Fix Hwloc API Incompatibility

[jlpeyton]

Summary

Older Hwloc's (< 1.10.0) don't offer the HWLOC_OBJ_NUMANODE nor HWLOC_OBJ_PACKAGE types. Instead they are named HWLOC_OBJ_NODE and HWLOC_OBJ_SOCKET instead. This patch just defines the newer names when using an older Hwloc. Tested with 1.9.1, 1.10.1, and 1.11.6.

[AndreyChurbanov]

runtime/src/kmp.h

91

Both HWLOC 1.10 and 1.10.1 have version 0x00010a00; in 10.0 there is no HWLOC_OBJ_NUMANODE definition, but in 10.1 this definition is provided in deprecate.h header file (probably as a preparation for future renaming?).

Is it safer to use:

#ifndef HWLOC_OBJ_NUMANODE
#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
#endif
#ifndef HWLOC_OBJ_PACKAGE
#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
#endif

as opposed to version checking?

[jlpeyton]

Updated the diff. I had gotten errors when I tried that first, but now I realize that the SOCKET name is used in address2os map creation. I've changed all of them to HWLOC_OBJ_PACKAGE now and added the #ifndef method.

[AndreyChurbanov]

LGTM

runtime/src/kmp.h

 
 #include "kmp_safe_c_api.h"
 
 #if KMP_STATS_ENABLED
 class kmp_stats_list;
 #endif
 
 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
-# include "hwloc.h"
+#include "hwloc.h"
+#ifndef HWLOC_OBJ_NUMANODE
+#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
+#endif
+#ifndef HWLOC_OBJ_PACKAGE
+#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
+#endif
 #endif
 
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
 #include <xmmintrin.h>
 #endif
 
 #include "kmp_version.h"
 #include "kmp_debug.h"

runtime/src/kmp_affinity.cpp

for (level = depth-1; level >= 0; --level) {
             }
             *coreLevel = -1;
         }
     }
     return address2os[0].first.depth;
 }
 
 // Returns the number of objects of type 'type' below 'obj' within the topology tree structure.
-// e.g., if obj is a HWLOC_OBJ_SOCKET object, and type is HWLOC_OBJ_PU, then
+// e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is HWLOC_OBJ_PU, then
 //  this will return the number of PU's under the SOCKET object.
 static int
 __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj, hwloc_obj_type_t type) {
     int retval = 0;
     hwloc_obj_t first;
     for(first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type, obj->logical_index, type, 0);
         first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) == obj;
         first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type, first))
kmp_i18n_id_t *const msg_id)
     if (! KMP_AFFINITY_CAPABLE())
     {
         //
         // Hack to try and infer the machine topology using only the data
         // available from cpuid on the current thread, and __kmp_xproc.
         //
         KMP_ASSERT(__kmp_affinity_type == affinity_none);
 
-        nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(hwloc_get_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_SOCKET, 0), HWLOC_OBJ_CORE);
+        nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(hwloc_get_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_PACKAGE, 0), HWLOC_OBJ_CORE);
         __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(hwloc_get_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_CORE, 0), HWLOC_OBJ_PU);
         __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
         nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
         if (__kmp_affinity_verbose) {
             KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY");
             KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
             if (__kmp_affinity_uniform_topology()) {
                 KMP_INFORM(Uniform, "KMP_AFFINITY");
kmp_i18n_id_t *const msg_id)
 
     hwloc_obj_t pu;
     hwloc_obj_t core;
     hwloc_obj_t socket;
     int nActiveThreads = 0;
     int socket_identifier = 0;
     // re-calculate globals to count only accessible resources
     __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0;
-    for(socket = hwloc_get_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_SOCKET, 0);
+    for(socket = hwloc_get_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_PACKAGE, 0);
         socket != NULL;
-        socket = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_SOCKET, socket),
+        socket = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_PACKAGE, socket),
         socket_identifier++)
     {
         int core_identifier = 0;
         int num_active_cores = 0;
         for(core = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, socket->type, socket->logical_index, HWLOC_OBJ_CORE, 0);
             core != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, socket->type, core) == socket;
             core = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_CORE, core),
             core_identifier++)