Skip to content

Axom-v0.4.0

Compare
Choose a tag to compare
@gzagaris gzagaris released this 23 Sep 20:17
· 7013 commits to main since this release
38c0d74

[Version 0.4.0] - Release date 2020-09-23

Please download the Axom-v0.4.0.tar.gz tarball below, which includes all of the Axom submodules as well.

Added

  • Exposed the tolerance parameter EPS that is used to determine intersections between
    triangles in primal:intersect() as an optional final parameter.
  • Added BVH spatial index option to the mesh_tester utility for calculating
    triangle-triangle intersection.
  • Added axom::execution_space< ExecSpace >::onDevice() to check if execution
    space is on device.
  • Added Axom macro AXOM_SUPPRESS_HD_WARN to silence host device compiler
    warnings.
  • Added option to quest's SignedDistance class and C API to toggle whether
    the distance query computes the sign.
  • Added a batched option to quest's signed distance query example application.
    This computes all distance queries on an array of points using a single call to computeDistance.
    The query uses OpenMP threading, when available.
  • Added new component, Inlet, to assist in retrieving and storing data from
    an input deck.
  • Added the ability to specify an [Umpire] allocator ID to use with the
    BVH. This allows the application to use a device allocator for the BVH and
    avoid use of Unified Memory (UM) on the GPU, which can hinder perfomrmance,
    or use a pool allocator to mitigate the latencies associated with allocation/deallocation.
    The allocator ID is specified as an optional argument to the BVH constructor.
  • Added new CMake option, AXOM_ENABLE_ANNOTATIONS, to enable/disable code
    annotations in Axom. Default is OFF.
  • Added Axom annotation macros. The macros can be used to annotate functions,
    using the AXOM_PERF_MARK_FUNCTION macro, or at a more fine grain level,
    different sections of code can be annotated by wrapping them within an
    AXOM_PERF_MARK_SECTION block. As a first cut, this works with NVTX tools.
    However, the hooks are in place to add support for Caliper in the future.
  • Added a simple interface to NVTX that allows an application to set the color
    and category for NVTX ranges corresponding to annotated code in Axom. The
    application can now call axom::nvtx:set_color() and
    axom::nvtx::set_category() to set the corresponding parameters respectively.
    This facilitates in the performance evaluation by allowing developers to easily
    filter out calls by category or visually by setting a different color to use
    in GUI tools, such as, NVVP and NSight.
  • Added a portable floating_point_limits traits class, to return min(), max(), lowest()
    and epsilon() of a float or double type. The functionality is equivalent to that provided by
    std::numeric_limits, but, the code is host/device decorated accordingly such that it
    can also be called on the device.
  • Added initial support for ray queries using the BVH. The caller may now supply a set of rays to
    a BVH and the BVH will return a set of candidate BVH bins that intersect each ray.
  • Added initial support for bounding box queries using the BVH. The caller may
    now supply a set of bounding boxes to a BVH and the BVH will return a set of
    candidate BVH bins that intersect each bounding box.
  • Added an axom-config.cmake file to axom's installation to streamline incorporating axom
    into user applications. See <axom-install>/examples/axom for example usages.
  • Added [Sol] as a built-in TPL for fast and simple C++ and Lua binding.
    Sol is automatically enabled when LUA_DIR is found.
    The version of Sol used in this release is v2.20.6, which requires C++14.

Removed

  • Removed the AXOM_ENABLE_CUB option, since CUB is no lonher used directly in
    Axom code. Instead, we use RAJA::stable_sort with RAJA-v0.12.1 and fallback
    to std::stable_sort with older versions of RAJA and when the code is built
    without RAJA.

Changed

  • Updated Axom to support RAJA-v0.12.1 and Umpire-v4.01, but the code remains
    backwards compatible with previous versions of RAJA and Umpire.
  • Transitioned Axom's code formatting tool from Uncrustify to [clang-format].
    Axom's clang-format rules depend on clang 10.
  • Modified the command line interface for mesh_tester utility. Interface
    now uses a -m, --method option to select the spatial index, and -p, policy
    option now accepts a string or integer value.
  • Renamed the AXOM_USE_MPI3option to AXOM_ENABLE_MPI3 for consistency.
  • Modified the API for the BVH to accomodate different query types. The queries are now
    more explicitly called BVH::findPoints() and BVH::findRays().
  • Modified the API of Axom's memory management routines to not leak usage of Umpire. Instead of
    passing an umpire::Allocator object to specify an allocator, we now use the corresponding
    integer ID associated with the allocator.
  • All names in the C API now preserve the case of the C++ function.
    ex. SIDRE_datastore_new is now SIDRE_DataStore_new.
  • Fortran API in slic module. axom::slic::message Level enums are changed
    from enum-name_enumerator to namespace_enumerator.
    ex. level_error is now message_error.
  • Fortran derived-type constructors are now generic functions named afer the derived type.
    datastore_new is now SidreDataStore
    iomanager_new is now IOManager

Fixed

  • Fixed a bug in primal::intersect(Segment, BoundingBox) and added regression tests.
  • Spin's octrees can now be used with 64-bit indexes. This allows octrees
    with up to 64 levels of resolution when using a 64-bit index type.
  • Resolved issue with AXOM_USE_64BIT_INDEXTYPE configurations. Axom can once again
    be configured with 64-bit index types.
  • Fixed a triangle-triangle intersection case in primal that produced inconsistent results
    depending on the order of the triangle's vertices.
  • Fixed issue in the parallel construction of the BVH on GPUs, due to incoherent
    L1 cache that could result in some data corruption in the BVH. The code now
    calls __threadfence_system() after the parent is computed and stored back
    to global memory to ensure that the write is visible to all threads.
  • Fixed issue in Mint that would cause the [email protected] compiler to segfault. The
    mint_cell_types.cpp test was causing a segfault in the compiler. The main
    issue triggering this compiler bug was the use of constexpr when defining the
    static cell_info array of structs. The code has been modified to use const
    instead.
  • Fixed issue in Quest's Signed Distance query that would prevent consecutive
    calls to Quest when MPI-3 shared memory is enabled due to not properly
    nullifying internal pointers when finalize is called.
  • Fixed issue where the BVH would dispatch to the CPU sort() routine when the
    specified execution policy was CUDA_EXEC async. Now, when the execution policy
    is CUDA_EXEC the code would correctly dispatch to the GPU sort, using CUB
    (when CUB is enabled), regardless of whether it's synchronous or asynchronous.
  • Fixed issue with missing the bvh_traverse.hpp from the install prefix, which was preventing
    applications from using the BVH when pointing to an Axom install prefix.
  • Fixed usage of cuda kernel policies in Mint. Raja v0.11.0 changed the way max threads
    launch bounds is calculated. Consequently, a large number of threads was being launched
    leading to max registry count violation when linking. We are now using fixed kernel size
    of 256 threads (16x16 in 2D and 8x8x4 in 3D).
  • Third-party libraries can now build on the Windows platform through uberenv using vcpkg
    ("zero-to-axom support on Windows")

Known Bugs

  • Encountered a compiler bug on IBM LC platforms when using the IBM XL C/C++
    compiler. The issue is manifested in the generate_aabbs_and_centroids method
    in the spin_bvh.cpp unit test. It seems that the compiler does not handle
    the lambda capture of the arrays correctly which leads to a segfault. A
    workaround for the IBM XL compiler is provided.
  • There is a known bug in MVAPICH that prevents consecutive creation/deletion
    of MPI windows. This was encountered on LC platforms when enabling shared
    memory in the Signed Distance Query. See the corresponding
    Github Issue for details.