Extension of HWLOC topology discovery with NUMA nodes and tiles

[omalyshe]

Summary

This patch implements granularity=tile and OMP_PLACES=tiles.

[omalyshe]

Fix for non-hwloc build

[AndreyChurbanov]

LGTM

runtime/src/i18n/en_US.txt

 # --------------------------------------------------------------------------------------------------
 
 # Meta information about message catalog.
 
 Language "English"
 Country  "USA"
 LangId   "1033"
 Version  "2"
-Revision "20170327"
+Revision "20170518"
 
 
 
 # --------------------------------------------------------------------------------------------------
 -*- STRINGS -*-
 # --------------------------------------------------------------------------------------------------
 
 # Strings are not complete messages, just fragments. We need to work on it and reduce number of

 ThreadIDsNotUnique           "thread ids not unique"
 UsingPthread                 "using pthread info"
 LegacyApicIDsNotUnique       "legacy APIC ids not unique"
 x2ApicIDsNotUnique           "x2APIC ids not unique"
 DisplayEnvBegin        "OPENMP DISPLAY ENVIRONMENT BEGIN"
 DisplayEnvEnd        "OPENMP DISPLAY ENVIRONMENT END"
 Device           "[device]"
 Host           "[host]"
+Tile                         "tile"
 
 
 
 # --------------------------------------------------------------------------------------------------
 -*- FORMATS -*-
 # --------------------------------------------------------------------------------------------------
 
 Info                         "OMP: Info #%1$d: %2$s\n"

 BarriersInDifferentOrder     "Threads encountered barriers in different order. "
 FunctionError                "Function %1$s failed:"
 TopologyExtra                "%1$s: %2$s packages x %3$d cores/pkg x %4$d threads/core (%5$d total cores)"
 WrongMessageCatalog          "Incompatible message catalog \"%1$s\": Version \"%2$s\" found, version \"%3$s\" expected."
 StgIgnored                   "%1$s: ignored because %2$s has been defined"
                                  # %1, -- name of ignored variable, %2 -- name of variable with higher priority.
 OBSOLETE                     "%1$s: overrides %3$s specified before"
                                  # %1, %2 -- name and value of the overriding variable, %3 -- name of overriden variable.
+TopologyExtraTile            "%1$s: %2$d packages x %3$d tiles/pkg x %4$d cores/tile x %5$d threads/core (%6$d total cores)"
+TopologyExtraNode            "%1$s: %2$d packages x %3$d nodes/pkg x %4$d cores/node x %5$d threads/core (%6$d total cores)"
+TopologyExtraNoTi            "%1$s: %2$d packages x %3$d nodes/pkg x %4$d tiles/node x %5$d cores/tile x %6$d threads/core (%7$d total cores)"
 OmptOutdatedWorkshare        "OMPT: Cannot determine workshare type; using the default (loop) instead. "
                              "This issue is fixed in an up-to-date compiler."
 
 # --- OpenMP errors detected at runtime ---
 #
 #    %1 is the name of OpenMP construct (formatted with "Pragma" format).
 #
 CnsBoundToWorksharing        "%1$s must be bound to a work-sharing or work-queuing construct with an \"ordered\" clause"

runtime/src/kmp.h

 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
                                              GROUP_AFFINITY *);
 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
 #endif /* KMP_OS_WINDOWS */
 
 #if KMP_USE_HWLOC
 extern hwloc_topology_t __kmp_hwloc_topology;
 extern int __kmp_hwloc_error;
+extern int __kmp_numa_detected;
+extern int __kmp_tile_depth;
 #endif
 
 extern size_t __kmp_affin_mask_size;
 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
   for (i = (mask)->begin(); i != (mask)->end(); i = (mask)->next(i))
enum affinity_type {
   affinity_disabled, // not used outsize the env var parser
   affinity_default
 };
 
 enum affinity_gran {
   affinity_gran_fine = 0,
   affinity_gran_thread,
   affinity_gran_core,
+  affinity_gran_tile,
+  affinity_gran_numa,
   affinity_gran_package,
   affinity_gran_node,
 #if KMP_GROUP_AFFINITY
   // The "group" granularity isn't necesssarily coarser than all of the
   // other levels, but we put it last in the enum.
   affinity_gran_group,
 #endif /* KMP_GROUP_AFFINITY */
   affinity_gran_default

runtime/src/kmp_affinity.cpp

for (proc = 0; proc < len; proc++) {
     KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
                buf.str);
     __kmp_str_buf_free(&buf);
   }
 }
 
 #if KMP_USE_HWLOC
 
+static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len,
+                                          int depth, int *levels) {
+  int proc;
+  kmp_str_buf_t buf;
+  __kmp_str_buf_init(&buf);
+  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
+  for (proc = 0; proc < len; proc++) {
+    __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package),
+                        addrP[proc].first.labels[0]);
+    if (depth > 1) {
+      int level = 1; // iterate over levels
+      int label = 1; // iterate over labels
+      if (__kmp_numa_detected)
+        // node level follows package
+        if (levels[level++] > 0)
+          __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node),
+                              addrP[proc].first.labels[label++]);
+      if (__kmp_tile_depth > 0)
+        // tile level follows node if any, or package
+        if (levels[level++] > 0)
+          __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile),
+                              addrP[proc].first.labels[label++]);
+      if (levels[level++] > 0)
+        // core level follows
+        __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core),
+                            addrP[proc].first.labels[label++]);
+      if (levels[level++] > 0)
+        // thread level is the latest
+        __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread),
+                            addrP[proc].first.labels[label++]);
+      KMP_DEBUG_ASSERT(label == depth);
+    }
+    KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
+    __kmp_str_buf_clear(&buf);
+  }
+  __kmp_str_buf_free(&buf);
+}
+
+static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile;
+
 // This function removes the topology levels that are radix 1 and don't offer
 // further information about the topology.  The most common example is when you
 // have one thread context per core, we don't want the extra thread context
 // level if it offers no unique labels.  So they are removed.
 // return value: the new depth of address2os
-static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *address2os,
-                                                  int nActiveThreads, int depth,
-                                                  int *pkgLevel, int *coreLevel,
-                                                  int *threadLevel) {
+static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh,
+                                                  int depth, int *levels) {
   int level;
   int i;
   int radix1_detected;
-
-  for (level = depth - 1; level >= 0; --level) {
-    // Always keep the package level
-    if (level == *pkgLevel)
-      continue;
+  int new_depth = depth;
+  for (level = depth - 1; level > 0; --level) {
     // Detect if this level is radix 1
     radix1_detected = 1;
-    for (i = 1; i < nActiveThreads; ++i) {
-      if (address2os[0].first.labels[level] !=
-          address2os[i].first.labels[level]) {
+    for (i = 1; i < nTh; ++i) {
+      if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) {
         // There are differing label values for this level so it stays
         radix1_detected = 0;
         break;
       }
     }
     if (!radix1_detected)
       continue;
     // Radix 1 was detected
-    if (level == *threadLevel) {
-      // If only one thread per core, then just decrement
-      // the depth which removes the threadlevel from address2os
-      for (i = 0; i < nActiveThreads; ++i) {
-        address2os[i].first.depth--;
-      }
-      *threadLevel = -1;
-    } else if (level == *coreLevel) {
-      // For core level, we move the thread labels over if they are still
-      // valid (*threadLevel != -1), and also reduce the depth another level
-      for (i = 0; i < nActiveThreads; ++i) {
-        if (*threadLevel != -1) {
-          address2os[i].first.labels[*coreLevel] =
-              address2os[i].first.labels[*threadLevel];
+    --new_depth;
+    levels[level] = -1; // mark level as not present in address2os array
+    if (level == new_depth) {
+      // "turn off" deepest level, just decrement the depth that removes
+      // the level from address2os array
+      for (i = 0; i < nTh; ++i) {
+        addrP[i].first.depth--;
+      }
+    } else {
+      // For other levels, we move labels over and also reduce the depth
+      int j;
+      for (j = level; j < new_depth; ++j) {
+        for (i = 0; i < nTh; ++i) {
+          addrP[i].first.labels[j] = addrP[i].first.labels[j + 1];
+          addrP[i].first.depth--;
         }
-        address2os[i].first.depth--;
+        levels[j + 1] -= 1;
       }
-      *coreLevel = -1;
     }
   }
-  return address2os[0].first.depth;
+  return new_depth;
 }
 
 // Returns the number of objects of type 'type' below 'obj' within the topology
 // tree structure. 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)) {
     ++retval;
   }
   return retval;
 }
 
+static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
+                                               hwloc_obj_t o, unsigned depth,
+                                               hwloc_obj_t *f) {
+  if (o->depth == depth) {
+    if (*f == NULL)
+      *f = o; // output first descendant found
+    return 1;
+  }
+  int sum = 0;
+  for (unsigned i = 0; i < o->arity; i++)
+    sum += __kmp_hwloc_count_children_by_depth(t, o->children[i], depth, f);
+  return sum; // will be 0 if no one found (as PU arity is 0)
+}
+
+static int __kmp_hwloc_count_children_by_type(hwloc_topology_t t, hwloc_obj_t o,
+                                              hwloc_obj_type_t type,
+                                              hwloc_obj_t *f) {
+  if (!hwloc_compare_types(o->type, type)) {
+    if (*f == NULL)
+      *f = o; // output first descendant found
+    return 1;
+  }
+  int sum = 0;
+  for (unsigned i = 0; i < o->arity; i++)
+    sum += __kmp_hwloc_count_children_by_type(t, o->children[i], type, f);
+  return sum; // will be 0 if no one found (as PU arity is 0)
+}
+
+static int __kmp_hwloc_process_obj_core_pu(AddrUnsPair *addrPair,
+                                           int &nActiveThreads,
+                                           int &num_active_cores,
+                                           hwloc_obj_t obj, int depth,
+                                           int *labels) {
+  hwloc_obj_t core = NULL;
+  hwloc_topology_t &tp = __kmp_hwloc_topology;
+  int NC = __kmp_hwloc_count_children_by_type(tp, obj, HWLOC_OBJ_CORE, &core);
+  for (int core_id = 0; core_id < NC; ++core_id, core = core->next_cousin) {
+    hwloc_obj_t pu = NULL;
+    KMP_DEBUG_ASSERT(core != NULL);
+    int num_active_threads = 0;
+    int NT = __kmp_hwloc_count_children_by_type(tp, core, HWLOC_OBJ_PU, &pu);
+    // int NT = core->arity; pu = core->first_child; // faster?
+    for (int pu_id = 0; pu_id < NT; ++pu_id, pu = pu->next_cousin) {
+      KMP_DEBUG_ASSERT(pu != NULL);
+      if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask))
+        continue; // skip inactive (inaccessible) unit
+      Address addr(depth + 2);
+      KA_TRACE(20, ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n",
+                    obj->os_index, obj->logical_index, core->os_index,
+                    core->logical_index, pu->os_index, pu->logical_index));
+      for (int i = 0; i < depth; ++i)
+        addr.labels[i] = labels[i]; // package, etc.
+      addr.labels[depth] = core_id; // core
+      addr.labels[depth + 1] = pu_id; // pu
+      addrPair[nActiveThreads] = AddrUnsPair(addr, pu->os_index);
+      __kmp_pu_os_idx[nActiveThreads] = pu->os_index;
+      nActiveThreads++;
+      ++num_active_threads; // count active threads per core
+    }
+    if (num_active_threads) { // were there any active threads on the core?
+      ++__kmp_ncores; // count total active cores
+      ++num_active_cores; // count active cores per socket
+      if (num_active_threads > __kmp_nThreadsPerCore)
+        __kmp_nThreadsPerCore = num_active_threads; // calc maximum
+    }
+  }
+  return 0;
+}
+
+// Check if NUMA node detected below the package,
+// and if tile object is detected and return its depth
+static int __kmp_hwloc_check_numa() {
+  hwloc_topology_t &tp = __kmp_hwloc_topology;
+  hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
+  int depth;
+
+  // Get some PU
+  hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, 0);
+  if (hT == NULL) // something has gone wrong
+    return 1;
+
+  // check NUMA node below PACKAGE
+  hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
+  hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
+  KMP_DEBUG_ASSERT(hS != NULL);
+  if (hN != NULL && hN->depth > hS->depth) {
+    __kmp_numa_detected = TRUE; // socket includes node(s)
+    if (__kmp_affinity_gran == affinity_gran_node) {
+      __kmp_affinity_gran == affinity_gran_numa;
+    }
+  }
+
+  // check tile, get object by depth because of multiple caches possible
+  depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
+  hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT);
+  hC = NULL; // not used, but reset it here just in case
+  if (hL != NULL &&
+      __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1)
+    __kmp_tile_depth = depth; // tile consists of multiple cores
+  return 0;
+}
+
 static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
                                            kmp_i18n_id_t *const msg_id) {
+  hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name
   *address2os = NULL;
   *msg_id = kmp_i18n_null;
 
   // Save the affinity mask for the current thread.
   kmp_affin_mask_t *oldMask;
   KMP_CPU_ALLOC(oldMask);
   __kmp_get_system_affinity(oldMask, TRUE);
-
-  int depth = 3;
-  int pkgLevel = 0;
-  int coreLevel = 1;
-  int threadLevel = 2;
+  __kmp_hwloc_check_numa();
 
   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_PACKAGE, 0),
-        HWLOC_OBJ_CORE);
+        hwloc_get_obj_by_type(tp, 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);
+        hwloc_get_obj_by_type(tp, 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");
       } else {
         KMP_INFORM(NonUniform, "KMP_AFFINITY");
       }
       KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
                  __kmp_nThreadsPerCore, __kmp_ncores);
     }
     KMP_CPU_FREE(oldMask);
     return 0;
   }
 
+  int depth = 3;
+  int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread
+  int labels[3] = {0}; // package [,node] [,tile] - head of lables array
+  if (__kmp_numa_detected)
+    ++depth;
+  if (__kmp_tile_depth)
+    ++depth;
+
   // Allocate the data structure to be returned.
   AddrUnsPair *retval =
       (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
+  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
   __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
 
   // When affinity is off, this routine will still be called to set
   // __kmp_ncores, as well as __kmp_nThreadsPerCore,
   // nCoresPerPkg, & nPackages.  Make sure all these vars are set
   // correctly, and return if affinity is not enabled.
 
-  hwloc_obj_t pu;
-  hwloc_obj_t core;
-  hwloc_obj_t socket;
+  hwloc_obj_t socket, node, tile;
   int nActiveThreads = 0;
-  int socket_identifier = 0;
+  int socket_id = 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_PACKAGE, 0);
-       socket != NULL; 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++) {
-      int pu_identifier = 0;
-      int num_active_threads = 0;
-      for (pu = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, core->type,
-                                            core->logical_index, HWLOC_OBJ_PU,
-                                            0);
-           pu != NULL &&
-           hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, core->type,
-                                          pu) == core;
-           pu = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, HWLOC_OBJ_PU,
-                                           pu),
-          pu_identifier++) {
-        Address addr(3);
-        if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask))
-          continue; // skip inactive (inaccessible) unit
-        KA_TRACE(20,
-                 ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n",
-                  socket->os_index, socket->logical_index, core->os_index,
-                  core->logical_index, pu->os_index, pu->logical_index));
-        addr.labels[0] = socket_identifier; // package
-        addr.labels[1] = core_identifier; // core
-        addr.labels[2] = pu_identifier; // pu
-        retval[nActiveThreads] = AddrUnsPair(addr, pu->os_index);
-        __kmp_pu_os_idx[nActiveThreads] =
-            pu->os_index; // keep os index for each active pu
-        nActiveThreads++;
-        ++num_active_threads; // count active threads per core
+  nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0;
+  for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL;
+       socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket),
+      socket_id++) {
+    labels[0] = socket_id;
+    if (__kmp_numa_detected) {
+      int NN;
+      int n_active_nodes = 0;
+      node = NULL;
+      NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE,
+                                              &node);
+      for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) {
+        labels[1] = node_id;
+        if (__kmp_tile_depth) {
+          // NUMA + tiles
+          int NT;
+          int n_active_tiles = 0;
+          tile = NULL;
+          NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth,
+                                                   &tile);
+          for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
+            labels[2] = tl_id;
+            int n_active_cores = 0;
+            __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
+                                            n_active_cores, tile, 3, labels);
+            if (n_active_cores) { // were there any active cores on the socket?
+              ++n_active_tiles; // count active tiles per node
+              if (n_active_cores > nCorePerTile)
+                nCorePerTile = n_active_cores; // calc maximum
+            }
+          }
+          if (n_active_tiles) { // were there any active tiles on the socket?
+            ++n_active_nodes; // count active nodes per package
+            if (n_active_tiles > nTilePerNode)
+              nTilePerNode = n_active_tiles; // calc maximum
+          }
+        } else {
+          // NUMA, no tiles
+          int n_active_cores = 0;
+          __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
+                                          n_active_cores, node, 2, labels);
+          if (n_active_cores) { // were there any active cores on the socket?
+            ++n_active_nodes; // count active nodes per package
+            if (n_active_cores > nCorePerNode)
+              nCorePerNode = n_active_cores; // calc maximum
           }
-      if (num_active_threads) { // were there any active threads on the core?
-        ++__kmp_ncores; // count total active cores
-        ++num_active_cores; // count active cores per socket
-        if (num_active_threads > __kmp_nThreadsPerCore)
-          __kmp_nThreadsPerCore = num_active_threads; // calc maximum
         }
       }
-    if (num_active_cores) { // were there any active cores on the socket?
+      if (n_active_nodes) { // were there any active nodes on the socket?
         ++nPackages; // count total active packages
-      if (num_active_cores > nCoresPerPkg)
-        nCoresPerPkg = num_active_cores; // calc maximum
+        if (n_active_nodes > nNodePerPkg)
+          nNodePerPkg = n_active_nodes; // calc maximum
+      }
+    } else {
+      if (__kmp_tile_depth) {
+        // no NUMA, tiles
+        int NT;
+        int n_active_tiles = 0;
+        tile = NULL;
+        NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth,
+                                                 &tile);
+        for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) {
+          labels[1] = tl_id;
+          int n_active_cores = 0;
+          __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads,
+                                          n_active_cores, tile, 2, labels);
+          if (n_active_cores) { // were there any active cores on the socket?
+            ++n_active_tiles; // count active tiles per package
+            if (n_active_cores > nCorePerTile)
+              nCorePerTile = n_active_cores; // calc maximum
+          }
+        }
+        if (n_active_tiles) { // were there any active tiles on the socket?
+          ++nPackages; // count total active packages
+          if (n_active_tiles > nTilePerPkg)
+            nTilePerPkg = n_active_tiles; // calc maximum
+        }
+      } else {
+        // no NUMA, no tiles
+        int n_active_cores = 0;
+        __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores,
+                                        socket, 1, labels);
+        if (n_active_cores) { // were there any active cores on the socket?
+          ++nPackages; // count total active packages
+          if (n_active_cores > nCoresPerPkg)
+            nCoresPerPkg = n_active_cores; // calc maximum
+        }
+      }
     }
   }
 
   // If there's only one thread context to bind to, return now.
   KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc);
   KMP_ASSERT(nActiveThreads > 0);
   if (nActiveThreads == 1) {
     __kmp_ncores = nPackages = 1;
if (nActiveThreads == 1) {
     if (__kmp_affinity_type == affinity_none) {
       __kmp_free(retval);
       KMP_CPU_FREE(oldMask);
       return 0;
     }
 
     // Form an Address object which only includes the package level.
     Address addr(1);
-    addr.labels[0] = retval[0].first.labels[pkgLevel];
+    addr.labels[0] = retval[0].first.labels[0];
     retval[0].first = addr;
 
     if (__kmp_affinity_gran_levels < 0) {
       __kmp_affinity_gran_levels = 0;
     }
 
     if (__kmp_affinity_verbose) {
       __kmp_affinity_print_topology(retval, 1, 1, 0, -1, -1);
     }
 
     *address2os = retval;
     KMP_CPU_FREE(oldMask);
     return 1;
   }
 
   // Sort the table by physical Id.
   qsort(retval, nActiveThreads, sizeof(*retval),
         __kmp_affinity_cmp_Address_labels);
 
   // Check to see if the machine topology is uniform
-  unsigned uniform =
-      (nPackages * nCoresPerPkg * __kmp_nThreadsPerCore == nActiveThreads);
+  int nPUs = nPackages * __kmp_nThreadsPerCore;
+  if (__kmp_numa_detected) {
+    if (__kmp_tile_depth) { // NUMA + tiles
+      nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile);
+    } else { // NUMA, no tiles
+      nPUs *= (nNodePerPkg * nCorePerNode);
+    }
+  } else {
+    if (__kmp_tile_depth) { // no NUMA, tiles
+      nPUs *= (nTilePerPkg * nCorePerTile);
+    } else { // no NUMA, no tiles
+      nPUs *= nCoresPerPkg;
+    }
+  }
+  unsigned uniform = (nPUs == nActiveThreads);
 
   // Print the machine topology summary.
   if (__kmp_affinity_verbose) {
     char mask[KMP_AFFIN_MASK_PRINT_LEN];
     __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask);
-
-    KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
     if (__kmp_affinity_respect_mask) {
       KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask);
     } else {
       KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask);
     }
     KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
     if (uniform) {
       KMP_INFORM(Uniform, "KMP_AFFINITY");
     } else {
       KMP_INFORM(NonUniform, "KMP_AFFINITY");
     }
-
+    if (__kmp_numa_detected) {
+      if (__kmp_tile_depth) { // NUMA + tiles
+        KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg,
+                   nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore,
+                   __kmp_ncores);
+      } else { // NUMA, no tiles
+        KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg,
+                   nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores);
+        nPUs *= (nNodePerPkg * nCorePerNode);
+      }
+    } else {
+      if (__kmp_tile_depth) { // no NUMA, tiles
+        KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg,
+                   nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores);
+      } else { // no NUMA, no tiles
         kmp_str_buf_t buf;
         __kmp_str_buf_init(&buf);
-
         __kmp_str_buf_print(&buf, "%d", nPackages);
-    // for (level = 1; level <= pkgLevel; level++) {
-    //    __kmp_str_buf_print(&buf, " x %d", maxCt[level]);
-    // }
         KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg,
                    __kmp_nThreadsPerCore, __kmp_ncores);
-
         __kmp_str_buf_free(&buf);
       }
+    }
+  }
 
   if (__kmp_affinity_type == affinity_none) {
     __kmp_free(retval);
     KMP_CPU_FREE(oldMask);
     return 0;
   }
 
+  int depth_full = depth; // number of levels before compressing
   // Find any levels with radiix 1, and remove them from the map
   // (except for the package level).
-  depth = __kmp_affinity_remove_radix_one_levels(
-      retval, nActiveThreads, depth, &pkgLevel, &coreLevel, &threadLevel);
-
+  depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth,
+                                                 levels);
+  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
   if (__kmp_affinity_gran_levels < 0) {
     // Set the granularity level based on what levels are modeled
     // in the machine topology map.
-    __kmp_affinity_gran_levels = 0;
-    if ((threadLevel >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
-      __kmp_affinity_gran_levels++;
-    }
-    if ((coreLevel >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
+    __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
+    if (__kmp_affinity_gran > affinity_gran_thread) {
+      for (int i = 1; i <= depth_full; ++i) {
+        if (__kmp_affinity_gran <= i) // only count deeper levels
+          break;
+        if (levels[depth_full - i] > 0)
           __kmp_affinity_gran_levels++;
       }
-    if (__kmp_affinity_gran > affinity_gran_package) {
-      __kmp_affinity_gran_levels++;
     }
+    if (__kmp_affinity_gran > affinity_gran_package)
+      __kmp_affinity_gran_levels++; // e.g. granularity = group
   }
 
-  if (__kmp_affinity_verbose) {
-    __kmp_affinity_print_topology(retval, nActiveThreads, depth, pkgLevel,
-                                  coreLevel, threadLevel);
-  }
+  if (__kmp_affinity_verbose)
+    __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels);
 
   KMP_CPU_FREE(oldMask);
   *address2os = retval;
   return depth;
 }
 #endif // KMP_USE_HWLOC
 
 // If we don't know how to retrieve the machine's processor topology, or
if (new_depth != depth) {
     int newThreadLevel = -1;
     int i;
     for (level = 0; level < depth; level++) {
       if ((maxCt[level] == 1) && (level != pkgLevel)) {
         // Remove this level. Never remove the package level
         continue;
       }
       if (level == pkgLevel) {
-        newPkgLevel = level;
+        newPkgLevel = new_level;
       }
       if (level == coreLevel) {
-        newCoreLevel = level;
+        newCoreLevel = new_level;
       }
       if (level == threadLevel) {
-        newThreadLevel = level;
+        newThreadLevel = new_level;
       }
       for (proc = 0; (int)proc < nApics; proc++) {
         new_retval[proc].first.labels[new_level] =
             retval[proc].first.labels[level];
       }
       new_level++;
     }
 
static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
   unsigned depth;
   unsigned maxOsId;
   unsigned i;
 
   KMP_ASSERT(numAddrs > 0);
   depth = address2os[0].first.depth;
 
   maxOsId = 0;
-  for (i = 0; i < numAddrs; i++) {
+  for (i = numAddrs - 1;; --i) {
     unsigned osId = address2os[i].second;
     if (osId > maxOsId) {
       maxOsId = osId;
     }
+    if (i == 0)
+      break;
   }
   kmp_affin_mask_t *osId2Mask;
   KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1));
 
   // Sort the address2os table according to physical order. Doing so will put
   // all threads on the same core/package/node in consecutive locations.
   qsort(address2os, numAddrs, sizeof(*address2os),
         __kmp_affinity_cmp_Address_labels);

 }
 
 #endif /* OMP_40_ENABLED */
 
 #undef ADD_MASK
 #undef ADD_MASK_OSID
 
 #if KMP_USE_HWLOC
-static int __kmp_hwloc_count_children_by_type(hwloc_topology_t t, hwloc_obj_t o,
-                                              hwloc_obj_type_t type,
-                                              hwloc_obj_t *f) {
-  if (!hwloc_compare_types(o->type, type)) {
-    if (*f == NULL)
-      *f = o; // output first descendant found
-    return 1;
-  }
-  int sum = 0;
-  for (unsigned i = 0; i < o->arity; i++)
-    sum += __kmp_hwloc_count_children_by_type(t, o->children[i], type, f);
-  return sum; // will be 0 if no one found (as PU arity is 0)
-}
-
-static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t,
-                                               hwloc_obj_t o, unsigned depth,
-                                               hwloc_obj_t *f) {
-  if (o->depth == depth) {
-    if (*f == NULL)
-      *f = o; // output first descendant found
-    return 1;
-  }
-  int sum = 0;
-  for (unsigned i = 0; i < o->arity; i++)
-    sum += __kmp_hwloc_count_children_by_depth(t, o->children[i], depth, f);
-  return sum; // will be 0 if no one found (as PU arity is 0)
-}
-
 static int __kmp_hwloc_skip_PUs_obj(hwloc_topology_t t, hwloc_obj_t o) {
   // skip PUs descendants of the object o
   int skipped = 0;
   hwloc_obj_t hT = NULL;
   int N = __kmp_hwloc_count_children_by_type(t, o, HWLOC_OBJ_PU, &hT);
   for (int i = 0; i < N; ++i) {
     KMP_DEBUG_ASSERT(hT);
     unsigned idx = hT->os_index;
if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
     // fill in unset items, validate settings -----------------------
     if (__kmp_hws_socket.num == 0)
       __kmp_hws_socket.num = nPackages; // use all available sockets
     if (__kmp_hws_socket.offset >= nPackages) {
       KMP_WARNING(AffHWSubsetManySockets);
       goto _exit;
     }
     if (numa_support) {
+      hN = NULL;
       int NN = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE,
                                                   &hN); // num nodes in socket
       if (__kmp_hws_node.num == 0)
         __kmp_hws_node.num = NN; // use all available nodes
       if (__kmp_hws_node.offset >= NN) {
         KMP_WARNING(AffHWSubsetManyNodes);
         goto _exit;
       }
if (depth < 0) {
       if (depth == 0) {
         KMP_EXIT_AFF_NONE;
       }
       KMP_ASSERT(depth > 0);
       KMP_ASSERT(address2os != NULL);
     }
   }
 
+#if KMP_USE_HWLOC
+  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
+    KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
+    if (__kmp_affinity_verbose) {
+      KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
+    }
+    depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
+    if (depth == 0) {
+      KMP_EXIT_AFF_NONE;
+    }
+  }
+#endif // KMP_USE_HWLOC
+
 // If the user has specified that a paricular topology discovery method is to be
 // used, then we abort if that method fails. The exception is group affinity,
 // which might have been implicitly set.
 
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
 
   else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
     if (__kmp_affinity_verbose) {
else if (__kmp_affinity_top_method == affinity_top_method_flat) {
     if (depth == 0) {
       KMP_EXIT_AFF_NONE;
     }
     // should not fail
     KMP_ASSERT(depth > 0);
     KMP_ASSERT(address2os != NULL);
   }
 
-#if KMP_USE_HWLOC
-  else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
-    KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
-    }
-    depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
-    if (depth == 0) {
-      KMP_EXIT_AFF_NONE;
-    }
-  }
-#endif // KMP_USE_HWLOC
-
   if (address2os == NULL) {
     if (KMP_AFFINITY_CAPABLE() &&
         (__kmp_affinity_verbose ||
          (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
       KMP_WARNING(ErrorInitializeAffinity);
     }
     __kmp_affinity_type = affinity_none;
     KMP_AFFINITY_DISABLE();

runtime/src/kmp_global.cpp

 
 #if KMP_AFFINITY_SUPPORTED
 
 KMPAffinity *__kmp_affinity_dispatch = NULL;
 
 #if KMP_USE_HWLOC
 int __kmp_hwloc_error = FALSE;
 hwloc_topology_t __kmp_hwloc_topology = NULL;
+int __kmp_numa_detected = FALSE;
+int __kmp_tile_depth = 0;
 #endif
 
 #if KMP_OS_WINDOWS
 #if KMP_GROUP_AFFINITY
 int __kmp_num_proc_groups = 1;
 #endif /* KMP_GROUP_AFFINITY */
 kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount = NULL;
 kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount = NULL;

runtime/src/kmp_settings.cpp

#endif
         set_gran(affinity_gran_fine, -1);
         buf = next;
       } else if (__kmp_match_str("thread", buf, CCAST(const char **, &next))) {
         set_gran(affinity_gran_thread, -1);
         buf = next;
       } else if (__kmp_match_str("core", buf, CCAST(const char **, &next))) {
         set_gran(affinity_gran_core, -1);
         buf = next;
+#if KMP_USE_HWLOC
+      } else if (__kmp_match_str("tile", buf, CCAST(const char **, &next))) {
+        set_gran(affinity_gran_tile, -1);
+        buf = next;
+#endif
       } else if (__kmp_match_str("package", buf, CCAST(const char **, &next))) {
         set_gran(affinity_gran_package, -1);
         buf = next;
       } else if (__kmp_match_str("node", buf, CCAST(const char **, &next))) {
         set_gran(affinity_gran_node, -1);
         buf = next;
 #if KMP_GROUP_AFFINITY
       } else if (__kmp_match_str("group", buf, CCAST(const char **, &next))) {
if (__kmp_match_str("threads", scan, &next)) {
     __kmp_affinity_dups = FALSE;
     kind = "\"threads\"";
   } else if (__kmp_match_str("cores", scan, &next)) {
     scan = next;
     __kmp_affinity_type = affinity_compact;
     __kmp_affinity_gran = affinity_gran_core;
     __kmp_affinity_dups = FALSE;
     kind = "\"cores\"";
+#if KMP_USE_HWLOC
+  } else if (__kmp_match_str("tiles", scan, &next)) {
+    scan = next;
+    __kmp_affinity_type = affinity_compact;
+    __kmp_affinity_gran = affinity_gran_tile;
+    __kmp_affinity_dups = FALSE;
+    kind = "\"tiles\"";
+#endif
   } else if (__kmp_match_str("sockets", scan, &next)) {
     scan = next;
     __kmp_affinity_type = affinity_compact;
     __kmp_affinity_gran = affinity_gran_package;
     __kmp_affinity_dups = FALSE;
     kind = "\"sockets\"";
   } else {
     if (__kmp_affinity_proclist != NULL) {
if (__kmp_affinity_gran == affinity_gran_thread) {
         __kmp_str_buf_print(buffer, "='threads'\n");
       }
     } else if (__kmp_affinity_gran == affinity_gran_core) {
       if (num > 0) {
         __kmp_str_buf_print(buffer, "='cores(%d)' \n", num);
       } else {
         __kmp_str_buf_print(buffer, "='cores'\n");
       }
+#if KMP_USE_HWLOC
+    } else if (__kmp_affinity_gran == affinity_gran_tile) {
+      if (num > 0) {
+        __kmp_str_buf_print(buffer, "='tiles(%d)' \n", num);
+      } else {
+        __kmp_str_buf_print(buffer, "='tiles'\n");
+      }
+#endif
     } else if (__kmp_affinity_gran == affinity_gran_package) {
       if (num > 0) {
         __kmp_str_buf_print(buffer, "='sockets(%d)'\n", num);
       } else {
         __kmp_str_buf_print(buffer, "='sockets'\n");
       }
     } else {
       __kmp_str_buf_print(buffer, ": %s\n", KMP_I18N_STR(NotDefined));

 
 #endif /* if (! OMP_40_ENABLED) */
 
 static void __kmp_stg_parse_topology_method(char const *name, char const *value,
                                             void *data) {
   if (__kmp_str_match("all", 1, value)) {
     __kmp_affinity_top_method = affinity_top_method_all;
   }
+#if KMP_USE_HWLOC
+  else if (__kmp_str_match("hwloc", 1, value)) {
+    __kmp_affinity_top_method = affinity_top_method_hwloc;
+  }
+#endif
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
   else if (__kmp_str_match("x2apic id", 9, value) ||
            __kmp_str_match("x2apic_id", 9, value) ||
            __kmp_str_match("x2apic-id", 9, value) ||
            __kmp_str_match("x2apicid", 8, value) ||
            __kmp_str_match("cpuid leaf 11", 13, value) ||
            __kmp_str_match("cpuid_leaf_11", 13, value) ||
            __kmp_str_match("cpuid-leaf-11", 13, value) ||
#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
   }
 #if KMP_GROUP_AFFINITY
   else if (__kmp_str_match("group", 1, value)) {
     __kmp_affinity_top_method = affinity_top_method_group;
   }
 #endif /* KMP_GROUP_AFFINITY */
   else if (__kmp_str_match("flat", 1, value)) {
     __kmp_affinity_top_method = affinity_top_method_flat;
-  }
-#if KMP_USE_HWLOC
-  else if (__kmp_str_match("hwloc", 1, value)) {
-    __kmp_affinity_top_method = affinity_top_method_hwloc;
-  }
-#endif
-  else {
+  } else {
     KMP_WARNING(StgInvalidValue, name, value);
   }
 } // __kmp_stg_parse_topology_method
 
 static void __kmp_stg_print_topology_method(kmp_str_buf_t *buffer,
                                             char const *name, void *data) {
   char const *value = NULL;
 
#endif
               str = "core";
               break;
             case affinity_gran_package:
               str = "package";
               break;
             case affinity_gran_node:
               str = "node";
               break;
+            case affinity_gran_tile:
+              str = "tile";
+              break;
             default:
               KMP_DEBUG_ASSERT(0);
             }
             KMP_WARNING(AffGranTopGroup, var, str);
             __kmp_affinity_gran = affinity_gran_fine;
           }
         } else {
           if (__kmp_affinity_gran == affinity_gran_default) {