Hello, We currently have a single object type for caches (HWLOC_OBJ_CACHE) but all machines have multiple levels of caches (usually, L1i, L1d, L2 and L3). Functions like hwloc_get_obj_by_type() don't work for caches because of these multiple levels. We got multiple complaints about this because this requires special cases like: https://github.com/open-mpi/ompi/blob/master/orte/mca/rmaps/base/rmaps_base_binding.c#L910
The new "cachetypes" branch splits the single HWLOC_OBJ_CACHE into multiple types, one per level: * HWLOC_OBJ_L1CACHE, ..., HWLOC_OBJ_L5CACHE (those are either Data or Unified caches, and we have never seen anything above L4). * HWLOC_OBJ_L1ICACHE, ..., HWLOC_OBJ_L3ICACHE (modern processors have a L1i, and I have never seen anything above L2i) Random notes: * Your switch cases on object types will need 8 lines to match all cache types: https://github.com/open-mpi/hwloc/blob/cachetypes/hwloc/topology-xml.c#L173 Ranges in switch cases are not standard C yet unfortunately. * There is a hwloc_obj_type_is_cache(type) for testing if a given type is a cache (and variants for data and instruction caches). https://github.com/open-mpi/hwloc/blob/cachetypes/include/hwloc/helper.h#L453 * If you want to ignore caches entirely in your topology, you need to call the ignore function 8 times :/ I am planning to rework the ignoring API in v2.0, we'll see if we can make that easier. * Contrary to the renaming of SOCKET into PACKAGE or NODE into NUMANODE, we can't #define the old HWLOC_OBJ_CACHE to a new type to ease forward-compatibility. I think it's better to force users to update their code anyway. In the vast majority of cases, the new code will be shorter thanks to no cache-specific hacks anymore. * There will be XML backward-compatiblity: old XML will be converted to new cache types. * You may still have multiple levels of the same type in case of "Groups of Groups" (usually on large SGI UV machines). That's why HWLOC_TYPE_DEPTH_MULTIPLE will still exist unfortunately. Fortunately, you probably never lookup "Group" objects "by_type". If you have comments, let me know. Brice
