v2.2.2: Merge pull request #20 from eth-cscs/cmake_fix

• remove the -mtune=native option from the CXXFLAGS

COSTA-v2.2.1

• fix a cmake bug when linking to cray libsci.

v2.2

• Update of cmake build system
• All optional sub-modules are treated as dependency

COSTA-v2.1

This version brings the following improvements/features:

• scalapack-wrappers: COSTA now implements all scalapack pxgemr2d, pxtran (transpose), pxtranu (complex transpose) and pxtranc (conjugate, complex transpose) routines, i.e.
  • pdgemr2d, psgemr2d, pcgemr2d, pzgemr2d
  • pstran, pdtran, pctranu, pztranu, pctranc, pztranc
    and their prefixed versions:
  • costa_pdgemr2d, costa_psgemr2d, costa_pcgemr2d, costa_pzgemr2d
  • costa_pstran, costa_pdtran, costa_pctranu, costa_pztranu, costa_pctranc, costa_pztranc
• code refactoring
• performance improvements

COSTA-v2.0

This release brings the following improvements:

• [row-major blocks]: COSTA now supports both row-major and col-major ordering of blocks. This is more general than scalapack which supports only the col-major ordering.
• [performance-improvements]: the multithreaded backends are improved to avoid cores oversubscription, resulting in a performance boost.
• [bugfixes]: a thourough testing has been performed during the integration of COSTA into COSMA and CP2K.

COSTA v1.0

This is the very first release of COSTA, bringing the following features:

• scalapack wrappers: for redistribute (pxgemr2d) and transpose (pxtran(u)).
• different layouts support: added representation for block-cyclic and arbitrary matrix layouts.
• multiple layouts: can transform multiple layouts at once, i.e. in the same communication round.
• comm-optimal: can minimize the communication volume.
• scaling & transpose: in addition to redistributing the matrix, can also scale initial and final layouts and also transpose them.
• highly optimized: optimized for distributed and multithreaded settings.