/*! @file dgsrfs.c * \brief Improves computed solution to a system of inear equations * *
 * -- SuperLU routine (version 3.0) --
 * Univ. of California Berkeley, Xerox Palo Alto Research Center,
 * and Lawrence Berkeley National Lab.
 * October 15, 2003
 *
 * Modified from lapack routine DGERFS
 * 
*/ /* * File name: dgsrfs.c * History: Modified from lapack routine DGERFS */ #include #include "slu_ddefs.h" /*! \brief * *
 *   Purpose   
 *   =======   
 *
 *   DGSRFS 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_D, 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_D, Mtype = SLU_TRLU.
 * 
 *   U       (input) SuperMatrix*
 *           The factor U from the factorization Pr*A*Pc=L*U as computed by
 *           dgstrf(). Use column-wise storage scheme, 
 *           i.e., U has types: Stype = SLU_NC, Dtype = SLU_D, 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_D, 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_D, Mtype = SLU_GE.
 *           On entry, the solution matrix X, as computed by dgstrs().
 *           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.   
 *
 * 
*/ void dgsrfs(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; double ndone = -1.; double done = 1.; /* Local variables */ NCformat *Astore; double *Aval; SuperMatrix Bjcol; DNformat *Bstore, *Xstore, *Bjcol_store; double *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; double *work; double *rwork; int *iwork; extern int dlacon_(int *, double *, double *, int *, double *, int *); #ifdef _CRAY extern int SCOPY(int *, double *, int *, double *, int *); extern int SSAXPY(int *, double *, double *, int *, double *, int *); #else extern int dcopy_(int *, double *, int *, double *, int *); extern int daxpy_(int *, double *, double *, int *, double *, 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_D || A->Mtype != SLU_GE ) *info = -2; else if ( L->nrow != L->ncol || L->nrow < 0 || L->Stype != SLU_SC || L->Dtype != SLU_D || L->Mtype != SLU_TRLU ) *info = -3; else if ( U->nrow != U->ncol || U->nrow < 0 || U->Stype != SLU_NC || U->Dtype != SLU_D || U->Mtype != SLU_TRU ) *info = -4; else if ( ldb < SUPERLU_MAX(0, A->nrow) || B->Stype != SLU_DN || B->Dtype != SLU_D || B->Mtype != SLU_GE ) *info = -10; else if ( ldx < SUPERLU_MAX(0, A->nrow) || X->Stype != SLU_DN || X->Dtype != SLU_D || X->Mtype != SLU_GE ) *info = -11; if (*info != 0) { i = -(*info); xerbla_("dgsrfs", &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 = doubleMalloc(2*A->nrow); rwork = (double *) SUPERLU_MALLOC( A->nrow * sizeof(double) ); iwork = intMalloc(2*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 SCOPY(&A->nrow, Bptr, &ione, work, &ione); #else dcopy_(&A->nrow, Bptr, &ione, work, &ione); #endif sp_dgemv(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] = fabs( Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if (notran) { for (k = 0; k < A->ncol; ++k) { xk = fabs( Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += fabs(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 += fabs(Aval[i]) * fabs(Xptr[irow]); } rwork[k] += s; } } s = 0.; for (i = 0; i < A->nrow; ++i) { if (rwork[i] > safe2) { s = SUPERLU_MAX( s, fabs(work[i]) / rwork[i] ); } else if ( rwork[i] != 0.0 ) { /* Adding SAFE1 to the numerator guards against spuriously zero residuals (underflow). */ s = SUPERLU_MAX( s, (safe1 + fabs(work[i])) / 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. */ dgstrs (trans, L, U, perm_c, perm_r, &Bjcol, stat, info); #ifdef _CRAY SAXPY(&A->nrow, &done, work, &ione, &Xmat[j*ldx], &ione); #else daxpy_(&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 DLACON 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] = fabs( Bptr[i] ); /* Compute abs(op(A))*abs(X) + abs(B). */ if ( notran ) { for (k = 0; k < A->ncol; ++k) { xk = fabs( Xptr[k] ); for (i = Astore->colptr[k]; i < Astore->colptr[k+1]; ++i) rwork[Astore->rowind[i]] += fabs(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 = fabs( Xptr[irow] ); s += fabs(Aval[i]) * xk; } rwork[k] += s; } } for (i = 0; i < A->nrow; ++i) if (rwork[i] > safe2) rwork[i] = fabs(work[i]) + (iwork[i]+1)*eps*rwork[i]; else rwork[i] = fabs(work[i])+(iwork[i]+1)*eps*rwork[i]+safe1; kase = 0; do { dlacon_(&A->nrow, &work[A->nrow], work, &iwork[A->nrow], &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) work[i] *= C[i]; else if ( !notran && rowequ ) for (i = 0; i < A->nrow; ++i) work[i] *= R[i]; dgstrs (transt, L, U, perm_c, perm_r, &Bjcol, stat, info); for (i = 0; i < A->nrow; ++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) work[i] *= rwork[i]; dgstrs (trans, L, U, perm_c, perm_r, &Bjcol, stat, info); if ( notran && colequ ) for (i = 0; i < A->ncol; ++i) work[i] *= C[i]; else if ( !notran && rowequ ) for (i = 0; i < A->ncol; ++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] * fabs( Xptr[i]) ); } else if ( !notran && rowequ ) { for (i = 0; i < A->nrow; ++i) lstres = SUPERLU_MAX( lstres, R[i] * fabs( Xptr[i]) ); } else { for (i = 0; i < A->nrow; ++i) lstres = SUPERLU_MAX( lstres, fabs( 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; } /* dgsrfs */