baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.10.0:
  Heterogeneous Memory core improvements
  + Better heuristics to identify the subtype of memory such
    as HBM, DRAM, NVM, CXL-DRAM, etc.
  + Build memory tiers, i.e. sets of NUMA nodes with the same
    subtype and similar performance.
    - NUMA node tier ranks are exposed in the new MemoryTier
    info attribute (starts from 0 for highest bandwidth tier).
  + Add hwloc_topology_free_group_object() to discard a Group
    created by hwloc_topology_alloc_group_object().
  + Fix cpukinds on NVIDIA Grace to report identical cores
    even if they actually have very small frequency differences.
  + Add CXLDevice attributes to CXL DAX objects and NUMA
    nodes to show which PCI device implements which window.
  + Ignore buggy memory-side caches and memory attributes
    when fake NUMA emulation is enabled on the Linux kernel
    command-line.
  + Add more info attributes in MemoryModule Misc objects,
  + Get CPUModel and CPUFamily info attributes on LoongArch
    platforms.
  + Add support for new AMD CPUID leaf 0x80000026 for better
    detection of Core Complex and Die on Zen4 processors.
  + Improve Zhaoxin CPU topology detection.
  + Input locations and many command-line options (e.g.
    hwloc-calc -I -N -H, lstopo --only) now accept filters such as
    "NUMA[HBM]" so that only objects are that type and subtype are
    considered.
  - NUMA[tier=1] is also accepted for selecting NUMA nodes
    depending on their MemoryTier info attribute.
  + Add --object-output to hwloc-calc to report the type as a

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.9.2:
  * Don't forget L3i when defining filters for multiple levels of
    caches with hwloc_topology_set_cache/icache_types_filter().
  * Fix object total_memory after hwloc_topology_insert_group_object().
  * Fix the (non-yet) exporting in synthetic description for
    complex memory hierarchies with memory-side caches, etc.
  * Fix some default size attributes when building synthetic
    topologies.
  * Fix size units in hwloc-annotate.
  * Improve bitmap reallocation error management in many functions.
  * Documentation improvements

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.9.1:
  * Fix a failed assertion in hwloc_topology_restrict() when some
    NUMA nodes are removed because of
    HWLOC_RESTRICT_FLAG_REMOVE_CPULESS but no PUs are.
  * Mark HPE Cray Slingshot NICs with subtype "Slingshot".

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.9.0:
  + Expose the memory size of CXL memory devices (Type 3) on Linux.
  + The LevelZero backend now reports the "XeLinkBandwidth" distance
    matrix between L0 devices (and subdevices) when available.
  + Add support for CUDA compute capability up to 9.0.
  + lstopo now switches to console mode when its output is redirected.
    Graphical window mode may be forced back with --of window.
  + hwloc-calc now accepts "numa" in -H, and I/O subtypes such as "gpu"
    in -I and -N.

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.8.0:
  * API
    + Add HWLOC_TOPOLOGY_FLAG_NO_DISTANCES, _NO_MEMATTRS and _NO_CPUKINDS
      to reduce the overhead when unneeded.
    + Add separate Read/Write Bandwidth/Latency memory attributes and
      implement them on Linux.
  * Backends
    + NUMA nodes may now have a subtype such as DRAM, HBM, SPM, or NVM
      on heterogeneous memory platforms on Linux.
      - Add DAXType and DAXParent attributes on Linux to tell where a
	DAX device or its corresponding NUMA node come from (SPM for
	Specific-Purpose or NVM for Non-Volatile Memory).
    + Detect heterogeneous caches in hybrid CPUs on MacOS X,
      thanks to Paul Bone for the help.
    + Max frequencies are not ignored in Linux cpukinds anymore (they were
      ignored in hwloc 2.7.0), but they may be slightly adjusted to avoid
      reporting hybrid CPUs because Intel Turbo Boost Max 3.0.
      - See the documentation of environment variable HWLOC_CPUKINDS_MAXFREQ.
    + Hardwire the PCI locality of HPE Cray EX235a nodes.
  * Tools
    + lstopo and other tools may now load Linux and x86 cpuid topology files
      from a tarball.
    + lstopo may now replace the P# and L# index prefixes with custom strings
      thanks to --os-index-prefix and --logical-index-prefix options.
  * Misc
    + Add --disable-readme to avoid regenerating the top-level hwloc README
      file from the documentation.

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.7.1:
  * Workaround crashes when virtual machines report incoherent x86 CPUID
    information about numbers of cores and threads.
    Thanks to Peter Bense for the report.
  * Use setenv() instead of putenv() when trying to force enable oneAPI L0
    support, to avoid issues with applications that touch the environment,
    thanks to Josh Hursey for the patch.
  * Add some warnings at the end of configure when GPU libraries are
    missing on the system or their path is missing in the environment.
  * Backends
    + Add support for NUMA nodes and caches with more than 64 PUs across
      multiple processor groups on Windows 11 and Windows Server 2022.
    + Group objects are not created for Windows processor groups anymore,
      except if HWLOC_WINDOWS_PROCESSOR_GROUP_OBJS=1 in the environment.
    + Expose "Cluster" group objects on Linux kernel 5.16+ for CPUs
      that share some internal cache or bus. This can be equivalent
      to the L2 Cache level on some platforms (e.g. x86) or a specific
      level between L2 and L3 on others (e.g. ARM Kungpeng 920).
      Thanks to Jonathan Cameron for the help.
      - HWLOC_DONT_MERGE_CLUSTER_GROUPS=1 may be set in the environment
	to prevent these groups from being merged with identical caches, etc.
    + Improve the oneAPI LevelZero backend:
      - Expose subdevices such as "ze0.1" inside root OS devices ("ze0")
	when the hardware contains multiple subdevices.
      - Add many new attributes to describe device type, and the
	numbers of slices, subslices, execution units and threads.
      - Expose the memory information as LevelZeroHBM/DDR/MemorySize infos.
    + Ignore the max frequencies of cores in Linux cpukinds when the
      base frequencies are available (to avoid exposing hybrid CPUs
      when Intel Turbo Boost Max 3.0 gives slightly different max

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.6.0:
  * Backends
    + Expose two cpukinds for energy-efficient cores (icestorm) and
      high-performance cores (firestorm) on Apple M1 on Mac OS X.
    + Use sysfs CPU "capacity" to rank hybrid cores by efficiency
      on Linux when available (mostly on recent ARM platforms for now).
    + Improve HWLOC_MEMBIND_BIND (without the STRICT flag) on Linux kernel
      >= 5.15: If more than one node is given, the kernel may now use all
      of them instead of only the first one before falling back to others.
    + Expose cache os_index when available on Linux, it may be needed
      when using resctrl to configure cache partitioning, memory bandwidth
      monitoring, etc.
    + Add a "XGMIHops" distances matrix in the RSMI backend for AMD GPU
      interconnected through XGMI links.
    + Expose AMD GPU memory information (VRAM and GTT) in the RSMI backend.
    + Add OS devices such as "bxi0" for Atos/Bull BXI HCAs on Linux.
  * Tools
    + lstopo has a better placement algorithm with respect to I/O
      objects, see --children-order in the manpage for details.
    + hwloc-annotate may now change object subtypes and cache or memory
      sizes.
  * Build
    + Allow to specify the ROCm installation for building the RSMI backend:
      - Use a custom installation path if specified with --with-rocm=<dir>.
      - Use /opt/rocm-<version> if specified with --with-rocm-version=<version>
	or the ROCM_VERSION environment variable.
      - Try /opt/rocm if it exists.
      - See "How do I enable ROCm SMI and select which version to use?"
	in the FAQ for details.
    + Add a CMakeLists for Windows under contrib/windows-cmake/ .

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.5.0: 
  + Add hwloc/windows.h to query Windows processor groups.
  + Add hwloc_get_obj_with_same_locality() to convert between objects
    with same locality, for instance NUMA nodes and Packages,
    or OS devices within a PCI device.
  + Add hwloc_distances_transform() to modify distances structures.
    - hwloc-annotate and lstopo have new distances-transform options.
  + hwloc_distances_add() is replaced with _add_create() followed by
    _add_values() and _add_commit(). See hwloc/distances.h for details.
  + Add topology flags to mitigate binding modifications during
    hwloc discovery, especially on Windows:
    - HWLOC_TOPOLOGY_FLAG_RESTRICT_TO_CPUBINDING and _MEMBINDING
      restrict discovery to PUs and NUMA nodes inside the binding.
    - HWLOC_TOPOLOGY_FLAG_DONT_CHANGE_BINDING prevents from ever
      changing the binding during discovery.
  + Add a levelzero backend for oneAPI L0 devices, exposed as OS devices
    of subtype "LevelZero" and name such as "ze0".
    - Add hwloc/levelzero.h for interoperability between converting
      between L0 API devices and hwloc cpusets or OS devices.
  + Expose NEC Vector Engine cards on Linux as OS devices of subtype
    "VectorEngine" and name "ve0", etc.
    Thanks to Anara Kozhokanova, Tim Cramer and Erich Focht for the help.
  + Add a NVLinkBandwidth distances structure between NVIDIA GPUs
    (and POWER processor or NVSwitches) in the NVML backend,
    and a XGMIBandwidth distances structure between AMD GPUs
    in the RSMI backends.
    - See "Topology Attributes: Distances, Memory Attributes and CPU Kinds"
      in the documentation for details about these new distances.
  + Add support for NUMA node 0 being offline in Linux, thanks to Jirka Hladky.
  + Add --with-cuda-version=<version> or look at the CUDA_VERSION

• Files Changed
• Browse Source

baserev update by copy to link target

• Files Changed
• Browse Source

- update to 2.4.1:
  * Fix AMD OpenCL device locality when PCI bus or device number >= 128.
    Thanks to Edgar Leon for reporting the issue.
    + Applications using any of the following inline functions must
      be recompiled to get the fix: hwloc_opencl_get_device_pci_busid()
      hwloc_opencl_get_device_cpuset(), hwloc_opencl_get_device_osdev().
  * Fix the ranking of cpukinds on non-Windows systems,
    thanks to Ivan Kochin for the report.
  * Fix the insertion of custom Groups after loading the topology,
    thanks to Scott Hicks.
  * Add support for CPU0 being offline in Linux, thanks to Garrett Clay.
  * Fix missing x86 Package and Core objects FreeBSD/NetBSD.
    Thanks to Thibault Payet and Yuri Victorovich for the report.
  * Fix the import of very large distances with heterogeneous object types.
  * Fix a memory leak in the Linux backend,
    thanks to Perceval Anichini.

• Files Changed
• Browse Source

• Browse Source