2008-01-25:
	- started ChangeLog
        - added copying of horizontal diagonal ghost cells
          to fd4_iter_get_ghost, enable with -DFD4_DIAGONAL_HORIZONTAL

2008-02-07:
	- added block%cells - an array for fast direct access to neighbor cells
        - added fd4_iter_get_ghost3 & fd4_iter_get_ghost4 - versions which use block%cells
        - use -DFD4_BLOCK_CELLS to enable

2008-02-15:
	- started implementing a version that allocates ghost cells within blocks
        - use -DFD4_BLOCKS_HAVE_GHOSTS to enable
        - NOT DONE: boundary exchange

2008-02-17:
	- now calculating maxnsteps, field_len, and field indices in a new
          fd4_vartab routine, which is called once in fd4_domain_create
        - new type fd4_varinfo in fd4_vartab_mod holds this information
        - this new type is used in fd4_domain and in fd4_block_allocate

2008-02-18:
        - Tested a version of fd4_iter which uses varinfo%fields in fd4_iter to get/put
          directly from/to sdata(:) without using the fd4_block%fields pointers.
          But it seemed to be slower than the original version. See fd4_iter_SDATA.F90.

2008-02-21:
	- added routine fd4_iter_ghost_copy which copies boundary from neighbors blocks to current block
          when using the FD4_BLOCKS_HAVE_GHOSTS version 
        - added version FD4_BLOCKS_NOSDATA, where block%sdata is not present. instead the data is directly
          allocated to block%fields(idx,st)%l
        - the versions FD4_BLOCK_CELLS, FD4_BLOCKS_HAVE_GHOSTS, FD4_BLOCKS_NOSDATA are running (do not mix!)

2008-02-27:
	- Bugfix: used optional variable idx in fd4_vis5d_write
        - Workaround for PGI: using additional real_variable when definining FD4_NOTHRES = huge(real_variable)

2008-03-19:
	- Testing some versions with constant domain parameters (fd4_parameter_test)
	- Advections test was not faster than before

2008-03-26:
	- Implemented general boundary condition support via boundary ghost blocks
        - Periodic boundary still handled by domain through periodic neighbor pointers (trivial)
        - New module fd4_ghostlist to iterate over boundary ghost blocks
        
2008-04-02:
        - New module fd4_boundary which implements boundary condition handling (currently only zero gradient)
        
2008-04-03:
	- Modified version of hilbert_mod (hilbert1.F90) uses start indices at 1 instead of 0
        - Starting with MPI parallelization

2008-04-09:
        - Bugfix in fd4_domain (concerning usage of hilbert1.F90)

2008-06-04:
	- fd4_balance_readjust working, starting implementing comm. ghost blocks

2008-06-20:
	- simple implementation of fd4_ghostcomm_exch working

2008-06-23:
	- vthres is now checked against abs(value) so that negative values are not considered null
          (changed fd4_block_empty_dataloop and fd4_util_put_array)

2008-06-26:
	- parallel advection tests performed with current simple ghostcomm

2008-07-07:
        - improved ghostcomm to send one message (mpi datatype) for each neighbor, thus reducing number of messages

2008-07-09:
        - improved ghostcomm to save the mpi datatype for each neighbor, thus reducing number of mpi_type_create_struct's

2008-07-14:
        - fixed bugs in fd4_ghostcomm and fd4_balance interaction, minor performance improvements
        - removed BLOCK_CELL, BLOCKS_HAVE_GHOSTS, and NO_SDATA versions

2008-07-17:
        - fd4_mpitype.F90 for better construction of complex mpi datatypes, used in fd4_ghostcomm and fd4_util_get_array
        - new version of fd4_util_get_array, much faster
        - FD4_COMM_EMPTY now default, because of some consistency problems

2008-07-23:
	- faster version of fd4_balance_readjust (faster method to compute new partition with less comm.,
          leave routine if nothing has changed, send only one message to neighbor when transferring blocks)

2008-07-28:
	- tiny serial optimizations in fd4_balance_readjust
        - option to use only a subset of the array in fd4_util_put_array

2008-07-21:
	- using a faster mapping 3D <--> 1D for indexing blocks (no SFC), hilbert is only used in fd4_part_sfc
        - new stack3_mod for a stack of 3D coordinates
        - the index (for sorting) of fd4_mpitype is now a 8 byte integer to prevent integer overflows
          - this required to add a 8 byte integer sorting function to mrgrnk.F90
	- added .incf-files for stack, stack3, and fd4_domain_i2c/c2i functions to enable inlining
        - small bugfix in fd4_vis5d when using FD4_VIS5D_SINGLE_BLOCKSTRUCTURE
        

2008-08-20:
	- fd4_part_metis.F90: first implementation of ParMetis repartitioning function working

2008-08-28:
	- implemented handling of processes without blocks after ParMetis repartitioning (only if there are
          enough blocks for all processes)
        - ParMetis needs some fixes to work properly (but not sure if this is enough for #blocks<#procs cases):
          - iidxsort.c:144:    volatile idxtype *mptr;       (volatile added)
          - wave.c:50:         solution = fmalloc(4*nparts+3*nedges, "WavefrontDiffusion: solution");      (more memory)
          - wave.c:57:         perm = idxmalloc(2*nvtxs+2*nparts+2*nedges+1, "WavefrontDiffusion: perm");  (more memory)
          - I think there is still a memory leak in ParMetis

2008-09-09:
        - allocation and deallocation of boundary ghost blocks in fd4_balance_readjust instead of fd4_domain_add_block
          (use FD4_INSTANT_BGB to disable)
	- implemented handling of processes without blocks *before* ParMetis repartitioning

2008-09-11:
        - fixed problem in fd4_vis5d: col and row were swapped in call to v5dcreate
        - fd4_vis5d now uses generic coordinates

2008-09-26:
        - new build system with config files in config/
        - advection test program in test/

2008-11-04:
	- workaround for gfortran bug in fd4_mpitype
        
2008-11-27:
        - added subroutine fd4_iter_empty to start empty check without using fd4_iter_put_ghost
        - support for cases where all variables are FD4_NOTHRES, i.e. no adapting block allocation
          - option FD4_NO_COMM_THRES to make ghost block allocation independent from vthres
          - parameter opt_ignore_vthres for fd4_util_put_array to enable block allocation for FD4_NOTHRES variables

2008-12-05:
        - added fd4_boundary_spec routines for special boundary conditions 
        
2008-12-11:
        - new module fd4_couple for coupling to other codes
        - fd4_couple_put as a first simple implementation to put data to the framework

2008-12-15:
        - small bugfixes (compiler compatibility, 32bit MPI)

2008-12-19:
        - added optional arguments to fd4_vis5d_open to enable nonequidistant and reversed height levels as in COSMO

2009-04-20:
	- added experimental fd4_part_metis_subset which calls ParMETIS only on a subset of ranks

2009-05-14:
        - introduced block weights (block%weight)
        - use them in fd4_part_metis and fd4_part_sfc
        - migration-optimization for fd4_part_sfc (new module partition1d)

2009-05-20:
        - coupling routine now handles multiple blocks per process
        - reverse coupling implemented (fd4_couple_get)

2009-05-25:
        - optional parameter startdispl for fd4_mpitype_commit 

2009-06-03:
        - first version of NetCDF output (works, but slow)       

2009-07-09:
        - replaced all block-index calculations by renewed fd4_domain_c2i routines (3*10bit index)
        - started fd4_couple_mark_blocks mechanism to allow effective coupling of vthres vars (still very slow)

2009-07-15:
        - tuned fd4_couple_mark_blocks and couple_put

2009-09-11:
        - applied couple_put optimizations to couple_get (couple context)
        - multiple typestores per block: this enables usage of multiple couple contexts

2009-09-18:
	- many fixes in fd4_couple, now allowing overlapping couple arrays for face variables in fd4_couple_get
        - added balance method FD4_BALANCE_SERIAL which moves all blocks to rank 0 (for debug)
        - new routine fd4_util_check_facevar which checks consistency of face variable values

2009-10-01:
        - fd4_couple can now handle 4D variables (for get, put, and mark_blocks)
        - fd4_couple_mark_blocks now uses the couple context

2010-01-06:
        - new subroutine fd4_iter_get_ghost_multi to get multiple variables at a time ("expert mode")
        - fd4_couple optimization: matching of blocks with couple arrays now faster using union of local blocks
          and union of couple arrays
           - fd4_domain:   domain%blockunion - spatial union of all local blocks bounds
           - fd4_balance:  blockunion is updated in ghost update routine 

2010-01-14:
        - fd4_balance: introduced new balance method FD4_BALANCE_HILBERT2 to swith between simple
          Hilbert SFC (FD4_BALANCE_HILBERT) and alternative version FD4_BALANCE_HILBERT2 which uses
          partition1d_mod for optimizing balance and migration costs.

2010-01-20:
        - fd4_comm: preprocessor switch FD4_MPI_TYPE_CREATE_RESIZED:
          Use MPI_Type_create_resized (MPI-2) instead of deprecated MPI_LB and MPI_UB.

2010-01-25:
        - optimization in fd4_part_sfc: if same blocks are present as in last call, reuse last hilbert sfc ranking
        - new util module section_mod, used to create clusters of rect. block unions for coupling (domain%blockunion)

2010-01-28:
        - added constants FD4_GHOST_BND and FD4_GHOST_CMM to identify ghost block types
        - new routine fd4_domain_delete_ghost (called by fd4_balance_update_ghosts, fd4_domain_delete_block, fd4_domain_add_block)
        - implemented pool for reusing removed (ghost) blocks without deallocation and allocation (domain%pool)
        - added statistics about real block (de)allocation to the fd4_balance_statistics

2010-02-02:
        - optimization for block allocation and reset in the case all variables have same vnull
        - fd4_balance_readjust:
           - parameter opt_ech to disable global empty check for cases when empty did not change
           - reusing allblocks array when empty did not change (allblocks is now part of domain%balance)

2010-02-03:
        - changed block%weight from integer to real, introduced a factor ParMetis (needs int) in fd4_part_metis

2010-02-04:
        - implemented smart "auto" mode for load balancing (time lost due to imbalance vs. time lost due to fd4_balance_readjust)
        - fd4_balance_statistics%status tells us what fd4_balance_readjust did and why

2010-02-05:
        - removed FD4_BLOCK_WEIGHTS, using weights is now default, can be disabled with parameter opt_weight=.false.
          of fd4_balance_readjust
        - scalability optimzation in fd4_couple: using mycplarrays when running over local couple arrays only

2010-02-08:
        - scalability optimzation in fd4_couple: using cpldescbnd when running over couple desc of specific rank
        - added opt_commit parameter to fd4_comm_build_* subroutines and using it wherever possible to disable
          commiting the type when it is used as a component of a bigger type:
            fd4_ghostcomm, fd4_balance, fd4_util, fd4_couple

2010-02-10:
        - replaced fd4block%typestore (stored in each block) by a global typestore in the domain: cpltypestore
        - cpltypestore is a new rbtree_int8, an rbtree which stores integers with an 64 bit integer key
        - new routine fd4_domain_block_encode to generate the key for the cpltypestore

2010-02-17:
        - fd4_part_sfc completety revisited: new algorithms, partition1D not longer used
        - trivial parallel SFC algorithm
        - using real(r8k) in all partitioning routines for weights/loads (blocks still store weights as r4k)
        - exchanging weights is now part of fd4_part_sfc (moved from fd4_balance)
        - started math.incf for basic math functions (currently only log2i)
        
2010-02-17:
        - added free_types parameter to fd4_mpitype_free and used it where neccessary
        - minor fixes found by compiling with gfortran

2010-02-19:
        - preprocessor macro FD4_WAIT_BEFORE_TYPE_FREE to force FD4 to keep the types until mpi_waitall finished

2010-03-01:
        - bugfix for IBM: using first element of array ("array(1,1,1)" instead of "array") for MPI_Irecv,
          MPI_Isend, MPI_Get_address, etc. - changes in various files

2010-03-08:
        - using FD4_WAIT_BEFORE_TYPE_FREE in fd4_util_propagate_facevar

2010-03-24:
        - support for serial output via NetCDF: fd4_netcdf_serial.F90
        - works with NetCDF3 and NetCDF4

2010-04-07:
        - bugfix in fd4_domain_block_encode: ishft requires i8k arguments, solved severe problems with fd4_couple

2010-04-15:
        - new implementation of fd4_comm_build_array_type based on MPI_Type_create_subarray
        - dropped fd4_comm_build_sdata_type, using fd4_comm_build_array_type instead (ghostcomm, couple)

2010-04-16: some clean-up
        - removed opt_commit parameter for fd4_comm routines
        - removed compile time options: FD4_NO_MPI_STATUSES_IGNORE, FD4_USE_SDATA, FD4_VIS5D_SINGLE_BLOCKSTRUCTURE
        
2010-04-21:
        - introduced the term "adaptive block mode", which is implicitely enabled if any variable has a threshold
        - load balancing never removes/adds blocks if not in adaptive block mode
        - new subroutine fd4_util_allocate_all_blocks to force allocation of all blocks (mark + fd4_balance_readjust)
        
2010-05-05:
        - bext return argument of iter_get_* routines is now optional

2010-05-06:
        - fd4_iter_start_clock and fd4_iter_stop_clock to time the block computation and set block weight
        - moved definition of fd4 preprocessor macros to config/fd4flags.in

2010-06-15:
        - bugfix in fd4_balance_update_ghost: forgot to set neighbor pointer to null when removing ghost block

2010-06-15:
        - bugfix in fd4_iter: opt_bext(0) was not set properly (2010-05-05)

2010-07-01:
	- added GNU GPLv3 copyright - FD4 is free software!

2010-07-09:
        - bugfix: fixed wrong intent in fd4_iter_stop_clock
        - bugfix: domain%balance%rnk and domain%balance%hidx need to be initialized with NULL()

2010-07-21:
        - fd4_ghostcomm_create: new optional parameter opt_exchdim to decide which dimensions to exchange for which variable

2010-07-27:
        - util/section: severe bug in section_intersect fixed
        - fd4_couple: new internal data structures and new interface for coupling, solved some bugs/problems

2010-07-30:
        - bugfix in fd4_balance: balance tolerance ignored if opt_auto was set to false

2010-08-19:
        - updates in tutorial, README, and manual
        - fd4_vartab: introduced parameters for vartab%face (FD4_CELLC, FD4_FACEX, FD4_FACEY, and FD4_FACEZ)
        - fd4_domain_create_specific: fixed little bug with the err argument
        - fd4_mod: MPI_I4K, MPI_I8K, MPI_I_K, MPI_R4K, MPI_R8K, MPI_R_K, MPI_R_K_size are now part of the user interface

2010-11-03:
        - type(fd4_domain) declared as target in all example programs
        - modified mkpackage to include only some of the config files (only conf.default, conf.doc, conf.example.*)
        - added target attribute to some (dummy) variables in couple, netcdf4, netcdf_serial, iter, util
        - added writing of block weights per grid cell to netcdf4 output (serial and parallel) and vis5d

2010-11-16:
        - little optimization in fd4_part_sfc (clever search strategy, enabled by default with CLEVER_SEARCH)

2010-11-29:
        - fd4_netcdf4 and fd4_netcdf_serial: append "_<timestep>" to variable names when same variable index appears
        - fd4_netcdf4: bugfix (used i instead nfcomm%idx(i) as variable index)
        - fd4_boundary: new routine fd4_boundary_spec_block, a little more flexibility, changed interface
        - updates in tutorial and manual

2011-03-11:
        - added util/hilbert2d1.F90 to calculate 2D Hilbert SFC
        - use 2D Hilbert SFC in fd4_part_sfc in case block decomposition is less than 3-dimensional

2011-04-01:
        - added util/morton1.F90 to calculate 3D and 2D Morton SFC
        - Morton SFC can be used via opt_sfctype switch of fd4_balance_readjust
        - added estimated_balance to fd4_balance_statistics (only filled when calling fd4_part_sfc_weight)
        - preprocessor option FD4_COMM_XFER_WGHT for fd4_comm: also transfer current block weight to new destination
        - implemented recursive coordinate bisection (testing version, do not use!) (FD4_BALANCE_RCB, new fd4_part_rcb.F90)
        - bugfix in log2i (math.incf)

2011-04-13:
        - bugfix in fd4_domain (cplblkenc generation, coupling did not work correctly since 2011-04-01)
        - introduced default lb params in fd4_domain_lb_params and subroutine fd4_balance_params to
          set the defaults, can be overriden when calling fd4_balance_readjust

2011-04-14:
        - fd4_part_metis can now be used for initial partitioning (fd4_util_allocate_all_blocks, fd4_util_put_array)
        - fd4_part_rcb prototype works in adaptive block mode - don't use for production!

2011-04-19:
        - bugfix in fd4_couple: ranks without own couple arrays crashed
        - fd4_netcdf4 completely redesigned, now using fd4_couple to transform data to IO-friendly decomposition

2011-05-02:
        - bugfix in fd4_balance: fd4_mpitype was used without "%datatype" in MPI_Isend and MPI_Irecv

2011-05-16:
        - fd4_couple: new parameter opt_cpldir for fd4_couple_create to restrict usage of a couple context
          to put or get and allow some memory optimizations (prepare less mpi types).

2011-05-18:
        - added util/rbtree_fd4_neigh.F90
        - fd4_ghostcomm: using rbtree_fd4_neigh to allocate only fd4_neigh for neighbors and not for all ranks

2011-05-19:
        - fd4_balance: using rbtree_fd4_neigh to allocate fd4_mpi_type for migration destinations only and not for all ranks
        - fd4_comm_build_type: bugfix: FD4_COMM_XFER_WGHT did not work properly when not all sdata(0:3) arrays where allocated
        - fd4_balance, fd4_block: mpi type for block migration now stored in the block and reused in fd4_balance
        - moved fd4_comm_build_type to fd4_block, now called in fd4_block_allocate only

2011-05-22:
        - fd4_block: added option to disable allocation of face variables
        - fd4_domain: block pools changed from arrays to linked lists
        - fd4_domain: separated pools for ghost blocks and normal blocks, number of ghost block pool lists reduced by gpmap
        - fd4_domain: ghost blocks do not allocate face variables and do not create an MPI type for migration
        - fd4_stat: added module for collecting statistics, output is dumped in fd4_domain_delete
        - added statistics to fd4_domain, fd4_balance, fd4_ghostcomm, and fd4_couple
        - fd4_couple: removed global commtype array, now using rbtree_fd4_neigh
        - fd4_util_propagate_facevar: removed global commtype arrays, now using rbtree_fd4_neigh
        - makro FD4_WAIT_BEFORE_TYPE_FREE removed (not necessary any longer)
        - fd4_stats: added min and max SFC index, dumping to file is enabled with makro FD4_STAT_DUMP_SFCBND

2011-05-27:
        - fd4_couple: makro FD4_COUPLE_FACE_ARRAYS_OVERLAP is now a run-time argument of fd4_couple_commit
        - makro FD4_INSTANT_BGB changed to FD4_ADAPT_BGB (inverted), default is not FD4_ADAPT_BGB

2011-05-30:
        - util/timing: added gettime_papi, general timing routine gettime with timer argument
        - fd4_iter: general timing routine used in fd4_iter_start_clock and fd4_iter_stop_clock
        - fd4_domain: the timer used in fd4_iter is defined in domain%timer

2011-06-05:
        - util/section: new clustering algorithm, better quality, less cluster size, a bit slower
        - fd4_domain: max blockpool size raised
        - util: purified some routines

2011-06-20:
        - util/section: bugfix in section_reset
        - fd4_stat: bugfix in fd4_stat_add

2011-07-14:
        - removed mpi%recvint, mpi%sendint, mpi%sendcnt

2011-08-01:
        - fd4_couple: bugfix in fd4_couple_commit, wrong types were constructed for FD4_CPL_BOTH

2011-08-02:
        - fd4_balance: removed needless rbtree_fd4_block_get in block migration

2011-08-14:
        - fd4_part_sfc: number of binary search steps now determined by required accuracy (exact method for accuracy=1.0)
        - fd4_part_sfc: added max. required number of binary search steps to statistics
        - fd4_balance, fd4_part_sfc: added accuracy and binsteps sfc load balancing parameters

2011-08-28:
        - fd4_balance, fd4_part_*, fd4_util: myblocks is now a stack3 and not a large array anymore

2011-09-09:
        - fd4_balance, fd4_domain: domain%empties is only allocated in adaptive block mode
        - fd4_balance, fd4_domain, fd4_part_*: domain%allblocks is only allocated in adaptive block mode

2011-09-12:
        - fd4_domain, fd4_part_*, fd4_netcdf*, fd4_vis5d: blkweights only allocated if required by partitioning method
        - fd4_part_1d: new module which contains the 1d-partitioning routines usind in fd4_part_sfc
        - fd4_part_sfc: new internal helper subroutines fd4_part_sfc_index and fd4_part_sfc_set_blkmap_boundary
        - fd4_part_sfc: fd4_part_new renamed to fd4_part_sfc_noweight, added sfctype argument
        - fd4_part_sfc: started new fd4_part_sfc_weight2
        - fd4_domain.incf: new routine fd4_domain_blext

2011-09-13:
        - util: morton and hilbert SFC routines with given query region size for contiguous SFC indexes
        - fd4_part_sfc: fd4_part_sfc_weight2 working - needs much less global memory than fd4_part_sfc_weight

2011-09-15:
        - util: morton and hilbert SFC routines to iterate over curve with query region size
        - fd4_part_sfc_weight2 uses SFC iterator to speed-up generation of block map

2011-09-24:
        - fd4_md_subdomain renamed to fd4_mdsd
        - fd4_couple: couple arrays for face variables must be defined without face extension
          (saves memory for array bounds)
        - fd4_couple, fd4_mdsd: new coupling meta data structures:
          requires much less memory, all ranks must define same set of variables, desc_var removed from mdsd
        - fd4_couple: new interface $fd4_couple_add_partition and $fd4_couple_add_var

2011-09-27:
        - fd4_mdsd: changed the stack3s to stack3
        - re-using the mpi type from last coupling if possible (stored in couple%neighstore)
        - statistics for number of couple types created

2011-09-29:
        - rbtree: renamed _iterator to _iter to circumvent problem of possibly too long symbol names
        - fd4_metadata: unused mdsds can be buffered with fd4_metadata_adapt
        - fd4_domain: user interface routine fd4_domain_adapt_metadata to enable mdsd buffering added

2011-10-06:
        - bugfix in couple statistics: "Couple put Bytes" were not advanced if no new type was created

2011-10-20:
        - added several statistics for timings in fd4_couple and fd4_balance
          note: time for partitioning calculation (stat_part_us) includes exchange_weights (stat_exwgt_us)
        - added FD4_BALANCE_HSFC2_BINSRCH
        - load balancing statistics are created in initialization of default load balance parameters (not domain_create)
        - added fd4_part_1d_simple_h2 (Heuristic 2 from Miguet and Pierson)
        - improved performance of fd4_part_1d_simple_h1 and fd4_part_1d_simple_h2 by using fd4_part_1d_scan
        - fd4_metadata: mdsd buffer now implemented as 3D array instead rbtree

2011-10-24:
        - added timing statistics for 1d partitioning method domain%balance%stat_p1d_us
        - bugfix in fd4_part_rcb
        - fd4_block, fd4_domain: allocating block%empty in adaptive block mode only
        - using b%sempty of ghost blocks to store the owner of the remote block (reducing accesses to blkmap)
        - moved partition vector from fd4_part_sfc_weight/fd4_part_sfc_weight2 to domain%balance%partition
        - added partition vector from last partitioning
        - added load balancing parameter opt_blkmap_update to disable filling of complete domain%blkmap
          in fd4_part_sfc_weight2 (FD4_BALANCE_HSFC2_PARALLEL and FD4_BALANCE_HSFC2_BINSRCH)
        - added statistics counter for number of full blkmap updates in case of opt_blkmap_update=.false.
        - fd4_balance: removed SECTION_ADD_DEBUG

2011-10-28:
        - fd4_part_metis: statistics for number of emergency block migrations, some optimizations for non-adaptive block mode
        - fd4_part_metis, fd4_balance: new balance parameters use_adjwgt and use_vsize
        - fd4_part_metis, fd4_part_rcb: bugfix for all_blocks (pointer argument)
        - modifications for ParMetis 4.0.1
        - fd4_part_sfc: bugfix in fd4_part_sfc_update_blkmap_bnd (neighbor blocks must also be updated)
        - fd4_couple: bugfix for dims with periodic bounds with only one grid cell size (2D grid)

2011-11-03:
        - default value of FD4_MD_MAX_PART reduced to 2 (from 8)
        - macro FD4_COUPLE_MAX_PARTITIONS to set FD4_MD_MAX_PART in fd4_mdsd
        - macro FD4_ESTIMATE_NEW_BALANCE to enable measurement of achieved balance after load balancing
        - fd4_part_1d: optimzed fd4_part_1d_simple_h1 and fd4_part_1d_simple_h2

2011-11-18:
        - added FD4_MULTIPLE_COMMS for extra MPI_Comm_dups for ghostcomm, coupling and migration

2011-11-25:
        - added fd4_balance_set_partition to set a partition explicitly (for experts)
        - added timers for stat_pfxs_us (prefix sum in sfc methods) and stat_blkm_us (blkmap update after repartitioning)

2012-01-10:
        - fd4_part_sfcd: distributed and hierarchical SFC load balancing method
        - fd4_part_sfc: moved small internal routines to fd4_part_sfc.incf for usage in fd4_part_sfcd
        - fd4_balance: added method FD4_BALANCE_HSFC_DISTR (for fd4_part_sfcd)
        - fd4_balance: added balance parameters groupsz and wgtxmeth (for fd4_part_sfcd)
        - fd4_stat: added option to dump a counter only if sum>0

2012-01-17:
        - shift lengths for fd4_domain_c2i now compile-time parameter DFD4_DOMAIN_C2I_SHFT3 and DFD4_DOMAIN_C2I_SHFT2

2013-03-14:
        - many changes in the 1D partitioning subroutines:
          - new methods (recursive bisection, exact bisect)
          - parallel prefix sum for all methods
          - serial methods: computation of 1d partition on one rank or all
          - distributed method with serial computation within group
          - distributed method with heuristic within group
          - and many more changes

2013-06-04:
        - bugfix in fd4_domain: reserved insufficient bits for bnd in cplblkenc

2014-04-03:
        - fixed problem with large number of blocks in fd4_domain_create 
        - bugfix in fd4_balance for nghosts=0, partition_id was not incremented
        - added option FD4_SAVE_SFC_IDX: save SFC index in blocks, speeds-up prefix sum of block weights

2014-11-05:
        - changed some default config parameters
        - fd4_iter_stop_clock: fact is now optional argument opt_fact
        - new release

2015-11-12:
        - fd4_block: bugfix in fd4_block_reset: resetting face variables was broken
        - fd4_netcdf*: added argument fillvalue_opt to set fillvalues other than vnull

2016-01-21:
        - added fd4_part_zoltan: uses zoltan to compute (geometric) partitioning
        - fd4_balance: new subroutine fd4_balance_blkmap_update_nosfc to update the full blockmap if ranks know
          only their local changes (and no partition vector from SFC methods)

2016-03-04:
        - added fd4_part_dir: distributed partition directory to avoid communicating partitioning to all
          when running distributed SFC methods (does not work with ghostcomm and coupling)
        - fd4_balance: migration stats are now local
        - fd4_part_sfc: added FD4_PART_1D_BINSRCH_INI_RB
        - various small changes
        - now using fixed MPI tags everywhere

2016-06-09:
        - fd4_netcdf_serial: enabling 64bit offset for large file support
        - may be disabled with fd4_netcdf4_open(..., offset64_opt=.false.)
        - new release
