ); if ( !xact ) { ABORT ( " Malloc fails for rhs[]." ); } ldx = n; sGenXtrue ( n, nrhs, xact, ldx ); sFillRHS ( options.Trans, nrhs, xact, ldx, &A, &B ); perm_c = intMalloc ( n ); if ( !perm_c ) { ABORT ( "Malloc fails for perm_c[]." ); } perm_r = intMalloc ( m ); if ( !perm_r ) { ABORT ( "Malloc fails for perm_r[]." ); } /* Initialize the statistics variables. */ StatInit ( &stat ); /* Call SGSSV to factor the matrix and solve the linear system. */ sgssv ( &options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info ); if ( info == 0 ) { /* To conveniently access the solution matrix, you need to get a pointer to it. */ sol = (float*) ((DNformat*) B.Store)->nzval; /* Compute the infinity norm of the error. */ sinf_norm_error ( nrhs, &B, xact ); Lstore = (SCformat *) L.Store; Ustore = (NCformat *) U.Store; printf ( "\n" ); printf ( " Number of nonzeros in factor L = %d\n", Lstore->nnz ); printf ( " Number of nonzeros in factor U = %d\n", Ustore->nnz ); printf ( " Number of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n ); sQuerySpace ( &L, &U, &mem_usage ); printf ( "\n" ); 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 ( "\n" ); printf ( " SGSSV error returns INFO= %d\n", info ); if ( info <= n ) { sQuerySpace ( &L, &U, &mem_usage ); printf ( "\n" ); 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 ); } } if ( options.PrintStat ) { StatPrint ( &stat ); } StatFree ( &stat ); /* Free the memory. */ 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 ); /* Say goodbye. */ printf ( "\n" ); printf ( "SUPER_LU_S2:\n" ); printf ( " Normal end of execution.\n"); return 0; }