void psgssvx(int nprocs, superlumt_options_t *superlumt_options, SuperMatrix *A, int *perm_c, int *perm_r, equed_t *equed, float *R, float *C, SuperMatrix *L, SuperMatrix *U, SuperMatrix *B, SuperMatrix *X, float *recip_pivot_growth, float *rcond, float *ferr, float *berr, superlu_memusage_t *superlu_memusage, int *info) { /* * -- SuperLU MT routine (version 2.0) -- * Lawrence Berkeley National Lab, Univ. of California Berkeley, * and Xerox Palo Alto Research Center. * September 10, 2007 * * Purpose * ======= * * psgssvx() solves the system of linear equations A*X=B or A'*X=B, using * the LU factorization from sgstrf(). Error bounds on the solution and * a condition estimate are also provided. It performs the following steps: * * 1. If A is stored column-wise (A->Stype = NC): * * 1.1. If fact = EQUILIBRATE, scaling factors are computed to equilibrate * the system: * trans = NOTRANS: diag(R)*A*diag(C)*inv(diag(C))*X = diag(R)*B * trans = TRANS: (diag(R)*A*diag(C))**T *inv(diag(R))*X = diag(C)*B * trans = CONJ: (diag(R)*A*diag(C))**H *inv(diag(R))*X = diag(C)*B * Whether or not the system will be equilibrated depends on the * scaling of the matrix A, but if equilibration is used, A is * overwritten by diag(R)*A*diag(C) and B by diag(R)*B * (if trans = NOTRANS) or diag(C)*B (if trans = TRANS or CONJ). * * 1.2. Permute columns of A, forming A*Pc, where Pc is a permutation matrix * that usually preserves sparsity. * For more details of this step, see ssp_colorder.c. * * 1.3. If fact = DOFACT or EQUILIBRATE, the LU decomposition is used to * factor the matrix A (after equilibration if fact = EQUILIBRATE) as * Pr*A*Pc = L*U, with Pr determined by partial pivoting. * * 1.4. Compute the reciprocal pivot growth factor. * * 1.5. If some U(i,i) = 0, so that U is exactly singular, then the routine * returns with info = i. Otherwise, the factored form of A is used to * estimate the condition number of the matrix A. If the reciprocal of * the condition number is less than machine precision, * info = A->ncol+1 is returned as a warning, but the routine still * goes on to solve for X and computes error bounds as described below. * * 1.6. The system of equations is solved for X using the factored form * of A. * * 1.7. Iterative refinement is applied to improve the computed solution * matrix and calculate error bounds and backward error estimates * for it. * * 1.8. If equilibration was used, the matrix X is premultiplied by * diag(C) (if trans = NOTRANS) or diag(R) (if trans = TRANS or CONJ) * so that it solves the original system before equilibration. * * 2. If A is stored row-wise (A->Stype = NR), apply the above algorithm * to the tranpose of A: * * 2.1. If fact = EQUILIBRATE, scaling factors are computed to equilibrate * the system: * trans = NOTRANS:diag(R)*A'*diag(C)*inv(diag(C))*X = diag(R)*B * trans = TRANS: (diag(R)*A'*diag(C))**T *inv(diag(R))*X = diag(C)*B * trans = CONJ: (diag(R)*A'*diag(C))**H *inv(diag(R))*X = diag(C)*B * Whether or not the system will be equilibrated depends on the * scaling of the matrix A, but if equilibration is used, A' is * overwritten by diag(R)*A'*diag(C) and B by diag(R)*B * (if trans = NOTRANS) or diag(C)*B (if trans = TRANS or CONJ). * * 2.2. Permute columns of transpose(A) (rows of A), * forming transpose(A)*Pc, where Pc is a permutation matrix that * usually preserves sparsity. * For more details of this step, see ssp_colorder.c. * * 2.3. If fact = DOFACT or EQUILIBRATE, the LU decomposition is used to * factor the matrix A (after equilibration if fact = EQUILIBRATE) as * Pr*transpose(A)*Pc = L*U, with the permutation Pr determined by * partial pivoting. * * 2.4. Compute the reciprocal pivot growth factor. * * 2.5. If some U(i,i) = 0, so that U is exactly singular, then the routine * returns with info = i. Otherwise, the factored form of transpose(A) * is used to estimate the condition number of the matrix A. * If the reciprocal of the condition number is less than machine * precision, info = A->nrow+1 is returned as a warning, but the * routine still goes on to solve for X and computes error bounds * as described below. * * 2.6. The system of equations is solved for X using the factored form * of transpose(A). * * 2.7. Iterative refinement is applied to improve the computed solution * matrix and calculate error bounds and backward error estimates * for it. * * 2.8. If equilibration was used, the matrix X is premultiplied by * diag(C) (if trans = NOTRANS) or diag(R) (if trans = TRANS or CONJ) * so that it solves the original system before equilibration. * * See supermatrix.h for the definition of 'SuperMatrix' structure. * * Arguments * ========= * * nprocs (input) int * Number of processes (or threads) to be spawned and used to perform * the LU factorization by psgstrf(). There is a single thread of * control to call psgstrf(), and all threads spawned by psgstrf() * are terminated before returning from psgstrf(). * * superlumt_options (input) superlumt_options_t* * The structure defines the input parameters and data structure * to control how the LU factorization will be performed. * The following fields should be defined for this structure: * * o fact (fact_t) * Specifies whether or not the factored form of the matrix * A is supplied on entry, and if not, whether the matrix A should * be equilibrated before it is factored. * = FACTORED: On entry, L, U, perm_r and perm_c contain the * factored form of A. If equed is not NOEQUIL, the matrix A has * been equilibrated with scaling factors R and C. * A, L, U, perm_r are not modified. * = DOFACT: The matrix A will be factored, and the factors will be * stored in L and U. * = EQUILIBRATE: The matrix A will be equilibrated if necessary, * then factored into L and U. * * o trans (trans_t) * Specifies the form of the system of equations: * = NOTRANS: A * X = B (No transpose) * = TRANS: A**T * X = B (Transpose) * = CONJ: A**H * X = B (Transpose) * * o refact (yes_no_t) * Specifies whether this is first time or subsequent factorization. * = NO: this factorization is treated as the first one; * = YES: it means that a factorization was performed prior to this * one. Therefore, this factorization will re-use some * existing data structures, such as L and U storage, column * elimination tree, and the symbolic information of the * Householder matrix. * * o panel_size (int) * A panel consists of at most panel_size consecutive columns. * * o relax (int) * To control degree of relaxing supernodes. If the number * of nodes (columns) in a subtree of the elimination tree is less * than relax, this subtree is considered as one supernode, * regardless of the row structures of those columns. * * o diag_pivot_thresh (float) * Diagonal pivoting threshold. At step j of the Gaussian * elimination, if * abs(A_jj) >= diag_pivot_thresh * (max_(i>=j) abs(A_ij)), * use A_jj as pivot, else use A_ij with maximum magnitude. * 0 <= diag_pivot_thresh <= 1. The default value is 1, * corresponding to partial pivoting. * * o usepr (yes_no_t) * Whether the pivoting will use perm_r specified by the user. * = YES: use perm_r; perm_r is input, unchanged on exit. * = NO: perm_r is determined by partial pivoting, and is output. * * o drop_tol (double) (NOT IMPLEMENTED) * Drop tolerance parameter. At step j of the Gaussian elimination, * if abs(A_ij)/(max_i abs(A_ij)) < drop_tol, drop entry A_ij. * 0 <= drop_tol <= 1. The default value of drop_tol is 0, * corresponding to not dropping any entry. * * o work (void*) of size lwork * User-supplied work space and space for the output data structures. * Not referenced if lwork = 0; * * o lwork (int) * Specifies the length of work array. * = 0: allocate space internally by system malloc; * > 0: use user-supplied work array of length lwork in bytes, * returns error if space runs out. * = -1: the routine guesses the amount of space needed without * performing the factorization, and returns it in * superlu_memusage->total_needed; no other side effects. * * A (input/output) SuperMatrix* * Matrix A in A*X=B, of dimension (A->nrow, A->ncol), where * A->nrow = A->ncol. Currently, the type of A can be: * Stype = NC or NR, Dtype = _D, Mtype = GE. In the future, * more general A will be handled. * * On entry, If superlumt_options->fact = FACTORED and equed is not * NOEQUIL, then A must have been equilibrated by the scaling factors * in R and/or C. On exit, A is not modified * if superlumt_options->fact = FACTORED or DOFACT, or * if superlumt_options->fact = EQUILIBRATE and equed = NOEQUIL. * * On exit, if superlumt_options->fact = EQUILIBRATE and equed is not * NOEQUIL, A is scaled as follows: * If A->Stype = NC: * equed = ROW: A := diag(R) * A * equed = COL: A := A * diag(C) * equed = BOTH: A := diag(R) * A * diag(C). * If A->Stype = NR: * equed = ROW: transpose(A) := diag(R) * transpose(A) * equed = COL: transpose(A) := transpose(A) * diag(C) * equed = BOTH: transpose(A) := diag(R) * transpose(A) * diag(C). * * perm_c (input/output) int* * If A->Stype = NC, Column permutation vector of size A->ncol, * which defines the permutation matrix Pc; perm_c[i] = j means * column i of A is in position j in A*Pc. * On exit, perm_c may be overwritten by the product of the input * perm_c and a permutation that postorders the elimination tree * of Pc'*A'*A*Pc; perm_c is not changed if the elimination tree * is already in postorder. * * If A->Stype = NR, column permutation vector of size A->nrow, * which describes permutation of columns of tranpose(A) * (rows of A) as described above. * * perm_r (input/output) int* * If A->Stype = NC, row permutation vector of size A->nrow, * which defines the permutation matrix Pr, and is determined * by partial pivoting. perm_r[i] = j means row i of A is in * position j in Pr*A. * * If A->Stype = NR, permutation vector of size A->ncol, which * determines permutation of rows of transpose(A) * (columns of A) as described above. * * If superlumt_options->usepr = NO, perm_r is output argument; * If superlumt_options->usepr = YES, the pivoting routine will try * to use the input perm_r, unless a certain threshold criterion * is violated. In that case, perm_r is overwritten by a new * permutation determined by partial pivoting or diagonal * threshold pivoting. * * equed (input/output) equed_t* * Specifies the form of equilibration that was done. * = NOEQUIL: No equilibration. * = ROW: Row equilibration, i.e., A was premultiplied by diag(R). * = COL: Column equilibration, i.e., A was postmultiplied by diag(C). * = BOTH: Both row and column equilibration, i.e., A was replaced * by diag(R)*A*diag(C). * If superlumt_options->fact = FACTORED, equed is an input argument, * otherwise it is an output argument. * * R (input/output) double*, dimension (A->nrow) * The row scale factors for A or transpose(A). * If equed = ROW or BOTH, A (if A->Stype = NC) or transpose(A) * (if A->Stype = NR) is multiplied on the left by diag(R). * If equed = NOEQUIL or COL, R is not accessed. * If fact = FACTORED, R is an input argument; otherwise, R is output. * If fact = FACTORED and equed = ROW or BOTH, each element of R must * be positive. * * C (input/output) double*, dimension (A->ncol) * The column scale factors for A or transpose(A). * If equed = COL or BOTH, A (if A->Stype = NC) or trnspose(A) * (if A->Stype = NR) is multiplied on the right by diag(C). * If equed = NOEQUIL or ROW, C is not accessed. * If fact = FACTORED, C is an input argument; otherwise, C is output. * If fact = FACTORED and equed = COL or BOTH, each element of C must * be positive. * * L (output) SuperMatrix* * The factor L from the factorization * Pr*A*Pc=L*U (if A->Stype = NC) or * Pr*transpose(A)*Pc=L*U (if A->Stype = NR). * Uses compressed row subscripts storage for supernodes, i.e., * L has types: Stype = SCP, Dtype = _D, Mtype = TRLU. * * U (output) SuperMatrix* * The factor U from the factorization * Pr*A*Pc=L*U (if A->Stype = NC) or * Pr*transpose(A)*Pc=L*U (if A->Stype = NR). * Uses column-wise storage scheme, i.e., U has types: * Stype = NCP, Dtype = _D, Mtype = TRU. * * B (input/output) SuperMatrix* * B has types: Stype = DN, Dtype = _D, Mtype = GE. * On entry, the right hand side matrix. * On exit, * if equed = NOEQUIL, B is not modified; otherwise * if A->Stype = NC: * if trans = NOTRANS and equed = ROW or BOTH, B is overwritten * by diag(R)*B; * if trans = TRANS or CONJ and equed = COL of BOTH, B is * overwritten by diag(C)*B; * if A->Stype = NR: * if trans = NOTRANS and equed = COL or BOTH, B is overwritten * by diag(C)*B; * if trans = TRANS or CONJ and equed = ROW of BOTH, B is * overwritten by diag(R)*B. * * X (output) SuperMatrix* * X has types: Stype = DN, Dtype = _D, Mtype = GE. * If info = 0 or info = A->ncol+1, X contains the solution matrix * to the original system of equations. Note that A and B are modified * on exit if equed is not NOEQUIL, and the solution to the * equilibrated system is inv(diag(C))*X if trans = NOTRANS and * equed = COL or BOTH, or inv(diag(R))*X if trans = TRANS or CONJ * and equed = ROW or BOTH. * * recip_pivot_growth (output) float* * The reciprocal pivot growth factor computed as * max_j ( max_i(abs(A_ij)) / max_i(abs(U_ij)) ). * If recip_pivot_growth is much less than 1, the stability of the * LU factorization could be poor. * * rcond (output) float* * The estimate of the reciprocal condition number of the matrix A * after equilibration (if done). If rcond is less than the machine * precision (in particular, if rcond = 0), the matrix is singular * to working precision. This condition is indicated by a return * code of info > 0. * * ferr (output) float*, 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) float*, 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). * * superlu_memusage (output) superlu_memusage_t* * Record the memory usage statistics, consisting of following fields: * - for_lu (float) * The amount of space used in bytes for L\U data structures. * - total_needed (float) * The amount of space needed in bytes to perform factorization. * - expansions (int) * The number of memory expansions during the LU factorization. * * info (output) int* * = 0: successful exit * < 0: if info = -i, the i-th argument had an illegal value * > 0: if info = i, and i is * <= A->ncol: U(i,i) is exactly zero. The factorization has * been completed, but the factor U is exactly * singular, so the solution and error bounds * could not be computed. * = A->ncol+1: U is nonsingular, but RCOND is less than machine * precision, meaning that the matrix is singular to * working precision. Nevertheless, the solution and * error bounds are computed because there are a number * of situations where the computed solution can be more * accurate than the value of RCOND would suggest. * > A->ncol+1: number of bytes allocated when memory allocation * failure occurred, plus A->ncol. * */ NCformat *Astore; DNformat *Bstore, *Xstore; float *Bmat, *Xmat; int ldb, ldx, nrhs; SuperMatrix *AA; /* A in NC format used by the factorization routine.*/ SuperMatrix AC; /* Matrix postmultiplied by Pc */ int colequ, equil, dofact, notran, rowequ; char norm[1]; trans_t trant; int i, j, info1; float amax, anorm, bignum, smlnum, colcnd, rowcnd, rcmax, rcmin; int n, relax, panel_size; Gstat_t Gstat; double t0; /* temporary time */ double *utime; flops_t *ops, flopcnt; /* External functions */ extern float slangs(char *, SuperMatrix *); extern double slamch_(char *); Astore = A->Store; Bstore = B->Store; Xstore = X->Store; Bmat = Bstore->nzval; Xmat = Xstore->nzval; n = A->ncol; ldb = Bstore->lda; ldx = Xstore->lda; nrhs = B->ncol; superlumt_options->perm_c = perm_c; superlumt_options->perm_r = perm_r; *info = 0; dofact = (superlumt_options->fact == DOFACT); equil = (superlumt_options->fact == EQUILIBRATE); notran = (superlumt_options->trans == NOTRANS); if (dofact || equil) { *equed = NOEQUIL; rowequ = FALSE; colequ = FALSE; } else { rowequ = (*equed == ROW) || (*equed == BOTH); colequ = (*equed == COL) || (*equed == BOTH); smlnum = slamch_("Safe minimum"); bignum = 1. / smlnum; } /* ------------------------------------------------------------ Test the input parameters. ------------------------------------------------------------*/ if ( nprocs <= 0 ) *info = -1; else if ( (!dofact && !equil && (superlumt_options->fact != FACTORED)) || (!notran && (superlumt_options->trans != TRANS) && (superlumt_options->trans != CONJ)) || (superlumt_options->refact != YES && superlumt_options->refact != NO) || (superlumt_options->usepr != YES && superlumt_options->usepr != NO) || superlumt_options->lwork < -1 ) *info = -2; 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 = -3; else if ((superlumt_options->fact == FACTORED) && !(rowequ || colequ || (*equed == NOEQUIL))) *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 ( B->ncol < 0 || Bstore->lda < SUPERLU_MAX(0, A->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_S || B->Mtype != SLU_GE ) *info = -11; else if ( X->ncol < 0 || Xstore->lda < SUPERLU_MAX(0, A->nrow) || B->ncol != X->ncol || X->Stype != SLU_DN || X->Dtype != SLU_S || X->Mtype != SLU_GE ) *info = -12; } } if (*info != 0) { i = -(*info); xerbla_("psgssvx", &i); return; } /* ------------------------------------------------------------ Allocate storage and initialize statistics variables. ------------------------------------------------------------*/ panel_size = superlumt_options->panel_size; relax = superlumt_options->relax; StatAlloc(n, nprocs, panel_size, relax, &Gstat); StatInit(n, nprocs, &Gstat); utime = Gstat.utime; ops = Gstat.ops; /* ------------------------------------------------------------ Convert A to 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 == NC */ trant = superlumt_options->trans; AA = A; } /* ------------------------------------------------------------ Diagonal scaling to equilibrate the matrix. ------------------------------------------------------------*/ if ( 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 = (*equed == ROW) || (*equed == BOTH); colequ = (*equed == COL) || (*equed == BOTH); } utime[EQUIL] = SuperLU_timer_() - t0; } /* ------------------------------------------------------------ Scale the right hand side. ------------------------------------------------------------*/ 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]; } } /* ------------------------------------------------------------ Perform the LU factorization. ------------------------------------------------------------*/ if ( dofact || equil ) { /* Obtain column etree, the column count (colcnt_h) and supernode partition (part_super_h) for the Householder matrix. */ t0 = SuperLU_timer_(); sp_colorder(AA, perm_c, superlumt_options, &AC); utime[ETREE] = SuperLU_timer_() - t0; #if ( PRNTlevel >= 2 ) printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n", relax, panel_size, sp_ienv(3), sp_ienv(4)); fflush(stdout); #endif /* Compute the LU factorization of A*Pc. */ t0 = SuperLU_timer_(); psgstrf(superlumt_options, &AC, perm_r, L, U, &Gstat, info); utime[FACT] = SuperLU_timer_() - t0; flopcnt = 0; for (i = 0; i < nprocs; ++i) flopcnt += Gstat.procstat[i].fcops; ops[FACT] = flopcnt; if ( superlumt_options->lwork == -1 ) { superlu_memusage->total_needed = *info - A->ncol; return; } } 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); } } else { /* ------------------------------------------------------------ Compute the reciprocal pivot growth factor *recip_pivot_growth. ------------------------------------------------------------*/ *recip_pivot_growth = sPivotGrowth(A->ncol, AA, perm_c, L, U); /* ------------------------------------------------------------ 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, info); utime[RCOND] = SuperLU_timer_() - t0; /* ------------------------------------------------------------ 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_r, perm_c, X, &Gstat, info); utime[SOLVE] = SuperLU_timer_() - t0; ops[SOLVE] = ops[TRISOLVE]; /* ------------------------------------------------------------ Use iterative refinement to improve the computed solution and compute error bounds and backward error estimates for it. ------------------------------------------------------------*/ t0 = SuperLU_timer_(); sgsrfs(trant, AA, L, U, perm_r, perm_c, *equed, R, C, B, X, ferr, berr, &Gstat, info); 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]; } } /* Set INFO = A->ncol+1 if the matrix is singular to working precision.*/ if ( *rcond < slamch_("E") ) *info = A->ncol + 1; } superlu_sQuerySpace(nprocs, L, U, panel_size, superlu_memusage); /* ------------------------------------------------------------ Deallocate storage after factorization. ------------------------------------------------------------*/ if ( superlumt_options->refact == NO ) { SUPERLU_FREE(superlumt_options->etree); SUPERLU_FREE(superlumt_options->colcnt_h); SUPERLU_FREE(superlumt_options->part_super_h); } if ( dofact || equil ) { Destroy_CompCol_Permuted(&AC); } if ( A->Stype == SLU_NR ) { Destroy_SuperMatrix_Store(AA); SUPERLU_FREE(AA); } /* ------------------------------------------------------------ Print timings, then deallocate statistic variables. ------------------------------------------------------------*/ #ifdef PROFILE { SCPformat *Lstore = (SCPformat *) L->Store; ParallelProfile(n, Lstore->nsuper+1, Gstat.num_panels, nprocs, &Gstat); } #endif PrintStat(&Gstat); StatFree(&Gstat); }
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); } }
main(int argc, char *argv[]) { /* * Purpose * ======= * * SDRIVE is the main test program for the FLOAT linear * equation driver routines SGSSV and SGSSVX. * * The program is invoked by a shell script file -- stest.csh. * The output from the tests are written into a file -- stest.out. * * ===================================================================== */ float *a, *a_save; int *asub, *asub_save; int *xa, *xa_save; SuperMatrix A, B, X, L, U; SuperMatrix ASAV, AC; GlobalLU_t Glu; /* Not needed on return. */ mem_usage_t mem_usage; int *perm_r; /* row permutation from partial pivoting */ int *perm_c, *pc_save; /* column permutation */ int *etree; float zero = 0.0; float *R, *C; float *ferr, *berr; float *rwork; float *wwork; void *work; int info, lwork, nrhs, panel_size, relax; int m, n, nnz; float *xact; float *rhsb, *solx, *bsav; int ldb, ldx; float rpg, rcond; int i, j, k1; float 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; float anorm, cndnum; float *Afull; float 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]; FILE *fp; /* 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 sgst01(int, int, SuperMatrix *, SuperMatrix *, SuperMatrix *, int *, int *, float *); extern int sgst02(trans_t, int, int, int, SuperMatrix *, float *, int, float *, int, float *resid); extern int sgst04(int, int, float *, int, float *, int, float rcond, float *resid); extern int sgst07(trans_t, int, int, SuperMatrix *, float *, int, float *, int, float *, int, float *, float *, float *); extern int slatb4_slu(char *, int *, int *, int *, char *, int *, int *, float *, int *, float *, char *); extern int slatms_slu(int *, int *, char *, int *, char *, float *d, int *, float *, float *, int *, int *, char *, float *, int *, float *, int *); extern int sp_sconvert(int, int, float *, int, int, int, float *a, int *, int *, int *); /* Executable statements */ strcpy(path, "SGE"); 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, &fp); 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 = SLU_SINGLE; 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 = floatCalloc(lda * n); sallocateA(n, nnz, &a, &asub, &xa); } else { /* Read a sparse matrix */ fimat = nimat = 0; sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa); } sallocateA(n, nnz, &a_save, &asub_save, &xa_save); rhsb = floatMalloc(m * nrhs); bsav = floatMalloc(m * nrhs); solx = floatMalloc(n * nrhs); ldb = m; ldx = n; sCreate_Dense_Matrix(&B, m, nrhs, rhsb, ldb, SLU_DN, SLU_S, SLU_GE); sCreate_Dense_Matrix(&X, n, nrhs, solx, ldx, SLU_DN, SLU_S, SLU_GE); xact = floatMalloc(n * nrhs); etree = intMalloc(n); perm_r = intMalloc(n); perm_c = intMalloc(n); pc_save = intMalloc(n); R = (float *) SUPERLU_MALLOC(m*sizeof(float)); C = (float *) SUPERLU_MALLOC(n*sizeof(float)); ferr = (float *) SUPERLU_MALLOC(nrhs*sizeof(float)); berr = (float *) SUPERLU_MALLOC(nrhs*sizeof(float)); j = SUPERLU_MAX(m,n) * SUPERLU_MAX(4,nrhs); rwork = (float *) SUPERLU_MALLOC(j*sizeof(float)); 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 = floatCalloc( 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 SLATB4 and generate a test matrix with SLATMS. */ slatb4_slu(path, &imat, &n, &n, sym, &kl, &ku, &anorm, &mode, &cndnum, dist); slatms_slu(&n, &n, dist, iseed, sym, &rwork[0], &mode, &cndnum, &anorm, &kl, &ku, "No packing", Afull, &lda, &wwork[0], &info); if ( info ) { printf(FMT3, "SLATMS", 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_sconvert(n, n, Afull, lda, kl, ku, a, asub, xa, &nnz); } else { izero = 0; zerot = 0; } sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE); /* Save a copy of matrix A in ASAV */ sCreate_CompCol_Matrix(&ASAV, m, n, nnz, a_save, asub_save, xa_save, SLU_NC, SLU_S, SLU_GE); sCopy_CompCol_Matrix(&A, &ASAV); /* Form exact solution. */ sGenXtrue(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. */ sCopy_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. */ sgsequ(&A, R, C, &rowcnd, &colcnd, &amax, &info); /* Force equilibration. */ if ( !info && n > 0 ) { if ( strncmp(equed, "R", 1)==0 ) { rowcnd = 0.; colcnd = 1.; } else if ( strncmp(equed, "C", 1)==0 ) { rowcnd = 1.; colcnd = 0.; } else if ( strncmp(equed, "B", 1)==0 ) { rowcnd = 0.; colcnd = 0.; } } /* Equilibrate the matrix. */ slaqgs(&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. */ sgstrf(&options, &AC, relax, panel_size, etree, work, lwork, perm_c, perm_r, &L, &U, &Glu, &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. */ sCopy_CompCol_Matrix(&ASAV, &A); /* Set the right hand side. */ sFillRHS(trans, nrhs, xact, ldx, &A, &B); sCopy_Dense_Matrix(m, nrhs, rhsb, ldb, bsav, ldb); /*---------------- * Test sgssv *----------------*/ if ( options.Fact == DOFACT && itran == 0) { /* Not yet factored, and untransposed */ sCopy_Dense_Matrix(m, nrhs, rhsb, ldb, solx, ldx); sgssv(&options, &A, perm_c, perm_r, &L, &U, &X, &stat, &info); if ( info && info != izero ) { printf(FMT3, "sgssv", info, izero, n, nrhs, imat, nfail); } else { /* Reconstruct matrix from factors and compute residual. */ sgst01(m, n, &A, &L, &U, perm_c, perm_r, &result[0]); nt = 1; if ( izero == 0 ) { /* Compute residual of the computed solution. */ sCopy_Dense_Matrix(m, nrhs, rhsb, ldb, wwork, ldb); sgst02(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, "sgssv", 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 sgssv */ /*---------------- * Test sgssvx *----------------*/ /* Equilibrate the matrix if fact = FACTORED and equed = 'R', 'C', or 'B'. */ if ( options.Fact == FACTORED && (equil || iequed) && n > 0 ) { slaqgs(&A, R, C, rowcnd, colcnd, amax, equed); } /* Solve the system and compute the condition number and error bounds using sgssvx. */ sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &Glu, &mem_usage, &stat, &info); if ( info && info != izero ) { printf(FMT3, "sgssvx", 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. */ sgst01(m, n, &A, &L, &U, perm_c, perm_r, &result[0]); k1 = 0; } else { k1 = 1; } if ( !info ) { /* Compute residual of the computed solution.*/ sCopy_Dense_Matrix(m, nrhs, bsav, ldb, wwork, ldb); sgst02(trans, m, n, nrhs, &ASAV, solx, ldx, wwork, ldb, &result[1]); /* Check solution from generated exact solution. */ sgst04(n, nrhs, solx, ldx, xact, ldx, rcond, &result[2]); /* Check the error bounds from iterative refinement. */ sgst07(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, "sgssvx", options.Fact, trans, *equed, n, imat, i, result[i]); ++nfail; } } nrun += NTESTS; } /* if .. info == 0 */ } /* else .. end of testing sgssvx */ } /* 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 Destroy_SuperMatrix_Store(&A); Destroy_SuperMatrix_Store(&ASAV); StatFree(&stat); } /* for imat ... */ /* Print a summary of the results. */ PrintSumm("SGE", nfail, nrun, nerrs); if ( strcmp(matrix_type, "LA") == 0 ) SUPERLU_FREE (Afull); 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); #if 0 Destroy_CompCol_Matrix(&A); Destroy_CompCol_Matrix(&ASAV); #else SUPERLU_FREE(a); SUPERLU_FREE(asub); SUPERLU_FREE(xa); SUPERLU_FREE(a_save); SUPERLU_FREE(asub_save); SUPERLU_FREE(xa_save); #endif if ( lwork > 0 ) { SUPERLU_FREE (work); Destroy_SuperMatrix_Store(&L); Destroy_SuperMatrix_Store(&U); } return 0; }
void sgsisx(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, 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, mc64; 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; int *perm = NULL; /* 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); mc64 = (options->RowPerm == LargeDiag); 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; } /* Test the input parameters */ if (!nofact && options->Fact != DOFACT && options->Fact != SamePattern && options->Fact != SamePattern_SameRowPerm && !notran && options->Trans != TRANS && options->Trans != CONJ && !equil && options->Equil != NO) *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 || Bstore->lda < SUPERLU_MAX(0, A->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_S || B->Mtype != SLU_GE ) *info = -13; else if ( X->ncol < 0 || 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_("sgsisx", &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 ) { register int i, j; NCformat *Astore = AA->Store; int nnz = Astore->nnz; int *colptr = Astore->colptr; int *rowind = Astore->rowind; float *nzval = (float *)Astore->nzval; int n = AA->nrow; if ( mc64 ) { *equed = 'B'; /*rowequ = colequ = 1;*/ t0 = SuperLU_timer_(); if ((perm = intMalloc(n)) == NULL) ABORT("SUPERLU_MALLOC fails for perm[]"); info1 = sldperm(5, n, nnz, colptr, rowind, nzval, perm, R, C); if (info1 > 0) { /* MC64 fails, call sgsequ() later */ mc64 = 0; SUPERLU_FREE(perm); perm = NULL; } else { rowequ = colequ = 1; for (i = 0; i < n; i++) { R[i] = exp(R[i]); C[i] = exp(C[i]); } /* permute and scale the matrix */ for (j = 0; j < n; j++) { for (i = colptr[j]; i < colptr[j + 1]; i++) { nzval[i] *= R[rowind[i]] * C[j]; rowind[i] = perm[rowind[i]]; } } } utime[EQUIL] = SuperLU_timer_() - t0; } if ( !mc64 & 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; /* Compute the LU factorization of A*Pc. */ t0 = SuperLU_timer_(); sgsitrf(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 ) 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, &info1); utime[RCOND] = SuperLU_timer_() - t0; } if ( nrhs > 0 ) { /* Solve the system */ float *tmp, *rhs_work; int n = A->nrow; if ( mc64 ) { if ((tmp = floatMalloc(n)) == NULL) ABORT("SUPERLU_MALLOC fails for tmp[]"); } /* Scale and permute the right-hand side if equilibration and permutation from MC64 were performed. */ if ( notran ) { if ( rowequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < n; ++i) Bmat[i + j*ldb] *= R[i]; } if ( mc64 ) { for (j = 0; j < nrhs; ++j) { rhs_work = &Bmat[j*ldb]; for (i = 0; i < n; i++) tmp[perm[i]] = rhs_work[i]; for (i = 0; i < n; i++) rhs_work[i] = tmp[i]; } } } else if ( colequ ) { for (j = 0; j < nrhs; ++j) for (i = 0; i < n; ++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, &info1); utime[SOLVE] = 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 < n; ++i) { Xmat[i + j*ldx] *= C[i]; } } } else { /* transposed system */ if ( rowequ ) { if ( mc64 ) { for (j = 0; j < nrhs; j++) { for (i = 0; i < n; i++) tmp[i] = Xmat[i + j * ldx]; /*dcopy*/ for (i = 0; i < n; i++) Xmat[i + j * ldx] = R[i] * tmp[perm[i]]; } } else { for (j = 0; j < nrhs; ++j) for (i = 0; i < A->nrow; ++i) { Xmat[i + j*ldx] *= R[i]; } } } } if ( mc64 ) SUPERLU_FREE(tmp); } /* 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 == 0) *info = A->ncol + 1; } if (perm) SUPERLU_FREE(perm); if ( nofact ) { ilu_sQuerySpace(L, U, mem_usage); Destroy_CompCol_Permuted(&AC); } if ( A->Stype == SLU_NR ) { Destroy_SuperMatrix_Store(AA); SUPERLU_FREE(AA); } }