Jan. 28, 2005: v4.4 released

    * bug fix:  when Qinit is provided to umfpack_*_qsymbolic,
	only the symmetric and unsymmetric strategies are now permitted.
	The auto and 2-by-2 strategies are not allowed.  In v4.3 and
	earlier, providing Qinit and requesting the symmetric strategy
	did not always work (you got the unsymmetric strategy instead).
	This does not affect umfpack_*_symbolic, which computes its own
	ordering and can use all 4 strategies (auto, symmetric, unsymmetric,
	and 2-by-2).

    * umfpack_get_determinant added.

    * packed complex case added for all routines (previously only used in
	umfpack_report_vector).  This allows arrays of ANSI C/C++ complex
	type to be passed directly to UMFPACK.

    * added umf_multicomple.c to assist in the compilation of UMFPACK
	in Microsoft Visual Studio, which does not have the required
	flexibility of the Unix "make" command.

    * local variable declarations reordered to encourage double-word
	alignment of double's and Entry's, for better performance.

    * note that with the exception of the behavior when a user-provided
	ordering is passed to umfpack_*_qsymbolic, versions 4.1 through 4.4
	have comparable performance (ordering quality, memory usage,
	and run time).  v4.1 is much better than v4.0 in performance.

Jan. 11, 2005: v4.3.1 released

    * bug fix in umf_solve.  This bug is only the 4th one found in the C
	versions of UMFPACK to date (Version 3.0 to 4.3.1, from March 2001 to
	Jan. 2005, excluding workarounds for quirky compilers).  No bugs have
	been reported in the last Fortran version of UMFPACK (MA38, or UMFPACK
	V2.2.1) since its release in Jan. 1998.

	In Version 4.3, a bug in umf_solve caused iterative refinement
	to be disabled when solving A'x=b or A.'x=b after factorizing A.
	Modified the umfpack mexFunction to factorize A and then solve A'x=b
	when performing the operation x=b/A (as "umfpack(b,'/',A).  Note that
	this has no effect on the use of UMFPACK in MATLAB itself, since MATLAB
	does not use the umfpack mexFunction for x=b/A.  When computing x=b/A,
	MATLAB factorizes A' and computes x=(A'\b')' instead. The following
	source code files changed:

	    UMFPACK/MATLAB/umfpackmex.c	 (see above)
	    UMFPACK/Source/umf_solve.c	 (see source code: 2 lines changed)
	    UMFPACK/Include/umfpack.h    (version and date changed)
	    UMFPACK/MATLAB/umfpack_test.m   (new file)

Jan. 16, 2004: v4.3 released.

    * user interface of v4.3 is upwardly-compatible with v4.2 and v4.1.
	No bugs found in v4.1 (except for one workaround for an old compiler).
	These changes add features only.

    * Note that v4.0 has a bug in umf_scale_column.c.  The bug was patched
	in that version on Jan. 12, 2004.  The bug does not appear in v4.1
	and later.  The bug is thus present in MATLAB 6.5, but it occurs
	very rarely, fortunately.  It can occur when dividing a nonzero entry
	in the pivot column by the pivot value results in an underflow.

    * <float.h> added to umfpackmex.c, for DBL_EPSILON.  Some non-standard
	compilers (Microsoft Visual C++) require this.

    * #pragma added to umf_analyze.c, as a workaround around a bug in an
	old Intel compiler.

    * mexFunction interface to MATLAB modified.  Call to mexCallMATLAB removed,
	which can be slow.  In V4.1 it was used only to get MATLAB's
	spparms ('spumoni') value.

    * The AMD mexFunction was also modified in the same way (v1.1), with
	the call to mexCallMATLAB removed.  Note that UMFPACK v4.1 through
	v4.3 can use either AMD v1.0 or AMD v1.1.

    * -DNO_DIVIDE_BY_ZERO option added.  If this non-default option is enabled
	at compile time, and if the pivot value is zero, then no division
	occurs (zeros on the diagonal of U are treated as if they were equal
	to one).  By default, the division by zero does occur.

    * -DNO_TIMER option added.  If this non-default option is enabled at
	compile time, then no timers (times ( ), clock ( ), getrusage ( ))
	are used.

V4.2:  A special release for COMSOL, Inc., only (FEMLAB)

    * drop tolerance added.  A few new parameters in the Control array are used,
	and a few new Info entries.

May 6, 2003:  V4.1 released.

    * No bugs were found in the prior version, Version 4.0.  New features
	added only.  Major changes throughout the code.  User interface
	nearly unchanged, however.

    * Version 4.1 is upward-compatible with Version 4.0.  The calling
	sequence of some user-callable routines in Version 4.0 have changed
	in this version.  The routines umfpack_*_symbolic, umfpack_*_qsymbolic,
	umfpack_*_get_symbolic, and umfpack_*_get_numeric have new arguments
	added to them.  The new arguments are optional.  If you want to use
	a calling sequence similar to v4.0, simply pass NULL pointers in
	place of the new arguments.  There are two new timing routines,
	umfpack_tic and umfpack_toc.  A new user-callable routine, 
	umfpack_*_scale, has been added.

    *	"auto", "unsymmetric", "symmetric", and "2-by-2" strategies added.
	The symmetric strategy uses AMD on A+A' as the column preordering,
	followed by a postorder of the assembly tree of A+A'.  Column ordering
	refinement is turned off, and diagonal entries are prefered as pivots.
	V4.0 only had the unsymmetric strategy.  The 2-by-2 strategy does row
	permutations and attempts to find a zero-free diagonal while at the
	same time maintaining structural symmetry, and then uses the 
	symmetric strategy on the permuted matrix.

    * row-scaling added.  The default is to divide each row by the sum of
	the absolute values of each row.  Other options are no scaling,
	and to divide each row by the max abs value in each row.

    * Matrices with upper bound memory usage greater than the maximum integer
	(2GB for 32-bit int's) can now be factorized (assuming the actual
	memory usage is still less than the maximum integer).  With this change,
	the UMFPACK_ERROR_problem_too_large error code is no longer returned.

    * The current frontal matrix (Work->Fx) is no longer allocated as a
	static size, via malloc.  It can grow and shrink, and is allocated
	from Numeric->Memory.

    * The AMD (Version 1.0) package is now required.  It is available
	separately.  To compile UMFPACK, it must appear as ../AMD if you are
	in the main UMFPACK directory.

    * The UMFPACK mexFunction now uses the internal utMalloc, utRealloc,
	and utFree routines, by default (except on Windows).

    * Three control parameters for modifying relaxed amalgamation removed.
	These values are now fixed at compile-time.

    * Many new statistics added to Info, and new control parameters added.

    * The umfpack mexFunction now returns permutation matrices for P and Q,
	not permutation vectors.  It also returns the scale factors as a
	diagonal matrix.  The factorization is now L*U = P*(R\A)*Q.

    * Option added for controlling the initial allocation of the workspace for
	the current frontal matrix.

    * pivot tolerance of zero treated differently.  symmetric pivot tolerance
	added.

    * Makefile and GNUmakefile changed.  umf_* routines with no double or
	complex values are now compiled just twice (int and long versions)
	rather than 4 times.

    * New routines added to save and load the Numeric and Symbolic objects
	to/from binary files.

    * Simple Fortran interface added.

Apr 11, 2002:

    * Version 4.0 released.

    * bug fix:  the Microsoft compiler doesn't handle NaN's properly.
	utIsNaN, and other ut* routines, added for MathWorks version
	to handle this properly.

Apr 1, 2002:

    * bug fix:  if a column was all NaN's, then UMFPACK would fail
	to find a pivot row.  umf_row_search.c and umf_internal.h
	modified to fix this problem.

Mar 9, 2002:  V4.0beta released

    * Map argument added to umfpack_*_triplet_to_col.  New files
	(umf_triplet.[ch]) added.
    * minor changes made so that UMFPACK can be compiled with g++ 
    * additional error checking added to umfpack_*_numeric, for
	detecting more changes in pattern (Ap, Ai) since last
	call to umfpack_*_symbolic

Feb 21, 2002:

    * User Guide explains the Makefile vs. GNUmakefile

    * umf_config.h modified, so that the complex SCSL C-BLAS uses
	(void *) arguments instead of (scsl_zomplex *).  gcc generates
	some spurious warnings (cc doesn't complain).  Affects the SGI
	IRIX only.

    * ported to Compaq Alpha

Feb 20, 2002: V4.0 (alpha) released.

    * V4.0 not yet ported to the Compaq Alpha (V3.2 was ported).

Feb 6 to Feb 19, 2002:

    * Relaxed restrictions on sizes of arrays for umfpack_*_transpose and
	umfpack_*_triplet_to_col.  Size of "max(n,nz)" now just size nz.

    * workspace for umfpack_*_wsolve increased in size.

    * two user arrays for umfpack_*_get_symbolic increased in size,
	by 1 (Chain_maxrows, Chain_maxcols).

    * lu_normest.m added.

Jan 18 to Feb 5, 2002:

    * The matrix A can be complex, singular, and/or rectangular.
	The solve step that uses the LU factors can only handle
	matrices that are complex or real, singuluar or non-singular, 
	and *** square ***, however.

    * Estimate of the condition number computed:
	(min (abs (diag (U))) / (max (abs (diag (U)))))

    * Forward/backsolves can solve with A.' as well as A'.

    * char * arguments removed from user-callable routines to make it
	easier for Fortran to call UMFPACK.  No Fortran interface is (yet)
	provided, however.

	The solve codes for umfpack_*_*solve changed to #define'd
	integers:

	    UMFPACK_A       Ax=b
	    UMFPACK_At      A'x=b
	    UMFPACK_Aat     A.'x=b
	    UMFPACK_Pt_L    P'Lx=b
	    UMFPACK_L       Lx=b
	    UMFPACK_Lt_P    L'Px=b
	    UMFPACK_Lat_P   L.'Px=b
	    UMFPACK_Lt      L'x=b
	    UMFPACK_U_Qt    UQ'x=b
	    UMFPACK_U       Ux=b
	    UMFPACK_Q_Ut    QU'x=b
	    UMFPACK_Q_Uat   QU.'x=b
	    UMFPACK_Ut      U'x=b
	    UMFPACK_Uat     U.'x=b

	All arguments are now either int, long scalars (pass by value),
	or int, long, double arrays (pass by reference), or void * pointers
	(pass by value or reference).  A void * pointer is of size 32 or 64
	bits on most machines.  There is no need for the caller (C or Fortran)
	to dereference the void * pointers, so these can be treated as
	integer*4 or integer*8 in Fortran.  A Fortran interface would have to
	have all arguments passed by reference.

    * All user-callable routine names changed.  The four sets are now:
	umfpack_di_*	real (double precision), int's as integers
	umfpack_dl_*	real (double precision), longs's as integers
	umfpack_zi_*	real (double precision), int's as integers
	umfpack_zl_*	real (double precision), longs's as integers

    * Ptree (row preordering) and info on pivotal rows for each front
	added to Symbolic object (extracted by umfpack_*_get_symbolic).
	Ptree added as output argument to "umfpack (A, 'symbolic')"
	mexFunction.

    * umfpack_*_transpose can do A' or A.'

    * umfpack_wsolve.c file removed (now generated from umfpack_solve.c).

    * Can now extract just the diagonal of U with umfpack_*_get_numeric,
	without having to extract the entire matrix U.

    * UMFPACK_ERROR_singular_matrix (-2) removed.

    * UMFPACK_WARNING_singular_matrix (1) added.

    * Control [UMFPACK_PIVOT_OPTION] removed.  No longer any symmetric
	pivot option (conflicts with the handling of singular and
	rectangular matrices).

    * Iterative refinement can do Ax=b, A'x=b, or A.'x=b.

    * Most floating-point operations done in macros, to support the complex
	versions.

    * Info [UMFPACK_N] is now Info [UMFPACK_NROW]

    * Info [UMFPACK_NCOL], Info [UMFPACK_UDIAG_NZ], Info [UMFPACK_UDIAG_NZ]
	added.

    * umfpack_* routines with "n" as input now use two arguments,
	n_row and n_col.

    * umfpack mexFunction now explicitly transposes A for b/A.  It computes
	it using the array transpose as (A.'\b.').'

January 1, 2002:  UMFPACK Version 3.2 released.  Submitted to ACM Trans.
	on Mathematical Software.

    * The umfpack mexFunction now returns the Info array when the matrix
	is singular.  Returned an empty array prior to this change.

    * Renamed variable that conflicted with system library routines
    	(system and j1).

    * Added a #ifdef MATHWORKS definition, so the built-in UMFPACK routine
	(in a future release of MATLAB) can use the internal ut* memory
	allocation routines, ut* assertion routine, and utPrintf.

    * MAX and MIN are not defined if they are already defined.

    * A bug fix in umf_kernel_init (a variable was not properly initialized).

    * Removed unused variables.

October 8, 2001:  UMFPACK Version 3.1 released.

August-October, 2001:

    * added umfpack_btf M-file.

    * modified the BLAS update in the frontal matrix.  If there are only
	a few pivots in remaining in the current front, then the BLAS3 update
	is delayed to include pivots in the next front.

    * Removed the special-case handling of dense columns from the numerical
	factorization (kept it in the colamd preordering).  This improves the
	performance of UMFPACK on dense matrices by a factor of 5 or so, and
	simplifies the code.

    * Added a symmetric-preference pivoting option.  The option slightly
	(but uniformly) improves the ordering when factorizing matrices with
	symmetric nonzero pattern.  That class of matrix is better handled by
	the symmetric-pattern multifrontal method (MA41 in the Harwell
	Subroutine Library), however.

    * Fixed the detection of integer overflow.  The 32-bit version cannot
	make use of more than 2GB of main memory (use the 64-bit version
	in that case, instead).  The 32-bit version did not correctly detect
	when it was trying to factorize too large of a matrix.

May 4, 2001:

    * SGI port extended.  It can now call the SCSL Scientific Library, with
	64-bit BLAS.  Make.sgi and umf_config.h modified.

April 30, 2001:  UMFPACK Version 3.0 released.  Changes since 3.0Beta release:

    * Long integer version added (umfpack_l_* user-callable routines).

    * Peak memory usage in the numerical factorization reduced by a total of
	12n integers (8n temporary workspace used during numerical factorization,
	and 4n for the permanent LU factors which was allocated
	at the beginning of factorization).

    * Ported to the IBM RS 6000 and Compaq Alpha, with help from Anshul Gupta
	and Friedrich Grund, respectively.

    * 64-bit version added.  Uses dgemm_64, dgemv_64, and dger_64 in the Sun
	Performance Library.  64-bit versions with the BLAS might not work on
	any other platform, because they take int's as their integer input
	arguments instead of long's.  Unfortunately, the proposed ANSI
	definition of the C-BLAS also uses int's as input integer arguments.
	It ought to use long's, or include a version that uses long's, just
	like the Sun Performance Library BLAS.

    * Additional statistics returned in Info:
	Info [UMFPACK_SIZE_OF_INT]	sizeof (int)
	Info [UMFPACK_SIZE_OF_LONG]	sizeof (long)
	Info [UMFPACK_SIZE_OF_POINTER]	sizeof (void *)
	Info [UMFPACK_SIZE_OF_ENTRY]	(was Info [UMFPACK_WORD])
	Info [UMFPACK_MAX_FRONT_SIZE_ESTIMATE]	est. front matrix size
	Info [UMFPACK_MAX_FRONT_SIZE]	actual max frontal matrix size.
	Contents of Info rearranged.

    * UMFPACK_ERROR_bad_configurution error code replaced with
	UMFPACK_ERROR_problem_too_large error code.  The "bad configuration"
	error occured when sizeof (int) < sizeof (size_t).  Now, the int
	version of UMFPACK can use 32-bit int's and 64-bit pointers, and the
	long version can use 64-bit long's and 64-bit pointers.  Both versions
	check to see if the array sizes allocated are larger than what can be
	accessed by an integer index variable (int or long, depending on the
	version), and returns UMFPACK_ERROR_problem_too_large if they become
	too large.

March 15, 2001:  UMFPACK Version 3.0Beta released.