[OpenMP] Refactor/Rework topology discovery code

[jlpeyton]

Summary

This patch does the following:

1. Introduces kmp_topology_t as the runtime-friendly structure (the corresponding global variable is __kmp_topology) to determine the exact topology structure which can vary widely among current and future machine architectures. For example, an SNC-4 enabled KNL machine can have 5 topology layers with Hwloc. The current design is not easy to expand beyond the assumed three layer topology: socket, core, and thread so a rework capable of using the existing KMP_AFFINITY mechanisms was required.

This new topology structure has:

• The depth and types of the topology
• Ratio count for each consecutive level (e.g., number of cores per socket, number of threads per core)
• Absolute count for each level (e.g., 2 sockets, 16 cores, 32 threads)
• Equivalent topology layer map (e.g., Numa domain is equivalent to socket, L1/L2 cache equivalent to core)
• Whether it is uniform or not

The hardware threads are represented with the kmp_hw_thread_t structure. This structure contains the ids (e.g., socket 0, core 1, thread 0) and other information grabbed from the previous Address structure. The kmp_topology_t structure contains an array of these.

2. Generalizes the KMP_HW_SUBSET envirable for the new kmp_topology_t structure. The algorithm doesn't assume any order with tiles/numa domains/sockets/cores/threads. Instead it just parses the envirable, makes sure it is consistent with the detected topology (including taking into account equivalent layers) and then trims away the unneeded subset of hardware threads. To enable this, a new kmp_hw_subset_t structure is introduced which contains a vector of items (hardware type, number user wants, offset). Any keyword within __kmp_hw_get_keyword() can be used as a name and can be shortened as well. e.g., KMP_HW_SUBSET=1s,2numa,4tile,2c,3t

3. Rework topology detection functions to be simpler and only do a singular task of detecting the topology. Printing, and all canonicalizing functionality is now done afterwards so many lines of duplicated code are reduced.

4. New TR8 ll_caches and numa_domains are added to OMP_PLACES, and consequently, KMP_AFFINITY's granularity setting. In fact, all the names within __kmp_hw_get_keyword() are available for use in OMP_PLACES or KMP_AFFINITY's granularity setting.

5. A lot of places where explicit listing of allowed names in affinity settings inside if() conditions is made more general so expanding the topology names is less burdensome in the future.

6. CPUID leaf 4 cache detection was added to existing x2apic id method so equivalent caches could be detected (in particular for the ll_caches place).

[tlwilmar]

LGTM

[protze.joachim]

openmp/runtime/src/kmp_affinity.cpp

2602–2605

This patch leaves back prod unused. Is it ok to just remove the variable? I'd include this into a cleanup unused variable patch.

openmp/runtime/src/i18n/en_US.txt

 L3Cache                      "L3 cache"
 L3Caches                     "L3 caches"
 NumaDomain                   "NUMA domain"
 NumaDomains                  "NUMA domains"
 ProcGroup                    "processor group"
 ProcGroups                   "processor groups"
 Unknown                      "unknown"
 NoLeaf31Support              "cpuid leaf 31 not supported"
+HwlocFailed                  "Hwloc api failure"
+LLCache                      "LL cache"
+LLCaches                     "LL caches"
 
 
 
 # --------------------------------------------------------------------------------------------------
 -*- FORMATS -*-
 # --------------------------------------------------------------------------------------------------
 
 Info                         "OMP: Info #%1$d: %2$s\n"

 AffTilesNoTiles              "%1$s: Tiles requested but were not detected on this HW, using granularity=package instead"
 TopologyExtraTile            "%1$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."
 OmpNoAllocator               "Allocator %1$s is not available, will use default allocator."
 TopologyGeneric              "%1$s: %2$s (%3$d total cores)"
+AffGranularityBad            "%1$s: granularity setting: %2$s does not exist in topology.  Using granularity=%3$s instead."
 
 # --- OpenMP errors detected at runtime ---
 #
 #    %1 is the name of OpenMP construct (formatted with "Pragma" format).
 #
 CnsBoundToWorksharing        "%1$s must be bound to a work-sharing or work-queuing construct with an \"ordered\" clause"
 CnsDetectedEnd               "Detected end of %1$s without first executing a corresponding beginning."
 CnsIterationRangeTooLarge    "Iteration range too large in %1$s."

 GompFeatureNotSupported      "libgomp compatibility layer does not support OpenMP feature: %1$s"
 AffHWSubsetManyDies          "KMP_HW_SUBSET ignored: too many Dies requested."
 AffUseGlobCpuidL             "%1$s: Affinity capable, using global cpuid leaf %2$d info"
 AffNotCapableUseLocCpuidL    "%1$s: Affinity not capable, using local cpuid leaf %2$d info"
 AffNotUsingHwloc             "%1$s: Affinity not capable, using hwloc."
 UserDirectedError            "%1$s: Encountered user-directed error: %2$s."
 UserDirectedWarning          "%1$s: Encountered user-directed warning: %2$s."
 FailedToCreateTeam           "Failed to create teams between lower bound (%1$d) and upper bound (%2$d)."
+AffHWSubsetManyGeneric       "KMP_HW_SUBSET ignored: %1$s: too many requested."
+AffHWSubsetNotExistGeneric   "KMP_HW_SUBSET ignored: %1$s: level not detected in machine topology."
+AffHWSubsetEqvLayers         "KMP_HW_SUBSET ignored: %1$s, %2$s: layers are equivalent, please only specify one."
+AffHWSubsetOutOfOrder        "KMP_HW_SUBSET ignored: %1$s layer should come after %2$s."
+AffEqualTopologyTypes        "%1$s: topology layer \"%2$s\" is equivalent to \"%3$s\"."
 
 # --------------------------------------------------------------------------------------------------
 -*- HINTS -*-
 # --------------------------------------------------------------------------------------------------
 
 # Hints. Hint may be printed after a message. Usually it is longer explanation text or suggestion.
 # To maintain hint numbers (they are visible to customers), add new hints to the end.
 

openmp/runtime/src/kmp.h

 
 #if KMP_OS_UNIX
 #include <dlfcn.h>
 #include <pthread.h>
 #endif
 
 enum kmp_hw_t : int {
   KMP_HW_UNKNOWN = -1,
-  KMP_HW_MACHINE = 0,
-  KMP_HW_SOCKET,
+  KMP_HW_SOCKET = 0,
   KMP_HW_PROC_GROUP,
   KMP_HW_NUMA,
   KMP_HW_DIE,
+  KMP_HW_LLC,
   KMP_HW_L3,
   KMP_HW_TILE,
   KMP_HW_MODULE,
   KMP_HW_L2,
   KMP_HW_L1,
   KMP_HW_CORE,
   KMP_HW_THREAD,
   KMP_HW_LAST
 };
 
-#define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
+#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
+#define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
+  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
 
 #define KMP_FOREACH_HW_TYPE(type)                                              \
   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
        type = (kmp_hw_t)((int)type + 1))
 
+const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
 
 /* Only Linux* OS and Windows* OS support thread affinity. */
 #if KMP_AFFINITY_SUPPORTED
 
 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
 #if KMP_OS_WINDOWS
 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT

 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(); (int)i != (mask)->end(); i = (mask)->next(i))
enum affinity_type {
   affinity_compact,
   affinity_scatter,
   affinity_explicit,
   affinity_balanced,
   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_die,
-  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
-};
-
 enum affinity_top_method {
   affinity_top_method_all = 0, // try all (supported) methods, in order
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
   affinity_top_method_apicid,
   affinity_top_method_x2apicid,
   affinity_top_method_x2apicid_1f,
 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
 #if KMP_GROUP_AFFINITY
   affinity_top_method_group,
 #endif /* KMP_GROUP_AFFINITY */
   affinity_top_method_flat,
 #if KMP_USE_HWLOC
   affinity_top_method_hwloc,
 #endif
   affinity_top_method_default
 };
 
 #define affinity_respect_mask_default (-1)
 
 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
-extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
+extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
 extern int __kmp_affinity_gran_levels; /* corresponding int value */
 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
 extern enum affinity_top_method __kmp_affinity_top_method;
 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
 extern int __kmp_affinity_offset; /* Affinity offset value  */
 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?

openmp/runtime/src/kmp_affinity.h

KMPAffinity::Mask *index_mask_array(KMPAffinity::Mask *array,
     Mask *windows_array = static_cast<Mask *>(array);
     return &(windows_array[index]);
   }
   api_type get_api_type() const override { return NATIVE_OS; }
 };
 #endif /* KMP_OS_WINDOWS */
 #endif /* KMP_AFFINITY_SUPPORTED */
 
-class Address {
+class kmp_hw_thread_t {
 public:
-  static const unsigned maxDepth = 32;
-  unsigned labels[maxDepth];
-  unsigned childNums[maxDepth];
-  unsigned depth;
-  unsigned leader;
-  Address(unsigned _depth) : depth(_depth), leader(FALSE) {}
-  Address &operator=(const Address &b) {
-    depth = b.depth;
-    for (unsigned i = 0; i < depth; i++) {
-      labels[i] = b.labels[i];
-      childNums[i] = b.childNums[i];
-    }
-    leader = FALSE;
-    return *this;
-  }
-  bool operator==(const Address &b) const {
-    if (depth != b.depth)
-      return false;
-    for (unsigned i = 0; i < depth; i++)
-      if (labels[i] != b.labels[i])
-        return false;
-    return true;
-  }
-  bool isClose(const Address &b, int level) const {
-    if (depth != b.depth)
-      return false;
-    if ((unsigned)level >= depth)
-      return true;
-    for (unsigned i = 0; i < (depth - level); i++)
-      if (labels[i] != b.labels[i])
-        return false;
-    return true;
-  }
-  bool operator!=(const Address &b) const { return !operator==(b); }
-  void print() const {
-    unsigned i;
-    printf("Depth: %u --- ", depth);
-    for (i = 0; i < depth; i++) {
-      printf("%u ", labels[i]);
-    }
+  static const int UNKNOWN_ID = -1;
+  static int compare_ids(const void *a, const void *b);
+  static int compare_compact(const void *a, const void *b);
+  int ids[KMP_HW_LAST];
+  int sub_ids[KMP_HW_LAST];
+  bool leader;
+  int os_id;
+  void print() const;
+  void clear() {
+    for (int i = 0; i < (int)KMP_HW_LAST; ++i)
+      ids[i] = UNKNOWN_ID;
+    leader = false;
   }
 };
 
-class AddrUnsPair {
+class kmp_topology_t {
+
+  struct flags_t {
+    int uniform : 1;
+    int reserved : 31;
+  };
+
+  int depth;
+
+  // The following arrays are all 'depth' long
+
+  // Orderd array of the types in the topology
+  kmp_hw_t *types;
+
+  // Keep quick topology ratios, for non-uniform topologies,
+  // this ratio holds the max number of itemAs per itemB
+  // e.g., [ 4 packages | 6 cores / package | 2 threads / core ]
+  int *ratio;
+
+  // Storage containing the absolute number of each topology layer
+  int *count;
+
+  // The hardware threads array
+  // hw_threads is num_hw_threads long
+  // Each hw_thread's ids and sub_ids are depth deep
+  int num_hw_threads;
+  kmp_hw_thread_t *hw_threads;
+
+  // Equivalence hash where the key is the hardware topology item
+  // and the value is the equivalent hardware topology type in the
+  // types[] array, if the value is KMP_HW_UNKNOWN, then there is no
+  // known equivalence for the topology type
+  kmp_hw_t equivalent[KMP_HW_LAST];
+
+  // Flags describing the topology
+  flags_t flags;
+
+  // Count each item & get the num x's per y
+  // e.g., get the number of cores and the number of threads per core
+  // for each (x, y) in (KMP_HW_* , KMP_HW_*)
+  void _gather_enumeration_information();
+
+  // Remove layers that don't add information to the topology.
+  // This is done by having the layer take on the id = UNKNOWN_ID (-1)
+  void _remove_radix1_layers();
+
+  // Find out if the topology is uniform
+  void _discover_uniformity();
+
+  // Set all the sub_ids for each hardware thread
+  void _set_sub_ids();
+
+  // Set global affinity variables describing the number of threads per
+  // core, the number of packages, the number of cores per package, and
+  // the number of cores.
+  void _set_globals();
+
+  // Set the last level cache equivalent type
+  void _set_last_level_cache();
+
 public:
-  Address first;
-  unsigned second;
-  AddrUnsPair(Address _first, unsigned _second)
-      : first(_first), second(_second) {}
-  AddrUnsPair &operator=(const AddrUnsPair &b) {
-    first = b.first;
-    second = b.second;
-    return *this;
-  }
-  void print() const {
-    printf("first = ");
-    first.print();
-    printf(" --- second = %u", second);
-  }
-  bool operator==(const AddrUnsPair &b) const {
-    if (first != b.first)
-      return false;
-    if (second != b.second)
-      return false;
-    return true;
-  }
-  bool operator!=(const AddrUnsPair &b) const { return !operator==(b); }
-};
-
-static int __kmp_affinity_cmp_Address_labels(const void *a, const void *b) {
-  const Address *aa = &(((const AddrUnsPair *)a)->first);
-  const Address *bb = &(((const AddrUnsPair *)b)->first);
-  unsigned depth = aa->depth;
-  unsigned i;
-  KMP_DEBUG_ASSERT(depth == bb->depth);
-  for (i = 0; i < depth; i++) {
-    if (aa->labels[i] < bb->labels[i])
-      return -1;
-    if (aa->labels[i] > bb->labels[i])
-      return 1;
+  // Force use of allocate()/deallocate()
+  kmp_topology_t() = delete;
+  kmp_topology_t(const kmp_topology_t &t) = delete;
+  kmp_topology_t(kmp_topology_t &&t) = delete;
+  kmp_topology_t &operator=(const kmp_topology_t &t) = delete;
+  kmp_topology_t &operator=(kmp_topology_t &&t) = delete;
+
+  static kmp_topology_t *allocate(int nproc, int ndepth, const kmp_hw_t *types);
+  static void deallocate(kmp_topology_t *);
+
+  // Functions used in create_map() routines
+  kmp_hw_thread_t &at(int index) {
+    KMP_DEBUG_ASSERT(index >= 0 && index < num_hw_threads);
+    return hw_threads[index];
+  }
+  const kmp_hw_thread_t &at(int index) const {
+    KMP_DEBUG_ASSERT(index >= 0 && index < num_hw_threads);
+    return hw_threads[index];
+  }
+  int get_num_hw_threads() const { return num_hw_threads; }
+  void sort_ids() {
+    qsort(hw_threads, num_hw_threads, sizeof(kmp_hw_thread_t),
+          kmp_hw_thread_t::compare_ids);
+  }
+  // Check if the hardware ids are unique, if they are
+  // return true, otherwise return false
+  bool check_ids() const;
+
+  // Function to call after the create_map() routine
+  void canonicalize();
+  void canonicalize(int pkgs, int cores_per_pkg, int thr_per_core, int cores);
+
+  // Functions used after canonicalize() called
+  bool filter_hw_subset();
+  bool is_close(int hwt1, int hwt2, int level) const;
+  bool is_uniform() const { return flags.uniform; }
+  // Tell whether a type is a valid type in the topology
+  // returns KMP_HW_UNKNOWN when there is no equivalent type
+  kmp_hw_t get_equivalent_type(kmp_hw_t type) const { return equivalent[type]; }
+  // Set type1 = type2
+  void set_equivalent_type(kmp_hw_t type1, kmp_hw_t type2) {
+    KMP_DEBUG_ASSERT_VALID_HW_TYPE(type1);
+    KMP_DEBUG_ASSERT_VALID_HW_TYPE(type2);
+    kmp_hw_t real_type2 = equivalent[type2];
+    if (real_type2 == KMP_HW_UNKNOWN)
+      real_type2 = type2;
+    equivalent[type1] = real_type2;
+    // This loop is required since any of the types may have been set to
+    // be equivalent to type1.  They all must be checked and reset to type2.
+    KMP_FOREACH_HW_TYPE(type) {
+      if (equivalent[type] == type1) {
+        equivalent[type] = real_type2;
+      }
+    }
+  }
+  // Calculate number of types corresponding to level1
+  // per types corresponding to level2 (e.g., number of threads per core)
+  int calculate_ratio(int level1, int level2) const {
+    KMP_DEBUG_ASSERT(level1 >= 0 && level1 < depth);
+    KMP_DEBUG_ASSERT(level2 >= 0 && level2 < depth);
+    int r = 1;
+    for (int level = level1; level > level2; --level)
+      r *= ratio[level];
+    return r;
+  }
+  int get_ratio(int level) const {
+    KMP_DEBUG_ASSERT(level >= 0 && level < depth);
+    return ratio[level];
+  }
+  int get_depth() const { return depth; };
+  kmp_hw_t get_type(int level) const {
+    KMP_DEBUG_ASSERT(level >= 0 && level < depth);
+    return types[level];
+  }
+  int get_level(kmp_hw_t type) const {
+    KMP_DEBUG_ASSERT_VALID_HW_TYPE(type);
+    int eq_type = equivalent[type];
+    if (eq_type == KMP_HW_UNKNOWN)
+      return -1;
+    for (int i = 0; i < depth; ++i)
+      if (types[i] == eq_type)
+        return i;
   }
-  return 0;
+  int get_count(int level) const {
+    KMP_DEBUG_ASSERT(level >= 0 && level < depth);
+    return count[level];
+  }
+#if KMP_AFFINITY_SUPPORTED
+  void sort_compact() {
+    qsort(hw_threads, num_hw_threads, sizeof(kmp_hw_thread_t),
+          kmp_hw_thread_t::compare_compact);
+  }
+#endif
+  void print(const char *env_var = "KMP_AFFINITY") const;
+  void dump() const;
+};
+
+class kmp_hw_subset_t {
+public:
+  struct item_t {
+    int num;
+    kmp_hw_t type;
+    int offset;
+  };
+
+private:
+  int depth;
+  int capacity;
+  item_t *items;
+  kmp_uint64 set;
+  bool absolute;
+  // The set must be able to handle up to KMP_HW_LAST number of layers
+  KMP_BUILD_ASSERT(sizeof(set) * 8 >= KMP_HW_LAST);
+
+public:
+  // Force use of allocate()/deallocate()
+  kmp_hw_subset_t() = delete;
+  kmp_hw_subset_t(const kmp_hw_subset_t &t) = delete;
+  kmp_hw_subset_t(kmp_hw_subset_t &&t) = delete;
+  kmp_hw_subset_t &operator=(const kmp_hw_subset_t &t) = delete;
+  kmp_hw_subset_t &operator=(kmp_hw_subset_t &&t) = delete;
+
+  static kmp_hw_subset_t *allocate() {
+    int initial_capacity = 5;
+    kmp_hw_subset_t *retval =
+        (kmp_hw_subset_t *)__kmp_allocate(sizeof(kmp_hw_subset_t));
+    retval->depth = 0;
+    retval->capacity = initial_capacity;
+    retval->set = 0ull;
+    retval->absolute = false;
+    retval->items = (item_t *)__kmp_allocate(sizeof(item_t) * initial_capacity);
+    return retval;
+  }
+  static void deallocate(kmp_hw_subset_t *subset) {
+    __kmp_free(subset->items);
+    __kmp_free(subset);
+  }
+  void set_absolute() { absolute = true; }
+  bool is_absolute() const { return absolute; }
+  void push_back(int num, kmp_hw_t type, int offset) {
+    if (depth == capacity - 1) {
+      capacity *= 2;
+      item_t *new_items = (item_t *)__kmp_allocate(sizeof(item_t) * capacity);
+      for (int i = 0; i < depth; ++i)
+        new_items[i] = items[i];
+      __kmp_free(items);
+      items = new_items;
+    }
+    items[depth].num = num;
+    items[depth].type = type;
+    items[depth].offset = offset;
+    depth++;
+    set |= (1ull << type);
+  }
+  int get_depth() const { return depth; }
+  const item_t &at(int index) const {
+    KMP_DEBUG_ASSERT(index >= 0 && index < depth);
+    return items[index];
+  }
+  item_t &at(int index) {
+    KMP_DEBUG_ASSERT(index >= 0 && index < depth);
+    return items[index];
+  }
+  void remove(int index) {
+    KMP_DEBUG_ASSERT(index >= 0 && index < depth);
+    set &= ~(1ull << items[index].type);
+    for (int j = index + 1; j < depth; ++j) {
+      items[j - 1] = items[j];
+    }
+    depth--;
+  }
+  bool specified(kmp_hw_t type) const { return ((set & (1ull << type)) > 0); }
+  void dump() const {
+    printf("**********************\n");
+    printf("*** kmp_hw_subset: ***\n");
+    printf("* depth: %d\n", depth);
+    printf("* items:\n");
+    for (int i = 0; i < depth; ++i) {
+      printf("num: %d, type: %s, offset: %d\n", items[i].num,
+             __kmp_hw_get_keyword(items[i].type), items[i].offset);
+    }
+    printf("* set: 0x%llx\n", set);
+    printf("* absolute: %d\n", absolute);
+    printf("**********************\n");
   }
+};
+
+extern kmp_topology_t *__kmp_topology;
+extern kmp_hw_subset_t *__kmp_hw_subset;
 
 /* A structure for holding machine-specific hierarchy info to be computed once
    at init. This structure represents a mapping of threads to the actual machine
    hierarchy, or to our best guess at what the hierarchy might be, for the
    purpose of performing an efficient barrier. In the worst case, when there is
    no machine hierarchy information, it produces a tree suitable for a barrier,
    similar to the tree used in the hyper barrier. */
 class hierarchy_info {
public:
 
   /** Level 0 corresponds to leaves. numPerLevel[i] is the number of children
       the parent of a node at level i has. For example, if we have a machine
       with 4 packages, 4 cores/package and 2 HT per core, then numPerLevel =
       {2, 4, 4, 1, 1}. All empty levels are set to 1. */
   kmp_uint32 *numPerLevel;
   kmp_uint32 *skipPerLevel;
 
-  void deriveLevels(AddrUnsPair *adr2os, int num_addrs) {
-    int hier_depth = adr2os[0].first.depth;
-    int level = 0;
-    for (int i = hier_depth - 1; i >= 0; --i) {
-      int max = -1;
-      for (int j = 0; j < num_addrs; ++j) {
-        int next = adr2os[j].first.childNums[i];
-        if (next > max)
-          max = next;
-      }
-      numPerLevel[level] = max + 1;
-      ++level;
+  void deriveLevels() {
+    int hier_depth = __kmp_topology->get_depth();
+    for (int i = hier_depth - 1, level = 0; i >= 0; --i, ++level) {
+      numPerLevel[level] = __kmp_topology->get_ratio(i);
     }
   }
 
   hierarchy_info()
       : maxLevels(7), depth(1), uninitialized(not_initialized), resizing(0) {}
 
   void fini() {
     if (!uninitialized && numPerLevel) {
       __kmp_free(numPerLevel);
       numPerLevel = NULL;
       uninitialized = not_initialized;
     }
   }
 
-  void init(AddrUnsPair *adr2os, int num_addrs) {
+  void init(int num_addrs) {
     kmp_int8 bool_result = KMP_COMPARE_AND_STORE_ACQ8(
         &uninitialized, not_initialized, initializing);
     if (bool_result == 0) { // Wait for initialization
       while (TCR_1(uninitialized) != initialized)
         KMP_CPU_PAUSE();
       return;
     }
     KMP_DEBUG_ASSERT(bool_result == 1);
void init(int num_addrs) {
     skipPerLevel = &(numPerLevel[maxLevels]);
     for (kmp_uint32 i = 0; i < maxLevels;
          ++i) { // init numPerLevel[*] to 1 item per level
       numPerLevel[i] = 1;
       skipPerLevel[i] = 1;
     }
 
     // Sort table by physical ID
-    if (adr2os) {
-      qsort(adr2os, num_addrs, sizeof(*adr2os),
-            __kmp_affinity_cmp_Address_labels);
-      deriveLevels(adr2os, num_addrs);
+    if (__kmp_topology && __kmp_topology->get_depth() > 0) {
+      deriveLevels();
     } else {
       numPerLevel[0] = maxLeaves;
       numPerLevel[1] = num_addrs / maxLeaves;
       if (num_addrs % maxLeaves)
         numPerLevel[1]++;
     }
 
     base_num_threads = num_addrs;

openmp/runtime/src/kmp_affinity.cpp

 #include "kmp_io.h"
 #include "kmp_str.h"
 #include "kmp_wrapper_getpid.h"
 #if KMP_USE_HIER_SCHED
 #include "kmp_dispatch_hier.h"
 #endif
 #if KMP_USE_HWLOC
 // Copied from hwloc
+#define HWLOC_GROUP_KIND_INTEL_MODULE 102
+#define HWLOC_GROUP_KIND_INTEL_TILE 103
 #define HWLOC_GROUP_KIND_INTEL_DIE 104
+#define HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP 220
 #endif
 
+// The machine topology
+kmp_topology_t *__kmp_topology = nullptr;
+// KMP_HW_SUBSET environment variable
+kmp_hw_subset_t *__kmp_hw_subset = nullptr;
+
 // Store the real or imagined machine hierarchy here
 static hierarchy_info machine_hierarchy;
 
 void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
 
 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
   kmp_uint32 depth;
   // The test below is true if affinity is available, but set to "none". Need to
   // init on first use of hierarchical barrier.
   if (TCR_1(machine_hierarchy.uninitialized))
-    machine_hierarchy.init(NULL, nproc);
+    machine_hierarchy.init(nproc);
 
   // Adjust the hierarchy in case num threads exceeds original
   if (nproc > machine_hierarchy.base_num_threads)
     machine_hierarchy.resize(nproc);
 
   depth = machine_hierarchy.depth;
   KMP_DEBUG_ASSERT(depth > 0);
 
   thr_bar->depth = depth;
   __kmp_type_convert(machine_hierarchy.numPerLevel[0] - 1,
                      &(thr_bar->base_leaf_kids));
   thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
 }
 
-#if KMP_AFFINITY_SUPPORTED
+static int nCoresPerPkg, nPackages;
+static int __kmp_nThreadsPerCore;
+#ifndef KMP_DFLT_NTH_CORES
+static int __kmp_ncores;
+#endif
 
 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
   switch (type) {
   case KMP_HW_SOCKET:
     return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
   case KMP_HW_DIE:
     return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
   case KMP_HW_MODULE:
     return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
   case KMP_HW_TILE:
     return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
   case KMP_HW_NUMA:
     return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
   case KMP_HW_L3:
     return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
   case KMP_HW_L2:
     return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
   case KMP_HW_L1:
     return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
+  case KMP_HW_LLC:
+    return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
   case KMP_HW_CORE:
     return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
   case KMP_HW_THREAD:
     return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
   case KMP_HW_PROC_GROUP:
     return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
   }
   return KMP_I18N_STR(Unknown);
 }
 
-// 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 *addrP, int nTh,
-                                                  int depth, kmp_hw_t *types) {
+const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
+  switch (type) {
+  case KMP_HW_SOCKET:
+    return ((plural) ? "sockets" : "socket");
+  case KMP_HW_DIE:
+    return ((plural) ? "dice" : "die");
+  case KMP_HW_MODULE:
+    return ((plural) ? "modules" : "module");
+  case KMP_HW_TILE:
+    return ((plural) ? "tiles" : "tile");
+  case KMP_HW_NUMA:
+    return ((plural) ? "numa_domains" : "numa_domain");
+  case KMP_HW_L3:
+    return ((plural) ? "l3_caches" : "l3_cache");
+  case KMP_HW_L2:
+    return ((plural) ? "l2_caches" : "l2_cache");
+  case KMP_HW_L1:
+    return ((plural) ? "l1_caches" : "l1_cache");
+  case KMP_HW_LLC:
+    return ((plural) ? "ll_caches" : "ll_cache");
+  case KMP_HW_CORE:
+    return ((plural) ? "cores" : "core");
+  case KMP_HW_THREAD:
+    return ((plural) ? "threads" : "thread");
+  case KMP_HW_PROC_GROUP:
+    return ((plural) ? "proc_groups" : "proc_group");
+  }
+  return ((plural) ? "unknowns" : "unknown");
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// kmp_hw_thread_t methods
+int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
+  const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
+  const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
+  int depth = __kmp_topology->get_depth();
+  for (int level = 0; level < depth; ++level) {
+    if (ahwthread->ids[level] < bhwthread->ids[level])
+      return -1;
+    else if (ahwthread->ids[level] > bhwthread->ids[level])
+      return 1;
+  }
+  if (ahwthread->os_id < bhwthread->os_id)
+    return -1;
+  else if (ahwthread->os_id > bhwthread->os_id)
+    return 1;
+  return 0;
+}
+
+#if KMP_AFFINITY_SUPPORTED
+int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
+  int i;
+  const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
+  const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
+  int depth = __kmp_topology->get_depth();
+  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
+  KMP_DEBUG_ASSERT(__kmp_affinity_compact <= depth);
+  for (i = 0; i < __kmp_affinity_compact; i++) {
+    int j = depth - i - 1;
+    if (aa->sub_ids[j] < bb->sub_ids[j])
+      return -1;
+    if (aa->sub_ids[j] > bb->sub_ids[j])
+      return 1;
+  }
+  for (; i < depth; i++) {
+    int j = i - __kmp_affinity_compact;
+    if (aa->sub_ids[j] < bb->sub_ids[j])
+      return -1;
+    if (aa->sub_ids[j] > bb->sub_ids[j])
+      return 1;
+  }
+  return 0;
+}
+#endif
+
+void kmp_hw_thread_t::print() const {
+  int depth = __kmp_topology->get_depth();
+  printf("%4d ", os_id);
+  for (int i = 0; i < depth; ++i) {
+    printf("%4d ", ids[i]);
+  }
+  printf("\n");
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// kmp_topology_t methods
+
+// Remove layers that don't add information to the topology.
+// This is done by having the layer take on the id = UNKNOWN_ID (-1)
+void kmp_topology_t::_remove_radix1_layers() {
   int preference[KMP_HW_LAST];
   int top_index1, top_index2;
   // Set up preference associative array
   preference[KMP_HW_PROC_GROUP] = 110;
   preference[KMP_HW_SOCKET] = 100;
   preference[KMP_HW_CORE] = 95;
   preference[KMP_HW_THREAD] = 90;
-  preference[KMP_HW_DIE] = 85;
-  preference[KMP_HW_NUMA] = 80;
+  preference[KMP_HW_NUMA] = 85;
+  preference[KMP_HW_DIE] = 80;
   preference[KMP_HW_TILE] = 75;
   preference[KMP_HW_MODULE] = 73;
   preference[KMP_HW_L3] = 70;
   preference[KMP_HW_L2] = 65;
   preference[KMP_HW_L1] = 60;
+  preference[KMP_HW_LLC] = 5;
   top_index1 = 0;
   top_index2 = 1;
   while (top_index1 < depth - 1 && top_index2 < depth) {
-    KMP_DEBUG_ASSERT(top_index1 >= 0 && top_index1 < depth);
-    KMP_DEBUG_ASSERT(top_index2 >= 0 && top_index2 < depth);
     kmp_hw_t type1 = types[top_index1];
     kmp_hw_t type2 = types[top_index2];
-    if (type1 == KMP_HW_SOCKET && type2 == KMP_HW_CORE) {
+    KMP_ASSERT_VALID_HW_TYPE(type1);
+    KMP_ASSERT_VALID_HW_TYPE(type2);
+    // Do not allow the three main topology levels (sockets, cores, threads) to
+    // be compacted down
+    if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
+         type1 == KMP_HW_SOCKET) &&
+        (type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
+         type2 == KMP_HW_SOCKET)) {
       top_index1 = top_index2++;
       continue;
     }
     bool radix1 = true;
     bool all_same = true;
-    unsigned id1 = addrP[0].first.labels[top_index1];
-    unsigned id2 = addrP[0].first.labels[top_index2];
+    int id1 = hw_threads[0].ids[top_index1];
+    int id2 = hw_threads[0].ids[top_index2];
     int pref1 = preference[type1];
     int pref2 = preference[type2];
-    for (int hwidx = 1; hwidx < nTh; ++hwidx) {
-      if (addrP[hwidx].first.labels[top_index1] == id1 &&
-          addrP[hwidx].first.labels[top_index2] != id2) {
+    for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
+      if (hw_threads[hwidx].ids[top_index1] == id1 &&
+          hw_threads[hwidx].ids[top_index2] != id2) {
         radix1 = false;
         break;
       }
-      if (addrP[hwidx].first.labels[top_index2] != id2)
+      if (hw_threads[hwidx].ids[top_index2] != id2)
         all_same = false;
-      id1 = addrP[hwidx].first.labels[top_index1];
-      id2 = addrP[hwidx].first.labels[top_index2];
+      id1 = hw_threads[hwidx].ids[top_index1];
+      id2 = hw_threads[hwidx].ids[top_index2];
     }
     if (radix1) {
       // Select the layer to remove based on preference
       kmp_hw_t remove_type, keep_type;
       int remove_layer, remove_layer_ids;
       if (pref1 > pref2) {
         remove_type = type2;
         remove_layer = remove_layer_ids = top_index2;
         keep_type = type1;
       } else {
         remove_type = type1;
         remove_layer = remove_layer_ids = top_index1;
         keep_type = type2;
       }
       // If all the indexes for the second (deeper) layer are the same.
       // e.g., all are zero, then make sure to keep the first layer's ids
       if (all_same)
         remove_layer_ids = top_index2;
       // Remove radix one type by setting the equivalence, removing the id from
       // the hw threads and removing the layer from types and depth
-      for (int idx = 0; idx < nTh; ++idx) {
-        Address &hw_thread = addrP[idx].first;
+      set_equivalent_type(remove_type, keep_type);
+      for (int idx = 0; idx < num_hw_threads; ++idx) {
+        kmp_hw_thread_t &hw_thread = hw_threads[idx];
         for (int d = remove_layer_ids; d < depth - 1; ++d)
-          hw_thread.labels[d] = hw_thread.labels[d + 1];
-        hw_thread.depth--;
+          hw_thread.ids[d] = hw_thread.ids[d + 1];
       }
       for (int idx = remove_layer; idx < depth - 1; ++idx)
         types[idx] = types[idx + 1];
       depth--;
     } else {
       top_index1 = top_index2++;
     }
   }
   KMP_ASSERT(depth > 0);
-  return depth;
 }
+
+void kmp_topology_t::_set_last_level_cache() {
+  if (get_equivalent_type(KMP_HW_L3) != KMP_HW_UNKNOWN)
+    set_equivalent_type(KMP_HW_LLC, KMP_HW_L3);
+  else if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
+    set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
+#if KMP_MIC_SUPPORTED
+  else if (__kmp_mic_type == mic3) {
+    if (get_equivalent_type(KMP_HW_L2) != KMP_HW_UNKNOWN)
+      set_equivalent_type(KMP_HW_LLC, KMP_HW_L2);
+    else if (get_equivalent_type(KMP_HW_TILE) != KMP_HW_UNKNOWN)
+      set_equivalent_type(KMP_HW_LLC, KMP_HW_TILE);
+    // L2/Tile wasn't detected so just say L1
+    else
+      set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
+  }
+#endif
+  else if (get_equivalent_type(KMP_HW_L1) != KMP_HW_UNKNOWN)
+    set_equivalent_type(KMP_HW_LLC, KMP_HW_L1);
+  // Fallback is to set last level cache to socket or core
+  if (get_equivalent_type(KMP_HW_LLC) == KMP_HW_UNKNOWN) {
+    if (get_equivalent_type(KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
+      set_equivalent_type(KMP_HW_LLC, KMP_HW_SOCKET);
+    else if (get_equivalent_type(KMP_HW_CORE) != KMP_HW_UNKNOWN)
+      set_equivalent_type(KMP_HW_LLC, KMP_HW_CORE);
+  }
+  KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
+}
+
 // Gather the count of each topology layer and the ratio
-// ratio contains the number of types[i] / types[i+1] and so forth
-// count contains the absolute number of types[i]
-static void __kmp_affinity_gather_enumeration_information(AddrUnsPair *addrP,
-                                                          int nTh, int depth,
-                                                          kmp_hw_t *types,
-                                                          int *ratio,
-                                                          int *count) {
+void kmp_topology_t::_gather_enumeration_information() {
   int previous_id[KMP_HW_LAST];
   int max[KMP_HW_LAST];
 
   for (int i = 0; i < depth; ++i) {
-    previous_id[i] = -1;
+    previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
     max[i] = 0;
     count[i] = 0;
     ratio[i] = 0;
   }
-  for (int i = 0; i < nTh; ++i) {
-    Address &hw_thread = addrP[i].first;
+  for (int i = 0; i < num_hw_threads; ++i) {
+    kmp_hw_thread_t &hw_thread = hw_threads[i];
     for (int layer = 0; layer < depth; ++layer) {
-      int id = hw_thread.labels[layer];
+      int id = hw_thread.ids[layer];
       if (id != previous_id[layer]) {
         // Add an additional increment to each count
         for (int l = layer; l < depth; ++l)
           count[l]++;
         // Keep track of topology layer ratio statistics
         max[layer]++;
         for (int l = layer + 1; l < depth; ++l) {
           if (max[l] > ratio[l])
             ratio[l] = max[l];
           max[l] = 1;
         }
         break;
       }
     }
     for (int layer = 0; layer < depth; ++layer) {
-      previous_id[layer] = hw_thread.labels[layer];
+      previous_id[layer] = hw_thread.ids[layer];
     }
   }
   for (int layer = 0; layer < depth; ++layer) {
     if (max[layer] > ratio[layer])
       ratio[layer] = max[layer];
   }
 }
 
 // Find out if the topology is uniform
-static bool __kmp_affinity_discover_uniformity(int depth, int *ratio,
-                                               int *count) {
+void kmp_topology_t::_discover_uniformity() {
   int num = 1;
   for (int level = 0; level < depth; ++level)
     num *= ratio[level];
-  return (num == count[depth - 1]);
+  flags.uniform = (num == count[depth - 1]);
 }
 
-// calculate the number of X's per Y
-static inline int __kmp_affinity_calculate_ratio(int *ratio, int deep_level,
-                                                 int shallow_level) {
-  int retval = 1;
-  if (deep_level < 0 || shallow_level < 0)
-    return retval;
-  for (int level = deep_level; level > shallow_level; --level)
-    retval *= ratio[level];
+// Set all the sub_ids for each hardware thread
+void kmp_topology_t::_set_sub_ids() {
+  int previous_id[KMP_HW_LAST];
+  int sub_id[KMP_HW_LAST];
+
+  for (int i = 0; i < depth; ++i) {
+    previous_id[i] = -1;
+    sub_id[i] = -1;
+  }
+  for (int i = 0; i < num_hw_threads; ++i) {
+    kmp_hw_thread_t &hw_thread = hw_threads[i];
+    // Setup the sub_id
+    for (int j = 0; j < depth; ++j) {
+      if (hw_thread.ids[j] != previous_id[j]) {
+        sub_id[j]++;
+        for (int k = j + 1; k < depth; ++k) {
+          sub_id[k] = 0;
+        }
+        break;
+      }
+    }
+    // Set previous_id
+    for (int j = 0; j < depth; ++j) {
+      previous_id[j] = hw_thread.ids[j];
+    }
+    // Set the sub_ids field
+    for (int j = 0; j < depth; ++j) {
+      hw_thread.sub_ids[j] = sub_id[j];
+    }
+  }
+}
+
+void kmp_topology_t::_set_globals() {
+  // Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
+  int core_level, thread_level, package_level;
+  package_level = get_level(KMP_HW_SOCKET);
+#if KMP_GROUP_AFFINITY
+  if (package_level == -1)
+    package_level = get_level(KMP_HW_PROC_GROUP);
+#endif
+  core_level = get_level(KMP_HW_CORE);
+  thread_level = get_level(KMP_HW_THREAD);
+
+  KMP_ASSERT(core_level != -1);
+  KMP_ASSERT(thread_level != -1);
+
+  __kmp_nThreadsPerCore = calculate_ratio(thread_level, core_level);
+  if (package_level != -1) {
+    nCoresPerPkg = calculate_ratio(core_level, package_level);
+    nPackages = get_count(package_level);
+  } else {
+    // assume one socket
+    nCoresPerPkg = get_count(core_level);
+    nPackages = 1;
+  }
+#ifndef KMP_DFLT_NTH_CORES
+  __kmp_ncores = get_count(core_level);
+#endif
+}
+
+kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
+                                         const kmp_hw_t *types) {
+  kmp_topology_t *retval;
+  // Allocate all data in one large allocation
+  size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
+                sizeof(int) * ndepth * 3;
+  char *bytes = (char *)__kmp_allocate(size);
+  retval = (kmp_topology_t *)bytes;
+  if (nproc > 0) {
+    retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
+  } else {
+    retval->hw_threads = nullptr;
+  }
+  retval->num_hw_threads = nproc;
+  retval->depth = ndepth;
+  int *arr =
+      (int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
+  retval->types = (kmp_hw_t *)arr;
+  retval->ratio = arr + ndepth;
+  retval->count = arr + 2 * ndepth;
+  KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
+  for (int i = 0; i < ndepth; ++i) {
+    retval->types[i] = types[i];
+    retval->equivalent[types[i]] = types[i];
+  }
   return retval;
 }
 
-static void __kmp_affinity_print_topology(AddrUnsPair *addrP, int len,
-                                          int depth, kmp_hw_t *types) {
-  int proc;
-  kmp_str_buf_t buf;
-  __kmp_str_buf_init(&buf);
-  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
-  for (proc = 0; proc < len; proc++) {
-    for (int i = 0; i < depth; ++i) {
-      __kmp_str_buf_print(&buf, "%s %d ", __kmp_hw_get_catalog_string(types[i]),
-                          addrP[proc].first.labels[i]);
-    }
-    KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str);
-    __kmp_str_buf_clear(&buf);
+void kmp_topology_t::deallocate(kmp_topology_t *topology) {
+  if (topology)
+    __kmp_free(topology);
+}
+
+bool kmp_topology_t::check_ids() const {
+  // Assume ids have been sorted
+  if (num_hw_threads == 0)
+    return true;
+  for (int i = 1; i < num_hw_threads; ++i) {
+    kmp_hw_thread_t &current_thread = hw_threads[i];
+    kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
+    bool unique = false;
+    for (int j = 0; j < depth; ++j) {
+      if (previous_thread.ids[j] != current_thread.ids[j]) {
+        unique = true;
+        break;
+      }
+    }
+    if (unique)
+      continue;
+    return false;
+  }
+  return true;
+}
+
+void kmp_topology_t::dump() const {
+  printf("***********************\n");
+  printf("*** __kmp_topology: ***\n");
+  printf("***********************\n");
+  printf("* depth: %d\n", depth);
+
+  printf("* types: ");
+  for (int i = 0; i < depth; ++i)
+    printf("%15s ", __kmp_hw_get_keyword(types[i]));
+  printf("\n");
+
+  printf("* ratio: ");
+    printf("%15d ", ratio[i]);
+  printf("\n");
+
+  printf("* count: ");
+  for (int i = 0; i < depth; ++i) {
+    printf("%15d ", count[i]);
   }
-  __kmp_str_buf_free(&buf);
+  printf("\n");
+
+  printf("* equivalent map:\n");
+  KMP_FOREACH_HW_TYPE(i) {
+    const char *key = __kmp_hw_get_keyword(i);
+    const char *value = __kmp_hw_get_keyword(equivalent[i]);
+    printf("%-15s -> %-15s\n", key, value);
   }
 
-// Print out the detailed machine topology map, i.e. the physical locations
-// of each OS proc.
-static void __kmp_affinity_print_topology(AddrUnsPair *address2os, int len,
-                                          int depth, int pkgLevel,
-                                          int coreLevel, int threadLevel) {
-  int proc;
+  printf("* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
+
+  printf("* num_hw_threads: %d\n", num_hw_threads);
+  printf("* hw_threads:\n");
+  for (int i = 0; i < num_hw_threads; ++i) {
+    hw_threads[i].print();
+  }
+  printf("***********************\n");
+}
 
-  KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY");
-  for (proc = 0; proc < len; proc++) {
-    int level;
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-    for (level = 0; level < depth; level++) {
-      if (level == threadLevel) {
-        __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Thread));
-      } else if (level == coreLevel) {
-        __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Core));
-      } else if (level == pkgLevel) {
-        __kmp_str_buf_print(&buf, "%s ", KMP_I18N_STR(Package));
-      } else if (level > pkgLevel) {
-        __kmp_str_buf_print(&buf, "%s_%d ", KMP_I18N_STR(Node),
-                            level - pkgLevel - 1);
-      } else {
-        __kmp_str_buf_print(&buf, "L%d ", level);
-      }
-      __kmp_str_buf_print(&buf, "%d ", address2os[proc].first.labels[level]);
-    }
-    KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", address2os[proc].second,
-               buf.str);
+void kmp_topology_t::print(const char *env_var) const {
+  int print_types_depth;
+  kmp_hw_t print_types[KMP_HW_LAST + 2];
+
+  // Num Available Threads
+  KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
+
+  // Uniform or not
+  if (is_uniform()) {
+    KMP_INFORM(Uniform, env_var);
+    KMP_INFORM(NonUniform, env_var);
+  }
+
+  // Equivalent types
+  KMP_FOREACH_HW_TYPE(type) {
+    kmp_hw_t eq_type = equivalent[type];
+    if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
+      KMP_INFORM(AffEqualTopologyTypes, env_var,
+                 __kmp_hw_get_catalog_string(type),
+                 __kmp_hw_get_catalog_string(eq_type));
+    }
+  }
+
+  // Quick topology
+  KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
+  // Create a print types array that always guarantees printing
+  // the core and thread level
+  print_types_depth = 0;
+  for (int level = 0; level < depth; ++level)
+    print_types[print_types_depth++] = types[level];
+  if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
+    // Force in the core level for quick topology
+    if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
+      // Force core before thread e.g., 1 socket X 2 threads/socket
+      // becomes 1 socket X 1 core/socket X 2 threads/socket
+      print_types[print_types_depth - 1] = KMP_HW_CORE;
+      print_types[print_types_depth++] = KMP_HW_THREAD;
+    } else {
+      print_types[print_types_depth++] = KMP_HW_CORE;
+  // Always put threads at very end of quick topology
+  if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
+    print_types[print_types_depth++] = KMP_HW_THREAD;
+
+  __kmp_str_buf_clear(&buf);
+  kmp_hw_t numerator_type;
+  kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
+  int core_level = get_level(KMP_HW_CORE);
+  int ncores = get_count(core_level);
+
+  for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
+    int c;
+    bool plural;
+    numerator_type = print_types[plevel];
+    KMP_ASSERT_VALID_HW_TYPE(numerator_type);
+    if (equivalent[numerator_type] != numerator_type)
+      c = 1;
+    else
+      c = get_ratio(level++);
+    plural = (c > 1);
+    if (plevel == 0) {
+      __kmp_str_buf_print(&buf, "%d %s", c,
+                          __kmp_hw_get_catalog_string(numerator_type, plural));
+    } else {
+      __kmp_str_buf_print(&buf, " x %d %s/%s", c,
+                          __kmp_hw_get_catalog_string(numerator_type, plural),
+                          __kmp_hw_get_catalog_string(denominator_type));
+    }
+    denominator_type = numerator_type;
+  }
+  KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
+
+  if (num_hw_threads <= 0) {
     __kmp_str_buf_free(&buf);
+    return;
+  }
+
+  // Full OS proc to hardware thread map
+  KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
+  for (int i = 0; i < num_hw_threads; i++) {
+    __kmp_str_buf_clear(&buf);
+    for (int level = 0; level < depth; ++level) {
+      kmp_hw_t type = types[level];
+      __kmp_str_buf_print(&buf, "%s ", __kmp_hw_get_catalog_string(type));
+      __kmp_str_buf_print(&buf, "%d ", hw_threads[i].ids[level]);
+    }
+    KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
+  }
+
+  __kmp_str_buf_free(&buf);
+}
+
+void kmp_topology_t::canonicalize() {
+  _remove_radix1_layers();
+  _gather_enumeration_information();
+  _discover_uniformity();
+  _set_sub_ids();
+  _set_globals();
+  _set_last_level_cache();
+
+  // Perform post canonicalization checking
+  KMP_ASSERT(depth > 0);
+  for (int level = 0; level < depth; ++level) {
+    // All counts, ratios, and types must be valid
+    KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
+    KMP_ASSERT_VALID_HW_TYPE(types[level]);
+    // Detected types must point to themselves
+    KMP_ASSERT(equivalent[types[level]] == types[level]);
+  }
+
+#if KMP_AFFINITY_SUPPORTED
+  // Set the number of affinity granularity levels
+  if (__kmp_affinity_gran_levels < 0) {
+    kmp_hw_t gran_type = get_equivalent_type(__kmp_affinity_gran);
+    // Check if user's granularity request is valid
+    if (gran_type == KMP_HW_UNKNOWN) {
+      // First try core, then thread, then package
+      kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
+      for (auto g : gran_types) {
+        if (__kmp_topology->get_equivalent_type(g) != KMP_HW_UNKNOWN) {
+          gran_type = g;
+          break;
+        }
       }
+      KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
+      // Warn user what granularity setting will be used instead
+      KMP_WARNING(AffGranularityBad, "KMP_AFFINITY",
+                  __kmp_hw_get_catalog_string(__kmp_affinity_gran),
+                  __kmp_hw_get_catalog_string(gran_type));
+      __kmp_affinity_gran = gran_type;
+    }
+    __kmp_affinity_gran_levels = 0;
+    for (int i = depth - 1; i >= 0 && get_type(i) != gran_type; --i)
+      __kmp_affinity_gran_levels++;
+  }
+#endif // KMP_AFFINITY_SUPPORTED
 }
 
+// Canonicalize an explicit packages X cores/pkg X threads/core topology
+void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
+                                  int nthreads_per_core, int ncores) {
+  int ndepth = 3;
+  depth = ndepth;
+  KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
+  for (int level = 0; level < depth; ++level) {
+    count[level] = 0;
+    ratio[level] = 0;
+  }
+  count[0] = npackages;
+  count[1] = ncores;
+  count[2] = __kmp_xproc;
+  ratio[0] = npackages;
+  ratio[1] = ncores_per_pkg;
+  ratio[2] = nthreads_per_core;
+  equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
+  equivalent[KMP_HW_CORE] = KMP_HW_CORE;
+  equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
+  types[0] = KMP_HW_SOCKET;
+  types[1] = KMP_HW_CORE;
+  types[2] = KMP_HW_THREAD;
+  //__kmp_avail_proc = __kmp_xproc;
+  _discover_uniformity();
+}
+
+// Apply the KMP_HW_SUBSET envirable to the topology
+// Returns true if KMP_HW_SUBSET filtered any processors
+// otherwise, returns false
+bool kmp_topology_t::filter_hw_subset() {
+  // If KMP_HW_SUBSET wasn't requested, then do nothing.
+  if (!__kmp_hw_subset)
+    return false;
+
+  // Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
+  int hw_subset_depth = __kmp_hw_subset->get_depth();
+  kmp_hw_t specified[KMP_HW_LAST];
+  KMP_ASSERT(hw_subset_depth > 0);
+  KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
+  for (int i = 0; i < hw_subset_depth; ++i) {
+    int max_count;
+    int num = __kmp_hw_subset->at(i).num;
+    int offset = __kmp_hw_subset->at(i).offset;
+    kmp_hw_t type = __kmp_hw_subset->at(i).type;
+    kmp_hw_t equivalent_type = equivalent[type];
+    int level = get_level(type);
+
+    // Check to see if current layer is in detected machine topology
+    if (equivalent_type != KMP_HW_UNKNOWN) {
+      __kmp_hw_subset->at(i).type = equivalent_type;
+    } else {
+      KMP_WARNING(AffHWSubsetNotExistGeneric,
+                  __kmp_hw_get_catalog_string(type));
+      return false;
+    }
+
+    // Check to see if current layer has already been specified
+    // either directly or through an equivalent type
+    if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
+      KMP_WARNING(AffHWSubsetEqvLayers, __kmp_hw_get_catalog_string(type),
+                  __kmp_hw_get_catalog_string(specified[equivalent_type]));
+      return false;
+    }
+    specified[equivalent_type] = type;
+
+    // Check to see if layers are in order
+    if (i + 1 < hw_subset_depth) {
+      kmp_hw_t next_type = get_equivalent_type(__kmp_hw_subset->at(i + 1).type);
+      if (next_type == KMP_HW_UNKNOWN) {
+        KMP_WARNING(
+            AffHWSubsetNotExistGeneric,
+            __kmp_hw_get_catalog_string(__kmp_hw_subset->at(i + 1).type));
+        return false;
+      }
+      int next_topology_level = get_level(next_type);
+      if (level > next_topology_level) {
+        KMP_WARNING(AffHWSubsetOutOfOrder, __kmp_hw_get_catalog_string(type),
+                    __kmp_hw_get_catalog_string(next_type));
+        return false;
+      }
+    }
+
+    // Check to see if each layer's num & offset parameters are valid
+    max_count = get_ratio(level);
+    if (max_count < 0 || num + offset > max_count) {
+      bool plural = (num > 1);
+      KMP_WARNING(AffHWSubsetManyGeneric,
+                  __kmp_hw_get_catalog_string(type, plural));
+      return false;
+    }
+  }
+
+  // Apply the filtered hardware subset
+  int new_index = 0;
+  for (int i = 0; i < num_hw_threads; ++i) {
+    kmp_hw_thread_t &hw_thread = hw_threads[i];
+    // Check to see if this hardware thread should be filtered
+    bool should_be_filtered = false;
+    for (int level = 0, hw_subset_index = 0;
+         level < depth && hw_subset_index < hw_subset_depth; ++level) {
+      kmp_hw_t topology_type = types[level];
+      auto hw_subset_item = __kmp_hw_subset->at(hw_subset_index);
+      kmp_hw_t hw_subset_type = hw_subset_item.type;
+      if (topology_type != hw_subset_type)
+        continue;
+      int num = hw_subset_item.num;
+      int offset = hw_subset_item.offset;
+      hw_subset_index++;
+      if (hw_thread.sub_ids[level] < offset ||
+          hw_thread.sub_ids[level] >= offset + num) {
+        should_be_filtered = true;
+        break;
+      }
+    }
+    if (!should_be_filtered) {
+      if (i != new_index)
+        hw_threads[new_index] = hw_thread;
+      new_index++;
+    } else {
+#if KMP_AFFINITY_SUPPORTED
+      KMP_CPU_CLR(hw_thread.os_id, __kmp_affin_fullMask);
+#endif
+      __kmp_avail_proc--;
+    }
+  }
+  KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
+  num_hw_threads = new_index;
+
+  // Post hardware subset canonicalization
+  _gather_enumeration_information();
+  _discover_uniformity();
+  _set_globals();
+  _set_last_level_cache();
+  return true;
+}
+
+bool kmp_topology_t::is_close(int hwt1, int hwt2, int hw_level) const {
+  if (hw_level >= depth)
+    return true;
+  bool retval = true;
+  const kmp_hw_thread_t &t1 = hw_threads[hwt1];
+  const kmp_hw_thread_t &t2 = hw_threads[hwt2];
+  for (int i = 0; i < (depth - hw_level); ++i) {
+    if (t1.ids[i] != t2.ids[i])
+      return false;
+  }
+  return retval;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+#if KMP_AFFINITY_SUPPORTED
+class kmp_affinity_raii_t {
+  kmp_affin_mask_t *mask;
+  bool restored;
+
+public:
+  kmp_affinity_raii_t() : restored(false) {
+    KMP_CPU_ALLOC(mask);
+    KMP_ASSERT(mask != NULL);
+    __kmp_get_system_affinity(mask, TRUE);
+  }
+  void restore() {
+    __kmp_set_system_affinity(mask, TRUE);
+    KMP_CPU_FREE(mask);
+    restored = true;
+  }
+  ~kmp_affinity_raii_t() {
+    if (!restored) {
+      __kmp_set_system_affinity(mask, TRUE);
+      KMP_CPU_FREE(mask);
+    }
+  }
+};
+
 bool KMPAffinity::picked_api = false;
 
 void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
 void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
 void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
 void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
 void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
 void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
#endif /* KMP_GROUP_AFFINITY */
   {
     int proc;
     for (proc = 0; proc < __kmp_xproc; proc++) {
       KMP_CPU_SET(proc, mask);
     }
   }
 }
 
-// When sorting by labels, __kmp_affinity_assign_child_nums() must first be
-// called to renumber the labels from [0..n] and place them into the child_num
-// vector of the address object.  This is done in case the labels used for
-// the children at one node of the hierarchy differ from those used for
-// another node at the same level.  Example:  suppose the machine has 2 nodes
-// with 2 packages each.  The first node contains packages 601 and 602, and
-// second node contains packages 603 and 604.  If we try to sort the table
-// for "scatter" affinity, the table will still be sorted 601, 602, 603, 604
-// because we are paying attention to the labels themselves, not the ordinal
-// child numbers.  By using the child numbers in the sort, the result is
-// {0,0}=601, {0,1}=603, {1,0}=602, {1,1}=604.
-static void __kmp_affinity_assign_child_nums(AddrUnsPair *address2os,
-                                             int numAddrs) {
-  KMP_DEBUG_ASSERT(numAddrs > 0);
-  int depth = address2os->first.depth;
-  unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
-  unsigned *lastLabel = (unsigned *)__kmp_allocate(depth * sizeof(unsigned));
-  int labCt;
-  for (labCt = 0; labCt < depth; labCt++) {
-    address2os[0].first.childNums[labCt] = counts[labCt] = 0;
-    lastLabel[labCt] = address2os[0].first.labels[labCt];
-  }
-  int i;
-  for (i = 1; i < numAddrs; i++) {
-    for (labCt = 0; labCt < depth; labCt++) {
-      if (address2os[i].first.labels[labCt] != lastLabel[labCt]) {
-        int labCt2;
-        for (labCt2 = labCt + 1; labCt2 < depth; labCt2++) {
-          counts[labCt2] = 0;
-          lastLabel[labCt2] = address2os[i].first.labels[labCt2];
-        }
-        counts[labCt]++;
-        lastLabel[labCt] = address2os[i].first.labels[labCt];
-        break;
-      }
-    }
-    for (labCt = 0; labCt < depth; labCt++) {
-      address2os[i].first.childNums[labCt] = counts[labCt];
-    }
-    for (; labCt < (int)Address::maxDepth; labCt++) {
-      address2os[i].first.childNums[labCt] = 0;
-    }
-  }
-  __kmp_free(lastLabel);
-  __kmp_free(counts);
-}
-
-// All of the __kmp_affinity_create_*_map() routines should set
-// __kmp_affinity_masks to a vector of affinity mask objects of length
-// __kmp_affinity_num_masks, if __kmp_affinity_type != affinity_none, and return
-// the number of levels in the machine topology tree (zero if
-// __kmp_affinity_type == affinity_none).
-//
-// All of the __kmp_affinity_create_*_map() routines should set
-// *__kmp_affin_fullMask to the affinity mask for the initialization thread.
-// They need to save and restore the mask, and it could be needed later, so
-// saving it is just an optimization to avoid calling kmp_get_system_affinity()
-// again.
+// All of the __kmp_affinity_create_*_map() routines should allocate the
+// internal topology object and set the layer ids for it.  Each routine
+// returns a boolean on whether it was successful at doing so.
 kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
 
-static int nCoresPerPkg, nPackages;
-static int __kmp_nThreadsPerCore;
-#ifndef KMP_DFLT_NTH_CORES
-static int __kmp_ncores;
-#endif
-static int *__kmp_pu_os_idx = NULL;
-static int nDiesPerPkg = 1;
-
-// __kmp_affinity_uniform_topology() doesn't work when called from
-// places which support arbitrarily many levels in the machine topology
-// map, i.e. the non-default cases in __kmp_affinity_create_cpuinfo_map()
-// __kmp_affinity_create_x2apicid_map().
-inline static bool __kmp_affinity_uniform_topology() {
-  return __kmp_avail_proc ==
-         (__kmp_nThreadsPerCore * nCoresPerPkg * nDiesPerPkg * nPackages);
-}
-
 #if KMP_USE_HWLOC
-
 static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
 #if HWLOC_API_VERSION >= 0x00020000
   return hwloc_obj_type_is_cache(obj->type);
 #else
   return obj->type == HWLOC_OBJ_CACHE;
 #endif
 }
 
case HWLOC_OBJ_NUMANODE:
     return KMP_HW_NUMA;
   case HWLOC_OBJ_CORE:
     return KMP_HW_CORE;
   case HWLOC_OBJ_PU:
     return KMP_HW_THREAD;
   case HWLOC_OBJ_GROUP:
     if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
       return KMP_HW_DIE;
+    else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
+      return KMP_HW_TILE;
+    else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
+      return KMP_HW_MODULE;
+    else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
+      return KMP_HW_PROC_GROUP;
+    return KMP_HW_UNKNOWN;
 #if HWLOC_API_VERSION >= 0x00020100
   case HWLOC_OBJ_DIE:
     return KMP_HW_DIE;
 #endif
   }
   return KMP_HW_UNKNOWN;
 }
 
for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
                                                        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,
-                                               kmp_hwloc_depth_t 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)
-}
-
 // This gets the sub_id for a lower object under a higher object in the
 // topology tree
 static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
                                   hwloc_obj_t lower) {
   hwloc_obj_t obj;
   hwloc_obj_type_t ltype = lower->type;
   int lindex = lower->logical_index - 1;
   int sub_id = 0;
while (obj && lindex >= 0 &&
     lindex--;
     obj = hwloc_get_obj_by_type(t, ltype, lindex);
   }
   // store sub_id + 1 so that 0 is differed from NULL
   lower->userdata = RCAST(void *, sub_id + 1);
   return sub_id;
 }
 
-static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os,
-                                           kmp_i18n_id_t *const msg_id) {
+static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
   kmp_hw_t type;
-  int hw_thread_index, sub_id, nActiveThreads;
+  int hw_thread_index, sub_id;
   int depth;
   hwloc_obj_t pu, obj, root, prev;
-  int ratio[KMP_HW_LAST];
-  int count[KMP_HW_LAST];
   kmp_hw_t types[KMP_HW_LAST];
+  hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
 
   hwloc_topology_t tp = __kmp_hwloc_topology;
   *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);
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
+  }
 
   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.
+    // available from hwloc on the current thread, and __kmp_xproc.
     KMP_ASSERT(__kmp_affinity_type == affinity_none);
     // hwloc only guarantees existance of PU object, so check PACKAGE and CORE
     hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
     if (o != NULL)
       nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
     else
       nCoresPerPkg = 1; // no PACKAGE found
     o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
     if (o != NULL)
       __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
     else
       __kmp_nThreadsPerCore = 1; // no CORE found
     __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
     if (nCoresPerPkg == 0)
       nCoresPerPkg = 1; // to prevent possible division by 0
     nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffNotUsingHwloc, "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;
+    return true;
   }
 
   root = hwloc_get_root_obj(tp);
 
   // Figure out the depth and types in the topology
   depth = 0;
   pu = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
+  KMP_ASSERT(pu);
   obj = pu;
   types[depth] = KMP_HW_THREAD;
+  hwloc_types[depth] = obj->type;
   depth++;
   while (obj != root && obj != NULL) {
     obj = obj->parent;
 #if HWLOC_API_VERSION >= 0x00020000
     if (obj->memory_arity) {
       hwloc_obj_t memory;
       for (memory = obj->memory_first_child; memory;
            memory = hwloc_get_next_child(tp, obj, memory)) {
         if (memory->type == HWLOC_OBJ_NUMANODE)
           break;
       }
       if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
         types[depth] = KMP_HW_NUMA;
+        hwloc_types[depth] = memory->type;
         depth++;
       }
     }
 #endif
     type = __kmp_hwloc_type_2_topology_type(obj);
     if (type != KMP_HW_UNKNOWN) {
       types[depth] = type;
+      hwloc_types[depth] = obj->type;
       depth++;
     }
   }
-  KMP_ASSERT(depth > 0 && depth <= KMP_HW_LAST);
+  KMP_ASSERT(depth > 0);
 
   // Get the order for the types correct
   for (int i = 0, j = depth - 1; i < j; ++i, --j) {
+    hwloc_obj_type_t hwloc_temp = hwloc_types[i];
     kmp_hw_t temp = types[i];
     types[i] = types[j];
     types[j] = temp;
+    hwloc_types[i] = hwloc_types[j];
+    hwloc_types[j] = hwloc_temp;
   }
 
   // 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);
+  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
 
   hw_thread_index = 0;
   pu = NULL;
-  nActiveThreads = 0;
   while (pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu)) {
     int index = depth - 1;
     bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
-    Address hw_thread(depth);
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
     if (included) {
-      hw_thread.labels[index] = pu->logical_index;
-      __kmp_pu_os_idx[hw_thread_index] = pu->os_index;
+      hw_thread.clear();
+      hw_thread.ids[index] = pu->logical_index;
+      hw_thread.os_id = pu->os_index;
       index--;
-      nActiveThreads++;
     }
     obj = pu;
     prev = obj;
     while (obj != root && obj != NULL) {
       obj = obj->parent;
 #if HWLOC_API_VERSION >= 0x00020000
       // NUMA Nodes are handled differently since they are not within the
       // parent/child structure anymore.  They are separate children
       // of obj (memory_first_child points to first memory child)
       if (obj->memory_arity) {
         hwloc_obj_t memory;
         for (memory = obj->memory_first_child; memory;
              memory = hwloc_get_next_child(tp, obj, memory)) {
           if (memory->type == HWLOC_OBJ_NUMANODE)
             break;
         }
         if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
           sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
           if (included) {
-            hw_thread.labels[index] = memory->logical_index;
-            hw_thread.labels[index + 1] = sub_id;
+            hw_thread.ids[index] = memory->logical_index;
+            hw_thread.ids[index + 1] = sub_id;
             index--;
           }
           prev = memory;
         }
+        prev = obj;
       }
 #endif
       type = __kmp_hwloc_type_2_topology_type(obj);
       if (type != KMP_HW_UNKNOWN) {
         sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
         if (included) {
-          hw_thread.labels[index] = obj->logical_index;
-          hw_thread.labels[index + 1] = sub_id;
+          hw_thread.ids[index] = obj->logical_index;
+          hw_thread.ids[index + 1] = sub_id;
           index--;
         }
         prev = obj;
       }
     }
-    if (included) {
-      retval[hw_thread_index] = AddrUnsPair(hw_thread, pu->os_index);
+    if (included)
       hw_thread_index++;
-    }
-  }
-
-  // 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;
-    __kmp_nThreadsPerCore = nCoresPerPkg = 1;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-      KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-                 __kmp_nThreadsPerCore, __kmp_ncores);
-    }
-
-    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[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);
-
-  // 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,
-                                                 types);
-
-  __kmp_affinity_gather_enumeration_information(retval, nActiveThreads, depth,
-                                                types, ratio, count);
-
-  for (int level = 0; level < depth; ++level) {
-    if ((types[level] == KMP_HW_L2 || types[level] == KMP_HW_L3))
-      __kmp_tile_depth = level;
-  }
-
-  // This routine should set __kmp_ncores, as well as
-  // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
-  int thread_level, core_level, tile_level, numa_level, socket_level;
-  thread_level = core_level = tile_level = numa_level = socket_level = -1;
-  for (int level = 0; level < depth; ++level) {
-    if (types[level] == KMP_HW_THREAD)
-      thread_level = level;
-    else if (types[level] == KMP_HW_CORE)
-      core_level = level;
-    else if (types[level] == KMP_HW_SOCKET)
-      socket_level = level;
-    else if (types[level] == KMP_HW_TILE)
-      tile_level = level;
-    else if (types[level] == KMP_HW_NUMA)
-      numa_level = level;
-  }
-  __kmp_nThreadsPerCore =
-      __kmp_affinity_calculate_ratio(ratio, thread_level, core_level);
-  nCoresPerPkg =
-      __kmp_affinity_calculate_ratio(ratio, core_level, socket_level);
-  if (socket_level >= 0)
-    nPackages = count[socket_level];
-  else
-    nPackages = 1;
-  if (core_level >= 0)
-    __kmp_ncores = count[core_level];
-  else
-    __kmp_ncores = 1;
-
-  unsigned uniform = __kmp_affinity_discover_uniformity(depth, ratio, count);
-
-  // Print the machine topology summary.
-  if (__kmp_affinity_verbose) {
-    kmp_hw_t numerator_type, denominator_type;
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-    KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-    if (uniform) {
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-    } else {
-      KMP_INFORM(NonUniform, "KMP_AFFINITY");
-    }
-
-    __kmp_str_buf_clear(&buf);
-
-    if (core_level < 0)
-      core_level = depth - 1;
-    int ncores = count[core_level];
-
-    denominator_type = KMP_HW_UNKNOWN;
-    for (int level = 0; level < depth; ++level) {
-      int c;
-      bool plural;
-      numerator_type = types[level];
-      c = ratio[level];
-      plural = (c > 1);
-      if (level == 0) {
-        __kmp_str_buf_print(
-            &buf, "%d %s", c,
-            __kmp_hw_get_catalog_string(numerator_type, plural));
-      } else {
-        __kmp_str_buf_print(&buf, " x %d %s/%s", c,
-                            __kmp_hw_get_catalog_string(numerator_type, plural),
-                            __kmp_hw_get_catalog_string(denominator_type));
-      }
-      denominator_type = numerator_type;
-    }
-    KMP_INFORM(TopologyGeneric, "KMP_AFFINITY", buf.str, ncores);
-    __kmp_str_buf_free(&buf);
-  }
-
-  if (__kmp_affinity_type == affinity_none) {
-    __kmp_free(retval);
-    KMP_CPU_FREE(oldMask);
-    return 0;
-  }
-
-  // Set the granularity level based on what levels are modeled
-  // in the machine topology map.
-  if (__kmp_affinity_gran == affinity_gran_node)
-    __kmp_affinity_gran = affinity_gran_numa;
-  KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default);
-  if (__kmp_affinity_gran_levels < 0) {
-    __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine)
-    if ((thread_level >= 0) && (__kmp_affinity_gran > affinity_gran_thread))
-      __kmp_affinity_gran_levels++;
-    if ((core_level >= 0) && (__kmp_affinity_gran > affinity_gran_core))
-      __kmp_affinity_gran_levels++;
-    if ((tile_level >= 0) && (__kmp_affinity_gran > affinity_gran_tile))
-      __kmp_affinity_gran_levels++;
-    if ((numa_level >= 0) && (__kmp_affinity_gran > affinity_gran_numa))
-      __kmp_affinity_gran_levels++;
-    if ((socket_level >= 0) && (__kmp_affinity_gran > affinity_gran_package))
-      __kmp_affinity_gran_levels++;
-  }
-
-  if (__kmp_affinity_verbose)
-    __kmp_affinity_print_topology(retval, nActiveThreads, depth, types);
-
-  KMP_CPU_FREE(oldMask);
-  *address2os = retval;
-  return depth;
+  __kmp_topology->sort_ids();
+  return true;
 }
 #endif // KMP_USE_HWLOC
 
 // If we don't know how to retrieve the machine's processor topology, or
 // encounter an error in doing so, this routine is called to form a "flat"
 // mapping of os thread id's <-> processor id's.
-static int __kmp_affinity_create_flat_map(AddrUnsPair **address2os,
-                                          kmp_i18n_id_t *const msg_id) {
-  *address2os = NULL;
+static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
   *msg_id = kmp_i18n_null;
+  int depth = 3;
+  kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
+
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
+  }
 
   // Even if __kmp_affinity_type == affinity_none, this routine might still
   // called to set __kmp_ncores, as well as
   // __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
   if (!KMP_AFFINITY_CAPABLE()) {
     KMP_ASSERT(__kmp_affinity_type == affinity_none);
     __kmp_ncores = nPackages = __kmp_xproc;
     __kmp_nThreadsPerCore = nCoresPerPkg = 1;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffFlatTopology, "KMP_AFFINITY");
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-      KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-                 __kmp_nThreadsPerCore, __kmp_ncores);
-    }
-    return 0;
+    return true;
   }
 
   // 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 now if affinity is
   // not enabled.
   __kmp_ncores = nPackages = __kmp_avail_proc;
   __kmp_nThreadsPerCore = nCoresPerPkg = 1;
-  if (__kmp_affinity_verbose) {
-    KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY");
-    KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-    KMP_INFORM(Uniform, "KMP_AFFINITY");
-    KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-               __kmp_nThreadsPerCore, __kmp_ncores);
-  }
-  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
-  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
-  if (__kmp_affinity_type == affinity_none) {
-    int avail_ct = 0;
-    int i;
-    KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
-      if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask))
-        continue;
-      __kmp_pu_os_idx[avail_ct++] = i; // suppose indices are flat
-    }
-    return 0;
-  }
 
   // Construct the data structure to be returned.
-  *address2os =
-      (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
+  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
   int avail_ct = 0;
   int i;
   KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
     // Skip this proc if it is not included in the machine model.
     if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
       continue;
     }
-    __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
-    Address addr(1);
-    addr.labels[0] = i;
-    (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct);
+    hw_thread.clear();
+    hw_thread.os_id = i;
+    hw_thread.ids[0] = i;
+    hw_thread.ids[1] = 0;
+    hw_thread.ids[2] = 0;
+    avail_ct++;
   }
   if (__kmp_affinity_verbose) {
     KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
   }
-
-  if (__kmp_affinity_gran_levels < 0) {
-    // Only the package level is modeled in the machine topology map,
-    // so the #levels of granularity is either 0 or 1.
-    if (__kmp_affinity_gran > affinity_gran_package) {
-      __kmp_affinity_gran_levels = 1;
-    } else {
-      __kmp_affinity_gran_levels = 0;
-    }
-  }
-  return 1;
+  return true;
 }
 
 #if KMP_GROUP_AFFINITY
-
 // If multiple Windows* OS processor groups exist, we can create a 2-level
 // topology map with the groups at level 0 and the individual procs at level 1.
 // This facilitates letting the threads float among all procs in a group,
 // if granularity=group (the default when there are multiple groups).
-static int __kmp_affinity_create_proc_group_map(AddrUnsPair **address2os,
-                                                kmp_i18n_id_t *const msg_id) {
-  *address2os = NULL;
+static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
   *msg_id = kmp_i18n_null;
+  int depth = 3;
+  kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
+  const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
 
-  // If we aren't affinity capable, then return now.
-  // The flat mapping will be used.
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
+  }
+
+  // If we aren't affinity capable, then use flat topology
   if (!KMP_AFFINITY_CAPABLE()) {
-    // FIXME set *msg_id
-    return -1;
+    KMP_ASSERT(__kmp_affinity_type == affinity_none);
+    nPackages = __kmp_num_proc_groups;
+    __kmp_nThreadsPerCore = 1;
+    __kmp_ncores = __kmp_xproc;
+    nCoresPerPkg = nPackages / __kmp_ncores;
+    return true;
   }
 
   // Construct the data structure to be returned.
-  *address2os =
-      (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc);
-  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
-  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
+  __kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
   int avail_ct = 0;
   int i;
   KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
     // Skip this proc if it is not included in the machine model.
     if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
       continue;
     }
-    __kmp_pu_os_idx[avail_ct] = i; // suppose indices are flat
-    Address addr(2);
-    addr.labels[0] = i / (CHAR_BIT * sizeof(DWORD_PTR));
-    addr.labels[1] = i % (CHAR_BIT * sizeof(DWORD_PTR));
-    (*address2os)[avail_ct++] = AddrUnsPair(addr, i);
-
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffOSProcToGroup, "KMP_AFFINITY", i, addr.labels[0],
-                 addr.labels[1]);
-    }
-  }
-
-  if (__kmp_affinity_gran_levels < 0) {
-    if (__kmp_affinity_gran == affinity_gran_group) {
-      __kmp_affinity_gran_levels = 1;
-    } else if ((__kmp_affinity_gran == affinity_gran_fine) ||
-               (__kmp_affinity_gran == affinity_gran_thread)) {
-      __kmp_affinity_gran_levels = 0;
-    } else {
-      const char *gran_str = NULL;
-      if (__kmp_affinity_gran == affinity_gran_core) {
-        gran_str = "core";
-      } else if (__kmp_affinity_gran == affinity_gran_package) {
-        gran_str = "package";
-      } else if (__kmp_affinity_gran == affinity_gran_node) {
-        gran_str = "node";
-      } else {
-        KMP_ASSERT(0);
-      }
-
-      // Warning: can't use affinity granularity \"gran\" with group topology
-      // method, using "thread"
-      __kmp_affinity_gran_levels = 0;
-    }
-  }
-  return 2;
-}
-
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
+    hw_thread.clear();
+    hw_thread.os_id = i;
+    hw_thread.ids[0] = i / BITS_PER_GROUP;
+    hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
+  return true;
 #endif /* KMP_GROUP_AFFINITY */
 
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
 
-/*
- * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
-    Bits            Bits            Bits           Bits
-    31-16           15-8            7-4            4-0
----+-----------+--------------+-------------+-----------------+
-EAX| reserved  |   reserved   |   reserved  |  Bits to Shift  |
----+-----------|--------------+-------------+-----------------|
-EBX| reserved  | Num logical processors at level (16 bits)    |
----+-----------|--------------+-------------------------------|
-ECX| reserved  |   Level Type |      Level Number (8 bits)    |
----+-----------+--------------+-------------------------------|
-EDX|                    X2APIC ID (32 bits)                   |
----+----------------------------------------------------------+
-*/
-
-enum {
-  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
-  INTEL_LEVEL_TYPE_SMT = 1,
-  INTEL_LEVEL_TYPE_CORE = 2,
-  INTEL_LEVEL_TYPE_TILE = 3,
-  INTEL_LEVEL_TYPE_MODULE = 4,
-  INTEL_LEVEL_TYPE_DIE = 5,
-  INTEL_LEVEL_TYPE_LAST = 6,
-};
-
-struct cpuid_level_info_t {
-  unsigned level_type, mask, mask_width, nitems, cache_mask;
-};
-
 template <kmp_uint32 LSB, kmp_uint32 MSB>
 static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
   const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
   const kmp_uint32 SHIFT_RIGHT = LSB;
   kmp_uint32 retval = v;
   retval <<= SHIFT_LEFT;
   retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
   return retval;
 }
 
-static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
-  switch (intel_type) {
-  case INTEL_LEVEL_TYPE_INVALID:
-    return KMP_HW_SOCKET;
-  case INTEL_LEVEL_TYPE_SMT:
-    return KMP_HW_THREAD;
-  case INTEL_LEVEL_TYPE_CORE:
-    return KMP_HW_CORE;
-  // TODO: add support for the tile and module
-  case INTEL_LEVEL_TYPE_TILE:
-    return KMP_HW_UNKNOWN;
-  case INTEL_LEVEL_TYPE_MODULE:
-    return KMP_HW_UNKNOWN;
-  case INTEL_LEVEL_TYPE_DIE:
-    return KMP_HW_DIE;
-  }
-  return KMP_HW_UNKNOWN;
-}
-
-// This function takes the topology leaf, a levels array to store the levels
-// detected and a bitmap of the known levels.
-// Returns the number of levels in the topology
-static unsigned
-__kmp_x2apicid_get_levels(int leaf,
-                          cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
-                          kmp_uint64 known_levels) {
-  unsigned level, levels_index;
-  unsigned level_type, mask_width, nitems;
-  kmp_cpuid buf;
-
-  // The new algorithm has known topology layers act as highest unknown topology
-  // layers when unknown topology layers exist.
-  // e.g., Suppose layers were SMT CORE <Y> <Z> PACKAGE
-  // Then CORE will take the characteristics (nitems and mask width) of <Z>.
-  // In developing the id mask for each layer, this eliminates unknown portions
-  // of the topology while still keeping the correct underlying structure.
-  level = levels_index = 0;
-  do {
-    __kmp_x86_cpuid(leaf, level, &buf);
-    level_type = __kmp_extract_bits<8, 15>(buf.ecx);
-    mask_width = __kmp_extract_bits<0, 4>(buf.eax);
-    nitems = __kmp_extract_bits<0, 15>(buf.ebx);
-    if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
-      return 0;
-
-    if (known_levels & (1ull << level_type)) {
-      // Add a new level to the topology
-      KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
-      levels[levels_index].level_type = level_type;
-      levels[levels_index].mask_width = mask_width;
-      levels[levels_index].nitems = nitems;
-      levels_index++;
-    } else {
-      // If it is an unknown level, then logically move the previous layer up
-      if (levels_index > 0) {
-        levels[levels_index - 1].mask_width = mask_width;
-        levels[levels_index - 1].nitems = nitems;
-      }
-    }
-    level++;
-  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
-
-  // Set the masks to & with apicid
-  for (unsigned i = 0; i < levels_index; ++i) {
-    if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
-      levels[i].mask = ~((-1) << levels[i].mask_width);
-      levels[i].cache_mask = (-1) << levels[i].mask_width;
-      for (unsigned j = 0; j < i; ++j)
-        levels[i].mask ^= levels[j].mask;
-    } else {
-      KMP_DEBUG_ASSERT(levels_index > 0);
-      levels[i].mask = (-1) << levels[i - 1].mask_width;
-      levels[i].cache_mask = 0;
-    }
-  }
-  return levels_index;
-}
-
 static int __kmp_cpuid_mask_width(int count) {
   int r = 0;
 
   while ((1 << r) < count)
     ++r;
   return r;
 }
 
if (aa->coreId > bb->coreId)
     return 1;
   if (aa->threadId < bb->threadId)
     return -1;
   if (aa->threadId > bb->threadId)
     return 1;
   return 0;
 }
 
+class kmp_cache_info_t {
+public:
+  struct info_t {
+    unsigned level, mask;
+  };
+  kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
+  size_t get_depth() const { return depth; }
+  info_t &operator[](size_t index) { return table[index]; }
+  const info_t &operator[](size_t index) const { return table[index]; }
+
+  static kmp_hw_t get_topology_type(unsigned level) {
+    KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
+    switch (level) {
+    case 1:
+      return KMP_HW_L1;
+    case 2:
+      return KMP_HW_L2;
+    case 3:
+      return KMP_HW_L3;
+    }
+    return KMP_HW_UNKNOWN;
+  }
+
+private:
+  static const int MAX_CACHE_LEVEL = 3;
+
+  size_t depth;
+  info_t table[MAX_CACHE_LEVEL];
+
+  void get_leaf4_levels() {
+    unsigned level = 0;
+    while (depth < MAX_CACHE_LEVEL) {
+      unsigned cache_type, max_threads_sharing;
+      unsigned cache_level, cache_mask_width;
+      kmp_cpuid buf2;
+      __kmp_x86_cpuid(4, level, &buf2);
+      cache_type = __kmp_extract_bits<0, 4>(buf2.eax);
+      if (!cache_type)
+        break;
+      // Skip instruction caches
+      if (cache_type == 2) {
+        level++;
+        continue;
+      }
+      max_threads_sharing = __kmp_extract_bits<14, 25>(buf2.eax) + 1;
+      cache_mask_width = __kmp_cpuid_mask_width(max_threads_sharing);
+      cache_level = __kmp_extract_bits<5, 7>(buf2.eax);
+      table[depth].level = cache_level;
+      table[depth].mask = ((-1) << cache_mask_width);
+      depth++;
+      level++;
+    }
+  }
+};
+
 // On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
 // an algorithm which cycles through the available os threads, setting
 // the current thread's affinity mask to that thread, and then retrieves
 // the Apic Id for each thread context using the cpuid instruction.
-static int __kmp_affinity_create_apicid_map(AddrUnsPair **address2os,
-                                            kmp_i18n_id_t *const msg_id) {
+static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
   kmp_cpuid buf;
-  *address2os = NULL;
   *msg_id = kmp_i18n_null;
 
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
+  }
+
   // Check if cpuid leaf 4 is supported.
   __kmp_x86_cpuid(0, 0, &buf);
   if (buf.eax < 4) {
     *msg_id = kmp_i18n_str_NoLeaf4Support;
-    return -1;
+    return false;
   }
 
   // The algorithm used starts by setting the affinity to each available thread
   // and retrieving info from the cpuid instruction, so if we are not capable of
   // calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
   // need to do something else - use the defaults that we calculated from
   // issuing cpuid without binding to each proc.
   if (!KMP_AFFINITY_CAPABLE()) {
if (!KMP_AFFINITY_CAPABLE()) {
     // - The performance penalty for mistakenly identifying a machine as HT when
     //   it isn't (which results in blocktime being incorrectly set to 0) is
     //   greater than the penalty when for mistakenly identifying a machine as
     //   being 1 thread/core when it is really HT enabled (which results in
     //   blocktime being incorrectly set to a positive value).
     __kmp_ncores = __kmp_xproc;
     nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
     __kmp_nThreadsPerCore = 1;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffNotCapableUseLocCpuid, "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);
-    }
-    return 0;
+    return true;
   }
 
   // From here on, we can assume that it is safe to call
   // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
   // __kmp_affinity_type = affinity_none.
 
   // Save the affinity mask for the current thread.
-  kmp_affin_mask_t *oldMask;
-  KMP_CPU_ALLOC(oldMask);
-  KMP_ASSERT(oldMask != NULL);
-  __kmp_get_system_affinity(oldMask, TRUE);
+  kmp_affinity_raii_t previous_affinity;
 
   // Run through each of the available contexts, binding the current thread
   // to it, and obtaining the pertinent information using the cpuid instr.
   //
   // The relevant information is:
   // - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
   //     has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
   // - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
     KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
 
     __kmp_affinity_dispatch->bind_thread(i);
     threadInfo[nApics].osId = i;
 
     // The apic id and max threads per pkg come from cpuid(1).
     __kmp_x86_cpuid(1, 0, &buf);
     if (((buf.edx >> 9) & 1) == 0) {
-      __kmp_set_system_affinity(oldMask, TRUE);
       __kmp_free(threadInfo);
-      KMP_CPU_FREE(oldMask);
       *msg_id = kmp_i18n_str_ApicNotPresent;
-      return -1;
+      return false;
     }
     threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
     threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
     if (threadInfo[nApics].maxThreadsPerPkg == 0) {
       threadInfo[nApics].maxThreadsPerPkg = 1;
     }
 
     // Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
     threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
 
     int widthC = __kmp_cpuid_mask_width(threadInfo[nApics].maxCoresPerPkg);
     int widthT = widthCT - widthC;
     if (widthT < 0) {
       // I've never seen this one happen, but I suppose it could, if the cpuid
       // instruction on a chip was really screwed up. Make sure to restore the
       // affinity mask before the tail call.
-      __kmp_set_system_affinity(oldMask, TRUE);
       __kmp_free(threadInfo);
-      KMP_CPU_FREE(oldMask);
       *msg_id = kmp_i18n_str_InvalidCpuidInfo;
-      return -1;
+      return false;
     }
 
     int maskC = (1 << widthC) - 1;
     threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
 
     int maskT = (1 << widthT) - 1;
     threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
 
     nApics++;
   }
 
   // We've collected all the info we need.
   // Restore the old affinity mask for this thread.
-  __kmp_set_system_affinity(oldMask, TRUE);
-
-  // If there's only one thread context to bind to, form an Address object
-  // with depth 1 and return immediately (or, if affinity is off, set
-  // address2os to NULL and return).
-  //
-  // If it is configured to omit the package level when there is only a single
-  // package, the logic at the end of this routine won't work if there is only
-  // a single thread - it would try to form an Address object with depth 0.
-  KMP_ASSERT(nApics > 0);
-  if (nApics == 1) {
-    __kmp_ncores = nPackages = 1;
-    __kmp_nThreadsPerCore = nCoresPerPkg = 1;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY");
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-      KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-                 __kmp_nThreadsPerCore, __kmp_ncores);
-    }
-
-    if (__kmp_affinity_type == affinity_none) {
-      __kmp_free(threadInfo);
-      KMP_CPU_FREE(oldMask);
-      return 0;
-    }
-
-    *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
-    Address addr(1);
-    addr.labels[0] = threadInfo[0].pkgId;
-    (*address2os)[0] = AddrUnsPair(addr, threadInfo[0].osId);
-
-    if (__kmp_affinity_gran_levels < 0) {
-      __kmp_affinity_gran_levels = 0;
-    }
-
-    if (__kmp_affinity_verbose) {
-      __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
-    }
-
-    __kmp_free(threadInfo);
-    KMP_CPU_FREE(oldMask);
-    return 1;
-  }
+  previous_affinity.restore();
 
   // Sort the threadInfo table by physical Id.
   qsort(threadInfo, nApics, sizeof(*threadInfo),
         __kmp_affinity_cmp_apicThreadInfo_phys_id);
 
   // The table is now sorted by pkgId / coreId / threadId, but we really don't
   // know the radix of any of the fields. pkgId's may be sparsely assigned among
   // the chips on a system. Although coreId's are usually assigned
if (threadInfo[i].coreId != lastCoreId) {
         __kmp_nThreadsPerCore = threadCt;
       threadCt = 1;
       lastThreadId = threadInfo[i].threadId;
     } else if (threadInfo[i].threadId != lastThreadId) {
       threadCt++;
       lastThreadId = threadInfo[i].threadId;
     } else {
       __kmp_free(threadInfo);
-      KMP_CPU_FREE(oldMask);
       *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
-      return -1;
+      return false;
     }
 
     // Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
     // fields agree between all the threads bounds to a given package.
     if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
         (prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
       __kmp_free(threadInfo);
-      KMP_CPU_FREE(oldMask);
       *msg_id = kmp_i18n_str_InconsistentCpuidInfo;
-      return -1;
+      return false;
     }
   }
+  // 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
   nPackages = pkgCt;
   if ((int)coreCt > nCoresPerPkg)
     nCoresPerPkg = coreCt;
   if ((int)threadCt > __kmp_nThreadsPerCore)
     __kmp_nThreadsPerCore = threadCt;
-
-  // 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 now if affinity is
-  // not enabled.
   __kmp_ncores = nCores;
-  if (__kmp_affinity_verbose) {
-    KMP_INFORM(AffUseGlobCpuid, "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_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
   KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
-  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
-  for (i = 0; i < nApics; ++i) {
-    __kmp_pu_os_idx[i] = threadInfo[i].osId;
-  }
-  if (__kmp_affinity_type == affinity_none) {
-    __kmp_free(threadInfo);
-    KMP_CPU_FREE(oldMask);
-    return 0;
-  }
 
   // Now that we've determined the number of packages, the number of cores per
   // package, and the number of threads per core, we can construct the data
   // structure that is to be returned.
+  int idx = 0;
   int pkgLevel = 0;
-  int coreLevel = (nCoresPerPkg <= 1) ? -1 : 1;
-  int threadLevel =
-      (__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
-  unsigned depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
+  int coreLevel = 1;
+  int threadLevel = 2;
+  //(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
+  int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
+  kmp_hw_t types[3];
+  if (pkgLevel >= 0)
+    types[idx++] = KMP_HW_SOCKET;
+  if (coreLevel >= 0)
+    types[idx++] = KMP_HW_CORE;
+  if (threadLevel >= 0)
+    types[idx++] = KMP_HW_THREAD;
 
   KMP_ASSERT(depth > 0);
-  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics);
+  __kmp_topology = kmp_topology_t::allocate(nApics, depth, types);
 
   for (i = 0; i < nApics; ++i) {
-    Address addr(depth);
+    idx = 0;
     unsigned os = threadInfo[i].osId;
-    int d = 0;
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
+    hw_thread.clear();
 
     if (pkgLevel >= 0) {
-      addr.labels[d++] = threadInfo[i].pkgId;
+      hw_thread.ids[idx++] = threadInfo[i].pkgId;
     }
     if (coreLevel >= 0) {
-      addr.labels[d++] = threadInfo[i].coreId;
+      hw_thread.ids[idx++] = threadInfo[i].coreId;
     }
     if (threadLevel >= 0) {
-      addr.labels[d++] = threadInfo[i].threadId;
+      hw_thread.ids[idx++] = threadInfo[i].threadId;
     }
-    (*address2os)[i] = AddrUnsPair(addr, os);
+    hw_thread.os_id = os;
   }
 
-  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++;
-    }
-    if ((pkgLevel >= 0) && (__kmp_affinity_gran > affinity_gran_package)) {
-      __kmp_affinity_gran_levels++;
-    }
-  }
+  __kmp_free(threadInfo);
+  __kmp_topology->sort_ids();
+  if (!__kmp_topology->check_ids()) {
+    kmp_topology_t::deallocate(__kmp_topology);
+    __kmp_topology = nullptr;
+    *msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
+    return false;
+  return true;
+
+// Intel(R) microarchitecture code name Nehalem, Dunnington and later
+// architectures support a newer interface for specifying the x2APIC Ids,
+// based on CPUID.B or CPUID.1F
+/*
+ * CPUID.B or 1F, Input ECX (sub leaf # aka level number)
+    Bits            Bits            Bits           Bits
+    31-16           15-8            7-4            4-0
+---+-----------+--------------+-------------+-----------------+
+EAX| reserved  |   reserved   |   reserved  |  Bits to Shift  |
+---+-----------|--------------+-------------+-----------------|
+EBX| reserved  | Num logical processors at level (16 bits)    |
+---+-----------|--------------+-------------------------------|
+ECX| reserved  |   Level Type |      Level Number (8 bits)    |
+---+-----------+--------------+-------------------------------|
+EDX|                    X2APIC ID (32 bits)                   |
+---+----------------------------------------------------------+
+*/
+
+enum {
+  INTEL_LEVEL_TYPE_INVALID = 0, // Package level
+  INTEL_LEVEL_TYPE_SMT = 1,
+  INTEL_LEVEL_TYPE_CORE = 2,
+  INTEL_LEVEL_TYPE_TILE = 3,
+  INTEL_LEVEL_TYPE_MODULE = 4,
+  INTEL_LEVEL_TYPE_DIE = 5,
+  INTEL_LEVEL_TYPE_LAST = 6,
+};
+
+struct cpuid_level_info_t {
+  unsigned level_type, mask, mask_width, nitems, cache_mask;
+};
+
+static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
+  switch (intel_type) {
+  case INTEL_LEVEL_TYPE_INVALID:
+    return KMP_HW_SOCKET;
+  case INTEL_LEVEL_TYPE_SMT:
+    return KMP_HW_THREAD;
+  case INTEL_LEVEL_TYPE_CORE:
+    return KMP_HW_CORE;
+  case INTEL_LEVEL_TYPE_TILE:
+    return KMP_HW_TILE;
+  case INTEL_LEVEL_TYPE_MODULE:
+    return KMP_HW_MODULE;
+  case INTEL_LEVEL_TYPE_DIE:
+    return KMP_HW_DIE;
+  return KMP_HW_UNKNOWN;
 
-  if (__kmp_affinity_verbose) {
-    __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel,
-                                  coreLevel, threadLevel);
+// This function takes the topology leaf, a levels array to store the levels
+// detected and a bitmap of the known levels.
+// Returns the number of levels in the topology
+static unsigned
+__kmp_x2apicid_get_levels(int leaf,
+                          cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
+                          kmp_uint64 known_levels) {
+  unsigned level, levels_index;
+  unsigned level_type, mask_width, nitems;
+  kmp_cpuid buf;
+
+  // New algorithm has known topology layers act as highest unknown topology
+  // layers when unknown topology layers exist.
+  // e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
+  // are unknown topology layers, Then SMT will take the characteristics of
+  // (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
+  // This eliminates unknown portions of the topology while still keeping the
+  // correct structure.
+  level = levels_index = 0;
+  do {
+    __kmp_x86_cpuid(leaf, level, &buf);
+    level_type = __kmp_extract_bits<8, 15>(buf.ecx);
+    mask_width = __kmp_extract_bits<0, 4>(buf.eax);
+    nitems = __kmp_extract_bits<0, 15>(buf.ebx);
+    if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
+      return 0;
+
+    if (known_levels & (1ull << level_type)) {
+      // Add a new level to the topology
+      KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
+      levels[levels_index].level_type = level_type;
+      levels[levels_index].mask_width = mask_width;
+      levels[levels_index].nitems = nitems;
+      levels_index++;
+    } else {
+      // If it is an unknown level, then logically move the previous layer up
+      if (levels_index > 0) {
+        levels[levels_index - 1].mask_width = mask_width;
+        levels[levels_index - 1].nitems = nitems;
       }
+    }
+    level++;
+  } while (level_type != INTEL_LEVEL_TYPE_INVALID);
 
-  __kmp_free(threadInfo);
-  KMP_CPU_FREE(oldMask);
-  return depth;
+  // Set the masks to & with apicid
+  for (unsigned i = 0; i < levels_index; ++i) {
+    if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
+      levels[i].mask = ~((-1) << levels[i].mask_width);
+      levels[i].cache_mask = (-1) << levels[i].mask_width;
+      for (unsigned j = 0; j < i; ++j)
+        levels[i].mask ^= levels[j].mask;
+    } else {
+      KMP_DEBUG_ASSERT(levels_index > 0);
+      levels[i].mask = (-1) << levels[i - 1].mask_width;
+      levels[i].cache_mask = 0;
+    }
+  }
+  return levels_index;
 }
 
-// Intel(R) microarchitecture code name Nehalem, Dunnington and later
-// architectures support a newer interface for specifying the x2APIC Ids,
-// based on CPUID.B or CPUID.1F
-static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os,
-                                              kmp_i18n_id_t *const msg_id) {
+static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
 
   cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
-  int ratio[KMP_HW_LAST];
-  int count[KMP_HW_LAST];
   kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
   unsigned levels_index;
   kmp_cpuid buf;
   kmp_uint64 known_levels;
   int topology_leaf, highest_leaf, apic_id;
   int num_leaves;
   static int leaves[] = {0, 0};
 
   kmp_i18n_id_t leaf_message_id;
 
   KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST);
 
   *msg_id = kmp_i18n_null;
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
+  }
 
   // Figure out the known topology levels
   known_levels = 0ull;
   for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
     if (__kmp_intel_type_2_topology_type(i) != KMP_HW_UNKNOWN) {
       known_levels |= (1ull << i);
     }
   }
for (int i = 0; i < num_leaves; ++i) {
     topology_leaf = leaf;
     levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
     if (levels_index == 0)
       continue;
     break;
   }
   if (topology_leaf == -1 || levels_index == 0) {
     *msg_id = leaf_message_id;
-    return -1;
+    return false;
   }
   KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
 
   // The algorithm used starts by setting the affinity to each available thread
   // and retrieving info from the cpuid instruction, so if we are not capable of
   // calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
   // we need to do something else - use the defaults that we calculated from
   // issuing cpuid without binding to each proc.
   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);
-
     for (unsigned i = 0; i < levels_index; ++i) {
       if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
         __kmp_nThreadsPerCore = levels[i].nitems;
       } else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
         nCoresPerPkg = levels[i].nitems;
-      } else if (levels[i].level_type == INTEL_LEVEL_TYPE_DIE) {
-        nDiesPerPkg = levels[i].nitems;
       }
     }
     __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
     nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffNotCapableUseLocCpuidL, "KMP_AFFINITY", topology_leaf);
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      if (__kmp_affinity_uniform_topology()) {
-        KMP_INFORM(Uniform, "KMP_AFFINITY");
-      } else {
-        KMP_INFORM(NonUniform, "KMP_AFFINITY");
+    return true;
+  }
+
+  // Allocate the data structure to be returned.
+  int depth = levels_index;
+  for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
+    types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
+  __kmp_topology =
+      kmp_topology_t::allocate(__kmp_avail_proc, levels_index, types);
+
+  // Insert equivalent cache types if they exist
+  kmp_cache_info_t cache_info;
+  for (size_t i = 0; i < cache_info.get_depth(); ++i) {
+    const kmp_cache_info_t::info_t &info = cache_info[i];
+    unsigned cache_mask = info.mask;
+    unsigned cache_level = info.level;
+    for (unsigned j = 0; j < levels_index; ++j) {
+      unsigned hw_cache_mask = levels[j].cache_mask;
+      kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(cache_level);
+      if (hw_cache_mask == cache_mask && j < levels_index - 1) {
+        kmp_hw_t type =
+            __kmp_intel_type_2_topology_type(levels[j + 1].level_type);
+        __kmp_topology->set_equivalent_type(cache_type, type);
       }
-      KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-                 __kmp_nThreadsPerCore, __kmp_ncores);
     }
-    return 0;
   }
 
   // From here on, we can assume that it is safe to call
   // __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
   // __kmp_affinity_type = affinity_none.
 
   // Save the affinity mask for the current thread.
-  kmp_affin_mask_t *oldMask;
-  KMP_CPU_ALLOC(oldMask);
-  __kmp_get_system_affinity(oldMask, TRUE);
-
-  // Allocate the data structure to be returned.
-  int depth = levels_index;
-  for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
-    types[j] = __kmp_intel_type_2_topology_type(levels[i].level_type);
-  AddrUnsPair *retval =
-      (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc);
+  kmp_affinity_raii_t previous_affinity;
 
   // Run through each of the available contexts, binding the current thread
   // to it, and obtaining the pertinent information using the cpuid instr.
   unsigned int proc;
-  int nApics = 0;
+  int hw_thread_index = 0;
   KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
     cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
     unsigned my_levels_index;
 
     // Skip this proc if it is not included in the machine model.
     if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
       continue;
     }
-    KMP_DEBUG_ASSERT(nApics < __kmp_avail_proc);
+    KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
 
     __kmp_affinity_dispatch->bind_thread(proc);
 
     // New algorithm
     __kmp_x86_cpuid(topology_leaf, 0, &buf);
     apic_id = buf.edx;
-    Address addr(depth);
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
     my_levels_index =
         __kmp_x2apicid_get_levels(topology_leaf, my_levels, known_levels);
     if (my_levels_index == 0 || my_levels_index != levels_index) {
-      KMP_CPU_FREE(oldMask);
       *msg_id = kmp_i18n_str_InvalidCpuidInfo;
-      return -1;
+      return false;
     }
+    hw_thread.clear();
+    hw_thread.os_id = proc;
     // Put in topology information
     for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
-      addr.labels[idx] = apic_id & my_levels[j].mask;
-      if (j > 0)
-        addr.labels[idx] >>= my_levels[j - 1].mask_width;
-    }
-    retval[nApics++] = AddrUnsPair(addr, proc);
-  }
-
-  // We've collected all the info we need.
-  // Restore the old affinity mask for this thread.
-  __kmp_set_system_affinity(oldMask, TRUE);
-
-  // If there's only one thread context to bind to, return now.
-  KMP_ASSERT(nApics > 0);
-  if (nApics == 1) {
-    int pkg_level;
-    __kmp_ncores = nPackages = 1;
-    __kmp_nThreadsPerCore = nCoresPerPkg = 1;
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffUseGlobCpuidL, "KMP_AFFINITY", topology_leaf);
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-      KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg,
-                 __kmp_nThreadsPerCore, __kmp_ncores);
-    }
-
-    if (__kmp_affinity_type == affinity_none) {
-      __kmp_free(retval);
-      KMP_CPU_FREE(oldMask);
-      return 0;
-    }
-
-    pkg_level = 0;
-    for (int i = 0; i < depth; ++i)
-      if (types[i] == KMP_HW_SOCKET) {
-        pkg_level = i;
-        break;
+      hw_thread.ids[idx] = apic_id & my_levels[j].mask;
+      if (j > 0) {
+        hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
       }
-    // Form an Address object which only includes the package level.
-    Address addr(1);
-    addr.labels[0] = retval[0].first.labels[pkg_level];
-    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, nApics, sizeof(*retval), __kmp_affinity_cmp_Address_labels);
-
-  __kmp_affinity_gather_enumeration_information(retval, nApics, depth, types,
-                                                ratio, count);
-
-  // 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.
-  int thread_level, core_level, socket_level, die_level;
-  thread_level = core_level = die_level = socket_level = -1;
-  for (int level = 0; level < depth; ++level) {
-    if (types[level] == KMP_HW_THREAD)
-      thread_level = level;
-    else if (types[level] == KMP_HW_CORE)
-      core_level = level;
-    else if (types[level] == KMP_HW_DIE)
-      die_level = level;
-    else if (types[level] == KMP_HW_SOCKET)
-      socket_level = level;
-  }
-  __kmp_nThreadsPerCore =
-      __kmp_affinity_calculate_ratio(ratio, thread_level, core_level);
-  if (die_level > 0) {
-    nDiesPerPkg =
-        __kmp_affinity_calculate_ratio(ratio, die_level, socket_level);
-    nCoresPerPkg = __kmp_affinity_calculate_ratio(ratio, core_level, die_level);
-  } else {
-    nCoresPerPkg =
-        __kmp_affinity_calculate_ratio(ratio, core_level, socket_level);
-  }
-  if (socket_level >= 0)
-    nPackages = count[socket_level];
-  else
-    nPackages = 1;
-  if (core_level >= 0)
-    __kmp_ncores = count[core_level];
-  else
-    __kmp_ncores = 1;
-
-  // Check to see if the machine topology is uniform
-  unsigned uniform = __kmp_affinity_discover_uniformity(depth, ratio, count);
-
-  // Print the machine topology summary.
-  if (__kmp_affinity_verbose) {
-    kmp_hw_t numerator_type, denominator_type;
-    KMP_INFORM(AffUseGlobCpuidL, "KMP_AFFINITY", topology_leaf);
-    KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-    if (uniform) {
-      KMP_INFORM(Uniform, "KMP_AFFINITY");
-    } else {
-      KMP_INFORM(NonUniform, "KMP_AFFINITY");
-    }
-
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-
-    if (core_level < 0)
-      core_level = depth - 1;
-    int ncores = count[core_level];
-
-    denominator_type = KMP_HW_UNKNOWN;
-    for (int level = 0; level < depth; ++level) {
-      int c;
-      bool plural;
-      numerator_type = types[level];
-      c = ratio[level];
-      plural = (c > 1);
-      if (level == 0) {
-        __kmp_str_buf_print(
-            &buf, "%d %s", c,
-            __kmp_hw_get_catalog_string(numerator_type, plural));
-      } else {
-        __kmp_str_buf_print(&buf, " x %d %s/%s", c,
-                            __kmp_hw_get_catalog_string(numerator_type, plural),
-                            __kmp_hw_get_catalog_string(denominator_type));
-      }
-      denominator_type = numerator_type;
-    }
-    KMP_INFORM(TopologyGeneric, "KMP_AFFINITY", buf.str, ncores);
-    __kmp_str_buf_free(&buf);
-  }
-
-  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
-  KMP_DEBUG_ASSERT(nApics == __kmp_avail_proc);
-  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
-  for (proc = 0; (int)proc < nApics; ++proc) {
-    __kmp_pu_os_idx[proc] = retval[proc].second;
-  }
-  if (__kmp_affinity_type == affinity_none) {
-    __kmp_free(retval);
-    KMP_CPU_FREE(oldMask);
-    return 0;
-  }
-
-  // Find any levels with radix 1, and remove them from the map
-  // (except for the package level).
-  depth = __kmp_affinity_remove_radix_one_levels(retval, nApics, depth, types);
-  thread_level = core_level = die_level = socket_level = -1;
-  for (int level = 0; level < depth; ++level) {
-    if (types[level] == KMP_HW_THREAD)
-      thread_level = level;
-    else if (types[level] == KMP_HW_CORE)
-      core_level = level;
-    else if (types[level] == KMP_HW_DIE)
-      die_level = level;
-    else if (types[level] == KMP_HW_SOCKET)
-      socket_level = level;
-  }
-
-  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 ((thread_level >= 0) && (__kmp_affinity_gran > affinity_gran_thread)) {
-      __kmp_affinity_gran_levels++;
-    }
-    if ((core_level >= 0) && (__kmp_affinity_gran > affinity_gran_core)) {
-      __kmp_affinity_gran_levels++;
-    }
-    if ((die_level >= 0) && (__kmp_affinity_gran > affinity_gran_die)) {
-      __kmp_affinity_gran_levels++;
-    }
-    if (__kmp_affinity_gran > affinity_gran_package) {
-      __kmp_affinity_gran_levels++;
     }
+    hw_thread_index++;
   }
-
-  if (__kmp_affinity_verbose) {
-    __kmp_affinity_print_topology(retval, nApics, depth, types);
+  KMP_ASSERT(hw_thread_index > 0);
+  __kmp_topology->sort_ids();
+  if (!__kmp_topology->check_ids()) {
+    kmp_topology_t::deallocate(__kmp_topology);
+    __kmp_topology = nullptr;
+    *msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
+    return false;
   }
-
-  KMP_CPU_FREE(oldMask);
-  *address2os = retval;
-  return depth;
+  return true;
 }
-
 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
 
 #define osIdIndex 0
 #define threadIdIndex 1
 #define coreIdIndex 2
 #define pkgIdIndex 3
 #define nodeIdIndex 4
 
int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
   KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
   KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
   KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
   // (nthreads/t2) / (nthreads/t1) = t1 / t2
   return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
 }
 #endif // KMP_USE_HIER_SCHED
 
+static inline const char *__kmp_cpuinfo_get_filename() {
+  const char *filename;
+  if (__kmp_cpuinfo_file != nullptr)
+    filename = __kmp_cpuinfo_file;
+  else
+    filename = "/proc/cpuinfo";
+  return filename;
+}
+
+static inline const char *__kmp_cpuinfo_get_envvar() {
+  const char *envvar = nullptr;
+  if (__kmp_cpuinfo_file != nullptr)
+    envvar = "KMP_CPUINFO_FILE";
+  return envvar;
+}
+
 // Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
 // affinity map.
-static int __kmp_affinity_create_cpuinfo_map(AddrUnsPair **address2os,
-                                             int *line,
-                                             kmp_i18n_id_t *const msg_id,
-                                             FILE *f) {
-  *address2os = NULL;
+static bool __kmp_affinity_create_cpuinfo_map(int *line,
+                                              kmp_i18n_id_t *const msg_id) {
+  const char *filename = __kmp_cpuinfo_get_filename();
+  const char *envvar = __kmp_cpuinfo_get_envvar();
   *msg_id = kmp_i18n_null;
 
+  if (__kmp_affinity_verbose) {
+    KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
+  }
+
+  kmp_safe_raii_file_t f(filename, "r", envvar);
+
   // Scan of the file, and count the number of "processor" (osId) fields,
   // and find the highest value of <n> for a node_<n> field.
   char buf[256];
   unsigned num_records = 0;
   while (!feof(f)) {
     buf[sizeof(buf) - 1] = 1;
     if (!fgets(buf, sizeof(buf), f)) {
       // Read errors presumably because of EOF
if (KMP_SSCANF(buf, "node_%u id", &level) == 1) {
       }
       continue;
     }
   }
 
   // Check for empty file / no valid processor records, or too many. The number
   // of records can't exceed the number of valid bits in the affinity mask.
   if (num_records == 0) {
-    *line = 0;
     *msg_id = kmp_i18n_str_NoProcRecords;
-    return -1;
+    return false;
   }
   if (num_records > (unsigned)__kmp_xproc) {
-    *line = 0;
     *msg_id = kmp_i18n_str_TooManyProcRecords;
-    return -1;
+    return false;
   }
 
   // Set the file pointer back to the beginning, so that we can scan the file
   // again, this time performing a full parse of the data. Allocate a vector of
   // ProcCpuInfo object, where we will place the data. Adding an extra element
   // at the end allows us to remove a lot of extra checks for termination
   // conditions.
   if (fseek(f, 0, SEEK_SET) != 0) {
-    *line = 0;
     *msg_id = kmp_i18n_str_CantRewindCpuinfo;
-    return -1;
+    return false;
   }
 
   // Allocate the array of records to store the proc info in.  The dummy
   // element at the end makes the logic in filling them out easier to code.
   unsigned **threadInfo =
       (unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
   unsigned i;
   for (i = 0; i <= num_records; i++) {
// jumping past an initialization to a target in the same block.
         // The line is longer than the buffer.  Set a flag and don't
         // emit an error if we were going to ignore the line, anyway.
         long_line = true;
 
 #define CHECK_LINE                                                             \
   if (long_line) {                                                             \
     CLEANUP_THREAD_INFO;                                                       \
     *msg_id = kmp_i18n_str_LongLineCpuinfo;                                    \
-    return -1;                                                                 \
+    return false;                                                              \
   }
       }
       (*line)++;
 
       char s1[] = "processor";
       if (strncmp(buf, s1, sizeof(s1) - 1) == 0) {
         CHECK_LINE;
         char *p = strchr(buf + sizeof(s1) - 1, ':');
#endif // KMP_OS_LINUX && USE_SYSFS_INFO
         continue;
       }
 
       // A newline has signalled the end of the processor record.
       // Check that there aren't too many procs specified.
       if ((int)num_avail == __kmp_xproc) {
         CLEANUP_THREAD_INFO;
         *msg_id = kmp_i18n_str_TooManyEntries;
-        return -1;
+        return false;
       }
 
       // Check for missing fields.  The osId field must be there, and we
       // currently require that the physical id field is specified, also.
       if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
         CLEANUP_THREAD_INFO;
         *msg_id = kmp_i18n_str_MissingProcField;
-        return -1;
+        return false;
       }
       if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
         CLEANUP_THREAD_INFO;
         *msg_id = kmp_i18n_str_MissingPhysicalIDField;
-        return -1;
+        return false;
       }
 
       // Skip this proc if it is not included in the machine model.
       if (!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
                          __kmp_affin_fullMask)) {
         INIT_PROC_INFO(threadInfo[num_avail]);
         continue;
       }
 
       // We have a successful parse of this proc's info.
       // Increment the counter, and prepare for the next proc.
       num_avail++;
       KMP_ASSERT(num_avail <= num_records);
       INIT_PROC_INFO(threadInfo[num_avail]);
     }
     continue;
 
   no_val:
     CLEANUP_THREAD_INFO;
     *msg_id = kmp_i18n_str_MissingValCpuinfo;
-    return -1;
+    return false;
 
   dup_field:
     CLEANUP_THREAD_INFO;
     *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
-    return -1;
+    return false;
   }
   *line = 0;
 
 #if KMP_MIC && REDUCE_TEAM_SIZE
   unsigned teamSize = 0;
 #endif // KMP_MIC && REDUCE_TEAM_SIZE
 
   // check for num_records == __kmp_xproc ???
 
-  // If there's only one thread context to bind to, form an Address object with
-  // depth 1 and return immediately (or, if affinity is off, set address2os to
-  // NULL and return).
-  //
   // If it is configured to omit the package level when there is only a single
   // package, the logic at the end of this routine won't work if there is only a
-  // single thread - it would try to form an Address object with depth 0.
+  // single thread
   KMP_ASSERT(num_avail > 0);
   KMP_ASSERT(num_avail <= num_records);
-  if (num_avail == 1) {
-    __kmp_ncores = 1;
-    __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
-    if (__kmp_affinity_verbose) {
-      if (!KMP_AFFINITY_CAPABLE()) {
-        KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
-        KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-        KMP_INFORM(Uniform, "KMP_AFFINITY");
-      } else {
-        KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
-        KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-        KMP_INFORM(Uniform, "KMP_AFFINITY");
-      }
-      int index;
-      kmp_str_buf_t buf;
-      __kmp_str_buf_init(&buf);
-      __kmp_str_buf_print(&buf, "1");
-      for (index = maxIndex - 1; index > pkgIdIndex; index--) {
-        __kmp_str_buf_print(&buf, " x 1");
-      }
-      KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, 1, 1, 1);
-      __kmp_str_buf_free(&buf);
-    }
-
-    if (__kmp_affinity_type == affinity_none) {
-      CLEANUP_THREAD_INFO;
-      return 0;
-    }
-
-    *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair));
-    Address addr(1);
-    addr.labels[0] = threadInfo[0][pkgIdIndex];
-    (*address2os)[0] = AddrUnsPair(addr, threadInfo[0][osIdIndex]);
-
-    if (__kmp_affinity_gran_levels < 0) {
-      __kmp_affinity_gran_levels = 0;
-    }
-
-    if (__kmp_affinity_verbose) {
-      __kmp_affinity_print_topology(*address2os, 1, 1, 0, -1, -1);
-    }
-
-    CLEANUP_THREAD_INFO;
-    return 1;
-  }
 
   // Sort the threadInfo table by physical Id.
   qsort(threadInfo, num_avail, sizeof(*threadInfo),
         __kmp_affinity_cmp_ProcCpuInfo_phys_id);
 
   // The table is now sorted by pkgId / coreId / threadId, but we really don't
   // know the radix of any of the fields. pkgId's may be sparsely assigned among
   // the chips on a system. Although coreId's are usually assigned
if (index < threadIdIndex) {
       // shouldn't need to).
       if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
         __kmp_free(lastId);
         __kmp_free(totals);
         __kmp_free(maxCt);
         __kmp_free(counts);
         CLEANUP_THREAD_INFO;
         *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
-        return -1;
+        return false;
       }
 
       // If the thread ids were not specified and we see entries entries that
       // are duplicates, start the loop over and assign the thread ids manually.
       assign_thread_ids = true;
       goto restart_radix_check;
     }
   }
for (index = threadIdIndex; index <= maxIndex; index++) {
     }
   }
 
   __kmp_nThreadsPerCore = maxCt[threadIdIndex];
   nCoresPerPkg = maxCt[coreIdIndex];
   nPackages = totals[pkgIdIndex];
 
   // Check to see if the machine topology is uniform
   unsigned prod = totals[maxIndex];
   for (index = threadIdIndex; index < maxIndex; index++) {
     prod *= maxCt[index];
   }
-  bool uniform = (prod == totals[threadIdIndex]);
 
   // 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 now if affinity is
   // not enabled.
   __kmp_ncores = totals[coreIdIndex];
-
-  if (__kmp_affinity_verbose) {
-    if (!KMP_AFFINITY_CAPABLE()) {
-      KMP_INFORM(AffNotCapableUseCpuinfo, "KMP_AFFINITY");
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      if (uniform) {
-        KMP_INFORM(Uniform, "KMP_AFFINITY");
-      } else {
-        KMP_INFORM(NonUniform, "KMP_AFFINITY");
-      }
-    } else {
-      KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY");
-      KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc);
-      if (uniform) {
-        KMP_INFORM(Uniform, "KMP_AFFINITY");
-      } else {
-        KMP_INFORM(NonUniform, "KMP_AFFINITY");
-      }
-    }
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-
-    __kmp_str_buf_print(&buf, "%d", totals[maxIndex]);
-    for (index = maxIndex - 1; index >= pkgIdIndex; index--) {
-      __kmp_str_buf_print(&buf, " x %d", maxCt[index]);
-    }
-    KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, maxCt[coreIdIndex],
-               maxCt[threadIdIndex], __kmp_ncores);
-
-    __kmp_str_buf_free(&buf);
+    KMP_ASSERT(__kmp_affinity_type == affinity_none);
+    return true;
   }
 
 #if KMP_MIC && REDUCE_TEAM_SIZE
   // Set the default team size.
   if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
     __kmp_dflt_team_nth = teamSize;
     KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
                   "__kmp_dflt_team_nth = %d\n",
                   __kmp_dflt_team_nth));
   }
 #endif // KMP_MIC && REDUCE_TEAM_SIZE
 
-  KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL);
   KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
-  __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc);
-  for (i = 0; i < num_avail; ++i) { // fill the os indices
-    __kmp_pu_os_idx[i] = threadInfo[i][osIdIndex];
-  }
-
-  if (__kmp_affinity_type == affinity_none) {
-    __kmp_free(lastId);
-    __kmp_free(totals);
-    __kmp_free(maxCt);
-    __kmp_free(counts);
-    CLEANUP_THREAD_INFO;
-    return 0;
-  }
 
   // Count the number of levels which have more nodes at that level than at the
   // parent's level (with there being an implicit root node of the top level).
   // This is equivalent to saying that there is at least one node at this level
   // which has a sibling. These levels are in the map, and the package level is
   // always in the map.
   bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
   for (index = threadIdIndex; index < maxIndex; index++) {
     KMP_ASSERT(totals[index] >= totals[index + 1]);
     inMap[index] = (totals[index] > totals[index + 1]);
   }
   inMap[maxIndex] = (totals[maxIndex] > 1);
   inMap[pkgIdIndex] = true;
+  inMap[coreIdIndex] = true;
+  inMap[threadIdIndex] = true;
 
   int depth = 0;
+  int idx = 0;
+  kmp_hw_t types[KMP_HW_LAST];
+  int pkgLevel = -1;
+  int coreLevel = -1;
+  int threadLevel = -1;
   for (index = threadIdIndex; index <= maxIndex; index++) {
     if (inMap[index]) {
       depth++;
     }
   }
+  if (inMap[pkgIdIndex]) {
+    pkgLevel = idx;
+    types[idx++] = KMP_HW_SOCKET;
+  }
+  if (inMap[coreIdIndex]) {
+    coreLevel = idx;
+    types[idx++] = KMP_HW_CORE;
+  }
+  if (inMap[threadIdIndex]) {
+    threadLevel = idx;
+    types[idx++] = KMP_HW_THREAD;
+  }
   KMP_ASSERT(depth > 0);
 
   // Construct the data structure that is to be returned.
-  *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * num_avail);
-  int pkgLevel = -1;
-  int coreLevel = -1;
-  int threadLevel = -1;
+  __kmp_topology = kmp_topology_t::allocate(num_avail, depth, types);
 
   for (i = 0; i < num_avail; ++i) {
-    Address addr(depth);
     unsigned os = threadInfo[i][osIdIndex];
     int src_index;
     int dst_index = 0;
+    kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
+    hw_thread.clear();
+    hw_thread.os_id = os;
 
+    idx = 0;
     for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
       if (!inMap[src_index]) {
         continue;
       }
-      addr.labels[dst_index] = threadInfo[i][src_index];
       if (src_index == pkgIdIndex) {
-        pkgLevel = dst_index;
+        hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
       } else if (src_index == coreIdIndex) {
-        coreLevel = dst_index;
+        hw_thread.ids[coreLevel] = threadInfo[i][src_index];
       } else if (src_index == threadIdIndex) {
-        threadLevel = dst_index;
+        hw_thread.ids[threadLevel] = threadInfo[i][src_index];
       }
       dst_index++;
     }
-    (*address2os)[i] = AddrUnsPair(addr, os);
-  }
-
-  if (__kmp_affinity_gran_levels < 0) {
-    // Set the granularity level based on what levels are modeled
-    // in the machine topology map.
-    unsigned src_index;
-    __kmp_affinity_gran_levels = 0;
-    for (src_index = threadIdIndex; src_index <= maxIndex; src_index++) {
-      if (!inMap[src_index]) {
-        continue;
-      }
-      switch (src_index) {
-      case threadIdIndex:
-        if (__kmp_affinity_gran > affinity_gran_thread) {
-          __kmp_affinity_gran_levels++;
-        }
-
-        break;
-      case coreIdIndex:
-        if (__kmp_affinity_gran > affinity_gran_core) {
-          __kmp_affinity_gran_levels++;
-        }
-        break;
-
-      case pkgIdIndex:
-        if (__kmp_affinity_gran > affinity_gran_package) {
-          __kmp_affinity_gran_levels++;
-        }
-        break;
-      }
-    }
-  }
-
-  if (__kmp_affinity_verbose) {
-    __kmp_affinity_print_topology(*address2os, num_avail, depth, pkgLevel,
-                                  coreLevel, threadLevel);
   }
 
   __kmp_free(inMap);
   __kmp_free(lastId);
   __kmp_free(totals);
   __kmp_free(maxCt);
   __kmp_free(counts);
   CLEANUP_THREAD_INFO;
-  return depth;
+  __kmp_topology->sort_ids();
+  if (!__kmp_topology->check_ids()) {
+    kmp_topology_t::deallocate(__kmp_topology);
+    __kmp_topology = nullptr;
+    *msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
+    return false;
+  }
+  return true;
 }
 
 // Create and return a table of affinity masks, indexed by OS thread ID.
 // This routine handles OR'ing together all the affinity masks of threads
 // that are sufficiently close, if granularity > fine.
 static kmp_affin_mask_t *__kmp_create_masks(unsigned *maxIndex,
-                                            unsigned *numUnique,
-                                            AddrUnsPair *address2os,
-                                            unsigned numAddrs) {
+                                            unsigned *numUnique) {
   // First form a table of affinity masks in order of OS thread id.
-  unsigned depth;
-  unsigned maxOsId;
-  unsigned i;
-
-  KMP_ASSERT(numAddrs > 0);
-  depth = address2os[0].first.depth;
+  int maxOsId;
+  int i;
+  int numAddrs = __kmp_topology->get_num_hw_threads();
+  int depth = __kmp_topology->get_depth();
+  KMP_ASSERT(numAddrs);
+  KMP_ASSERT(depth);
 
   maxOsId = 0;
   for (i = numAddrs - 1;; --i) {
-    unsigned osId = address2os[i].second;
+    int osId = __kmp_topology->at(i).os_id;
     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);
-
   KMP_ASSERT(__kmp_affinity_gran_levels >= 0);
   if (__kmp_affinity_verbose && (__kmp_affinity_gran_levels > 0)) {
     KMP_INFORM(ThreadsMigrate, "KMP_AFFINITY", __kmp_affinity_gran_levels);
   }
   if (__kmp_affinity_gran_levels >= (int)depth) {
     if (__kmp_affinity_verbose ||
         (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
       KMP_WARNING(AffThreadsMayMigrate);
     }
   }
 
   // Run through the table, forming the masks for all threads on each core.
-  // Threads on the same core will have identical "Address" objects, not
+  // Threads on the same core will have identical kmp_hw_thread_t objects, not
   // considering the last level, which must be the thread id. All threads on a
   // core will appear consecutively.
-  unsigned unique = 0;
-  unsigned j = 0; // index of 1st thread on core
-  unsigned leader = 0;
-  Address *leaderAddr = &(address2os[0].first);
+  int unique = 0;
+  int j = 0; // index of 1st thread on core
+  int leader = 0;
   kmp_affin_mask_t *sum;
   KMP_CPU_ALLOC_ON_STACK(sum);
   KMP_CPU_ZERO(sum);
-  KMP_CPU_SET(address2os[0].second, sum);
+  KMP_CPU_SET(__kmp_topology->at(0).os_id, sum);
   for (i = 1; i < numAddrs; i++) {
     // If this thread is sufficiently close to the leader (within the
     // granularity setting), then set the bit for this os thread in the
     // affinity mask for this group, and go on to the next thread.
-    if (leaderAddr->isClose(address2os[i].first, __kmp_affinity_gran_levels)) {
-      KMP_CPU_SET(address2os[i].second, sum);
+    if (__kmp_topology->is_close(leader, i, __kmp_affinity_gran_levels)) {
+      KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
       continue;
     }
 
     // For every thread in this group, copy the mask to the thread's entry in
     // the osId2Mask table.  Mark the first address as a leader.
     for (; j < i; j++) {
-      unsigned osId = address2os[j].second;
+      int osId = __kmp_topology->at(j).os_id;
       KMP_DEBUG_ASSERT(osId <= maxOsId);
       kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
       KMP_CPU_COPY(mask, sum);
-      address2os[j].first.leader = (j == leader);
+      __kmp_topology->at(j).leader = (j == leader);
     }
     unique++;
 
     // Start a new mask.
     leader = i;
-    leaderAddr = &(address2os[i].first);
     KMP_CPU_ZERO(sum);
-    KMP_CPU_SET(address2os[i].second, sum);
+    KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
   }
 
   // For every thread in last group, copy the mask to the thread's
   // entry in the osId2Mask table.
   for (; j < i; j++) {
-    unsigned osId = address2os[j].second;
+    int osId = __kmp_topology->at(j).os_id;
     KMP_DEBUG_ASSERT(osId <= maxOsId);
     kmp_affin_mask_t *mask = KMP_CPU_INDEX(osId2Mask, osId);
     KMP_CPU_COPY(mask, sum);
-    address2os[j].first.leader = (j == leader);
+    __kmp_topology->at(j).leader = (j == leader);
   }
   unique++;
   KMP_CPU_FREE_FROM_STACK(sum);
 
   *maxIndex = maxOsId;
   *numUnique = unique;
   return osId2Mask;
 }
for (i = 0; i < nextNewMask; i++) {
     KMP_CPU_COPY(dest, src);
   }
   KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
 }
 
 #undef ADD_MASK
 #undef ADD_MASK_OSID
 
-#if KMP_USE_HWLOC
-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_CPU_ISSET(idx, __kmp_affin_fullMask)) {
-      KMP_CPU_CLR(idx, __kmp_affin_fullMask);
-      KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
-      ++skipped;
-    }
-    hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
-  }
-  return skipped; // count number of skipped units
-}
-
-static int __kmp_hwloc_obj_has_PUs(hwloc_topology_t t, hwloc_obj_t o) {
-  // check if obj has PUs present in fullMask
-  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_CPU_ISSET(idx, __kmp_affin_fullMask))
-      return 1; // found PU
-    hT = hwloc_get_next_obj_by_type(t, HWLOC_OBJ_PU, hT);
-  }
-  return 0; // no PUs found
-}
-#endif // KMP_USE_HWLOC
-
-static void __kmp_apply_thread_places(AddrUnsPair **pAddr, int depth) {
-  AddrUnsPair *newAddr;
-  if (__kmp_hws_requested == 0)
-    goto _exit; // no topology limiting actions requested, exit
-#if KMP_USE_HWLOC
-  if (__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
-    // Number of subobjects calculated dynamically, this works fine for
-    // any non-uniform topology.
-    // L2 cache objects are determined by depth, other objects - by type.
-    hwloc_topology_t tp = __kmp_hwloc_topology;
-    int nS = 0, nN = 0, nL = 0, nC = 0,
-        nT = 0; // logical index including skipped
-    int nCr = 0, nTr = 0; // number of requested units
-    int nPkg = 0, nCo = 0, n_new = 0, n_old = 0, nCpP = 0, nTpC = 0; // counters
-    hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to)
-    int L2depth, idx;
-
-    // check support of extensions ----------------------------------
-    int numa_support = 0, tile_support = 0;
-    if (__kmp_pu_os_idx)
-      hT = hwloc_get_pu_obj_by_os_index(tp,
-                                        __kmp_pu_os_idx[__kmp_avail_proc - 1]);
-    else
-      hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, __kmp_avail_proc - 1);
-    if (hT == NULL) { // something's gone wrong
-      KMP_WARNING(AffHWSubsetUnsupported);
-      goto _exit;
-    }
-    // check NUMA node
-    hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT);
-    hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT);
-    if (hN != NULL && hN->depth > hS->depth) {
-      numa_support = 1; // 1 in case socket includes node(s)
-    } else if (__kmp_hws_node.num > 0) {
-      // don't support sockets inside NUMA node (no such HW found for testing)
-      KMP_WARNING(AffHWSubsetUnsupported);
-      goto _exit;
-    }
-    // check L2 cahce, get object by depth because of multiple caches
-    L2depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED);
-    hL = hwloc_get_ancestor_obj_by_depth(tp, L2depth, hT);
-    if (hL != NULL &&
-        __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) {
-      tile_support = 1; // no sense to count L2 if it includes single core
-    } else if (__kmp_hws_tile.num > 0) {
-      if (__kmp_hws_core.num == 0) {
-        __kmp_hws_core = __kmp_hws_tile; // replace L2 with core
-        __kmp_hws_tile.num = 0;
-      } else {
-        // L2 and core are both requested, but represent same object
-        KMP_WARNING(AffHWSubsetInvalid);
-        goto _exit;
-      }
-    }
-    // end of check of extensions -----------------------------------
-
-    // 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 (tile_support) {
-        // get num tiles in node
-        int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
-        if (__kmp_hws_tile.num == 0) {
-          __kmp_hws_tile.num = NL + 1;
-        } // use all available tiles, some node may have more tiles, thus +1
-        if (__kmp_hws_tile.offset >= NL) {
-          KMP_WARNING(AffHWSubsetManyTiles);
-          goto _exit;
-        }
-        int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
-                                                    &hC); // num cores in tile
-        if (__kmp_hws_core.num == 0)
-          __kmp_hws_core.num = NC; // use all available cores
-        if (__kmp_hws_core.offset >= NC) {
-          KMP_WARNING(AffHWSubsetManyCores);
-          goto _exit;
-        }
-      } else { // tile_support
-        int NC = __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE,
-                                                    &hC); // num cores in node
-        if (__kmp_hws_core.num == 0)
-          __kmp_hws_core.num = NC; // use all available cores
-        if (__kmp_hws_core.offset >= NC) {
-          KMP_WARNING(AffHWSubsetManyCores);
-          goto _exit;
-        }
-      } // tile_support
-    } else { // numa_support
-      if (tile_support) {
-        // get num tiles in socket
-        int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
-        if (__kmp_hws_tile.num == 0)
-          __kmp_hws_tile.num = NL; // use all available tiles
-        if (__kmp_hws_tile.offset >= NL) {
-          KMP_WARNING(AffHWSubsetManyTiles);
-          goto _exit;
-        }
-        int NC = __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE,
-                                                    &hC); // num cores in tile
-        if (__kmp_hws_core.num == 0)
-          __kmp_hws_core.num = NC; // use all available cores
-        if (__kmp_hws_core.offset >= NC) {
-          KMP_WARNING(AffHWSubsetManyCores);
-          goto _exit;
-        }
-      } else { // tile_support
-        int NC = __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE,
-                                                    &hC); // num cores in socket
-        if (__kmp_hws_core.num == 0)
-          __kmp_hws_core.num = NC; // use all available cores
-        if (__kmp_hws_core.offset >= NC) {
-          KMP_WARNING(AffHWSubsetManyCores);
-          goto _exit;
-        }
-      } // tile_support
-    }
-    if (__kmp_hws_proc.num == 0)
-      __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all available procs
-    if (__kmp_hws_proc.offset >= __kmp_nThreadsPerCore) {
-      KMP_WARNING(AffHWSubsetManyProcs);
-      goto _exit;
-    }
-    // end of validation --------------------------------------------
-
-    if (pAddr) // pAddr is NULL in case of affinity_none
-      newAddr = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) *
-                                              __kmp_avail_proc); // max size
-    // main loop to form HW subset ----------------------------------
-    hS = NULL;
-    int NP = hwloc_get_nbobjs_by_type(tp, HWLOC_OBJ_PACKAGE);
-    for (int s = 0; s < NP; ++s) {
-      // Check Socket -----------------------------------------------
-      hS = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hS);
-      if (!__kmp_hwloc_obj_has_PUs(tp, hS))
-        continue; // skip socket if all PUs are out of fullMask
-      ++nS; // only count objects those have PUs in affinity mask
-      if (nS <= __kmp_hws_socket.offset ||
-          nS > __kmp_hws_socket.num + __kmp_hws_socket.offset) {
-        n_old += __kmp_hwloc_skip_PUs_obj(tp, hS); // skip socket
-        continue; // move to next socket
-      }
-      nCr = 0; // count number of cores per socket
-      // socket requested, go down the topology tree
-      // check 4 cases: (+NUMA+Tile), (+NUMA-Tile), (-NUMA+Tile), (-NUMA-Tile)
-      if (numa_support) {
-        nN = 0;
-        hN = NULL;
-        // num nodes in current socket
-        int NN =
-            __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_NUMANODE, &hN);
-        for (int n = 0; n < NN; ++n) {
-          // Check NUMA Node ----------------------------------------
-          if (!__kmp_hwloc_obj_has_PUs(tp, hN)) {
-            hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
-            continue; // skip node if all PUs are out of fullMask
-          }
-          ++nN;
-          if (nN <= __kmp_hws_node.offset ||
-              nN > __kmp_hws_node.num + __kmp_hws_node.offset) {
-            // skip node as not requested
-            n_old += __kmp_hwloc_skip_PUs_obj(tp, hN); // skip node
-            hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
-            continue; // move to next node
-          }
-          // node requested, go down the topology tree
-          if (tile_support) {
-            nL = 0;
-            hL = NULL;
-            int NL = __kmp_hwloc_count_children_by_depth(tp, hN, L2depth, &hL);
-            for (int l = 0; l < NL; ++l) {
-              // Check L2 (tile) ------------------------------------
-              if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
-                hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-                continue; // skip tile if all PUs are out of fullMask
-              }
-              ++nL;
-              if (nL <= __kmp_hws_tile.offset ||
-                  nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
-                // skip tile as not requested
-                n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
-                hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-                continue; // move to next tile
-              }
-              // tile requested, go down the topology tree
-              nC = 0;
-              hC = NULL;
-              // num cores in current tile
-              int NC = __kmp_hwloc_count_children_by_type(tp, hL,
-                                                          HWLOC_OBJ_CORE, &hC);
-              for (int c = 0; c < NC; ++c) {
-                // Check Core ---------------------------------------
-                if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
-                  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                  continue; // skip core if all PUs are out of fullMask
-                }
-                ++nC;
-                if (nC <= __kmp_hws_core.offset ||
-                    nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
-                  // skip node as not requested
-                  n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
-                  hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                  continue; // move to next node
-                }
-                // core requested, go down to PUs
-                nT = 0;
-                nTr = 0;
-                hT = NULL;
-                // num procs in current core
-                int NT = __kmp_hwloc_count_children_by_type(tp, hC,
-                                                            HWLOC_OBJ_PU, &hT);
-                for (int t = 0; t < NT; ++t) {
-                  // Check PU ---------------------------------------
-                  idx = hT->os_index;
-                  if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
-                    hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                    continue; // skip PU if not in fullMask
-                  }
-                  ++nT;
-                  if (nT <= __kmp_hws_proc.offset ||
-                      nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
-                    // skip PU
-                    KMP_CPU_CLR(idx, __kmp_affin_fullMask);
-                    ++n_old;
-                    KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
-                    hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                    continue; // move to next node
-                  }
-                  ++nTr;
-                  if (pAddr) // collect requested thread's data
-                    newAddr[n_new] = (*pAddr)[n_old];
-                  ++n_new;
-                  ++n_old;
-                  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                } // threads loop
-                if (nTr > 0) {
-                  ++nCr; // num cores per socket
-                  ++nCo; // total num cores
-                  if (nTr > nTpC)
-                    nTpC = nTr; // calc max threads per core
-                }
-                hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-              } // cores loop
-              hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-            } // tiles loop
-          } else { // tile_support
-            // no tiles, check cores
-            nC = 0;
-            hC = NULL;
-            // num cores in current node
-            int NC =
-                __kmp_hwloc_count_children_by_type(tp, hN, HWLOC_OBJ_CORE, &hC);
-            for (int c = 0; c < NC; ++c) {
-              // Check Core ---------------------------------------
-              if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
-                hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                continue; // skip core if all PUs are out of fullMask
-              }
-              ++nC;
-              if (nC <= __kmp_hws_core.offset ||
-                  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
-                // skip node as not requested
-                n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
-                hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                continue; // move to next node
-              }
-              // core requested, go down to PUs
-              nT = 0;
-              nTr = 0;
-              hT = NULL;
-              int NT =
-                  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
-              for (int t = 0; t < NT; ++t) {
-                // Check PU ---------------------------------------
-                idx = hT->os_index;
-                if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
-                  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                  continue; // skip PU if not in fullMask
-                }
-                ++nT;
-                if (nT <= __kmp_hws_proc.offset ||
-                    nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
-                  // skip PU
-                  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
-                  ++n_old;
-                  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
-                  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                  continue; // move to next node
-                }
-                ++nTr;
-                if (pAddr) // collect requested thread's data
-                  newAddr[n_new] = (*pAddr)[n_old];
-                ++n_new;
-                ++n_old;
-                hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-              } // threads loop
-              if (nTr > 0) {
-                ++nCr; // num cores per socket
-                ++nCo; // total num cores
-                if (nTr > nTpC)
-                  nTpC = nTr; // calc max threads per core
-              }
-              hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-            } // cores loop
-          } // tiles support
-          hN = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hN);
-        } // nodes loop
-      } else { // numa_support
-        // no NUMA support
-        if (tile_support) {
-          nL = 0;
-          hL = NULL;
-          // num tiles in current socket
-          int NL = __kmp_hwloc_count_children_by_depth(tp, hS, L2depth, &hL);
-          for (int l = 0; l < NL; ++l) {
-            // Check L2 (tile) ------------------------------------
-            if (!__kmp_hwloc_obj_has_PUs(tp, hL)) {
-              hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-              continue; // skip tile if all PUs are out of fullMask
-            }
-            ++nL;
-            if (nL <= __kmp_hws_tile.offset ||
-                nL > __kmp_hws_tile.num + __kmp_hws_tile.offset) {
-              // skip tile as not requested
-              n_old += __kmp_hwloc_skip_PUs_obj(tp, hL); // skip tile
-              hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-              continue; // move to next tile
-            }
-            // tile requested, go down the topology tree
-            nC = 0;
-            hC = NULL;
-            // num cores per tile
-            int NC =
-                __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC);
-            for (int c = 0; c < NC; ++c) {
-              // Check Core ---------------------------------------
-              if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
-                hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                continue; // skip core if all PUs are out of fullMask
-              }
-              ++nC;
-              if (nC <= __kmp_hws_core.offset ||
-                  nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
-                // skip node as not requested
-                n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
-                hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-                continue; // move to next node
-              }
-              // core requested, go down to PUs
-              nT = 0;
-              nTr = 0;
-              hT = NULL;
-              // num procs per core
-              int NT =
-                  __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
-              for (int t = 0; t < NT; ++t) {
-                // Check PU ---------------------------------------
-                idx = hT->os_index;
-                if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
-                  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                  continue; // skip PU if not in fullMask
-                }
-                ++nT;
-                if (nT <= __kmp_hws_proc.offset ||
-                    nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
-                  // skip PU
-                  KMP_CPU_CLR(idx, __kmp_affin_fullMask);
-                  ++n_old;
-                  KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
-                  hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                  continue; // move to next node
-                }
-                ++nTr;
-                if (pAddr) // collect requested thread's data
-                  newAddr[n_new] = (*pAddr)[n_old];
-                ++n_new;
-                ++n_old;
-                hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-              } // threads loop
-              if (nTr > 0) {
-                ++nCr; // num cores per socket
-                ++nCo; // total num cores
-                if (nTr > nTpC)
-                  nTpC = nTr; // calc max threads per core
-              }
-              hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-            } // cores loop
-            hL = hwloc_get_next_obj_by_depth(tp, L2depth, hL);
-          } // tiles loop
-        } else { // tile_support
-          // no tiles, check cores
-          nC = 0;
-          hC = NULL;
-          // num cores in socket
-          int NC =
-              __kmp_hwloc_count_children_by_type(tp, hS, HWLOC_OBJ_CORE, &hC);
-          for (int c = 0; c < NC; ++c) {
-            // Check Core -------------------------------------------
-            if (!__kmp_hwloc_obj_has_PUs(tp, hC)) {
-              hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-              continue; // skip core if all PUs are out of fullMask
-            }
-            ++nC;
-            if (nC <= __kmp_hws_core.offset ||
-                nC > __kmp_hws_core.num + __kmp_hws_core.offset) {
-              // skip node as not requested
-              n_old += __kmp_hwloc_skip_PUs_obj(tp, hC); // skip core
-              hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-              continue; // move to next node
-            }
-            // core requested, go down to PUs
-            nT = 0;
-            nTr = 0;
-            hT = NULL;
-            // num procs per core
-            int NT =
-                __kmp_hwloc_count_children_by_type(tp, hC, HWLOC_OBJ_PU, &hT);
-            for (int t = 0; t < NT; ++t) {
-              // Check PU ---------------------------------------
-              idx = hT->os_index;
-              if (!KMP_CPU_ISSET(idx, __kmp_affin_fullMask)) {
-                hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                continue; // skip PU if not in fullMask
-              }
-              ++nT;
-              if (nT <= __kmp_hws_proc.offset ||
-                  nT > __kmp_hws_proc.num + __kmp_hws_proc.offset) {
-                // skip PU
-                KMP_CPU_CLR(idx, __kmp_affin_fullMask);
-                ++n_old;
-                KC_TRACE(200, ("KMP_HW_SUBSET: skipped proc %d\n", idx));
-                hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-                continue; // move to next node
-              }
-              ++nTr;
-              if (pAddr) // collect requested thread's data
-                newAddr[n_new] = (*pAddr)[n_old];
-              ++n_new;
-              ++n_old;
-              hT = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, hT);
-            } // threads loop
-            if (nTr > 0) {
-              ++nCr; // num cores per socket
-              ++nCo; // total num cores
-              if (nTr > nTpC)
-                nTpC = nTr; // calc max threads per core
-            }
-            hC = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_CORE, hC);
-          } // cores loop
-        } // tiles support
-      } // numa_support
-      if (nCr > 0) { // found cores?
-        ++nPkg; // num sockets
-        if (nCr > nCpP)
-          nCpP = nCr; // calc max cores per socket
-      }
-    } // sockets loop
-
-    // check the subset is valid
-    KMP_DEBUG_ASSERT(n_old == __kmp_avail_proc);
-    KMP_DEBUG_ASSERT(nPkg > 0);
-    KMP_DEBUG_ASSERT(nCpP > 0);
-    KMP_DEBUG_ASSERT(nTpC > 0);
-    KMP_DEBUG_ASSERT(nCo > 0);
-    KMP_DEBUG_ASSERT(nPkg <= nPackages);
-    KMP_DEBUG_ASSERT(nCpP <= nCoresPerPkg);
-    KMP_DEBUG_ASSERT(nTpC <= __kmp_nThreadsPerCore);
-    KMP_DEBUG_ASSERT(nCo <= __kmp_ncores);
-
-    nPackages = nPkg; // correct num sockets
-    nCoresPerPkg = nCpP; // correct num cores per socket
-    __kmp_nThreadsPerCore = nTpC; // correct num threads per core
-    __kmp_avail_proc = n_new; // correct num procs
-    __kmp_ncores = nCo; // correct num cores
-    // hwloc topology method end
-  } else
-#endif // KMP_USE_HWLOC
-  {
-    int n_old = 0, n_new = 0, proc_num = 0;
-    if (__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0) {
-      KMP_WARNING(AffHWSubsetNoHWLOC);
-      goto _exit;
-    }
-    if (__kmp_hws_socket.num == 0)
-      __kmp_hws_socket.num = nPackages; // use all available sockets
-    if (__kmp_hws_die.num == 0)
-      __kmp_hws_die.num = nDiesPerPkg; // use all available dies
-    if (__kmp_hws_core.num == 0)
-      __kmp_hws_core.num = nCoresPerPkg; // use all available cores
-    if (__kmp_hws_proc.num == 0 || __kmp_hws_proc.num > __kmp_nThreadsPerCore)
-      __kmp_hws_proc.num = __kmp_nThreadsPerCore; // use all HW contexts
-    if (!__kmp_affinity_uniform_topology()) {
-      KMP_WARNING(AffHWSubsetNonUniform);
-      goto _exit; // don't support non-uniform topology
-    }
-    if (depth > 4) {
-      KMP_WARNING(AffHWSubsetNonThreeLevel);
-      goto _exit; // don't support not-3-level topology
-    }
-    if (__kmp_hws_socket.offset + __kmp_hws_socket.num > nPackages) {
-      KMP_WARNING(AffHWSubsetManySockets);
-      goto _exit;
-    }
-    if (depth == 4 && __kmp_hws_die.offset + __kmp_hws_die.num > nDiesPerPkg) {
-      KMP_WARNING(AffHWSubsetManyDies);
-      goto _exit;
-    }
-    if (__kmp_hws_core.offset + __kmp_hws_core.num > nCoresPerPkg) {
-      KMP_WARNING(AffHWSubsetManyCores);
-      goto _exit;
-    }
-    // Form the requested subset
-    if (pAddr) // pAddr is NULL in case of affinity_none
-      newAddr = (AddrUnsPair *)__kmp_allocate(
-          sizeof(AddrUnsPair) * __kmp_hws_socket.num * __kmp_hws_die.num *
-          __kmp_hws_core.num * __kmp_hws_proc.num);
-    for (int i = 0; i < nPackages; ++i) {
-      if (i < __kmp_hws_socket.offset ||
-          i >= __kmp_hws_socket.offset + __kmp_hws_socket.num) {
-        // skip not-requested socket
-        n_old += nDiesPerPkg * nCoresPerPkg * __kmp_nThreadsPerCore;
-        if (__kmp_pu_os_idx != NULL) {
-          // walk through skipped socket
-          for (int l = 0; l < nDiesPerPkg; ++l) {
-            for (int j = 0; j < nCoresPerPkg; ++j) {
-              for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
-                KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
-                ++proc_num;
-              }
-            }
-          }
-        }
-      } else {
-        // walk through requested socket
-        for (int l = 0; l < nDiesPerPkg; ++l) {
-          // skip unwanted die
-          if (l < __kmp_hws_die.offset ||
-              l >= __kmp_hws_die.offset + __kmp_hws_die.num) {
-            n_old += nCoresPerPkg;
-            if (__kmp_pu_os_idx != NULL) {
-              for (int k = 0; k < nCoresPerPkg; ++k) {
-                KMP_CPU_CLR(__kmp_pu_os_idx[proc_num], __kmp_affin_fullMask);
-                ++proc_num;
-              }
-            }
-          } else {
-            for (int j = 0; j < nCoresPerPkg; ++j) {
-              if (j < __kmp_hws_core.offset ||
-                  j >= __kmp_hws_core.offset +
-                           __kmp_hws_core.num) { // skip not-requested core
-                n_old += __kmp_nThreadsPerCore;
-                if (__kmp_pu_os_idx != NULL) {
-                  for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
-                    KMP_CPU_CLR(__kmp_pu_os_idx[proc_num],
-                                __kmp_affin_fullMask);
-                    ++proc_num;
-                  }
-                }
-              } else {
-                // walk through requested core
-                for (int k = 0; k < __kmp_nThreadsPerCore; ++k) {
-                  if (k < __kmp_hws_proc.num) {
-                    if (pAddr) // collect requested thread's data
-                      newAddr[n_new] = (*pAddr)[n_old];
-                    n_new++;
-                  } else {
-                    if (__kmp_pu_os_idx != NULL)
-                      KMP_CPU_CLR(__kmp_pu_os_idx[proc_num],
-                                  __kmp_affin_fullMask);
-                  }
-                  n_old++;
-                  ++proc_num;
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-    KMP_DEBUG_ASSERT(n_old == nPackages * nDiesPerPkg * nCoresPerPkg *
-                                  __kmp_nThreadsPerCore);
-    KMP_DEBUG_ASSERT(n_new == __kmp_hws_socket.num * __kmp_hws_die.num *
-                                  __kmp_hws_core.num * __kmp_hws_proc.num);
-    nPackages = __kmp_hws_socket.num; // correct nPackages
-    nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg
-    nDiesPerPkg = __kmp_hws_die.num; // correct nDiesPerPkg
-    __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore
-    __kmp_avail_proc = n_new; // correct avail_proc
-    __kmp_ncores =
-        nPackages * nDiesPerPkg * __kmp_hws_core.num; // correct ncores
-  } // non-hwloc topology method
-  if (pAddr) {
-    __kmp_free(*pAddr);
-    *pAddr = newAddr; // replace old topology with new one
-  }
-  if (__kmp_affinity_verbose) {
-    KMP_INFORM(AvailableOSProc, "KMP_HW_SUBSET", __kmp_avail_proc);
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-    __kmp_str_buf_print(&buf, "%d", nPackages);
-    KMP_INFORM(TopologyExtra, "KMP_HW_SUBSET", buf.str, nCoresPerPkg,
-               __kmp_nThreadsPerCore, __kmp_ncores);
-    __kmp_str_buf_free(&buf);
-  }
-_exit:
-  if (__kmp_pu_os_idx != NULL) {
-    __kmp_free(__kmp_pu_os_idx);
-    __kmp_pu_os_idx = NULL;
-  }
-}
-
 // This function figures out the deepest level at which there is at least one
 // cluster/core with more than one processing unit bound to it.
-static int __kmp_affinity_find_core_level(const AddrUnsPair *address2os,
-                                          int nprocs, int bottom_level) {
+static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
   int core_level = 0;
 
   for (int i = 0; i < nprocs; i++) {
+    const kmp_hw_thread_t &hw_thread = __kmp_topology->at(i);
     for (int j = bottom_level; j > 0; j--) {
-      if (address2os[i].first.labels[j] > 0) {
+      if (hw_thread.ids[j] > 0) {
         if (core_level < (j - 1)) {
           core_level = j - 1;
         }
       }
     }
   }
   return core_level;
 }
 
 // This function counts number of clusters/cores at given level.
-static int __kmp_affinity_compute_ncores(const AddrUnsPair *address2os,
-                                         int nprocs, int bottom_level,
+static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
                                          int core_level) {
-  int ncores = 0;
-  int i, j;
-
-  j = bottom_level;
-  for (i = 0; i < nprocs; i++) {
-    for (j = bottom_level; j > core_level; j--) {
-      if ((i + 1) < nprocs) {
-        if (address2os[i + 1].first.labels[j] > 0) {
-          break;
-        }
-      }
-    }
-    if (j == core_level) {
-      ncores++;
+  return __kmp_topology->get_count(core_level);
+// This function finds to which cluster/core given processing unit is bound.
+static int __kmp_affinity_find_core(int proc, int bottom_level,
+                                    int core_level) {
+  int core = 0;
+  KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
+  for (int i = 0; i <= proc; ++i) {
+    if (i + 1 <= proc) {
+      for (int j = 0; j <= core_level; ++j) {
+        if (__kmp_topology->at(i + 1).sub_ids[j] !=
+            __kmp_topology->at(i).sub_ids[j]) {
+          core++;
+          break;
         }
       }
-  if (j > core_level) {
-    // In case of ( nprocs < __kmp_avail_proc ) we may end too deep and miss one
-    // core. May occur when called from __kmp_affinity_find_core().
-    ncores++;
     }
-  return ncores;
   }
-
-// This function finds to which cluster/core given processing unit is bound.
-static int __kmp_affinity_find_core(const AddrUnsPair *address2os, int proc,
-                                    int bottom_level, int core_level) {
-  return __kmp_affinity_compute_ncores(address2os, proc + 1, bottom_level,
-                                       core_level) -
-         1;
+  return core;
 }
 
 // This function finds maximal number of processing units bound to a
 // cluster/core at given level.
-static int __kmp_affinity_max_proc_per_core(const AddrUnsPair *address2os,
-                                            int nprocs, int bottom_level,
+static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
                                             int core_level) {
-  int maxprocpercore = 0;
-
-  if (core_level < bottom_level) {
-    for (int i = 0; i < nprocs; i++) {
-      int percore = address2os[i].first.labels[core_level + 1] + 1;
-
-      if (percore > maxprocpercore) {
-        maxprocpercore = percore;
-      }
-    }
-  } else {
-    maxprocpercore = 1;
-  }
-  return maxprocpercore;
+  if (core_level >= bottom_level)
+    return 1;
+  int thread_level = __kmp_topology->get_level(KMP_HW_THREAD);
+  return __kmp_topology->calculate_ratio(thread_level, core_level);
 }
 
-static AddrUnsPair *address2os = NULL;
 static int *procarr = NULL;
 static int __kmp_aff_depth = 0;
 
-#if KMP_USE_HIER_SCHED
-#define KMP_EXIT_AFF_NONE                                                      \
-  KMP_ASSERT(__kmp_affinity_type == affinity_none);                            \
-  KMP_ASSERT(address2os == NULL);                                              \
-  __kmp_apply_thread_places(NULL, 0);                                          \
-  __kmp_create_affinity_none_places();                                         \
-  __kmp_dispatch_set_hierarchy_values();                                       \
-  return;
-#else
-#define KMP_EXIT_AFF_NONE                                                      \
-  KMP_ASSERT(__kmp_affinity_type == affinity_none);                            \
-  KMP_ASSERT(address2os == NULL);                                              \
-  __kmp_apply_thread_places(NULL, 0);                                          \
-  __kmp_create_affinity_none_places();                                         \
-  return;
-#endif
-
 // Create a one element mask array (set of places) which only contains the
 // initial process's affinity mask
 static void __kmp_create_affinity_none_places() {
   KMP_ASSERT(__kmp_affin_fullMask != NULL);
   KMP_ASSERT(__kmp_affinity_type == affinity_none);
   __kmp_affinity_num_masks = 1;
   KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
   kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, 0);
   KMP_CPU_COPY(dest, __kmp_affin_fullMask);
 }
 
-static int __kmp_affinity_cmp_Address_child_num(const void *a, const void *b) {
-  const Address *aa = &(((const AddrUnsPair *)a)->first);
-  const Address *bb = &(((const AddrUnsPair *)b)->first);
-  unsigned depth = aa->depth;
-  unsigned i;
-  KMP_DEBUG_ASSERT(depth == bb->depth);
-  KMP_DEBUG_ASSERT((unsigned)__kmp_affinity_compact <= depth);
-  KMP_DEBUG_ASSERT(__kmp_affinity_compact >= 0);
-  for (i = 0; i < (unsigned)__kmp_affinity_compact; i++) {
-    int j = depth - i - 1;
-    if (aa->childNums[j] < bb->childNums[j])
-      return -1;
-    if (aa->childNums[j] > bb->childNums[j])
-      return 1;
-  }
-  for (; i < depth; i++) {
-    int j = i - __kmp_affinity_compact;
-    if (aa->childNums[j] < bb->childNums[j])
-      return -1;
-    if (aa->childNums[j] > bb->childNums[j])
-      return 1;
-  }
-  return 0;
-}
-
 static void __kmp_aux_affinity_initialize(void) {
   if (__kmp_affinity_masks != NULL) {
     KMP_ASSERT(__kmp_affin_fullMask != NULL);
     return;
   }
 
   // Create the "full" mask - this defines all of the processors that we
   // consider to be in the machine model. If respect is set, then it is the

 #if KMP_OS_WINDOWS
       // Set the process affinity mask since threads' affinity
       // masks must be subset of process mask in Windows* OS
       __kmp_affin_fullMask->set_process_affinity(true);
 #endif
     }
   }
 
-  if (__kmp_affinity_gran == affinity_gran_tile &&
-      // check if user's request is valid
-      __kmp_affinity_dispatch->get_api_type() == KMPAffinity::NATIVE_OS) {
-    KMP_WARNING(AffTilesNoHWLOC, "KMP_AFFINITY");
-    __kmp_affinity_gran = affinity_gran_package;
-  }
-
-  int depth = -1;
   kmp_i18n_id_t msg_id = kmp_i18n_null;
 
   // For backward compatibility, setting KMP_CPUINFO_FILE =>
   // KMP_TOPOLOGY_METHOD=cpuinfo
   if ((__kmp_cpuinfo_file != NULL) &&
       (__kmp_affinity_top_method == affinity_top_method_all)) {
     __kmp_affinity_top_method = affinity_top_method_cpuinfo;
   }
 
+  bool success = false;
   if (__kmp_affinity_top_method == affinity_top_method_all) {
-    // In the default code path, errors are not fatal - we just try using
-    // another method. We only emit a warning message if affinity is on, or the
-    // verbose flag is set, and the nowarnings flag was not set.
-    const char *file_name = NULL;
-    int line = 0;
+// verbose flag is set, an the nowarnings flag was not set.
 #if KMP_USE_HWLOC
-    if (depth < 0 &&
+    if (!success &&
         __kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
-      if (__kmp_affinity_verbose) {
-        KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
-      }
       if (!__kmp_hwloc_error) {
-        depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id);
-        if (depth == 0) {
-          KMP_EXIT_AFF_NONE;
+        success = __kmp_affinity_create_hwloc_map(&msg_id);
+        if (!success && __kmp_affinity_verbose) {
+          KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
         }
       } else if (__kmp_affinity_verbose) {
         KMP_INFORM(AffIgnoringHwloc, "KMP_AFFINITY");
       }
     }
 #endif
 
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
-
-    if (depth < 0) {
-      if (__kmp_affinity_verbose) {
-        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
+    if (!success) {
+      success = __kmp_affinity_create_x2apicid_map(&msg_id);
+      if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
+        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
       }
-
-      file_name = NULL;
-      depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
-      if (depth == 0) {
-        KMP_EXIT_AFF_NONE;
-      }
-
-      if (depth < 0) {
-        if (__kmp_affinity_verbose) {
-          if (msg_id != kmp_i18n_null) {
-            KMP_INFORM(AffInfoStrStr, "KMP_AFFINITY",
-                       __kmp_i18n_catgets(msg_id),
-                       KMP_I18N_STR(DecodingLegacyAPIC));
-          } else {
-            KMP_INFORM(AffInfoStr, "KMP_AFFINITY",
-                       KMP_I18N_STR(DecodingLegacyAPIC));
-          }
-        }
-
-        file_name = NULL;
-        depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
-        if (depth == 0) {
-          KMP_EXIT_AFF_NONE;
-        }
-      }
-    }
-
+    if (!success) {
+      success = __kmp_affinity_create_apicid_map(&msg_id);
+      if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
+        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
 
 #if KMP_OS_LINUX
-
-    if (depth < 0) {
-      if (__kmp_affinity_verbose) {
-        if (msg_id != kmp_i18n_null) {
-          KMP_INFORM(AffStrParseFilename, "KMP_AFFINITY",
-                     __kmp_i18n_catgets(msg_id), "/proc/cpuinfo");
-        } else {
-          KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "/proc/cpuinfo");
-        }
-      }
-
-      kmp_safe_raii_file_t f("/proc/cpuinfo", "r");
-      depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
-      if (depth == 0) {
-        KMP_EXIT_AFF_NONE;
+    if (!success) {
+      int line = 0;
+      success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
+      if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
+        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
       }
     }
-
 #endif /* KMP_OS_LINUX */
 
 #if KMP_GROUP_AFFINITY
-
-    if ((depth < 0) && (__kmp_num_proc_groups > 1)) {
-      if (__kmp_affinity_verbose) {
-        KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
+    if (!success && (__kmp_num_proc_groups > 1)) {
+      success = __kmp_affinity_create_proc_group_map(&msg_id);
+      if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
+        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
       }
-
-      depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
-      KMP_ASSERT(depth != 0);
     }
-
 #endif /* KMP_GROUP_AFFINITY */
 
-    if (depth < 0) {
-      if (__kmp_affinity_verbose && (msg_id != kmp_i18n_null)) {
-        if (file_name == NULL) {
-          KMP_INFORM(UsingFlatOS, __kmp_i18n_catgets(msg_id));
-        } else if (line == 0) {
-          KMP_INFORM(UsingFlatOSFile, file_name, __kmp_i18n_catgets(msg_id));
-        } else {
-          KMP_INFORM(UsingFlatOSFileLine, file_name, line,
-                     __kmp_i18n_catgets(msg_id));
-        }
-      }
-      // FIXME - print msg if msg_id = kmp_i18n_null ???
-
-      file_name = "";
-      depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
-      if (depth == 0) {
-        KMP_EXIT_AFF_NONE;
+    if (!success) {
+      success = __kmp_affinity_create_flat_map(&msg_id);
+      if (!success && __kmp_affinity_verbose && msg_id != kmp_i18n_null) {
+        KMP_INFORM(AffInfoStr, "KMP_AFFINITY", __kmp_i18n_catgets(msg_id));
       }
-      KMP_ASSERT(depth > 0);
-      KMP_ASSERT(address2os != NULL);
+      KMP_ASSERT(success);
     }
   }
 
+// 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_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;
+    success = __kmp_affinity_create_hwloc_map(&msg_id);
+    if (!success) {
+      KMP_ASSERT(msg_id != kmp_i18n_null);
+      KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
     }
   }
 #endif // KMP_USE_HWLOC
 
-  // If the user has specified that a particular 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 ||
            __kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
-    }
-
-    depth = __kmp_affinity_create_x2apicid_map(&address2os, &msg_id);
-    if (depth == 0) {
-      KMP_EXIT_AFF_NONE;
-    }
-    if (depth < 0) {
+    success = __kmp_affinity_create_x2apicid_map(&msg_id);
+    if (!success) {
       KMP_ASSERT(msg_id != kmp_i18n_null);
       KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
     }
   } else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
-    }
-
-    depth = __kmp_affinity_create_apicid_map(&address2os, &msg_id);
-    if (depth == 0) {
-      KMP_EXIT_AFF_NONE;
-    }
-    if (depth < 0) {
+    success = __kmp_affinity_create_apicid_map(&msg_id);
+    if (!success) {
       KMP_ASSERT(msg_id != kmp_i18n_null);
       KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
     }
   }
-
 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
 
   else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
-    const char *filename;
-    const char *env_var = nullptr;
-    if (__kmp_cpuinfo_file != NULL) {
-      filename = __kmp_cpuinfo_file;
-      env_var = "KMP_CPUINFO_FILE";
-    } else {
-      filename = "/proc/cpuinfo";
-    }
-
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
-    }
-
-    kmp_safe_raii_file_t f(filename, "r", env_var);
     int line = 0;
-    depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f);
-    if (depth < 0) {
+    success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
+    if (!success) {
       KMP_ASSERT(msg_id != kmp_i18n_null);
+      const char *filename = __kmp_cpuinfo_get_filename();
       if (line > 0) {
         KMP_FATAL(FileLineMsgExiting, filename, line,
                   __kmp_i18n_catgets(msg_id));
       } else {
         KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
       }
     }
-    if (__kmp_affinity_type == affinity_none) {
-      KMP_ASSERT(depth == 0);
-      KMP_EXIT_AFF_NONE;
-    }
   }
 
 #if KMP_GROUP_AFFINITY
-
   else if (__kmp_affinity_top_method == affinity_top_method_group) {
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
-    }
-
-    depth = __kmp_affinity_create_proc_group_map(&address2os, &msg_id);
-    KMP_ASSERT(depth != 0);
-    if (depth < 0) {
+    success = __kmp_affinity_create_proc_group_map(&msg_id);
+    KMP_ASSERT(success);
+    if (!success) {
       KMP_ASSERT(msg_id != kmp_i18n_null);
       KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
     }
   }
-
 #endif /* KMP_GROUP_AFFINITY */
 
   else if (__kmp_affinity_top_method == affinity_top_method_flat) {
-    if (__kmp_affinity_verbose) {
-      KMP_INFORM(AffUsingFlatOS, "KMP_AFFINITY");
-    }
-
-    depth = __kmp_affinity_create_flat_map(&address2os, &msg_id);
-    if (depth == 0) {
-      KMP_EXIT_AFF_NONE;
-    }
+    success = __kmp_affinity_create_flat_map(&msg_id);
     // should not fail
-    KMP_ASSERT(depth > 0);
-    KMP_ASSERT(address2os != NULL);
+    KMP_ASSERT(success);
   }
 
-#if KMP_USE_HIER_SCHED
-  __kmp_dispatch_set_hierarchy_values();
-#endif
-
-  if (address2os == NULL) {
+  // Early exit if topology could not be created
+  if (!__kmp_topology) {
     if (KMP_AFFINITY_CAPABLE() &&
         (__kmp_affinity_verbose ||
          (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) {
       KMP_WARNING(ErrorInitializeAffinity);
     }
+    if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
+        __kmp_ncores > 0) {
+      __kmp_topology = kmp_topology_t::allocate(0, 0, NULL);
+      __kmp_topology->canonicalize(nPackages, nCoresPerPkg,
+                                   __kmp_nThreadsPerCore, __kmp_ncores);
+      if (__kmp_affinity_verbose) {
+        __kmp_topology->print("KMP_AFFINITY");
+      }
+    }
     __kmp_affinity_type = affinity_none;
     __kmp_create_affinity_none_places();
+#if KMP_USE_HIER_SCHED
+    __kmp_dispatch_set_hierarchy_values();
+#endif
     KMP_AFFINITY_DISABLE();
     return;
   }
 
-  if (__kmp_affinity_gran == affinity_gran_tile
-#if KMP_USE_HWLOC
-      && __kmp_tile_depth == 0
+  // Canonicalize, print (if requested), apply KMP_HW_SUBSET, and
+  // initialize other data structures which depend on the topology
+  __kmp_topology->canonicalize();
+  if (__kmp_affinity_verbose)
+    __kmp_topology->print("KMP_AFFINITY");
+  bool filtered = __kmp_topology->filter_hw_subset();
+  if (filtered && __kmp_affinity_verbose)
+    __kmp_topology->print("KMP_HW_SUBSET");
+  machine_hierarchy.init(__kmp_topology->get_num_hw_threads());
+  KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
+  // If KMP_AFFINITY=none, then only create the single "none" place
+  // which is the process's initial affinity mask or the number of
+  // hardware threads depending on respect,norespect
+  if (__kmp_affinity_type == affinity_none) {
+    __kmp_create_affinity_none_places();
+#if KMP_USE_HIER_SCHED
+    __kmp_dispatch_set_hierarchy_values();
 #endif
-  ) {
-    // tiles requested but not detected, warn user on this
-    KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY");
+    return;
   }
-
-  __kmp_apply_thread_places(&address2os, depth);
+  int depth = __kmp_topology->get_depth();
 
   // Create the table of masks, indexed by thread Id.
   unsigned maxIndex;
   unsigned numUnique;
-  kmp_affin_mask_t *osId2Mask =
-      __kmp_create_masks(&maxIndex, &numUnique, address2os, __kmp_avail_proc);
+  kmp_affin_mask_t *osId2Mask = __kmp_create_masks(&maxIndex, &numUnique);
   if (__kmp_affinity_gran_levels == 0) {
     KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
   }
 
-  // Set the childNums vector in all Address objects. This must be done before
-  // we can sort using __kmp_affinity_cmp_Address_child_num(), which takes into
-  // account the setting of __kmp_affinity_compact.
-  __kmp_affinity_assign_child_nums(address2os, __kmp_avail_proc);
-
   switch (__kmp_affinity_type) {
 
   case affinity_explicit:
     KMP_DEBUG_ASSERT(__kmp_affinity_proclist != NULL);
     if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
       __kmp_affinity_process_proclist(
           &__kmp_affinity_masks, &__kmp_affinity_num_masks,
           __kmp_affinity_proclist, osId2Mask, maxIndex);
     } else {
       __kmp_affinity_process_placelist(
           &__kmp_affinity_masks, &__kmp_affinity_num_masks,
           __kmp_affinity_proclist, osId2Mask, maxIndex);
     }
     if (__kmp_affinity_num_masks == 0) {
       if (__kmp_affinity_verbose ||
           (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) {
         KMP_WARNING(AffNoValidProcID);
       }
       __kmp_affinity_type = affinity_none;
       __kmp_create_affinity_none_places();
       return;
     }
     break;
 
-    // The other affinity types rely on sorting the Addresses according to some
-    // permutation of the machine topology tree. Set __kmp_affinity_compact and
-    // __kmp_affinity_offset appropriately, then jump to a common code fragment
-    // to do the sort and create the array of affinity masks.
-
+  // The other affinity types rely on sorting the hardware threads according to
+  // some permutation of the machine topology tree. Set __kmp_affinity_compact
+  // and __kmp_affinity_offset appropriately, then jump to a common code
+  // fragment to do the sort and create the array of affinity masks.
   case affinity_logical:
     __kmp_affinity_compact = 0;
     if (__kmp_affinity_offset) {
       __kmp_affinity_offset =
           __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
     }
-    goto sortAddresses;
+    goto sortTopology;
 
   case affinity_physical:
     if (__kmp_nThreadsPerCore > 1) {
       __kmp_affinity_compact = 1;
       if (__kmp_affinity_compact >= depth) {
         __kmp_affinity_compact = 0;
       }
     } else {
       __kmp_affinity_compact = 0;
     }
     if (__kmp_affinity_offset) {
       __kmp_affinity_offset =
           __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc;
     }
-    goto sortAddresses;
+    goto sortTopology;
 
   case affinity_scatter:
     if (__kmp_affinity_compact >= depth) {
       __kmp_affinity_compact = 0;
     } else {
       __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact;
     }
-    goto sortAddresses;
+    goto sortTopology;
 
   case affinity_compact:
     if (__kmp_affinity_compact >= depth) {
       __kmp_affinity_compact = depth - 1;
     }
-    goto sortAddresses;
+    goto sortTopology;
 
   case affinity_balanced:
     if (depth <= 1) {
       if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
         KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
       }
       __kmp_affinity_type = affinity_none;
       __kmp_create_affinity_none_places();
       return;
-    } else if (!__kmp_affinity_uniform_topology()) {
+    } else if (!__kmp_topology->is_uniform()) {
       // Save the depth for further usage
       __kmp_aff_depth = depth;
 
-      int core_level = __kmp_affinity_find_core_level(
-          address2os, __kmp_avail_proc, depth - 1);
-      int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
-                                                 depth - 1, core_level);
+      int core_level =
+          __kmp_affinity_find_core_level(__kmp_avail_proc, depth - 1);
+      int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc, depth - 1,
+                                                 core_level);
       int maxprocpercore = __kmp_affinity_max_proc_per_core(
-          address2os, __kmp_avail_proc, depth - 1, core_level);
+          __kmp_avail_proc, depth - 1, core_level);
 
       int nproc = ncores * maxprocpercore;
       if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
         if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
           KMP_WARNING(AffBalancedNotAvail, "KMP_AFFINITY");
         }
         __kmp_affinity_type = affinity_none;
         return;
       }
 
       procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
       for (int i = 0; i < nproc; i++) {
         procarr[i] = -1;
       }
 
       int lastcore = -1;
       int inlastcore = 0;
       for (int i = 0; i < __kmp_avail_proc; i++) {
-        int proc = address2os[i].second;
-        int core =
-            __kmp_affinity_find_core(address2os, i, depth - 1, core_level);
+        int proc = __kmp_topology->at(i).os_id;
+        int core = __kmp_affinity_find_core(i, depth - 1, core_level);
 
         if (core == lastcore) {
           inlastcore++;
         } else {
           inlastcore = 0;
         }
         lastcore = core;
 
         procarr[core * maxprocpercore + inlastcore] = proc;
       }
     }
     if (__kmp_affinity_compact >= depth) {
       __kmp_affinity_compact = depth - 1;
     }
 
-  sortAddresses:
+  sortTopology:
     // Allocate the gtid->affinity mask table.
     if (__kmp_affinity_dups) {
       __kmp_affinity_num_masks = __kmp_avail_proc;
     } else {
       __kmp_affinity_num_masks = numUnique;
     }
 
     if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
         (__kmp_affinity_num_places > 0) &&
         ((unsigned)__kmp_affinity_num_places < __kmp_affinity_num_masks)) {
       __kmp_affinity_num_masks = __kmp_affinity_num_places;
     }
 
     KMP_CPU_ALLOC_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks);
 
-    // Sort the address2os table according to the current setting of
+    // Sort the topology table according to the current setting of
     // __kmp_affinity_compact, then fill out __kmp_affinity_masks.
-    qsort(address2os, __kmp_avail_proc, sizeof(*address2os),
-          __kmp_affinity_cmp_Address_child_num);
+    __kmp_topology->sort_compact();
     {
       int i;
       unsigned j;
-      for (i = 0, j = 0; i < __kmp_avail_proc; i++) {
-        if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) {
+      int num_hw_threads = __kmp_topology->get_num_hw_threads();
+      for (i = 0, j = 0; i < num_hw_threads; i++) {
+        if ((!__kmp_affinity_dups) && (!__kmp_topology->at(i).leader)) {
           continue;
         }
-        unsigned osId = address2os[i].second;
+        int osId = __kmp_topology->at(i).os_id;
+
         kmp_affin_mask_t *src = KMP_CPU_INDEX(osId2Mask, osId);
         kmp_affin_mask_t *dest = KMP_CPU_INDEX(__kmp_affinity_masks, j);
         KMP_ASSERT(KMP_CPU_ISSET(osId, src));
         KMP_CPU_COPY(dest, src);
         if (++j >= __kmp_affinity_num_masks) {
           break;
         }
       }
       KMP_DEBUG_ASSERT(j == __kmp_affinity_num_masks);
     }
+    // Sort the topology back using ids
+    __kmp_topology->sort_ids();
     break;
 
   default:
     KMP_ASSERT2(0, "Unexpected affinity setting");
   }
 
   KMP_CPU_FREE_ARRAY(osId2Mask, maxIndex + 1);
-  machine_hierarchy.init(address2os, __kmp_avail_proc);
 }
-#undef KMP_EXIT_AFF_NONE
 
 void __kmp_affinity_initialize(void) {
   // Much of the code above was written assuming that if a machine was not
   // affinity capable, then __kmp_affinity_type == affinity_none.  We now
   // explicitly represent this as __kmp_affinity_type == affinity_disabled.
   // There are too many checks for __kmp_affinity_type == affinity_none
   // in this code.  Instead of trying to change them all, check if
   // __kmp_affinity_type == affinity_disabled, and if so, slam it with
void __kmp_affinity_uninitialize(void) {
   }
   __kmp_affinity_num_masks = 0;
   __kmp_affinity_type = affinity_default;
   __kmp_affinity_num_places = 0;
   if (__kmp_affinity_proclist != NULL) {
     __kmp_free(__kmp_affinity_proclist);
     __kmp_affinity_proclist = NULL;
   }
-  if (address2os != NULL) {
-    __kmp_free(address2os);
-    address2os = NULL;
-  }
   if (procarr != NULL) {
     __kmp_free(procarr);
     procarr = NULL;
   }
 #if KMP_USE_HWLOC
   if (__kmp_hwloc_topology != NULL) {
     hwloc_topology_destroy(__kmp_hwloc_topology);
     __kmp_hwloc_topology = NULL;
   }
 #endif
+  if (__kmp_hw_subset) {
+    kmp_hw_subset_t::deallocate(__kmp_hw_subset);
+    __kmp_hw_subset = nullptr;
+  }
+  if (__kmp_topology) {
+    kmp_topology_t::deallocate(__kmp_topology);
+    __kmp_topology = nullptr;
+  }
   KMPAffinity::destroy_api();
 }
 
 void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
   if (!KMP_AFFINITY_CAPABLE()) {
     return;
   }
 

 
 // Dynamic affinity settings - Affinity balanced
 void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
   KMP_DEBUG_ASSERT(th);
   bool fine_gran = true;
   int tid = th->th.th_info.ds.ds_tid;
 
   switch (__kmp_affinity_gran) {
-  case affinity_gran_fine:
-  case affinity_gran_thread:
+  case KMP_HW_THREAD:
     break;
-  case affinity_gran_core:
+  case KMP_HW_CORE:
     if (__kmp_nThreadsPerCore > 1) {
       fine_gran = false;
     }
     break;
-  case affinity_gran_package:
+  case KMP_HW_SOCKET:
     if (nCoresPerPkg > 1) {
       fine_gran = false;
     }
     break;
   default:
     fine_gran = false;
   }
 
-  if (__kmp_affinity_uniform_topology()) {
+  if (__kmp_topology->is_uniform()) {
     int coreID;
     int threadID;
     // Number of hyper threads per core in HT machine
     int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
     // Number of cores
     int ncores = __kmp_ncores;
     if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
       __kmp_nth_per_core = __kmp_avail_proc / nPackages;
       ncores = nPackages;
     }
     // How many threads will be bound to each core
     int chunk = nthreads / ncores;
     // How many cores will have an additional thread bound to it - "big cores"
     int big_cores = nthreads % ncores;
     // Number of threads on the big cores
     int big_nth = (chunk + 1) * big_cores;
     if (tid < big_nth) {
       coreID = tid / (chunk + 1);
       threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
     } else { // tid >= big_nth
       coreID = (tid - big_cores) / chunk;
       threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
     }
-
     KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
                       "Illegal set affinity operation when not capable");
 
     kmp_affin_mask_t *mask = th->th.th_affin_mask;
     KMP_CPU_ZERO(mask);
 
     if (fine_gran) {
-      int osID = address2os[coreID * __kmp_nth_per_core + threadID].second;
+      int osID =
+          __kmp_topology->at(coreID * __kmp_nth_per_core + threadID).os_id;
       KMP_CPU_SET(osID, mask);
     } else {
       for (int i = 0; i < __kmp_nth_per_core; i++) {
         int osID;
-        osID = address2os[coreID * __kmp_nth_per_core + i].second;
+        osID = __kmp_topology->at(coreID * __kmp_nth_per_core + i).os_id;
         KMP_CPU_SET(osID, mask);
       }
     }
     if (__kmp_affinity_verbose) {
       char buf[KMP_AFFIN_MASK_PRINT_LEN];
       __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
       KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(),
                  __kmp_gettid(), tid, buf);
     }
     __kmp_set_system_affinity(mask, TRUE);
   } else { // Non-uniform topology
 
     kmp_affin_mask_t *mask = th->th.th_affin_mask;
     KMP_CPU_ZERO(mask);
 
-    int core_level = __kmp_affinity_find_core_level(
-        address2os, __kmp_avail_proc, __kmp_aff_depth - 1);
-    int ncores = __kmp_affinity_compute_ncores(address2os, __kmp_avail_proc,
+    int core_level =
+        __kmp_affinity_find_core_level(__kmp_avail_proc, __kmp_aff_depth - 1);
+    int ncores = __kmp_affinity_compute_ncores(__kmp_avail_proc,
                                                __kmp_aff_depth - 1, core_level);
     int nth_per_core = __kmp_affinity_max_proc_per_core(
-        address2os, __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
+        __kmp_avail_proc, __kmp_aff_depth - 1, core_level);
 
     // For performance gain consider the special case nthreads ==
     // __kmp_avail_proc
     if (nthreads == __kmp_avail_proc) {
       if (fine_gran) {
-        int osID = address2os[tid].second;
+        int osID = __kmp_topology->at(tid).os_id;
         KMP_CPU_SET(osID, mask);
       } else {
-        int core = __kmp_affinity_find_core(address2os, tid,
-                                            __kmp_aff_depth - 1, core_level);
+        int core =
+            __kmp_affinity_find_core(tid, __kmp_aff_depth - 1, core_level);
         for (int i = 0; i < __kmp_avail_proc; i++) {
-          int osID = address2os[i].second;
-          if (__kmp_affinity_find_core(address2os, i, __kmp_aff_depth - 1,
-                                       core_level) == core) {
+          int osID = __kmp_topology->at(i).os_id;
+          if (__kmp_affinity_find_core(i, __kmp_aff_depth - 1, core_level) ==
+              core) {
             KMP_CPU_SET(osID, mask);
           }
         }
       }
     } else if (nthreads <= ncores) {
 
       int core = 0;
       for (int i = 0; i < ncores; i++) {

openmp/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;
 kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity = NULL;
 kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity = NULL;
 #endif /* KMP_OS_WINDOWS */
 
 size_t __kmp_affin_mask_size = 0;
 enum affinity_type __kmp_affinity_type = affinity_default;
-enum affinity_gran __kmp_affinity_gran = affinity_gran_default;
+kmp_hw_t __kmp_affinity_gran = KMP_HW_UNKNOWN;
 int __kmp_affinity_gran_levels = -1;
 int __kmp_affinity_dups = TRUE;
 enum affinity_top_method __kmp_affinity_top_method =
     affinity_top_method_default;
 int __kmp_affinity_compact = 0;
 int __kmp_affinity_offset = 0;
 int __kmp_affinity_verbose = FALSE;
 int __kmp_affinity_warnings = TRUE;
 int __kmp_affinity_respect_mask = affinity_respect_mask_default;
 char *__kmp_affinity_proclist = NULL;
 kmp_affin_mask_t *__kmp_affinity_masks = NULL;
 unsigned __kmp_affinity_num_masks = 0;
 
 char *__kmp_cpuinfo_file = NULL;
 
 #endif /* KMP_AFFINITY_SUPPORTED */
 
 kmp_nested_proc_bind_t __kmp_nested_proc_bind = {NULL, 0, 0};
 int __kmp_affinity_num_places = 0;
 int __kmp_display_affinity = FALSE;
 char *__kmp_affinity_format = NULL;
 
-kmp_hws_item_t __kmp_hws_socket = {0, 0};
-kmp_hws_item_t __kmp_hws_die = {0, 0};
-kmp_hws_item_t __kmp_hws_node = {0, 0};
-kmp_hws_item_t __kmp_hws_tile = {0, 0};
-kmp_hws_item_t __kmp_hws_core = {0, 0};
-kmp_hws_item_t __kmp_hws_proc = {0, 0};
-int __kmp_hws_requested = 0;
-int __kmp_hws_abs_flag = 0; // absolute or per-item number requested
-
 kmp_int32 __kmp_default_device = 0;
 
 kmp_tasking_mode_t __kmp_tasking_mode = tskm_task_teams;
 kmp_int32 __kmp_max_task_priority = 0;
 kmp_uint64 __kmp_taskloop_min_tasks = 0;
 
 int __kmp_memkind_available = 0;
 omp_allocator_handle_t const omp_null_allocator = NULL;

openmp/runtime/src/kmp_settings.cpp

 // If KMP_AFFINITY is specified without a type, then
 // __kmp_affinity_notype should point to its setting.
 static kmp_setting_t *__kmp_affinity_notype = NULL;
 
 static void __kmp_parse_affinity_env(char const *name, char const *value,
                                      enum affinity_type *out_type,
                                      char **out_proclist, int *out_verbose,
                                      int *out_warn, int *out_respect,
-                                     enum affinity_gran *out_gran,
-                                     int *out_gran_levels, int *out_dups,
-                                     int *out_compact, int *out_offset) {
+                                     kmp_hw_t *out_gran, int *out_gran_levels,
+                                     int *out_dups, int *out_compact,
+                                     int *out_offset) {
   char *buffer = NULL; // Copy of env var value.
   char *buf = NULL; // Buffer for strtok_r() function.
   char *next = NULL; // end of token / start of next.
   const char *start; // start of current token (for err msgs)
   int count = 0; // Counter of parsed integer numbers.
   int number[2]; // Parsed numbers.
 
   // Guards.
   int type = 0;
   int proclist = 0;
   int verbose = 0;
   int warnings = 0;
   int respect = 0;
   int gran = 0;
   int dups = 0;
+  bool set = false;
 
   KMP_ASSERT(value != NULL);
 
   if (TCR_4(__kmp_init_middle)) {
     KMP_WARNING(EnvMiddleWarn, name);
     __kmp_env_toPrint(name, 0);
     return;
   }
if (__kmp_match_str("none", buf, CCAST(const char **, &next))) {
       if (*next != '=') {
         EMIT_WARN(TRUE, (AffInvalidParam, name, start));
         continue;
       }
       next++; // skip '='
       SKIP_WS(next);
 
       buf = next;
-      if (__kmp_match_str("fine", buf, CCAST(const char **, &next))) {
-        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);
+
+      // Try any hardware topology type for granularity
+      KMP_FOREACH_HW_TYPE(type) {
+        const char *name = __kmp_hw_get_keyword(type);
+        if (__kmp_match_str(name, buf, CCAST(const char **, &next))) {
+          set_gran(type, -1);
           buf = next;
-#endif
-      } else if (__kmp_match_str("die", buf, CCAST(const char **, &next))) {
-        set_gran(affinity_gran_die, -1);
+          set = true;
+          break;
+        }
+      }
+      if (!set) {
+        // Support older names for different granularity layers
+        if (__kmp_match_str("fine", buf, CCAST(const char **, &next))) {
+          set_gran(KMP_HW_THREAD, -1);
           buf = next;
-      } else if (__kmp_match_str("package", buf, CCAST(const char **, &next))) {
-        set_gran(affinity_gran_package, -1);
+          set = true;
+        } else if (__kmp_match_str("package", buf,
+                                   CCAST(const char **, &next))) {
+          set_gran(KMP_HW_SOCKET, -1);
           buf = next;
+          set = true;
         } else if (__kmp_match_str("node", buf, CCAST(const char **, &next))) {
-        set_gran(affinity_gran_node, -1);
+          set_gran(KMP_HW_NUMA, -1);
           buf = next;
+          set = true;
 #if KMP_GROUP_AFFINITY
         } else if (__kmp_match_str("group", buf, CCAST(const char **, &next))) {
-        set_gran(affinity_gran_group, -1);
+          set_gran(KMP_HW_PROC_GROUP, -1);
           buf = next;
+          set = true;
 #endif /* KMP_GROUP AFFINITY */
         } else if ((*buf >= '0') && (*buf <= '9')) {
           int n;
           next = buf;
           SKIP_DIGITS(next);
           n = __kmp_str_to_int(buf, *next);
           KMP_ASSERT(n >= 0);
           buf = next;
-        set_gran(affinity_gran_default, n);
+          set_gran(KMP_HW_UNKNOWN, n);
+          set = true;
         } else {
           EMIT_WARN(TRUE, (AffInvalidParam, name, start));
           continue;
         }
+      }
     } else if (__kmp_match_str("proclist", buf, CCAST(const char **, &next))) {
       char *temp_proclist;
 
       SKIP_WS(next);
       if (*next != '=') {
         EMIT_WARN(TRUE, (AffInvalidParam, name, start));
         continue;
       }
#undef set_granularity
   case affinity_balanced: {
     if (count > 0) {
       *out_compact = number[0];
     }
     if (count > 1) {
       *out_offset = number[1];
     }
 
-    if (__kmp_affinity_gran == affinity_gran_default) {
+    if (__kmp_affinity_gran == KMP_HW_UNKNOWN) {
 #if KMP_MIC_SUPPORTED
       if (__kmp_mic_type != non_mic) {
         if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
           KMP_WARNING(AffGranUsing, "KMP_AFFINITY", "fine");
         }
-        __kmp_affinity_gran = affinity_gran_fine;
+        __kmp_affinity_gran = KMP_HW_THREAD;
       } else
 #endif
       {
         if (__kmp_affinity_verbose || __kmp_affinity_warnings) {
           KMP_WARNING(AffGranUsing, "KMP_AFFINITY", "core");
         }
-        __kmp_affinity_gran = affinity_gran_core;
+        __kmp_affinity_gran = KMP_HW_CORE;
       }
     }
   } break;
   case affinity_scatter:
   case affinity_compact: {
     if (count > 0) {
       *out_compact = number[0];
     }
if (__kmp_affinity_warnings) {
     __kmp_str_buf_print(buffer, "%s,", "nowarnings");
   }
   if (KMP_AFFINITY_CAPABLE()) {
     if (__kmp_affinity_respect_mask) {
       __kmp_str_buf_print(buffer, "%s,", "respect");
     } else {
       __kmp_str_buf_print(buffer, "%s,", "norespect");
     }
-    switch (__kmp_affinity_gran) {
-    case affinity_gran_default:
-      __kmp_str_buf_print(buffer, "%s", "granularity=default,");
-      break;
-    case affinity_gran_fine:
-      __kmp_str_buf_print(buffer, "%s", "granularity=fine,");
-      break;
-    case affinity_gran_thread:
-      __kmp_str_buf_print(buffer, "%s", "granularity=thread,");
-      break;
-    case affinity_gran_core:
-      __kmp_str_buf_print(buffer, "%s", "granularity=core,");
-      break;
-    case affinity_gran_package:
-      __kmp_str_buf_print(buffer, "%s", "granularity=package,");
-      break;
-    case affinity_gran_node:
-      __kmp_str_buf_print(buffer, "%s", "granularity=node,");
-      break;
-#if KMP_GROUP_AFFINITY
-    case affinity_gran_group:
-      __kmp_str_buf_print(buffer, "%s", "granularity=group,");
-      break;
-#endif /* KMP_GROUP_AFFINITY */
-    }
+    __kmp_str_buf_print(buffer, "granularity=%s,",
+                        __kmp_hw_get_keyword(__kmp_affinity_gran, false));
   }
   if (!KMP_AFFINITY_CAPABLE()) {
     __kmp_str_buf_print(buffer, "%s", "disabled");
   } else
     switch (__kmp_affinity_type) {
     case affinity_none:
       __kmp_str_buf_print(buffer, "%s", "none");
       break;
static void __kmp_stg_parse_gomp_cpu_affinity(char const *name,
   __kmp_env_toPrint(name, 1);
 
   if (__kmp_parse_affinity_proc_id_list(name, value, &next, &temp_proclist)) {
     SKIP_WS(next);
     if (*next == '\0') {
       // GOMP_CPU_AFFINITY => granularity=fine,explicit,proclist=...
       __kmp_affinity_proclist = temp_proclist;
       __kmp_affinity_type = affinity_explicit;
-      __kmp_affinity_gran = affinity_gran_fine;
+      __kmp_affinity_gran = KMP_HW_THREAD;
       __kmp_nested_proc_bind.bind_types[0] = proc_bind_intel;
     } else {
       KMP_WARNING(AffSyntaxError, name);
       if (temp_proclist != NULL) {
         KMP_INTERNAL_FREE((void *)temp_proclist);
       }
     }
   } else {
for (;;) {
     retlist[len] = '\0';
     *place_list = retlist;
   }
   return TRUE;
 }
 
 static void __kmp_stg_parse_places(char const *name, char const *value,
                                    void *data) {
+  struct kmp_place_t {
+    const char *name;
+    kmp_hw_t type;
+  };
   int count;
+  bool set = false;
   const char *scan = value;
   const char *next = scan;
   const char *kind = "\"threads\"";
+  kmp_place_t std_places[] = {{"threads", KMP_HW_THREAD},
+                              {"cores", KMP_HW_CORE},
+                              {"numa_domains", KMP_HW_NUMA},
+                              {"ll_caches", KMP_HW_LLC},
+                              {"sockets", KMP_HW_SOCKET}};
   kmp_setting_t **rivals = (kmp_setting_t **)data;
   int rc;
 
   rc = __kmp_stg_check_rivals(name, value, rivals);
   if (rc) {
     return;
   }
 
-  if (__kmp_match_str("threads", scan, &next)) {
-    scan = next;
-    __kmp_affinity_type = affinity_compact;
-    __kmp_affinity_gran = affinity_gran_thread;
-    __kmp_affinity_dups = FALSE;
-    kind = "\"threads\"";
-  } else if (__kmp_match_str("cores", scan, &next)) {
+  // Standard choices
+  for (size_t i = 0; i < sizeof(std_places) / sizeof(std_places[0]); ++i) {
+    const kmp_place_t &place = std_places[i];
+    if (__kmp_match_str(place.name, scan, &next)) {
       scan = next;
       __kmp_affinity_type = affinity_compact;
-    __kmp_affinity_gran = affinity_gran_core;
+      __kmp_affinity_gran = place.type;
       __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("dice", scan, &next) ||
-             __kmp_match_str("dies", scan, &next)) {
-    scan = next;
-    __kmp_affinity_type = affinity_compact;
-    __kmp_affinity_gran = affinity_gran_die;
-    __kmp_affinity_dups = FALSE;
-    kind = "\"dice\"";
-  } else if (__kmp_match_str("sockets", scan, &next)) {
+      set = true;
+      break;
+    }
+  }
+  // Implementation choices for OMP_PLACES based on internal types
+  if (!set) {
+    KMP_FOREACH_HW_TYPE(type) {
+      const char *name = __kmp_hw_get_keyword(type, true);
+      if (__kmp_match_str("unknowns", scan, &next))
+        continue;
+      if (__kmp_match_str(name, scan, &next)) {
         scan = next;
         __kmp_affinity_type = affinity_compact;
-    __kmp_affinity_gran = affinity_gran_package;
+        __kmp_affinity_gran = type;
         __kmp_affinity_dups = FALSE;
-    kind = "\"sockets\"";
-  } else {
+        set = true;
+        break;
+      }
+    }
+  }
+  if (!set) {
     if (__kmp_affinity_proclist != NULL) {
       KMP_INTERNAL_FREE((void *)__kmp_affinity_proclist);
       __kmp_affinity_proclist = NULL;
     }
     if (__kmp_parse_place_list(name, value, &__kmp_affinity_proclist)) {
       __kmp_affinity_type = affinity_explicit;
-      __kmp_affinity_gran = affinity_gran_fine;
+      __kmp_affinity_gran = KMP_HW_THREAD;
       __kmp_affinity_dups = FALSE;
     } else {
       // Syntax error fallback
       __kmp_affinity_type = affinity_compact;
-      __kmp_affinity_gran = affinity_gran_core;
+      __kmp_affinity_gran = KMP_HW_CORE;
       __kmp_affinity_dups = FALSE;
     }
     if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_default) {
       __kmp_nested_proc_bind.bind_types[0] = proc_bind_true;
     }
     return;
   }
+  if (__kmp_affinity_gran != KMP_HW_UNKNOWN) {
+    kind = __kmp_hw_get_keyword(__kmp_affinity_gran);
+  }
 
   if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_default) {
     __kmp_nested_proc_bind.bind_types[0] = proc_bind_true;
   }
 
   SKIP_WS(scan);
   if (*scan == '\0') {
     return;
if ((__kmp_nested_proc_bind.used == 0) ||
     int num;
     if (__kmp_affinity_num_masks > 0) {
       num = __kmp_affinity_num_masks;
     } else if (__kmp_affinity_num_places > 0) {
       num = __kmp_affinity_num_places;
     } else {
       num = 0;
     }
-    if (__kmp_affinity_gran == affinity_gran_thread) {
-      if (num > 0) {
-        __kmp_str_buf_print(buffer, "='threads(%d)'\n", num);
-      } else {
-        __kmp_str_buf_print(buffer, "='threads'\n");
-      }
-    } else if (__kmp_affinity_gran == affinity_gran_core) {
+    if (__kmp_affinity_gran != KMP_HW_UNKNOWN) {
+      const char *name = __kmp_hw_get_keyword(__kmp_affinity_gran, true);
       if (num > 0) {
-        __kmp_str_buf_print(buffer, "='cores(%d)' \n", num);
+        __kmp_str_buf_print(buffer, "='%s(%d)'\n", name, 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");
+        __kmp_str_buf_print(buffer, "='%s'\n", name);
       }
     } else {
       __kmp_str_buf_print(buffer, ": %s\n", KMP_I18N_STR(NotDefined));
     }
   } else {
     __kmp_str_buf_print(buffer, ": %s\n", KMP_I18N_STR(NotDefined));
   }
 }
case affinity_top_method_default:
     value = "default";
     break;
 
   case affinity_top_method_all:
     value = "all";
     break;
 
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
+  case affinity_top_method_x2apicid_1f:
+    value = "x2APIC id leaf 0x1f";
+    break;
+
   case affinity_top_method_x2apicid:
-    value = "x2APIC id";
+    value = "x2APIC id leaf 0xb";
     break;
 
   case affinity_top_method_apicid:
     value = "APIC id";
     break;
 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
 
 #if KMP_USE_HWLOC

 } // __kmp_stg_print_speculative_statsfile
 
 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
 
 #endif // KMP_USE_ADAPTIVE_LOCKS
 
 // -----------------------------------------------------------------------------
 // KMP_HW_SUBSET (was KMP_PLACE_THREADS)
+// 2s16c,2t => 2S16C,2T => 2S16C \0 2T
+
+// Return KMP_HW_SUBSET preferred hardware type in case a token is ambiguously
+// short. The original KMP_HW_SUBSET environment variable had single letters:
+// s, c, t for sockets, cores, threads repsectively.
+static kmp_hw_t __kmp_hw_subset_break_tie(const kmp_hw_t *possible,
+                                          size_t num_possible) {
+  for (size_t i = 0; i < num_possible; ++i) {
+    if (possible[i] == KMP_HW_THREAD)
+      return KMP_HW_THREAD;
+    else if (possible[i] == KMP_HW_CORE)
+      return KMP_HW_CORE;
+    else if (possible[i] == KMP_HW_SOCKET)
+      return KMP_HW_SOCKET;
+  }
+  return KMP_HW_UNKNOWN;
+}
+
+// Return hardware type from string or HW_UNKNOWN if string cannot be parsed
+// This algorithm is very forgiving to the user in that, the instant it can
+// reduce the search space to one, it assumes that is the topology level the
+// user wanted, even if it is misspelled later in the token.
+static kmp_hw_t __kmp_stg_parse_hw_subset_name(char const *token) {
+  size_t index, num_possible, token_length;
+  kmp_hw_t possible[KMP_HW_LAST];
+  const char *end;
+
+  // Find the end of the hardware token string
+  end = token;
+  token_length = 0;
+  while (isalnum(*end) || *end == '_') {
+    token_length++;
+    end++;
+  }
+
+  // Set the possibilities to all hardware types
+  num_possible = 0;
+  KMP_FOREACH_HW_TYPE(type) { possible[num_possible++] = type; }
+
+  // Eliminate hardware types by comparing the front of the token
+  // with hardware names
+  // In most cases, the first letter in the token will indicate exactly
+  // which hardware type is parsed, e.g., 'C' = Core
+  index = 0;
+  while (num_possible > 1 && index < token_length) {
+    size_t n = num_possible;
+    char token_char = (char)toupper(token[index]);
+    for (size_t i = 0; i < n; ++i) {
+      const char *s;
+      kmp_hw_t type = possible[i];
+      s = __kmp_hw_get_keyword(type, false);
+      if (index < KMP_STRLEN(s)) {
+        char c = (char)toupper(s[index]);
+        // Mark hardware types for removal when the characters do not match
+        if (c != token_char) {
+          possible[i] = KMP_HW_UNKNOWN;
+          num_possible--;
+        }
+      }
+    }
+    // Remove hardware types that this token cannot be
+    size_t start = 0;
+    for (size_t i = 0; i < n; ++i) {
+      if (possible[i] != KMP_HW_UNKNOWN) {
+        kmp_hw_t temp = possible[i];
+        possible[i] = possible[start];
+        possible[start] = temp;
+        start++;
+      }
+    }
+    KMP_ASSERT(start == num_possible);
+    index++;
+  }
+
+  // Attempt to break a tie if user has very short token
+  // (e.g., is 'T' tile or thread?)
+  if (num_possible > 1)
+    return __kmp_hw_subset_break_tie(possible, num_possible);
+  if (num_possible == 1)
+    return possible[0];
+  return KMP_HW_UNKNOWN;
+}
 
-// The longest observable sequence of items is
-// Socket-Node-Tile-Core-Thread
-// So, let's limit to 5 levels for now
+// The longest observable sequence of items can only be HW_LAST length
 // The input string is usually short enough, let's use 512 limit for now
-#define MAX_T_LEVEL 5
+#define MAX_T_LEVEL KMP_HW_LAST
 #define MAX_STR_LEN 512
 static void __kmp_stg_parse_hw_subset(char const *name, char const *value,
                                       void *data) {
   // Value example: 1s,5c@3,2T
   // Which means "use 1 socket, 5 cores with offset 3, 2 threads per core"
   kmp_setting_t **rivals = (kmp_setting_t **)data;
   if (strcmp(name, "KMP_PLACE_THREADS") == 0) {
     KMP_INFORM(EnvVarDeprecated, name, "KMP_HW_SUBSET");
   }
   if (__kmp_stg_check_rivals(name, value, rivals)) {
     return;
   }
 
   char *components[MAX_T_LEVEL];
   char const *digits = "0123456789";
   char input[MAX_STR_LEN];
   size_t len = 0, mlen = MAX_STR_LEN;
   int level = 0;
-  // Canonize the string (remove spaces, unify delimiters, etc.)
+  bool absolute = false;
+  // Canonicalize the string (remove spaces, unify delimiters, etc.)
   char *pos = CCAST(char *, value);
   while (*pos && mlen) {
     if (*pos != ' ') { // skip spaces
       if (len == 0 && *pos == ':') {
-        __kmp_hws_abs_flag = 1; // if the first symbol is ":", skip it
+        absolute = true;
       } else {
         input[len] = (char)(toupper(*pos));
         if (input[len] == 'X')
           input[len] = ','; // unify delimiters of levels
         if (input[len] == 'O' && strchr(digits, *(pos + 1)))
           input[len] = '@'; // unify delimiters of offset
         len++;
       }
     }
     mlen--;
     pos++;
   }
-  if (len == 0 || mlen == 0)
+  if (len == 0 || mlen == 0) {
     goto err; // contents is either empty or too long
+  }
   input[len] = '\0';
-  __kmp_hws_requested = 1; // mark that subset requested
   // Split by delimiter
   pos = input;
   components[level++] = pos;
   while ((pos = strchr(pos, ','))) {
     if (level >= MAX_T_LEVEL)
       goto err; // too many components provided
     *pos = '\0'; // modify input and avoid more copying
     components[level++] = ++pos; // expect something after ","
   }
+
+  __kmp_hw_subset = kmp_hw_subset_t::allocate();
+  if (absolute)
+    __kmp_hw_subset->set_absolute();
+
   // Check each component
   for (int i = 0; i < level; ++i) {
     int offset = 0;
     int num = atoi(components[i]); // each component should start with a number
+    if (num <= 0) {
+      goto err; // only positive integers are valid for count
+    }
     if ((pos = strchr(components[i], '@'))) {
       offset = atoi(pos + 1); // save offset
       *pos = '\0'; // cut the offset from the component
     }
     pos = components[i] + strspn(components[i], digits);
-    if (pos == components[i])
+    if (pos == components[i]) {
       goto err;
+    }
     // detect the component type
-    switch (*pos) {
-    case 'S': // Socket
-      if (__kmp_hws_socket.num > 0)
-        goto err; // duplicate is not allowed
-      __kmp_hws_socket.num = num;
-      __kmp_hws_socket.offset = offset;
-      break;
-    case 'N': // NUMA Node
-      if (__kmp_hws_node.num > 0)
-        goto err; // duplicate is not allowed
-      __kmp_hws_node.num = num;
-      __kmp_hws_node.offset = offset;
-      break;
-    case 'D': // Die
-      if (__kmp_hws_die.num > 0)
-        goto err; // duplicate is not allowed
-      __kmp_hws_die.num = num;
-      __kmp_hws_die.offset = offset;
-      break;
-    case 'L': // Cache
-      if (*(pos + 1) == '2') { // L2 - Tile
-        if (__kmp_hws_tile.num > 0)
-          goto err; // duplicate is not allowed
-        __kmp_hws_tile.num = num;
-        __kmp_hws_tile.offset = offset;
-      } else if (*(pos + 1) == '3') { // L3 - Socket
-        if (__kmp_hws_socket.num > 0 || __kmp_hws_die.num > 0)
-          goto err; // duplicate is not allowed
-        __kmp_hws_socket.num = num;
-        __kmp_hws_socket.offset = offset;
-      } else if (*(pos + 1) == '1') { // L1 - Core
-        if (__kmp_hws_core.num > 0)
-          goto err; // duplicate is not allowed
-        __kmp_hws_core.num = num;
-        __kmp_hws_core.offset = offset;
-      }
-      break;
-    case 'C': // Core (or Cache?)
-      if (*(pos + 1) != 'A') {
-        if (__kmp_hws_core.num > 0)
-          goto err; // duplicate is not allowed
-        __kmp_hws_core.num = num;
-        __kmp_hws_core.offset = offset;
-      } else { // Cache
-        char *d = pos + strcspn(pos, digits); // find digit
-        if (*d == '2') { // L2 - Tile
-          if (__kmp_hws_tile.num > 0)
-            goto err; // duplicate is not allowed
-          __kmp_hws_tile.num = num;
-          __kmp_hws_tile.offset = offset;
-        } else if (*d == '3') { // L3 - Socket
-          if (__kmp_hws_socket.num > 0 || __kmp_hws_die.num > 0)
-            goto err; // duplicate is not allowed
-          __kmp_hws_socket.num = num;
-          __kmp_hws_socket.offset = offset;
-        } else if (*d == '1') { // L1 - Core
-          if (__kmp_hws_core.num > 0)
-            goto err; // duplicate is not allowed
-          __kmp_hws_core.num = num;
-          __kmp_hws_core.offset = offset;
-        } else {
-          goto err;
-        }
-      }
-      break;
-    case 'T': // Thread
-      if (__kmp_hws_proc.num > 0)
-        goto err; // duplicate is not allowed
-      __kmp_hws_proc.num = num;
-      __kmp_hws_proc.offset = offset;
-      break;
-    default:
+    kmp_hw_t type = __kmp_stg_parse_hw_subset_name(pos);
+    if (type == KMP_HW_UNKNOWN) {
+    if (__kmp_hw_subset->specified(type)) {
       goto err;
     }
+    __kmp_hw_subset->push_back(num, type, offset);
   }
   return;
 err:
   KMP_WARNING(AffHWSubsetInvalid, name, value);
-  __kmp_hws_requested = 0; // mark that subset not requested
+  if (__kmp_hw_subset) {
+    kmp_hw_subset_t::deallocate(__kmp_hw_subset);
+    __kmp_hw_subset = nullptr;
+  }
   return;
 }
 
 static void __kmp_stg_print_hw_subset(kmp_str_buf_t *buffer, char const *name,
                                       void *data) {
-  if (__kmp_hws_requested) {
-    int comma = 0;
-    kmp_str_buf_t buf;
-    __kmp_str_buf_init(&buf);
-    if (__kmp_env_format)
-      KMP_STR_BUF_PRINT_NAME_EX(name);
-    else
-      __kmp_str_buf_print(buffer, "   %s='", name);
-    if (__kmp_hws_socket.num) {
-      __kmp_str_buf_print(&buf, "%ds", __kmp_hws_socket.num);
-      if (__kmp_hws_socket.offset)
-        __kmp_str_buf_print(&buf, "@%d", __kmp_hws_socket.offset);
-      comma = 1;
-    }
-    if (__kmp_hws_die.num) {
-      __kmp_str_buf_print(&buf, "%s%dd", comma ? "," : "", __kmp_hws_die.num);
-      if (__kmp_hws_die.offset)
-        __kmp_str_buf_print(&buf, "@%d", __kmp_hws_die.offset);
-      comma = 1;
-    }
-    if (__kmp_hws_node.num) {
-      __kmp_str_buf_print(&buf, "%s%dn", comma ? "," : "", __kmp_hws_node.num);
-      if (__kmp_hws_node.offset)
-        __kmp_str_buf_print(&buf, "@%d", __kmp_hws_node.offset);
-      comma = 1;
-    }
-    if (__kmp_hws_tile.num) {
-      __kmp_str_buf_print(&buf, "%s%dL2", comma ? "," : "", __kmp_hws_tile.num);
-      if (__kmp_hws_tile.offset)
-        __kmp_str_buf_print(&buf, "@%d", __kmp_hws_tile.offset);
-      comma = 1;
-    }
-    if (__kmp_hws_core.num) {
-      __kmp_str_buf_print(&buf, "%s%dc", comma ? "," : "", __kmp_hws_core.num);
-      if (__kmp_hws_core.offset)
-        __kmp_str_buf_print(&buf, "@%d", __kmp_hws_core.offset);
-      comma = 1;
-    }
-    if (__kmp_hws_proc.num)
-      __kmp_str_buf_print(&buf, "%s%dt", comma ? "," : "", __kmp_hws_proc.num);
-    __kmp_str_buf_print(buffer, "%s'\n", buf.str);
-    __kmp_str_buf_free(&buf);
-  }
-}
+  int depth;
+  if (!__kmp_hw_subset)
+    return;
+
+  depth = __kmp_hw_subset->get_depth();
+  for (int i = 0; i < depth; ++i) {
+    const auto &item = __kmp_hw_subset->at(i);
+    __kmp_str_buf_print(&buf, "%s%d%s", (i > 0 ? "," : ""), item.num,
+                        __kmp_hw_get_keyword(item.type));
+    if (item.offset)
+      __kmp_str_buf_print(&buf, "@%d", item.offset);
 
 #if USE_ITT_BUILD
 // -----------------------------------------------------------------------------
 // KMP_FORKJOIN_FRAMES
 
 static void __kmp_stg_parse_forkjoin_frames(char const *name, char const *value,
                                             void *data) {
   __kmp_stg_parse_bool(name, value, &__kmp_forkjoin_frames);
if ((FIND(aff_str, "none") == NULL) &&
         (FIND(aff_str, "balanced") == NULL) &&
         (FIND(aff_str, "disabled") == NULL)) {
       __kmp_affinity_notype = __kmp_stg_find("KMP_AFFINITY");
     } else {
       // A new affinity type is specified.
       // Reset the affinity flags to their default values,
       // in case this is called from kmp_set_defaults().
       __kmp_affinity_type = affinity_default;
-      __kmp_affinity_gran = affinity_gran_default;
+      __kmp_affinity_gran = KMP_HW_UNKNOWN;
       __kmp_affinity_top_method = affinity_top_method_default;
       __kmp_affinity_respect_mask = affinity_respect_mask_default;
     }
 #undef FIND
 
     // Also reset the affinity flags if OMP_PROC_BIND is specified.
     aff_str = __kmp_env_blk_var(&block, "OMP_PROC_BIND");
     if (aff_str != NULL) {
       __kmp_affinity_type = affinity_default;
-      __kmp_affinity_gran = affinity_gran_default;
+      __kmp_affinity_gran = KMP_HW_UNKNOWN;
       __kmp_affinity_top_method = affinity_top_method_default;
       __kmp_affinity_respect_mask = affinity_respect_mask_default;
     }
   }
 
 #endif /* KMP_AFFINITY_SUPPORTED */
 
   // Set up the nested proc bind type vector.

 #endif
   }
 
 #if KMP_AFFINITY_SUPPORTED
 
   if (!TCR_4(__kmp_init_middle)) {
 #if KMP_USE_HWLOC
     // Force using hwloc when either tiles or numa nodes requested within
-    // KMP_HW_SUBSET and no other topology method is requested
-    if ((__kmp_hws_node.num > 0 || __kmp_hws_tile.num > 0 ||
-         __kmp_affinity_gran == affinity_gran_tile) &&
-        (__kmp_affinity_top_method == affinity_top_method_default)) {
+    // KMP_HW_SUBSET or granularity setting and no other topology method
+    // is requested
+    if (__kmp_hw_subset &&
+        __kmp_affinity_top_method == affinity_top_method_default)
+      if (__kmp_hw_subset->specified(KMP_HW_NUMA) ||
+          __kmp_hw_subset->specified(KMP_HW_TILE) ||
+          __kmp_affinity_gran == KMP_HW_TILE ||
+          __kmp_affinity_gran == KMP_HW_NUMA)
+        __kmp_affinity_top_method = affinity_top_method_hwloc;
+    // Force using hwloc when tiles or numa nodes requested for OMP_PLACES
+    if (__kmp_affinity_gran == KMP_HW_NUMA ||
+        __kmp_affinity_gran == KMP_HW_TILE)
       __kmp_affinity_top_method = affinity_top_method_hwloc;
-    }
 #endif
     // Determine if the machine/OS is actually capable of supporting
     // affinity.
     const char *var = "KMP_AFFINITY";
     KMPAffinity::pick_api();
 #if KMP_USE_HWLOC
     // If Hwloc topology discovery was requested but affinity was also disabled,
     // then tell user that Hwloc request is being ignored and use default
if (__kmp_affinity_type == affinity_disabled) {
             (__kmp_affinity_warnings &&
              (__kmp_affinity_type != affinity_default) &&
              (__kmp_affinity_type != affinity_none) &&
              (__kmp_affinity_type != affinity_disabled))) {
           KMP_WARNING(AffNotSupported, var);
         }
         __kmp_affinity_type = affinity_disabled;
         __kmp_affinity_respect_mask = 0;
-        __kmp_affinity_gran = affinity_gran_fine;
+        __kmp_affinity_gran = KMP_HW_THREAD;
       }
     }
 
     if (__kmp_affinity_type == affinity_disabled) {
       __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
     } else if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_true) {
       // OMP_PROC_BIND=true maps to OMP_PROC_BIND=spread.
       __kmp_nested_proc_bind.bind_types[0] = proc_bind_spread;
#if KMP_GROUP_AFFINITY
             exactly_one_group) {
           __kmp_affinity_respect_mask = FALSE;
         }
         if (__kmp_affinity_type == affinity_default) {
           __kmp_affinity_type = affinity_compact;
           __kmp_nested_proc_bind.bind_types[0] = proc_bind_intel;
         }
         if (__kmp_affinity_top_method == affinity_top_method_default) {
-          if (__kmp_affinity_gran == affinity_gran_default) {
+          if (__kmp_affinity_gran == KMP_HW_UNKNOWN) {
             __kmp_affinity_top_method = affinity_top_method_group;
-            __kmp_affinity_gran = affinity_gran_group;
-          } else if (__kmp_affinity_gran == affinity_gran_group) {
+            __kmp_affinity_gran = KMP_HW_PROC_GROUP;
+          } else if (__kmp_affinity_gran == KMP_HW_PROC_GROUP) {
             __kmp_affinity_top_method = affinity_top_method_group;
           } else {
             __kmp_affinity_top_method = affinity_top_method_all;
           }
         } else if (__kmp_affinity_top_method == affinity_top_method_group) {
-          if (__kmp_affinity_gran == affinity_gran_default) {
-            __kmp_affinity_gran = affinity_gran_group;
-          } else if ((__kmp_affinity_gran != affinity_gran_group) &&
-                     (__kmp_affinity_gran != affinity_gran_fine) &&
-                     (__kmp_affinity_gran != affinity_gran_thread)) {
-            const char *str = NULL;
-            switch (__kmp_affinity_gran) {
-            case affinity_gran_core:
-              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);
-            }
+          if (__kmp_affinity_gran == KMP_HW_UNKNOWN) {
+            __kmp_affinity_gran = KMP_HW_PROC_GROUP;
+          } else if ((__kmp_affinity_gran != KMP_HW_PROC_GROUP) &&
+                     (__kmp_affinity_gran != KMP_HW_THREAD)) {
+            const char *str = __kmp_hw_get_keyword(__kmp_affinity_gran);
             KMP_WARNING(AffGranTopGroup, var, str);
-            __kmp_affinity_gran = affinity_gran_fine;
+            __kmp_affinity_gran = KMP_HW_THREAD;
           }
         } else {
-          if (__kmp_affinity_gran == affinity_gran_default) {
-            __kmp_affinity_gran = affinity_gran_core;
-          } else if (__kmp_affinity_gran == affinity_gran_group) {
+          if (__kmp_affinity_gran == KMP_HW_UNKNOWN) {
+            __kmp_affinity_gran = KMP_HW_CORE;
+          } else if (__kmp_affinity_gran == KMP_HW_PROC_GROUP) {
             const char *str = NULL;
             switch (__kmp_affinity_type) {
             case affinity_physical:
               str = "physical";
               break;
             case affinity_logical:
               str = "logical";
               break;
             case affinity_compact:
               str = "compact";
               break;
             case affinity_scatter:
               str = "scatter";
               break;
             case affinity_explicit:
               str = "explicit";
               break;
             // No MIC on windows, so no affinity_balanced case
             default:
               KMP_DEBUG_ASSERT(0);
             }
             KMP_WARNING(AffGranGroupType, var, str);
-            __kmp_affinity_gran = affinity_gran_core;
+            __kmp_affinity_gran = KMP_HW_CORE;
           }
         }
       } else
 
 #endif /* KMP_GROUP_AFFINITY */
 
       {
         if (__kmp_affinity_respect_mask == affinity_respect_mask_default) {
#if KMP_MIC_SUPPORTED
           if (__kmp_mic_type != non_mic) {
             __kmp_affinity_type = affinity_scatter;
           } else
 #endif
           {
             __kmp_affinity_type = affinity_none;
           }
         }
-        if ((__kmp_affinity_gran == affinity_gran_default) &&
+        if ((__kmp_affinity_gran == KMP_HW_UNKNOWN) &&
             (__kmp_affinity_gran_levels < 0)) {
 #if KMP_MIC_SUPPORTED
           if (__kmp_mic_type != non_mic) {
-            __kmp_affinity_gran = affinity_gran_fine;
+            __kmp_affinity_gran = KMP_HW_THREAD;
           } else
 #endif
           {
-            __kmp_affinity_gran = affinity_gran_core;
+            __kmp_affinity_gran = KMP_HW_CORE;
           }
         }
         if (__kmp_affinity_top_method == affinity_top_method_default) {
           __kmp_affinity_top_method = affinity_top_method_all;
         }
       }
     }
 

openmp/runtime/test/affinity/kmp-affinity.c

+// RUN: %libomp-compile -D_GNU_SOURCE
+// RUN: env KMP_AFFINITY=granularity=thread,compact %libomp-run
+// RUN: env KMP_AFFINITY=granularity=core,compact %libomp-run
+// RUN: env KMP_AFFINITY=granularity=socket,compact %libomp-run
+// REQUIRES: linux
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "libomp_test_affinity.h"
+#include "libomp_test_topology.h"
+
+// Compare place lists. Make sure every place in p1 is in p2.
+static int compare_places(const place_list_t *p1, const place_list_t *p2) {
+  int i, j;
+  for (i = 0; i < p1->num_places; ++i) {
+    int found = 0;
+    for (j = 0; j < p2->num_places; ++j) {
+      if (affinity_mask_equal(p1->masks[i], p2->masks[j])) {
+        found = 1;
+        break;
+      }
+    }
+    if (!found) {
+      printf("Found place in p1 not in p2!\n");
+      printf("p1 places:\n");
+      topology_print_places(p1);
+      printf("\n");
+      printf("p2 places:\n");
+      topology_print_places(p1);
+      return EXIT_FAILURE;
+    }
+  }
+  return EXIT_SUCCESS;
+}
+
+static int check_places() {
+  int status;
+  const char *value = getenv("KMP_AFFINITY");
+  if (!value) {
+    fprintf(stderr, "error: must set OMP_PLACES envirable for this test!\n");
+    return EXIT_FAILURE;
+  }
+  place_list_t *places, *openmp_places;
+  if (strstr(value, "socket")) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_SOCKET);
+  } else if (strstr(value, "core")) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_CORE);
+  } else if (strstr(value, "thread")) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_THREAD);
+  } else {
+    fprintf(
+        stderr,
+        "error: KMP_AFFINITY granularity must be one of thread,core,socket!\n");
+    return EXIT_FAILURE;
+  }
+  openmp_places = topology_alloc_openmp_places();
+  status = compare_places(openmp_places, places);
+  topology_free_places(places);
+  topology_free_places(openmp_places);
+  return status;
+}
+
+int main() {
+  if (!topology_using_full_mask()) {
+    printf("Thread does not have access to all logical processors. Skipping "
+           "test.\n");
+    return EXIT_SUCCESS;
+  }
+  return check_places();
+}

openmp/runtime/test/affinity/kmp-hw-subset.c

+// RUN: %libomp-compile -D_GNU_SOURCE
+// RUN: env OMP_PLACES=threads %libomp-run
+// RUN: env OMP_PLACES=cores %libomp-run
+// RUN: env OMP_PLACES=sockets %libomp-run
+// REQUIRES: linux
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "libomp_test_affinity.h"
+#include "libomp_test_topology.h"
+
+// Check openmp place list to make sure it follow KMP_HW_SUBSET restriction
+static int compare_hw_subset_places(const place_list_t *openmp_places,
+                                    topology_obj_type_t type, int nsockets,
+                                    int ncores_per_socket,
+                                    int nthreads_per_core) {
+  int i, j, expected_total, expected_per_place;
+  if (type == TOPOLOGY_OBJ_THREAD) {
+    expected_total = nsockets * ncores_per_socket * nthreads_per_core;
+    expected_per_place = 1;
+  } else if (type == TOPOLOGY_OBJ_CORE) {
+    expected_total = nsockets * ncores_per_socket;
+    expected_per_place = nthreads_per_core;
+  } else {
+    expected_total = nsockets;
+    expected_per_place = ncores_per_socket;
+  }
+  if (openmp_places->num_places != expected_total) {
+    fprintf(stderr, "error: KMP_HW_SUBSET did not half each resource layer!\n");
+    printf("openmp_places places:\n");
+    topology_print_places(openmp_places);
+    printf("\n");
+    return EXIT_FAILURE;
+  }
+  for (i = 0; i < openmp_places->num_places; ++i) {
+    int count = affinity_mask_count(openmp_places->masks[i]);
+    if (count != expected_per_place) {
+      fprintf(stderr, "error: place %d has %d OS procs instead of %d\n", i,
+              count, expected_per_place);
+      return EXIT_FAILURE;
+    }
+  }
+  return EXIT_SUCCESS;
+}
+
+static int check_places() {
+  char buf[100];
+  topology_obj_type_t type;
+  const char *value;
+  int status = EXIT_SUCCESS;
+  place_list_t *threads, *cores, *sockets, *openmp_places;
+  threads = topology_alloc_type_places(TOPOLOGY_OBJ_THREAD);
+  cores = topology_alloc_type_places(TOPOLOGY_OBJ_CORE);
+  sockets = topology_alloc_type_places(TOPOLOGY_OBJ_SOCKET);
+
+  if (threads->num_places <= 1) {
+    printf("Only one hardware thread to execute on. Skipping test.\n");
+    return status;
+  }
+
+  value = getenv("OMP_PLACES");
+  if (!value) {
+    fprintf(stderr,
+            "error: OMP_PLACES must be set to one of threads,cores,sockets!\n");
+    return EXIT_FAILURE;
+  }
+  if (strcmp(value, "threads") == 0)
+    type = TOPOLOGY_OBJ_THREAD;
+  else if (strcmp(value, "cores") == 0)
+    type = TOPOLOGY_OBJ_CORE;
+  else if (strcmp(value, "sockets") == 0)
+    type = TOPOLOGY_OBJ_SOCKET;
+  else {
+    fprintf(stderr,
+            "error: OMP_PLACES must be one of threads,cores,sockets!\n");
+    return EXIT_FAILURE;
+  }
+
+  // Calculate of num threads per core, num cores per socket, & num sockets
+  if (cores->num_places <= 0) {
+    printf("Invalid number of cores (%d). Skipping test.\n", cores->num_places);
+    return status;
+  } else if (sockets->num_places <= 0) {
+    printf("Invalid number of sockets (%d). Skipping test.\n",
+           cores->num_places);
+    return status;
+  }
+  int nthreads_per_core = threads->num_places / cores->num_places;
+  int ncores_per_socket = cores->num_places / sockets->num_places;
+  int nsockets = sockets->num_places;
+
+  if (nsockets * ncores_per_socket * nthreads_per_core != threads->num_places) {
+    printf("Only uniform topologies can be tested. Skipping test.\n");
+    return status;
+  }
+
+  // Use half the resources of every level
+  if (nthreads_per_core > 1)
+    nthreads_per_core /= 2;
+  if (ncores_per_socket > 1)
+    ncores_per_socket /= 2;
+  if (nsockets > 1)
+    nsockets /= 2;
+
+  snprintf(buf, sizeof(buf), "%ds,%dc,%dt", nsockets, ncores_per_socket,
+           nthreads_per_core);
+  setenv("KMP_HW_SUBSET", buf, 1);
+
+  openmp_places = topology_alloc_openmp_places();
+  status = compare_hw_subset_places(openmp_places, type, nsockets,
+                                    ncores_per_socket, nthreads_per_core);
+  topology_free_places(threads);
+  topology_free_places(cores);
+  topology_free_places(sockets);
+  topology_free_places(openmp_places);
+  return status;
+}
+
+int main() {
+  if (!topology_using_full_mask()) {
+    printf("Thread does not have access to all logical processors. Skipping "
+           "test.\n");
+    return EXIT_SUCCESS;
+  }
+  return check_places();
+}

openmp/runtime/test/affinity/libomp_test_topology.h

+#ifndef LIBOMP_TEST_TOPOLOGY_H
+#define LIBOMP_TEST_TOPOLOGY_H
+
+#include "libomp_test_affinity.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <dirent.h>
+#include <errno.h>
+#include <ctype.h>
+#include <omp.h>
+
+typedef enum topology_obj_type_t {
+  TOPOLOGY_OBJ_THREAD,
+  TOPOLOGY_OBJ_CORE,
+  TOPOLOGY_OBJ_SOCKET,
+  TOPOLOGY_OBJ_MAX
+} topology_obj_type_t;
+
+typedef struct place_list_t {
+  int num_places;
+  affinity_mask_t **masks;
+} place_list_t;
+
+// Return the first character in file 'f' that is not a whitespace character
+// including newlines and carriage returns
+static int get_first_nonspace_from_file(FILE *f) {
+  int c;
+  do {
+    c = fgetc(f);
+  } while (c != EOF && (isspace(c) || c == '\n' || c == '\r'));
+  return c;
+}
+
+// Read an integer from file 'f' into 'number'
+// Return 1 on successful read of integer,
+//        0 on unsuccessful read of integer,
+//        EOF on end of file.
+static int get_integer_from_file(FILE *f, int *number) {
+  int n;
+  n = fscanf(f, "%d", number);
+  if (feof(f))
+    return EOF;
+  if (n != 1)
+    return 0;
+  return 1;
+}
+
+// Read a siblings list file from Linux /sys/devices/system/cpu/cpu?/topology/*
+static affinity_mask_t *topology_get_mask_from_file(const char *filename) {
+  int status = EXIT_SUCCESS;
+  FILE *f = fopen(filename, "r");
+  if (!f) {
+    perror(filename);
+    exit(EXIT_FAILURE);
+  }
+  affinity_mask_t *mask = affinity_mask_alloc();
+  while (1) {
+    int c, i, n, lower, upper;
+    // Read the first integer
+    n = get_integer_from_file(f, &lower);
+    if (n == EOF) {
+      break;
+    } else if (n == 0) {
+      fprintf(stderr, "syntax error: expected integer\n");
+      status = EXIT_FAILURE;
+      break;
+    }
+
+    // Now either a , or -
+    c = get_first_nonspace_from_file(f);
+    if (c == EOF || c == ',') {
+      affinity_mask_set(mask, lower);
+      if (c == EOF)
+        break;
+    } else if (c == '-') {
+      n = get_integer_from_file(f, &upper);
+      if (n == EOF || n == 0) {
+        fprintf(stderr, "syntax error: expected integer\n");
+        status = EXIT_FAILURE;
+        break;
+      }
+      for (i = lower; i <= upper; ++i)
+        affinity_mask_set(mask, i);
+      c = get_first_nonspace_from_file(f);
+      if (c == EOF) {
+        break;
+      } else if (c == ',') {
+        continue;
+      } else {
+        fprintf(stderr, "syntax error: unexpected character: '%c (%d)'\n", c,
+                c);
+        status = EXIT_FAILURE;
+        break;
+      }
+    } else {
+      fprintf(stderr, "syntax error: unexpected character: '%c (%d)'\n", c, c);
+      status = EXIT_FAILURE;
+      break;
+    }
+  }
+  fclose(f);
+  if (status == EXIT_FAILURE) {
+    affinity_mask_free(mask);
+    mask = NULL;
+  }
+  return mask;
+}
+
+static int topology_get_num_cpus() {
+  char buf[1024];
+  // Count the number of cpus
+  int cpu = 0;
+  while (1) {
+    snprintf(buf, sizeof(buf), "/sys/devices/system/cpu/cpu%d", cpu);
+    DIR *dir = opendir(buf);
+    if (dir) {
+      closedir(dir);
+      cpu++;
+    } else {
+      break;
+    }
+  }
+  if (cpu == 0)
+    cpu = 1;
+  return cpu;
+}
+
+// Return whether the current thread has access to all logical processors
+static int topology_using_full_mask() {
+  int cpu;
+  int has_all = 1;
+  int num_cpus = topology_get_num_cpus();
+  affinity_mask_t *mask = affinity_mask_alloc();
+  get_thread_affinity(mask);
+  for (cpu = 0; cpu < num_cpus; ++cpu) {
+    if (!affinity_mask_isset(mask, cpu)) {
+      has_all = 0;
+      break;
+    }
+  }
+  affinity_mask_free(mask);
+  return has_all;
+}
+
+// Return array of masks representing OMP_PLACES keyword (e.g., sockets, cores,
+// threads)
+static place_list_t *topology_alloc_type_places(topology_obj_type_t type) {
+  char buf[1024];
+  int i, cpu, num_places, num_unique;
+  int num_cpus = topology_get_num_cpus();
+  place_list_t *places = (place_list_t *)malloc(sizeof(place_list_t));
+  affinity_mask_t **masks =
+      (affinity_mask_t **)malloc(sizeof(affinity_mask_t *) * num_cpus);
+  num_unique = 0;
+  for (cpu = 0; cpu < num_cpus; ++cpu) {
+    affinity_mask_t *mask;
+    if (type == TOPOLOGY_OBJ_CORE) {
+      snprintf(buf, sizeof(buf),
+               "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
+               cpu);
+      mask = topology_get_mask_from_file(buf);
+    } else if (type == TOPOLOGY_OBJ_SOCKET) {
+      snprintf(buf, sizeof(buf),
+               "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list",
+               cpu);
+      mask = topology_get_mask_from_file(buf);
+    } else if (type == TOPOLOGY_OBJ_THREAD) {
+      mask = affinity_mask_alloc();
+      affinity_mask_set(mask, cpu);
+    } else {
+      fprintf(stderr, "Unknown topology type (%d)\n", (int)type);
+      exit(EXIT_FAILURE);
+    }
+    // Check for unique topology objects above the thread level
+    if (type != TOPOLOGY_OBJ_THREAD) {
+      for (i = 0; i < num_unique; ++i) {
+        if (affinity_mask_equal(masks[i], mask)) {
+          affinity_mask_free(mask);
+          mask = NULL;
+          break;
+        }
+      }
+    }
+    if (mask)
+      masks[num_unique++] = mask;
+  }
+  places->num_places = num_unique;
+  places->masks = masks;
+  return places;
+}
+
+static place_list_t *topology_alloc_openmp_places() {
+  int place, i;
+  int num_places = omp_get_num_places();
+  place_list_t *places = (place_list_t *)malloc(sizeof(place_list_t));
+  affinity_mask_t **masks =
+      (affinity_mask_t **)malloc(sizeof(affinity_mask_t *) * num_places);
+  for (place = 0; place < num_places; ++place) {
+    int num_procs = omp_get_place_num_procs(place);
+    int *ids = (int *)malloc(sizeof(int) * num_procs);
+    omp_get_place_proc_ids(place, ids);
+    affinity_mask_t *mask = affinity_mask_alloc();
+    for (i = 0; i < num_procs; ++i)
+      affinity_mask_set(mask, ids[i]);
+    masks[place] = mask;
+  }
+  places->num_places = num_places;
+  places->masks = masks;
+  return places;
+}
+
+// Free the array of masks from one of: topology_alloc_type_masks()
+// or topology_alloc_openmp_masks()
+static void topology_free_places(place_list_t *places) {
+  int i;
+  for (i = 0; i < places->num_places; ++i)
+    affinity_mask_free(places->masks[i]);
+  free(places->masks);
+  free(places);
+}
+
+static void topology_print_places(const place_list_t *p) {
+  int i;
+  char buf[1024];
+  for (i = 0; i < p->num_places; ++i) {
+    affinity_mask_snprintf(buf, sizeof(buf), p->masks[i]);
+    printf("Place %d: %s\n", i, buf);
+  }
+}
+
+#endif

openmp/runtime/test/affinity/omp-places.c

+// RUN: %libomp-compile -D_GNU_SOURCE
+// RUN: env OMP_PLACES=threads %libomp-run
+// RUN: env OMP_PLACES=cores %libomp-run
+// RUN: env OMP_PLACES=sockets %libomp-run
+// REQUIRES: linux
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "libomp_test_affinity.h"
+#include "libomp_test_topology.h"
+
+// Compare place lists. The order is not taken into consideration here.
+// The OS detection might have the cores/sockets in a different
+// order from the runtime.
+static int compare_places(const place_list_t *p1, const place_list_t *p2) {
+  int i, j;
+  if (p1->num_places != p2->num_places) {
+    fprintf(stderr, "error: places do not have same number of places! (p1 has "
+                    "%d, p2 has %d)\n",
+            p1->num_places, p2->num_places);
+    printf("p1 places:\n");
+    topology_print_places(p1);
+    printf("\n");
+    printf("p2 places:\n");
+    topology_print_places(p1);
+    return EXIT_FAILURE;
+  }
+  for (i = 0; i < p1->num_places; ++i) {
+    int found = 0;
+    for (j = 0; j < p2->num_places; ++j) {
+      if (affinity_mask_equal(p1->masks[i], p2->masks[j])) {
+        found = 1;
+        break;
+      }
+    }
+    if (!found) {
+      printf("Found difference in places!\n");
+      printf("p1 places:\n");
+      topology_print_places(p1);
+      printf("\n");
+      printf("p2 places:\n");
+      topology_print_places(p1);
+      return EXIT_FAILURE;
+    }
+  }
+  return EXIT_SUCCESS;
+}
+
+static int check_places() {
+  int status;
+  const char *value = getenv("OMP_PLACES");
+  if (!value) {
+    fprintf(stderr, "error: must set OMP_PLACES envirable for this test!\n");
+    return EXIT_FAILURE;
+  }
+  place_list_t *places, *openmp_places;
+  if (strcmp(value, "sockets") == 0) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_SOCKET);
+  } else if (strcmp(value, "cores") == 0) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_CORE);
+  } else if (strcmp(value, "threads") == 0) {
+    places = topology_alloc_type_places(TOPOLOGY_OBJ_THREAD);
+  } else {
+    fprintf(stderr,
+            "error: OMP_PLACES must be one of threads,cores,sockets!\n");
+    return EXIT_FAILURE;
+  }
+  openmp_places = topology_alloc_openmp_places();
+  status = compare_places(places, openmp_places);
+  topology_free_places(places);
+  topology_free_places(openmp_places);
+  return status;
+}
+
+int main() {
+  if (!topology_using_full_mask()) {
+    printf("Thread does not have access to all logical processors. Skipping "
+           "test.\n");
+    return EXIT_SUCCESS;
+  }
+  return check_places();
+}