Tags give the ability to mark specific points in history as being important
  • europar2020
  • v6.1.0
    Release v6.1.0
    • Add a new dynamic scheduler supported by the internal threads for numerical factorization
    • Add MPI support for the numerical factorization and solve
    • Available for all schedulers: sequential, internal threads (static, dynamic), StarPU, and PaRSEC
    • WARNING: The RHS is not distributed yet, and must be replicated on all nodes
    • WARNING: The low-rank and Schur functionalities are not available in distributed yet
    • Enable the use of an external SPM module
    • Improve splitting strategy:
    • avoid unnecessary splits when using K-Way
    • reduce the range of possible split to limit the apparition of small blocks
    • Change the preselected behavior to be:
    • never compressed in the JustInTime scenario
    • compress the preselected block just before applying the TRSM in the MinimalMemory scenario
    • the behavior can be change through IPARM_COMPRESS_PRESELECT
    • Add a cmake summary
    • Add coverity scan
    • Update README.md and documentation
  • v6.0.3
    50c92728 · Add missing line ·
    Release v6.0.3
    • Update spm module to ada4963
    • Update morse_cmake to ade4996
    • CMake: Update cmake_module to integrate the last version of the precision generator
    • Change StarPU requirement to >= 1.3
    • Refactor and extend the CI/CTests
    • Update documentation
    • Low-rank:
    • Add a new parameter IPARM_COMPRESS_RELTOL to switch between absolute and relative tolerance
    • Improved stability of low-rank kernels
    • Extend the number of tests
    • Add rotation QR kernels (unstable/work in progress)
    • Enable multiple low-rank factorization in a row
    • Supports compilation with mpicc (no distributed solver yet)
    • Add separated output directories for future distributed process, or for MPI multiple instances
    • Octave: Fix issue with number of threads larger than the number of columns
    • Octave: Fix compilation on Windows system
    • Documentation: add documentation on process binding
    • HwLoc: fix binding when already restricted through batch scheduler and/or MPI
    • Fix issue #35, make pastix_task_analyze thread safe
    • Add support for multi-dof in Fortran
    • Fix issue in simulation, and a switch between cost and tree levels
    • Refinement: Fix issue with gemv computation and PastixConjTrans
    • CMake: Enable a round-robin selection of CMAKE_BUILD_TYPE depending on the sanitizers provided by the compiler
    • Homebrew: update formula
  • v6.0.2
    4321e9a4 · Fix URL in curl request ·
    Release v6.0.2
    • Integrate the clusting strategies developped for low-rank (See https://hal.inria.fr/hal-01961675)
    • Restructure the ordering/symbolic factorization code to make sure with exit the ordering step with permutation, partition, and elimination tree.
    • Relook the splitting/proportional mapping strategy
    • Add new compression kernels: PQRCP, RQRCP, and TQRCP
    • Fix inplace compilation (Issue #36)
    • Fix issue when StarPU threads where fighting for ressources with PaStiX threads (Add pause/resume calls)
    • Handle multi-dof static and/or variable in the analysis steps
  • clustering
    4321e9a4 · Fix URL in curl request ·
  • 6.0.1
    Release 6.0.1
    • Final 6.0.1 release of the PaStiX solver

    • This release includes support for:

      • Support for HWLOC 2.0.0
      • Move the SPM library as a submodule
      • Parallel (multithreaded) version of the reordering step
      • Parallel (multithreaded) version of the refinement steps
      • StarPU version of the solve step
      • Fix Python/Fortran interface
      • Fix Schur functions
      • Fix multi-RHS solve
      • Fix for METIS
      • Update morse_cmake FindPACKAGE
      • Update PaRSEC for release 6.0.1
      • Improve PKG-CONFIG
      • Improve documentation
      • Better handle of disconnected graphs
      • Add optimal reordering for grids
      • Add more detailed statistics during analysis step
      • Add detailed statistics about memory gain for the low-rank solver
      • Add a function to compress the solver matrix outside the factorization step
      • Add an example to dump the symbol matrix including the ranks of the block
      • Add an refinement driver and testings
      • Add a subtask_refine which does not perform the vector ordering
      • Add a more complex testing based on example proposed by @andrea3.14
      • Add an iparm IPARM_APPLYPERM_WS to enable/disable the use of an extra workspace to make the functions bvec_xlapmr thread-safe (by default, it is enabled, if disabled, the functions have no memory overhead but loose the thread-safe property)
      • Remove the sparse-kit package to avoid conflict (the driver is replaced by HB)

    pastix-6.0.1.tar.gz (SHA256: 276fff4e95ef0bad1402a7ad40d558722c509e3cc0193fbe509887c639227a30)

  • 6.0.0
    Release 6.0.0
    • Final 6.0.0 release of the PaStiX solver

    • This release includes support for:

      • low-rank compression
      • static scheduler, PaRSEC and StarPU runtime support
      • GPUs (Kepler) and KNL support through runtime systems

    pastix-6.0.0.tar.gz (SHA256: 1dc37475964d7baa0df30c5909bfdf459dd0db66b866f9d9f566ab6fd7e12d7f)

  • 6.0.0alpha   This is the first alpha release of the 6.0.0. This includes a complete rewrite of the PaStiX code, inclusion of block low-rank solver, and integration of the PaRSEC runtime support.
    783dd718 · Update README.md ·
  • papers/pdsec17   Tag release of the version used for the PDSEC'17 paper.