
    Version history
    ---------------

  Version           Comments
     1.0	Initial release (08/15/93)

     1.1	Upgrade (10/15/93)

     1.11	Bug fixes (12/15/93)

     1.2        Upgrade (01/07/94)

     1.3	Upgrade (06/10/94)

     1.31	Bug fixes (10/25/94)

     1.32	Bug fixes (1/15/95)

     1.33	Bug fixes (2/20/96)

     1.34	Bug fixes (4/15/96)

    Changes listing
    ---------------

From v1.33 to v1.34:

- Corrected bug in ZSQMR in the computation of the indices for _S_ (pointed
  out by J. Shen).

From v1.32 to v1.33:

- Added test data for SGI running IRIS 5.3.

- Changed string delimiters in comments to stop complaints from Cray compilers.

- Changed the dimension of ZWK in zssys.src (pointed out by M. Schlarmann).

- Changed the SED command files to all have fewer than 100 commands in them
  (some `sed' programs choked on command files with more than 100 commands).

- Cleaned up the symmetric codes, removed stuff inherited from the
  nonsymmetric codes.

- Corrected a typo in the duqmr.doc (pointed out by Masha Sosonkina).

- Corrected the dimension of DWK in the example drivers dulal.f and sulal.f.

From v1.31 to v1.32:

- Changed the dimensions in the documentation for d?lal to match the internal
  dimensions (pointed out by Lawrence C. Cowsar).

- Changed the AWK script `convert.awk' to work with older versions of `awk'.

From v1.3 to v1.31:

- Added check for NLIM.LE.MAXN-2 in slal and ulal, and updated the
  documentation (pointed out by Anand V. Lakshmikumaran).

- Added comment about running the Setup script on a DEC-5000.

- Added test data for a DEC-5000.

- Changed the AWK script `convert.awk' so it won't split strings.

- Fixed bug in initialization of IDX (pointed out by Lawrence C. Cowsar).

- Fixed bug in counting the elements of L is sillt (pointed out by Paul
  Soudais).

From v1.2 to v1.3:

- Added version for symmetric matrices

- Changed the meaning of the two highest digits in INFO(1) for all algorithms
  that take an auxiliary starting vector

- Renamed the routines to make their names unique across precisions and matrix
  types (symmetric, unsymmetric)

From v1.11 to v1.2:

- Changed the logic used in the eigenvalue code to identify the good Lanczos
  eigenvalues.

- Corrected description of VECS in complex versions of lal.doc.

- Fixed bug in the initialization of HLAL in the real versions of lal.src.

- Saved the variable IERR across reverse communication calls in the algorithms
  without look-ahead.

From v1.1 to v1.11:

- Fixed bug that appeared when NLIM.EQ.MAXN-1 (NLIM must be LE.MAXN-2).

- Fixed bug that made the linear solvers stop when UNRM.LT.0.01 if RESN was not
  computed at every step.

From v1.0 to v1.1:

- Added double and single precision complex versions of the codes

- Added two-sided SSOR preconditioner with the Eisenstat trick for the CSR
  format.

- Changed the example linear systems driver to a driver that can select
  on-the-fly among the solvers and the preconditioners.  This eliminates the
  `select' script that was used in v1.0 to select a solver, a data format, and
  a preconditioner.

- Changed the routine names for the algorithms so they are all unique and thus
  can be grouped in a library.

- Deleted the dns subdirectory from fmts.

- Reduced the use of COMMON blocks in the codes; the only remaining COMMONs are
  in the look-ahead algorithms, for the inner recurrence coefficients.

- Use reverse communication for matrix-vector multiplies instead of requiring
  the user to provide routines AXB and ATXB with a given calling sequence.

