ReleaseNotes.txt

EigenExa Release Notes

------------------------------------------------------------------------
2.13a : July 24, 2024)
  * Experimental C APIs have been supported in C.
    But, not yet merged into libEigenExa.(a|so).

2.13 : July 23, 2024)
  * Repository release on the R-CCS github started
    https://github.com/RIKEN-RCCS/EigenExa
  * Fix some internal errors occurring in a >32bit integer expression.
    Still, unexpected overflow may happen in huge calculation.
  * Fix some binding mismatches or instantiation of the modules of the
    compiler versions in C, C++, F9X, and F2XXX.
  * Fix Intel oenAPI compiler name from mpiicc to mpiicx, even though 
    they are not called in this version...
  * Experimental implementation of a minimal set of caller APIs from
    C/C++ language environment (confirmed gcc, icc, and DPC++/c++).

2.12 : Octorber 25, 2022)
  * [Serious] Bug fix for a very rare condensed clustering case of
    eigenmodes. Modified the deflation couting correctly.
  * Modify the benchmark options to check the specific kernel.
  * Modify the collective communication algorithms.

2.11 : December 1, 2021)
  * [Upgrade] Experimental support for Hermite solver, namely, eigen_h.
  * [Serious] Fix the numerical error happened to be included in v2.10,
    where the tall-skinny QR coded in eigen_prd_t4x.F of eigen_sx was
    too sensitive to treat tiny values and forced-double truncation.
    But, it might depend on the compiler version and the code generator.
  * Modify pointer attribution to 'allocatable' to avoid automatic
    deallocation on the exit of callee routines.
  * Some code modifications are applied to pass strong debugging tools
    with respect to Fortran 95 and some of Fortran 2003 extensions.
  * Fix some bugs, for exmaple, missing private attribution to some
    variables for OpenMP.

2.10 : October 17, 2021)
  * [Serious] Bug fix for violation of the result of allreduce in DC.
    It happened very rarely when data to be transferred was shorter
    than the number of processes participating.
  * [Serious] Bug fix for inconsistent API interpretation of DLAED4,
    when K is less than or equal to 2. This bug happened when a lot
    of deflations are carried out, and sub-matrices are shrunk tiny as 1
    or 2. So, it is infrequent to see.
  * [Serious] Bug fix for non-deterministic behavior of the DC branch,
    which happened if an uninitialized variable referred in the brach
    condition, and is affected by the side-effects of other modules,
    etc. It was fixed when 2.9 was released but noted in the release.
  * Reduce the internal data capacity in the TRD and DC routines.
  * Fix the installation of Fortran modules.

2.9  : September 24, 2021)
  * Modify the flops count precise in DC kernels.
  * Modify trbak not to multiply D^{-1} and TRSM.
  * Add enable/disable-switch for building the shared library.
  * Modify to detect the memory allocation fault.

2.8  : August 20, 2021)
  * Modify the DC kernel to reduce intermediate buffer storage.
  * Bug fix on a t1 loop structure
  * Updated the error check routine
  * Fixed on Makefile to add the missing fortran module.

2.7 : April 1, 2021)
  * Modify the compilation rules corresponding to static/shared library
    defined in src/Makefile.am
  * Performance tweak with a modification of the compilation options
    not to use -fPIC when build a static library.
  * License document is packed as an independent file (the license
    notice was stated in User's manual for version 2.6).

2.6 : November 1, 2020
  * Reduced the generation of MPI communicator in the divide and conquer
    algorithm routine.
  * Changed calling MPI_Allreduce from the subroutine of MPI library to
    an in-house/hand-written subroutine equaling the MPI functions due to
    the violation of numerical reproducibility in some MPI implementations.
  * Fix module dependencies.

2.5 : August 1, 2019
  * Refine the data distribution in the divide and conquer algorithm routine.

2.4b : August 20, 2018
  * [Serious] Bug fix for incorrect data redistribution, which might
    violate allocated memory.
    The bug might have happened in the case that the number of processes,
    P=Px*Py, is large, and Px and Py are not equal but nearly equal.
  * This version is for only bug fix for the serious one.

2.4p1 : May 25, 2017
  * [Serious] Bug fix for incorrect data redistribution in eigen_s.
  * Major change with Autoconf -and- Automake framework

2.4 : April 18, 2017
  * Major change with Autoconf -and- Automake framework

2.3k2 : April 12, 2017
  * Communication Avoiding algorithms to the eigen_s driver.
  * The optional argument nvec is available, which specifies the number of
    eigenvectors to be computed from the smallest. This version does not
    employ the special algorithm to reduce the computational cost. It only
    drops off the unneccessary eigenmodes in the backtransformation.

2.3d : July 07, 2015
  * Tuned up the parameters according to target architectures.
  * Introduce a sort routine in bisect.F and bisect2.F for eigenvalues.
  * Modify the algorithm to create reflector vectors in eigen_prd_t4x.F
  * Modify the matrix setting routine to load the mtx (Matrix Market)
    format file via both 'A.mtx' and 'B.mtx'.
  * Re-format the source code by the fortran-mode of emacs and extra rules.

2.3c : April 23, 2015
  * Fix bug on flops count of eigen_s which returned incorrect value
    due to missing initialization in dc2.F.
    This bug is found in version 2.3a and version 2.3b.
  * Minor change on timer routines.
  * Minor change on broadcast algorithm in comm.F.

2.3b : April 15, 2015
  * Minor change to manage the real constants.
  * Minor change to use Level 1 and 2 BLAS routines.
  * Minor change to preserve invalid or oversized matrices.
  * Minor change of Makefile to allow '-j' option.

2.3a : April 14, 2015
  * Minor change on thread parallelization of eigne_s.
  * Minor change of the API's for timer routines.
  * Fix the unexpected optimization of rounding errors in eigen_dcx.

2.3 : April 12, 2015
  * Bug fix on the benchmark program.
  * Refine the race condition in the backtransformation routine.
  * Introduce Communication Avoiding algorithms to the eigen_s driver.

2.2d : March 20, 2015
  * Bug fix on the timer print part in trbakwy4.F not to do zero division.
  * Modify the synchronization point in eigen_s.
  * Modify thread parallelization in eigen_dc2 and eigen_dcx.

2.2c : March 10, 2015
  * Bug fix on the benchmark program.
  * Add the make_inc file for an NEC SX platform.

2.2b : October 30, 2014
  * Introduce new API to query the current version.
  * Introduce the constant eigen_NB=64, which refers to the block size
    for cooperative work with the ScaLAPACK routines.
  * Correct the requred array size in eigen_mat_dims().
  * Improve the performance of test matrix generator routine mat_set().
  * Add the listing option of test matrices in eigenexa_benchmark.

2.2a : June 20, 2014
  * Fix minor bug of Makefile, miscC.c and etc for BG/Q.
  * Modify the initialization process not to use invalid communicators.
  * Comment out the calling BLACS_EXIT in eigen_free().

2.2 : April 10, 2014
  * Arrange the structure of source directory.
  * Reversion of the DC routines back to version 1.3a to avoid bug.
  * Hack miscC.c to be called from IBM BG/Q.
  * Fix bug on the benchmark program for exceptional case of MPI_COMM_NULL.
  * Fix bug on eigen_s with splitted communicator.
  * Update machine depended configuration files.
  * Experimental support of building a shared library

2.1a : Feb 23, 2014
  * Fix bug on the benchmark program.

2.1 : Feb 10, 2014
  * Fix bug on eigen_sx: it gave wrong results when N=3.
  * Modify the bisect2 by a pivoting algorithm.
  * Update the test program 'eigenexa_benchmark' in order to check
    accuracy with several test matrices and computing modes.
  * Tune performance for K computer and Fujitsu FX10 platforms.
  * Add make_inc file for a BlueGeneQ platform, but it is not official
    support, just an experimental.

2.0 : Dec 13, 2013
  * Add eigen_s, which adopts the conventional 1-stage algorithm.
  * Add optional modes to compute only eigenvalues and to improve
    the accuracy of eigenvalues.
  * Modify to support a thread mode with any number of threads.
  * Tune performance for K computer and Fujitsu FX10 platforms.

1.3a : Sep 21, 2013
  * Fix bug on syncronization mechanism of eigen_trbakwyx().

1.3 : Sep 20, 2013
  * Fix bug on eigen_init() in initialization with MPI_Cart's or
    MPI_COMM_NULL's.
  * Add test programs to check several process patterns.

1.2 : Sep 17, 2013
  * Fix bug on benchmark code in making a random seed.
  * Modify to support upto 64-thread running.

1.1 : Aug 30, 2013
  * Fix bug on data-redistribution row vector to column vector
    when P=p*q and p and q have common divisor except themselves.
  * Optimize data redistribution algorithm in dc_redist[12].F

1.0 : Aug  1, 2013
  * This is the first release
  * Standard eigenvalue problem for a dense symmetric matrix by
    a novel one-stage algorithm