History of Changes


Version: NPB3.3-MZ

[27-Aug-07]
 - Inclusion of the nested OpenMP implementation (NPB3.3-MZ-OMP)
   to replace the SMP+OpenMP implementation.

 - Merge vector codes into the standard distribution.
   The vector codes can be selected with the VERSION=VEC option
   during compilation.

 - Performance improvement to SP-MZ.

 - Minor bug fix in mpi_setup.f for mismatched arguments


Version: NPB3.2-MZ

[02-Nov-05]
 - Introduce two new classes of problem: E and F
   Versions affected: all

 - Combine the 'timer.flag' file into 'input<benchmark>.data' so that
   parameters are more consistently handled.
   Versions affected: all

 - Add "SCHEDULE(STATIC)" to the OpenMP layer to ensure OpenMP codes
   won't break under different types of runtime scheduling.
   Versions affected: all parallel versions

 - Minor cache optimization in z_solve of BT-MZ.
   Versions affected: all

 - Speed optimization in setparams.c (mainly get_comm_index).
   Versions affected: all parallel versions

 - New environment variable NPB_VERBOSE to control verbose printing.
   Versions affected: all

 - Re-order communication buffers so that large messages can be sent.
   Versions affected: MPI version

 - Fix a dependency bug in LU-MZ/Makefile, add variable 'v' to
   "!$OMP FLUSH()" to ensure data properly flushed for pipeline.
   Versions affected: all parallel versions


Version: NPB3.1-MZ

[29-Mar-04]
 - Allow a variable number of threads for individual process and define
   a processor group for each process.  This is done by reading data 
   from a load data file, i.e. loadbt-mz.data, loadsp-mz.data, or 
   loadlu-mz.data.
   Versions affected: all parallel versions

 - Report the total number of threads (instead of threads-per-process).
   Versions affected: all parallel versions

 - Use a surface term in the MFLOPS calculation so that the results
   take into account the variable dimension sizes of multi-zones.
   Versions affected: all

 - Improvements for LU-MZ:
    * removed a redundant 'exch_qbc' call
    * add one SSOR iteraction before the timing loop to touch all data pages
    * improved the memory usage for array 'tv'
    * made rhs and erhs more cache-friendly
   Versions affected: all

 - Print built-in timers for all processes
   Versions affected: all parallel versions


Version: NPB3.0.2-MZ

[22-Dec-03]
 - Bug fix in NPB3.0-MZ-MPI: 'omp_set_num_threads()' is required
   for the MPI+OpenMP version.  Use broadcast for setting global parameters.

[02-Jan-04]
 - Minor improvement in sys/setparams.c: Calculate the exact value
   for problem_size, which could potentially improve BT-MZ.
   Versions modified: all

 - Added the initial data touching for 'tv' and 'tu' in LU-MZ.
   Versions modified: all


Version: NPB3.0.1-MZ

[15-Aug-03]
 - Improved the load balancing scheme with threads.  The new version
   has more balanced work distribution among threads.
   This involves modification to the following files:
      NPB3.0-MZ-MPI/*/mpi_setup.f
      NPB3.0-MZ-SMP/*/smp_setup.f
      NPB3.0-MZ-MLP/*/mlp_setup.f

 - Include 'mzone_bload' as part of NPB3.0-MZ to subdirectory 'utils'.
   The load balancing scheme was included as option [4].

 - Organized the testing scripts in 'tests'.

[27-Aug-03]
 - Fixed a timing bug in NPB3.0-MZ-SMP and NPB3.0-MZ-MLP.
   A barrier was added right before the total timer starts.

 - Fixed an incorrect report of num_threads when OpenMP was not compiled.
   Versions affected: NPB3.0-MZ-MPI, NPB3.0-MZ-SMP, NPB3.0-MZ-MLP

 - Slightly speed up the comm_index calculation.  It won't change any
   benchmark timing results.

[03-Sep-03] NPB3.0.1-MZ release
 - Made the use of THREADPRIVATE in the MPI+OpenMP version as the default
   option.  Included a note in the corresponding README that indicates
   a workaround for running on SGI Origins.


Version: NPB3.0-MZ

[06-Aug-03] NPB3.0-MZ release
   The first official release of NPB3.0-MZ.  
   It includes the following packages:
      NPB3.0-MZ-MPI
      NPB3.0-MZ-SMP
      NPB3.0-MZ-SER
   Description of these packages is in README.