-------------
= ChangeLog =
-------------

Changes from 5.6.1 to 5.6.2
* Fixed time for symbolic factorization
* Fixed parallel analysis when PAR=0, 1 MPI per node, #MPI not a power of 2
* -DAVOID_MPI_IN_PLACE avoids MPI_IN_PLACE usage (in case not supported by MPI)
* Fixed a bug in case of parallel analysis+analysis by blocks+scalapack off
* Added missing $(PLAT) for libpord in makefile
* Fixed message printing with error -17 

Changes from 5.6.0 to 5.6.1
* Fixed compatibility of parallel analysis with analysis by blocks/out-of-range
* Minor fix of INFO(2) value in a case of INFO(1)=-7 with parallel analysis
* WRITE_PROBLEM: avoid leading space when printing "%%MatrixMarket matrix [..]"
* Fix incorrect error code -74 instead of -76 (save-restore)
* Avoid a remaining hard-coded Fortran unit (save-restore)
* Avoid -Wundef warning in case __STDC_VERSION__ is not defined
* Avoid calling MPI_IPROBE in case of a single MPI process
  (performance improvement on matrices with small fronts)
* During analysis, write time for symbolic factorization
* Update GEMMT link in INSTALL file

Changes from 5.5.1 to 5.6.0
* Analysis by blocks and out-of-range entries compatible with parallel analysis 
* Compact workarray S before solution phase (ICNTL(49)=1,2)
* JOB=-4 frees data from factorization and keeps results from analysis
* NEC vector engine version: tuned block low rank and GEMMT usage
* Reduced memory for symbolic datastructures (order-N arrays, arrowheads)
* Use new symbolic factorization (column counts) by default (+allow forest
  in case of Schur)
* Improved amalgamation algorithm to reduce the number tiny nodes
* Discard factors option is now compatible with BLR 
* Discard L factors option in case of LU factorization available for in-core 
* Count null negative pivots (INFOG(50))
* Compilation with -DBLR_MT is no longer needed (see also -DBLR_NOOPENMP)
* Improve performance of A-1 entries computation when #MPI > 1
* Fixed risk of hanging in case of OOC error (e.g. disk full)
* Free internal data earlier (e.g., factors in case of failed factorization, etc.)
* Avoid setenv/unsetenv on MinGW environments (Scotch versions >= 7)
* Fix error during ordering when processing matrices of order 1
* Out-Of-Core, values of ICNTL(22) different from 1 are treated as zero (in-core)
* BLKPTR ignored when ICNTL(15) < 0 (INFO(1:2)=-57 4 no longer occurs)
* BLR OpenMP fac_lr.F fix for possible non-increasing scheduling of loop indices
* Fix: id%LRGROUPS was not nullified in CMUMPS_FREE_ONENTRY_ANA_DRIVER
* Fixed parameter error 1202 in PDGETRS after restoring a factorization
* Workaround MPI_SSEND intel MPI 2021.6 issue
* NEC: Fix MUMPS_WRAP_GINP94 offloading to host
* Shared libraries: minor update in Makefiles
* Fixed missing printings of preprocessing constants in mumps_print_defined.F
* -DNOSCALAPACK: compilation without BLACS/ScaLAPACK (reduced performance)

Changes from 5.5.0 to 5.5.1
* 64-bit default integer version: avoid MPI buffer size > 2^31
  (could lead to negative counts in MPI_Recv with Intel ilp64 MPI)
* Use ICNTL(3) unit instead of '*' to print parallel ordering time
* Output symbolic factorization choice (ICNTL(58)) on ICNTL(3) unit
* SCOTCH 7.x on Windows: use _putenv instead of setenv
  (unsymmetric case, #MPI >1)
* Fixed parallel analysis in case of 64-bit integers in parmetis and -i8 MUMPS
* Limit recursivity depth in mumps_static_mapping.F (limit stack
  usage on trees with small nodes at the top and large nodes below)
* Limit subroutine names to 32 characters
* -DWORKAROUNDILP64MPICUSTOMREDUCE introduced (platformMPI and 64-bit default integer)

Changes from 5.4.1 to 5.5.0
* New feature: analysis by blocks with compressed graph (ICNTL(15))
* New feature: symbolic factorization using column counts (ICNTL(58))
* Matrix scaling compatible with Schur complement feature
* Automatic setting of CNTL(3) and CNTL(5) compatible with Schur 
  complement feature
* Fixed possible error in memory (INFOG(3)) and FLOPS (RINFOG(1)) 
  estimations with Schur feature
* Memory allowed (ICNTL(23)) is per MPI process and compatible with WK_USER
* More compact storage of L factors for symmetric matrices
* Exploit multithreading in SCOTCH versions >= 7.0
* Support for NEC Aurora vector processors
* Possibility to build shared instead of static libraries
* Adjustment of BLR cluster size
* Shorter time for matrix redistribution 
* Avoid a possible case of overflow in computation of MAXIS
* Fixed memory estimate of PTRAIW/PTRARW arrays
* -DDPRINT_BACKTRACE_ON_ABORT added in case of call to MUMPS_ABORT()
* More usage of dynamic memory to avoid -9 errors
* Fixed a possible "Internal error 3 in ZMUMPS_BLR_RETRIEVE_PANEL_LORU"
* Change construction of mumps_int_def.h to allow for cross compilation
* Fixed a possible error in deficiency detection on unsymmetric matrices
* Duplicated few loops during solve (OMP / non-OMP) for gfortran performance
* Fixed compilation issue with -DUPPER and -DMUMPS_WIN32 (5.4.1 regression)
* Find free Fortan units instead of imposing a choice (WRITE_PROBLEM, 
  save/restore)
* Avoid freeing IRN/JCN when not allocated by MUMPS
* Fixed a potential memory leak in dynamic memory management
* -DALLOC_FROM_C introduced to fix dynamic memory management 
  with Cray compilers


Changes from 5.4.0 to 5.4.1
* Added feature to dump a matrix/rhs in binary form (see id%WRITE_PROBLEM)
* Fixed error during constrained ordering (ICNTL(12)=3) leading to a segfault
* Fixed type 2 assembly by pieces in case of huge blocks of delayed pivots
* Avoid repeated small allocations in MUMPS_BUILD_SORT_INDEX
* Fixed automatic choice for ICNTL(6) when values are not provided at analysis
* Avoid an access to an uninitialized value (FLOP_FRFRONTS, MPI only, LDLT)
* Fixed error printing in case of error -53
* Fixed an erroneous -9 error in case of large Schur + discard factors
* -DWORKAROUNDINTELILP64OPENMPLIMITATION avoids -i8 compilation warnings
* Limited repeated amalgamations of large child with small parent

Changes from 5.3.5 to 5.4.0
* Modified default threshold for null pivot detection (CNTL(3))
* Improved performance of Schur complement computations
* Reduced memory of centralized analysis with distributed matrix
* Improved BLR clustering time during analysis
* Improved BLR performance when CNTL(1)=0.0 and ICNTL(36)=0
* Improved OOC performance in case CNTL(1)=0.0 or SYM=1
* Improved multitreaded performance on matrices with many 2x2 pivots
* Fixed a problem with parallel compilation (in case of: make -j all)
* Fixed some timings during solve that were printed as 0 when PAR=0
* Avoid possible integer overflow on INFOG(10)
* Fixed type 1 assembly by pieces in case static CB moves to dynamic storage
* Use with ITAC: avoid errors (e.g. in DMUMPS_UPPER_PREDICT...) due to
  MPI returning nonzero error codes when traces are activated
* Add a JOB=-200 call (no MPI communications, clean local OOC files)
* Fixed "Start processing the root ..." printing in case of Schur
* Fixed SYM_PERM in case of Schur on reducible matrices

Changes from 5.3.4 to 5.3.5
* Fixed 2x2 pivots bug from 5.3.4 release in MPI LDLT factorization
* Fixed ICNTL(8)=-2 option during analysis (code and documentation)
* Include basic fix for make -j

Changes from 5.3.3 to 5.3.4
* Fixed a rare bug (segfault) related to dynamic storage management
  on numerically difficult matrices
* Fixed a rare deadlock in BLR for symmetric matrices
* Fixed an uninitialized variable (which could lead to incorrect -19 error)
* Minor fix in userguide (CNTL(1) vs. ICNTL(1) in ICNTL(36) description)
* Fixed a possible runtime issue during solve, related to "TO_PROCESS" array

Changes from 5.3.1 to 5.3.3
* Assume ilp64 MPI interface only applies to Fortran in c_example.c
* Note on gfortran-10 compilation added
* Avoid intent on pointers (F2003-only)
* More robust multithreading for matrix reformatting (arrowheads)
* Fixed ICNTL(31) interpretation in case of repeated analysis
* Fixed multiple mpif.h inclusion (distributed rhs, ifort+openmpi)
* Fixed computation of effectively used memory statistics
* Minor fix in userguide
* Suppressed a !$OMP CRITICAL from solve phase (introduced in 5.3.0)

Changes from 5.3.0 to 5.3.1
* Improved multithreaded performance of BLR backward solve
* Fixed return code in build_mumps_int_def.c + openmp compilation (pgi)
* Forbid a loop vectorization in [sdcz]sol_c.F (segfault with ifort)

Changes from 5.2.1 to 5.3.0
* New feature: distributed right-hand sides
* Improved time for arrowheads construction (single MPI case, mainly)
* C interface: ability to know if MUMPS_INT is 64-bit from include file
* Improved BLR performance when CNTL(1)=0.0 and ICNTL(36)=1
* Fixed INFO(34),INFO(35),INFO(37),INFO(38) on processes with rank > 0
* More portable MPI_IS_IN_PLACE feature in libseq
* Fixed determinant computation when Cholesky ScaLapack is used
* Information on advancement (flops done) on each MPI process
* Allow rhs_sparse and irhs_sparse to be unassociated if nz_rhs=0
* Fixed INFO(30) and INFO(31) computation on MPI processes with rank > 0
* OMP collapsed loops: avoid FIRSTPRIVATE on internal loop bound (for pgi)
* Fix for compilers not freeing local allocatable arrays (64-bit metis)
* Fixed RINFO(5-6) and RINFOG(15-16) metrics (#entries=>#bytes)
* C interface: A_ELT/SCHUR/RHS/REDRHS/RHS_loc/SOL_loc may exceed 2^31 entries
* Local Schur (ICNTL(19)=2 or 3) may now exceed 2^31 entries
* Fixed internal dynamic storage of blocks with more than 2^31 entries
* Fixed a bug in the parallel analysis that limited scalability

Changes from 5.2.0 to 5.2.1
* Fixed a minor "Internal error in CMUMPS_DM_FREEALLDYNAMICCB"
* Default value of ICNTL(14) for MPI executions independent 
  of SYM + slightly less aggressive than for 5.2.0
* Avoided accesses to uninitialized data in symmetric (2D root, BLR)
* Fixed some incorrect "out" intents for routine arguments
* Avoided CHUNK=0 in OMP loops even if loop not parallelized (pgi)
* Fixed COLSCA&ROWSCA declarations in [SDCZ]MUMPS_ANA_F
* Avoided a possible segfault in presence of NaN's in pivot search
* Minor update to userguide
* Fixed MPI_IN_PLACE usage in libseq (preventing compiler optimization)


Changes from 5.1.2 to 5.2.0
* Memory gains due to low-rank factorization are now effective, low-rank solve
* Internal dynamic storage possible in case static workspace too small
* Improved distributed memory usage and MPI granularity (some sym. matrices)
* Improved granularity (and performance) for symmetric matrices; ability to
  use [DSCZ]GEMMT kernel (BLAS extension) if available (see INSTALL)
* A-1 functionality: improved performance due to solution gathering 
* Memory peak for analysis reduced (distributed-entry, 64-bit orderings)
* Time for analysis reduced by avoiding some preprocessing (when possible) 
* More exploitation of RHS sparsity during forward substitution
* Ability to save/restore an instance to/from disk
* INFO and INFOG dimension extended from 40 to 80
* METIS_OPTIONS introduced for METIS users to define some specific Metis options
* MUMPS can be asked to call omp_set_num_threads with a value provided in ICNTL(16)
* Fixed: INFO(16)/INFOG(21)/INFOG(22) did not take into account the extra memory
  allocated due to memory allowed (ICNTL(23)>0); INFOG(8) was not correclty set
* Initialize only lower-diagonal part for workers in symmetric type 2 fronts
* Workaround a segfault at beg. of facto due to a gfortran-8 bug
* Fixed a bug in weighted matching algorithm when all matrix values are 0
* Portability: include stdint.h instead of int_types.h
* Forced some initializations to make C interface more valgrind-friendly
* Workaround intel 2017 vectorization bug in pivot search (symmetric+MPI+large matrices)
* Stop trying to send messages on COMM_LOAD in case of error (risk of deadlock)
* Avoided most array creation by compiler due to Fortran pointers
* Avoid two cases of int. overflow (KEEP(66), A-1 with large ICNTL(27))
* Fixed a bug with compressed ordering (ICNTL(12)=2) (regression from 5.0.0)
  and suppress compress ordering only in case of automatic setting


Changes from 5.1.1 to 5.1.2
* Corrected an overestimation of memory (regression from 5.1.0)
* Corrected/extended WORKAROUNDINTELILP64MPI2INTEGER mechanism (see INSTALL)
* Parallel analysis: fixed a bug, limited number of MPI processes on small
  problems, and reverted to sequential analysis on tiny problems. This is
  to avoid erroneous behavior and failures in the parallel ordering tools.
* Faster BLR clustering on matrices with quasi-dense rows (which are skipped)
* Improved performance of solve phase on very small matrices
* Solve phase with a single MPI process is more thread-safe
* Fixed compilation issue with opensolaris ([SDCZ]MUMPS_TRUNCATED_RRQR)
* Fixed minor bug in BLR factorization (uninitialized timer)
* Corrected minor compiler warnings
* Minor correction to userguide
* Add -DBLR_MT in Intel example Makefile


Changes from 5.1.0 to 5.1.1
* Fix in parallel analysis
* Stabilization of 5.1.0:
  - Improved stability of Block-Low-Rank feature
  - Corrected an incorrect deallocation of POSINRHSCOMP_COL
  - Correction of a case of uninitialized data access in type 2 pivoting
  - Suppressed occasional debug trace "write(6,*) " KEEP265= ", KEEP265"

Changes from 5.0.2 to 5.1.0
* New feature: selective 64-bit integers (introduced only where needed)
  to process matrices with more than 2^{31}-1 entries.
   -mixed 32/64 bit integers for API: NNZ/NNZ_loc 64-bit
    (NZ/NZ_LOC kept temporarily for backward compatibility)
   - both 32 or 64 bit integer versions of external orderings 
    (Metis/ParMetis, SCOTCH/pt-SCOTCH, PORD), can be used
   - Error -51 when a 32-bit external ordering is invoked on 
    a graph larger than  2^{31}-1
* New feature: (experimental) factorization based on Block-Low-Rank format,
  (ICNTL(35) to activate it and CNTL(7) for low-rank precision)
* Improved performance on numerically hard matrices (LU and LDLt)
* "-DALLOW_NON_INIT" flag has disappeared and needs no longer be used
* Fixed incorrect deallocation in case of JOB=3/ICNTL(26)=1 followed by JOB=2
* Fixed compilation problem with Intel2017 + openMPI in [sdcz]ana_aux_par.F
* Minor correction of memory statistics for solve
* Use 64-bit integers where needed during the solve phase to enable 
  large number of right-hand-sides (NRHS) in one block
  (i.e. ICNTL(27)xN can be larger than 2^{31}-1)
* Improved performance of solve phase
* Allow pivoting thresholds CNTL(1) equal to 1.0
* New error -52: when default Fortran integers are 64 bit, external
  orderings should also have 64-bit default integers
* New error -22, INFO(2)=16 when IRN_loc or JCN_loc not associated
  while ICNTL(18) is set to 3
* Missing O_BINARY flag was added to open binary files on MINGW systems
* New error -53 that could reflect a matrix structure change between
analysis and factorization

Changes from 5.0.1 to 5.0.2
* Suppress error on id%SCHUR_CINTERFACE in mumps_driver.F when bound
  check is enabled and when using 2D block cyclic Schur complement
  feature (ICNTL(19)=2 or 3) from C or Matlab interfaces
* Problem of failed assertion in [SDCZ]MUMPS_TREAT_DESCBAND solved (static variable
  INODE_WAITED_FOR was not initialized and was not detected by valgrind)
* Correction of very minor memory leaks and access to uninitialized data
* A setting of INFO(1)=-1-17 should have been INFO(1)=-17
* Some settings of INFO(1)=-17 should have been INFO(1)=-20
* Suppress absolute tolerance 10^-20 in pivot selection for SYM=2;
  skip 2x2 pivot search if only 1 pivot candidate, avoid pivots
  that are subnormal numbers (their inverse is equal to infinity)
* Warning +2 now only occurs when solution is really close to 0
* Occasional bug in OOC and multiple instances solved
* Better selection of equations for bwd errors (W1 and W2) and better
  forward error estimates on some machines with 80-bit registers
* Improved users' guide (OOC files cleaning, permutation details, usage
  of multithreading, clarification of MegaByte unit)
* Cleaning of asynchronous messages after facto/solve was
  revisited and is more robust
* More robust suppression of integer overflow risk during
  solve for huge ICNTL(23)
* Improved performance of symbolic factorization in case of matrices with
  relatively dense rows and/or with large number of Lagrange multipliers
* Improved performance of numerical factorization phase during 
  pivot search for symmetric indefinite matrices
* Use of -xcore-avx2 requires !DEC$ NOOPTIMIZE in MUMPS_BIT_GET4PROC
  with current versions of Intel compilers
* Suppressed some temporary array creation and implicit conversions

Changes from 5.0.0 to 5.0.1
* Iterative refinement convergence check corrected (problem
  introduced in 5.0.0)
* Used communicator provided by user instead of MPI_COMM_WORLD
  in two places (parallel analysis only)
* Matlab interface patched to avoid memory corruption in some
  situations (Schur, colsca/rowsca management)
* Corrected a case of error not properly processed which could
  cause a segfault instead of a standard "-9" error, or an
  abort on "ERR: ERROR: NBROWS > NBROWF"
* Amalgamation without fill forced for single children
* (rare) segfault related to assemblies of delayed columns in
  scalapack root node corrected
* Automatic strategy for ordering choice improved
* Further improvements to userguide (mainly iterative
  refinement, error analysis, discard factors and forward
  elimination during factorization)
* Error -51 also raised in case of integer overflow during
  parallel analysis

Changes from 4.10.0 to 5.0.0
* Userguide revisited
* Compatibility with Metis 5.1.0/ParMetis 4.0.3, and with
  SCOTCH/pt-SCOTCH 6.0
* Matlab interface updated (scaling vectors (COLSCA, ROWSCA) and
  A-1 feature ICNTL(30) are now available)
* Improved sequential and parallel performance for computing selected entries of 
  A-1 (ICNTL(30))
* Workspace for solve phase, of size B x N per processor (B: block
  size controlled by ICNTL(27)) divided by almost #procs. Default
  value of B increased.
* Parallel symmetric indefinite elemental matrices: improved numerical behaviour
* Performance of solve phase improved 
* Finer control of error analysis and iterative refinement (ICNTL(11))
* Memory for analysis phase (mapping) reduced.
* Better support for 64-bit integers (see INSTALL file)
* Error raised instead of silent integer overflow during analysis (but
  not during external orderings)
* Improvements and corrections to parallel analysis (ICNTL(28)),
  deterministic graph construction forced with -DDETERMINISTIC_PARALLEL_GRAPH
* Forward elimination (ICNTL(32)) can be done during factorization
* Possibility to use a workspace (WK_USER, LWK_USER) allocated by user
* Very occasional numerical bug in parallel out-of-core case
  corrected (thanks to EDF and Samtech for the validation)
* More efficient processing of sparse right-hand-sides (see ICNTL(20))
* Count for entries in factors now include parallel root node
* Amalgamation of the assembly tree revisited
* Scaling arrays (COLSCA, ROWSCA) also returned at C interface level
* OOC_NB_FILE_TYPE is part of the MUMPS structure, for
  a better management of multiple OOC instances
* Warning +2 set only once (could lead to incorrect +4 in
  case of iterative refinement + error analysis)
* Warning +4 has disappeared from documentation (since it
  was never occurring -- JCN never modified on exit)
* Error code -16 now raised for the case N=0 even on distributed
  matrices (thanks to P. Jolivet for noticing this)
* Use BLAS3 routines for efficiency even in case of BLAS2 operations
  (-DMUMPS_USE_BLAS2 allows the use of BLAS2 routines for such
  operations)
* Message "problem with NIV2_FLOPS message" should no more occur
  (there was still an occasional problem in 4.10.0)
* Improved determinant computation (ICNTL(33)) in case of singular
  matrix + scaling (where zero pivots are excluded)
* Trace ' PANEL: INIT and force STRAT_IO=' suppressed
* Some OpenMP directives added (multithreaded BLAS still needed)
* Later allocation of strips of distributed fronts with improved locality
* Front factorization algorithms redesigned (two levels of panels)
* Null pivot (ICNTL(24)) and null space detection ICNTL(25)) improved
  for unsymmetric matrices
* Fortran automatic arrays (e.g. in mumps_static_mapping.F) suppressed
  to avoid risks of stack overflows
* Routine names and filenames changed

Changes from 4.9.2 to 4.10.0
* Modified variable names and variable contents in Make.inc/Makefile*
  for Windows (Makefile.inc from an older version needs modifications,
  please do a diff)
* Option to discard factors during factorization when
  not needed (ICNTL(31))
* Option to compute the determinant (ICNTL(33)) 
* Experimental "A-1" functionality (ICNTL(30))
* Matlab interface updated for 64-bit machines
* Improved users' guide
* Suppressed a memory leak occurring when Scalapack is used
  and user does loops on JOB=6 without JOB=-2/JOB=-1 in-between
* Avoid occasional deadlock with huge values of ICNTL(14)
* Avoid problem of -17 error code during solve phase
* Avoid checking association of pointer arrays ISOL_loc and SOL_loc
  on procs with no components of solution (small problems)
* Some data structures were not free at the end of the parallel analysis. Bug fixed.	
* Fixed unsafe test of overflow "IF (WFLG+N .LE. WFLG)"
* Large Schur complements sent by blocks if ICNTL(19)=1 (but
  options ICNTL(19)=2 or 3 are recommended when Schur complement
  is large)
* Corrected problem with sparse RHS + unsymmetric permutation +
  transpose solve (problem appeared in 4.9)
* Case where ICNTL(19)=2 or 3 and small value of SIZE_SCHUR
  causing problems in parallel solved.
* In case an error is detected, solved occasional problem of
  deallocating non-allocated local array PERM.
* Correction in computation of matrix norm in complex arithmetic
  (MPI_COMPLEX was used in place of MPI_REAL in MPI_REDUCE)
* Scaling works on singular matrices
* Compilation problem with -i8 solved
* MUMPS_INT used in OOC layer to facilitate compilation with
  64 bit integers

Changes from 4.9.1 to 4.9.2
* Compressed orderings (ICNTL(12)=2) are now compatible with PORD
  and PT-Scotch
* Mapping problem on large numbers of MPI processes, leading to
  INFOG(1)=-135 on "special" matrices solved (problem appeared
  in 4.9.1)

Changes from 4.9 to 4.9.1
* Balancing on the processors of both work and memory improved.
  In a parallel environment memory consumption should be reduced
  and performance improved 
* Modification of the amalgamation to solve both the problem of 
  small root nodes and the problem of tiny nodes implying too many
  small MPI messages
* Corrected bug occurring on big-endian environments when passing
  a 64-bit integer argument in place of 32-bit one. This was causing
  problems in parallel, when ScaLAPACK is used, on IBM machines.
* Internal ERROR 2 in MUMPS_271 now impossible (was
  already not happening in practice)
* Solved compiler warnings (or even errors) related to the
  order of the declarations of arrays and array sizes
* Parallel analysis: fixed the problem due to the invocation of the size
  function on non-allocated pointers, corrected a bug due to initialization
  of pointers in the declaration statements, and improved the Makefiles
* Corrected bug in the reallocation of arrays
* Corrected several accesses to uninitialized variables
* Internal Error (4) in OOC (MUMPS_597) no more occurs
* Suppressed possible printing of "Internal WARNING 1 in CMUMPS_274"
* (Minor) numerical pivoting problem in parallel LDLt solved
* Estimated flops corrected when SYM=2 and Scalapack is used (because
  we use LU on the root node, not LDLt, in that case)
* Scaling option effectively used is now returned in INFOG(33) and
  ICNTL(8) is no more modified by the package
* INFO(25) is now correctly documented, new statistic INFO(27) added

Changes from 4.8.4 to 4.9
* Parallel analysis available
* Use of 64-bit integer addressing for large internal workarrays
* overflow in computation of INFO(9) in out-of-core corrected
* fixed Matlab and Scilab interfaces to sparse RHS functionality
* time cost of analysis reduced for "optimisation" matrices
* time to gather solution on processor 0 reduced and automatic copying
  of some routine arguments by some compilers resolved.
* extern "C" added to header file mpi.h of libseq for C++ compilers
* Problem with NZ_loc=0 and scaling with ifort 10 solved
* Statistics about current state of the factorization
  produced/printed even in case of error.
* Avoid using complex arrays as real workspace (complex versions)
* New error code -40 (instead of -10) when SYM=1 is used and ScaLAPACK
  detects a negative pivot
* Solved problem of "Internal error 1" in [SDCZ]MUMPS_264 and [SDCZ]MUMPS_274
* Solved undeterministic bug occurring with asynchronous OOC + panels
  when uninitialized memory access had value -7777
* Fixed a remaining problem with OOC filenames having more than 150 characters
* Fixed some problems related to the usage of intrinsic functions inside PARAMETER
  statements (HP-UX compilers)
* Fixed problem of explicit interface in [SDCZ]MUMPS_521
* Out-of-core strategy from 4.7.3 can be reactivated with -DOLD_OOC_NOPANEL
* Message "problem with NIV2_FLOPS message" should no more occur
* Avoid compilation problem with old versions of gfortran


Changes from 4.8.3 to 4.8.4
* Absolute threshold criterion for null pivot detection added to CNTL(3)
* Problems related to messages "Increase small buffer size ..." solved.
* New option for ICNTL(8) to scale matrices. Default scaling cheaper to
  compute 
* Problem of filename clash with unsymmetric matrices on Windows
  platforms solved 
* Allow for longer filenames for temporary OOC files
* Strategy to update blocksize during factorization of frontal
  matrices modified to avoid too large messages during pipelined
  factorization (that could lead to a -17 error code)
* Messages corresponding to delayed pivots can now be sent
  in several packets. This avoids some other cases of error -17
* One rare case of deadlock solved
* Corrected values and sign of INFO(8) and INFO(20)

Changes from 4.8.2 to 4.8.3
* Fix compilation issues on Windows platforms
* Fix ranlib issue with libseq on MacOSX platforms
* Fix a few problems of uninitialized variables

Changes from 4.8.1 to 4.8.2
* Problem of wrong argument in the call to [sdcz]mumps_246 solved
* Limit occurrence of error -11 in the in-core case
* Problem with the use of SIZE on an unassociated pointer solved
* Problem with distributed solution combined with non-working host solved
* Fix generation of MM matrices 
* Fix of a minor bug in OOC error management
* Fix portability issues on usleep

Changes from 4.8.0 to 4.8.1
* New distributed scaling is now on by default for distributed matrices
* Error management corrected in case of 32-bit overflow during factorization
* SEPARATOR is now defined as "\\" in Windows version
* Bug fix in OOC panel version

Changes from 4.7.3 to 4.8.0
* Parallel scalings algorithms available
* Possibility to dump a matrix in matrix-market format from both
  C and Fortran interfaces
* Correction when dumping a distributed matrix in matrix-market format
* Minor numerical stability problem in some LDL^t parallel
  factorizations corrected.
* Memory usage significantly reduced in both parallel and sequential 
  (limit communication buffers, in-place assembly for assembled matrices,
  overlapping during stack).
* Better alignment properties of mumps_struc.h
* Reduced time for static mapping during the analysis phase.
* Correction in dynamic scheduler
* "Internal error 2 in DMUMPS_26" no more occurs, even if SIZE_SCHUR=0
* Corrections in the management of ICNTL(25), some useful code was
  protected with -Dtry_null_space and not compiled.
* Scaling arrays are now declared real even in complex versions
* Out-of-core functionality storing factors on disk 
* Possibility to tell MUMPS how much memory the package is allowed
  to allocate (ICNTL(23))
* Estimated and effective number of entries in factors returned to user
* API change: MAXS and MAXIS have disappeared from the interface,
  please use ICNTL(14) and ICNTL(23) to control the memory usage
* Error code -11 raised less often, especially in out-of-core executions
* Error code -14 should no more occur
* Memory used at the solve phase is now returned to the user
* Possibility to control the blocking size for multiple right-hand sides
  (strong impact on performance, in particular for out-of-core executions)
* Solved problems of 32-bit integer overflows during analysis related
  to memory estimations.
* New error code -37 related to integer overflows during
  factorization
* Compile one single arithmetic with make s, make d, make c or make z,
  examples are now in examples/, test/ has disappeared.
* Arithmetic-independent parts are isolated into a libmumps_common.a, that
  must now be linked too (see examples/Makefile).

Changes from 4.7.2 to 4.7.3
* detection of null pivots for unsymmetric matrices corrected
* improved pivoting in parallel symmetric solver
* possible problem when Schur on and out-of-core : Schur was splitted
* type of parameters of intrinsic function MAX not compatible in 
  single precision arithmetic versions.
* minor changes for Windows
* correction with reduced RHS functionality in parallel case

Changes from 4.7.1 to 4.7.2
* negative loads suppressed in mumps distribution

Changes from 4.7 to 4.7.1
* Release number in Fortran interface corrected
* "Negative load !!" message replaced by a warning

Changes from 4.6.4 to 4.7
* New functionality: build reduced RHS / use partial solution
* New functionality: detection of zero pivots
* Memory reduced (especially communication buffers)
* Problem of integer overflow "MEMORY_SENT" corrected
* Error code -20 used when receive buffer too small
  (instead of -17 in some cases)
* Erroneous memory access with singular matrices (since 4.6.3) corrected
* Minor bug correction in hybrid scheduler
* Parallel solution step uses less memory
* Performance and memory usage of solution step improved
* String containing the version number now available as a
  component of the MUMPS structure
* Case of error "-9964" has disappeared

Changes from 4.6.3 to 4.6.4
* Avoid name clashes (F_INT, ...) when C interface is used and
  user wants to include, say, smumps_c.h, zmumps_c.h (etc.) at
  the same time
* Avoid large array copies (by some compilers) in distributed
  matrix entry functionality
* Default ordering less dependent on number of processors
* New garbage collector for contribution blocks
* Original matrix in "arrowhead form" on candidate processors
  only (assembled case)
* Corrected bug occurring rarely, on large number of
  processors, and that depended on value of uninitialized
  data
* Parallel LDL^t factorization numerically improved
* Less memory allocation in mapping phase (in some cases)

Changes from 4.6.2 to 4.6.3
* Reduced memory usage for symmetric matrices (compressed CB)
* Reduced memory allocation for parallel executions
* Scheduler parameters for parallel executions modified
* Memory estimates (that were too large) corrected with
  2Dcyclic Schur complement option
* Portability improved (C/Fortran interfacing for strings)
* The situation leading to Warning "RHS associated in MUMPS_301"
  no more occurs.
* Parameters INFO/RINFO from the Scilab/Matlab API are now called
  INFOG/RINFOG in order to match the MUMPS user's guide.

Changes from 4.6.1 to 4.6.2
* Metis ordering now available with Schur option
* Schur functionality correctly working with Scilab interface
* Occasional SIGBUS problem on single precision versions corrected

Changes from 4.6 to 4.6.1
* Problem with hybrid scheduler and elemental matrix entry corrected
* Improved numerical processing of symmetric matrices with quasi-dense rows
* Better use of Blacs/Scalapack on processor grids smaller than MPI_COMM_WORLD
* Block sizes improved for large symmetric matrices

Changes from 4.5.6 to 4.6
* Official release with Scilab and Matlab interfaces available
* Correction in 2x2 pivots for symmetric indefinite complex matrices
* New hybrid scheduler active by default

Changes from 4.5.5 to 4.5.6
* Preliminary developments for an out-of-core code (not yet available)
* Improvement in parallel symmetric indefinite solver
* Preliminary distribution of a SCILAB and a MATLAB interface
  to MUMPS.

Changes from 4.5.4 to 4.5.5
* Improved tree management
* Improved weighted matching preprocessing:
  duplicates allowed, overflow avoided, dense rows
* Improved strategy for selecting default ordering
* Improved node amalgamation

Changes from 4.5.3 to 4.5.4
* Double complex version no more depends on
  double precision version.
* Simplification of some complex indirections in
  mumps_cv.F that were causing difficultiels to
  some compilers.

Changes from 4.5.2 to 4.5.3
* Correction of a minor problem leading to
  INFO(1)=-135 in some cases.

Changes from 4.5.1 to 4.5.2
* correction of two uninitialized variables in
  proportional mapping

Changes from 4.5.0 to 4.5.1
* better management of contribution messages
* minor modifications in symmetric preprocessing step

Changes from 4.4.0 to 4.5.0
* improved numerical features for symmetric indefinite matrices
    - two-by-two pivots
    - symmetric scaling
    - ordering based on compressed graph preserving two by two pivots
    - constrained ordering
* 2D cyclic Schur better validated
* problems resulting from automatic array copies done by compiler corrected
* reduced memory requirement for maximum transversal features

Changes from 4.3.4 to 4.4.0
* 2D block cyclic Schur complement matrix
* symmetric indefinite matrices better handled
* Right-hand side vectors can be sparse
* Solution can be kept distributed on the processors
* Metis allowed for element-entry
* Parallel performance and memory usage improved:
   - load is updated more often for type 2 nodes
   - scheduling under memory constraints
   - reduced message sizes in symmetric case
   - some linear searches avoided when sending contributions
* Avoid array copies in the call to the partial mapping routine
(candidates); such copies appeared with intel compiler version 8.0.
* Workaround MPI_AllReduce problem with booleans if mpich
  and MUMPS are compiled with different compilers
* Reduced message sizes for CB blocks in symmetric case
* Various minor improvements

Changes from 4.3.3 to 4.3.4
* Copies of some large CB blocks suppressed
  in local assemblies from child to parent
* gathering of solution optimized in solve phase

Changes from 4.3.2 to 4.3.3
* Control parameters of symbolic factorization modified.
* Global distribution time and arrowheads computation
  slightly optimized.
* Multiple Right-Hand-Side implemented.

Changes from 4.3.1 to 4.3.2
* Thresholds for symbolic factorization modified.
* Use merge sort for candidates (faster)
* User's communicator copied when entering MUMPS
* Code to free CB areas factorized in various places
* One array suppressed in solve phase

Changes from 4.3 to 4.3.1
* Memory leaks in PORD corrected
* Minor compilation problem on T3E solved
* Avoid taking into account absolute criterion
  CNTL(3) for partial LDLt factorization when whole
  column is known (relative stability is enough).
* Symbol MPI_WTICK removed from mpif.h
* Bug wrt inertia computation INFOG(12) corrected

Changes from 4.2beta to 4.3
* C INTERFACE CHANGE: comm_fortran must be defined
  from the calling program, since MUMPS uses a Fortran
  communicator (see user guide).
* LAPACK library is no more required
* User guide improved
* Default ordering changed
* Return number of negative diagonal elements in LDLt
  factorization (except for root node if treated in parallel)
* Rank-revealing options no more available by default
* Improved parallel performance
    - new incremental mechanism for load information
    - new communicator dedicated to load information
    - improved candidate strategy
    - improved management of SMP platforms
* Include files can be used in both free and fixed forms
* Bug fixes:
    - some uninitialized values
    - pbs with size of data on t3e
    - minor problems corrected with distributed matrix entry
    - count of negative pivots corrected
    - AMD for element entries
    - symbolic factorization
    - memory leak in tree reordering and in solve step
* Solve step uses less memory (and should be more efficient)

Changes from 4.1.6 to 4.2beta
* More precisions available (single, double, complex, double complex).
* Uniprocessor version available (doesn't require MPI installed)
* Interface changes (Users of MUMPS 4.1.6 will have to slightly
  modify their codes):
     - MUMPS -> ZMUMPS, CMUMPS, SMUMPS, DMUMPS depending the precision
     - the Schur complement matrix should now be allocated by the
       user before the call to MUMPS
     - NEW: C interface available.
     - ICNTL(6)=6 in 4.1.6 (automatic choice) is now ICNTL(6)=7 in 4.2
* Tighter integration of new ordering packages (for assembled matrices),
  see the description of ICNTL(7):
     - AMF, 
     - Metis,
     - PORD,
* Memory usage decreased and memory scalability improved.
* Problem when using multiple instances solved.
* Various improvments and bug fixes.

Changes from 4.1.4 to 4.1.6
* Modifications/Tuning done by P.Amestoy during his
  visit at NERSC.
* Additional memory and communication statistics.
* minor pbs solved.

Changes from 4.0.4 to 4.1.4
* Tuning on Cray T3e (and minor debugging)
* Improved strategy for asynchronous 
  communications 
  (irecv during factorization) 
* Improved Dynamic scheduling 
  and splitting strategies
* New maximal transversal strategies
* New Option (default) automatic decision 
  for scaling and maximum transversal



-------------------
= Release history =
-------------------

Release 5.6.2            : October 2023
Release 5.6.1            : July 2023
Release 5.6.0            : April 2023
Release 5.5.1            : July 2022
Release 5.5.0            : April 2022
Release 5.4.1            : August 2021
Release 5.4.0            : April 2021
Release 5.3.5            : October 2020
Release 5.3.4            : September 2020
Release 5.3.3            : June 2020
Release 5.3.2[.x]        : May 2020, internal/experimental
Release 5.3.1            : April 2020
Release 5.3.0            : April 2020
Release 5.2.1            : June 2019
Release 5.2.0            : April 2019
Release 5.1.2            : October 2017
Release 5.1.1            : March 2017
Release 5.1.0            : Feb 2017, internal release (limited diffusion)
Release 5.0.2            : July 2016
Release 5.0.1            : July 2015
Release 5.0.0            : February 2015
Release 4.10.0           : May 2011
Release 4.9.2            : November 2009
Release 4.9.1            : October 2009
Release 4.9              : July 2009
Release 4.8.4            : December 2008
Release 4.8.3            : September 2008
Release 4.8.2            : September 2008
Release 4.8.1            : August 2008
Release 4.8.0            : July 2008
Release 4.7.3            : May 2007
Release 4.7.2            : April 2007
Release 4.7.1            : April 2007
Release 4.7              : April 2007
Release 4.6.4            : January 2007
Release 4.6.3            : June 2006
Release 4.6.2            : April 2006
Release 4.6.1            : February 2006
Release 4.6              : January 2006
Release 4.5.6            : December 2005, internal release
Release 4.5.5            : October 2005
Release 4.5.4            : September 2005
Release 4.5.3            : September 2005
Release 4.5.2            : September 2005
Release 4.5.1            : September 2005
Release 4.5.0            : July 2005
Releases 4.3.3 -- 4.4.3  : internal releases
Release 4.3.2            : November 2003
Release 4.3.1            : October 2003
Release 4.3              : July 2003
Release 4.2 (beta)       : December 2002
Release 4.1.6            : March  2000 
Release 4.0.4            : Wed Sept 22, 1999 <-- Final version from PARASOL

