/* * kmp_affinity.cpp -- affinity management */ //===----------------------------------------------------------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// #include "kmp.h" #include "kmp_affinity.h" #include "kmp_i18n.h" #include "kmp_io.h" #include "kmp_str.h" #include "kmp_wrapper_getpid.h" #if KMP_USE_HIER_SCHED #include "kmp_dispatch_hier.h" #endif // 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); // 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; thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0] - 1; thr_bar->skip_per_level = machine_hierarchy.skipPerLevel; } #if KMP_AFFINITY_SUPPORTED 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); } void KMPAffinity::pick_api() { KMPAffinity *affinity_dispatch; if (picked_api) return; #if KMP_USE_HWLOC // Only use Hwloc if affinity isn't explicitly disabled and // user requests Hwloc topology method if (__kmp_affinity_top_method == affinity_top_method_hwloc && __kmp_affinity_type != affinity_disabled) { affinity_dispatch = new KMPHwlocAffinity(); } else #endif { affinity_dispatch = new KMPNativeAffinity(); } __kmp_affinity_dispatch = affinity_dispatch; picked_api = true; } void KMPAffinity::destroy_api() { if (__kmp_affinity_dispatch != NULL) { delete __kmp_affinity_dispatch; __kmp_affinity_dispatch = NULL; picked_api = false; } } #define KMP_ADVANCE_SCAN(scan) \ while (*scan != '\0') { \ scan++; \ } // Print the affinity mask to the character array in a pretty format. // The format is a comma separated list of non-negative integers or integer // ranges: e.g., 1,2,3-5,7,9-15 // The format can also be the string "{}" if no bits are set in mask char *__kmp_affinity_print_mask(char *buf, int buf_len, kmp_affin_mask_t *mask) { int start = 0, finish = 0, previous = 0; bool first_range; KMP_ASSERT(buf); KMP_ASSERT(buf_len >= 40); KMP_ASSERT(mask); char *scan = buf; char *end = buf + buf_len - 1; // Check for empty set. if (mask->begin() == mask->end()) { KMP_SNPRINTF(scan, end - scan + 1, "{}"); KMP_ADVANCE_SCAN(scan); KMP_ASSERT(scan <= end); return buf; } first_range = true; start = mask->begin(); while (1) { // Find next range // [start, previous] is inclusive range of contiguous bits in mask for (finish = mask->next(start), previous = start; finish == previous + 1 && finish != mask->end(); finish = mask->next(finish)) { previous = finish; } // The first range does not need a comma printed before it, but the rest // of the ranges do need a comma beforehand if (!first_range) { KMP_SNPRINTF(scan, end - scan + 1, "%s", ","); KMP_ADVANCE_SCAN(scan); } else { first_range = false; } // Range with three or more contiguous bits in the affinity mask if (previous - start > 1) { KMP_SNPRINTF(scan, end - scan + 1, "%d-%d", static_cast(start), static_cast(previous)); } else { // Range with one or two contiguous bits in the affinity mask KMP_SNPRINTF(scan, end - scan + 1, "%d", static_cast(start)); KMP_ADVANCE_SCAN(scan); if (previous - start > 0) { KMP_SNPRINTF(scan, end - scan + 1, ",%d", static_cast(previous)); } } KMP_ADVANCE_SCAN(scan); // Start over with new start point start = finish; if (start == mask->end()) break; // Check for overflow if (end - scan < 2) break; } // Check for overflow KMP_ASSERT(scan <= end); return buf; } #undef KMP_ADVANCE_SCAN // Print the affinity mask to the string buffer object in a pretty format // The format is a comma separated list of non-negative integers or integer // ranges: e.g., 1,2,3-5,7,9-15 // The format can also be the string "{}" if no bits are set in mask kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf, kmp_affin_mask_t *mask) { int start = 0, finish = 0, previous = 0; bool first_range; KMP_ASSERT(buf); KMP_ASSERT(mask); __kmp_str_buf_clear(buf); // Check for empty set. if (mask->begin() == mask->end()) { __kmp_str_buf_print(buf, "%s", "{}"); return buf; } first_range = true; start = mask->begin(); while (1) { // Find next range // [start, previous] is inclusive range of contiguous bits in mask for (finish = mask->next(start), previous = start; finish == previous + 1 && finish != mask->end(); finish = mask->next(finish)) { previous = finish; } // The first range does not need a comma printed before it, but the rest // of the ranges do need a comma beforehand if (!first_range) { __kmp_str_buf_print(buf, "%s", ","); } else { first_range = false; } // Range with three or more contiguous bits in the affinity mask if (previous - start > 1) { __kmp_str_buf_print(buf, "%d-%d", static_cast(start), static_cast(previous)); } else { // Range with one or two contiguous bits in the affinity mask __kmp_str_buf_print(buf, "%d", static_cast(start)); if (previous - start > 0) { __kmp_str_buf_print(buf, ",%d", static_cast(previous)); } } // Start over with new start point start = finish; if (start == mask->end()) break; } return buf; } void __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) { KMP_CPU_ZERO(mask); #if KMP_GROUP_AFFINITY if (__kmp_num_proc_groups > 1) { int group; KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL); for (group = 0; group < __kmp_num_proc_groups; group++) { int i; int num = __kmp_GetActiveProcessorCount(group); for (i = 0; i < num; i++) { KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask); } } } else #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. 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; // __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 * nPackages); } // 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; 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); __kmp_str_buf_free(&buf); } } #if KMP_USE_HWLOC static void __kmp_affinity_print_hwloc_tp(AddrUnsPair *addrP, int len, int depth, int *levels) { int proc; kmp_str_buf_t buf; __kmp_str_buf_init(&buf); KMP_INFORM(OSProcToPhysicalThreadMap, "KMP_AFFINITY"); for (proc = 0; proc < len; proc++) { __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Package), addrP[proc].first.labels[0]); if (depth > 1) { int level = 1; // iterate over levels int label = 1; // iterate over labels if (__kmp_numa_detected) // node level follows package if (levels[level++] > 0) __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Node), addrP[proc].first.labels[label++]); if (__kmp_tile_depth > 0) // tile level follows node if any, or package if (levels[level++] > 0) __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Tile), addrP[proc].first.labels[label++]); if (levels[level++] > 0) // core level follows __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Core), addrP[proc].first.labels[label++]); if (levels[level++] > 0) // thread level is the latest __kmp_str_buf_print(&buf, "%s %d ", KMP_I18N_STR(Thread), addrP[proc].first.labels[label++]); KMP_DEBUG_ASSERT(label == depth); } KMP_INFORM(OSProcMapToPack, "KMP_AFFINITY", addrP[proc].second, buf.str); __kmp_str_buf_clear(&buf); } __kmp_str_buf_free(&buf); } static int nNodePerPkg, nTilePerPkg, nTilePerNode, nCorePerNode, nCorePerTile; // This function removes the topology levels that are radix 1 and don't offer // further information about the topology. The most common example is when you // have one thread context per core, we don't want the extra thread context // level if it offers no unique labels. So they are removed. // return value: the new depth of address2os static int __kmp_affinity_remove_radix_one_levels(AddrUnsPair *addrP, int nTh, int depth, int *levels) { int level; int i; int radix1_detected; int new_depth = depth; for (level = depth - 1; level > 0; --level) { // Detect if this level is radix 1 radix1_detected = 1; for (i = 1; i < nTh; ++i) { if (addrP[0].first.labels[level] != addrP[i].first.labels[level]) { // There are differing label values for this level so it stays radix1_detected = 0; break; } } if (!radix1_detected) continue; // Radix 1 was detected --new_depth; levels[level] = -1; // mark level as not present in address2os array if (level == new_depth) { // "turn off" deepest level, just decrement the depth that removes // the level from address2os array for (i = 0; i < nTh; ++i) { addrP[i].first.depth--; } } else { // For other levels, we move labels over and also reduce the depth int j; for (j = level; j < new_depth; ++j) { for (i = 0; i < nTh; ++i) { addrP[i].first.labels[j] = addrP[i].first.labels[j + 1]; addrP[i].first.depth--; } levels[j + 1] -= 1; } } } return new_depth; } // Returns the number of objects of type 'type' below 'obj' within the topology // tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is // HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET // object. static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj, hwloc_obj_type_t type) { int retval = 0; hwloc_obj_t first; for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type, obj->logical_index, type, 0); first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology, obj->type, first) == obj; first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type, first)) { ++retval; } return retval; } static int __kmp_hwloc_count_children_by_depth(hwloc_topology_t t, hwloc_obj_t o, 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) } static int __kmp_hwloc_process_obj_core_pu(AddrUnsPair *addrPair, int &nActiveThreads, int &num_active_cores, hwloc_obj_t obj, int depth, int *labels) { hwloc_obj_t core = NULL; hwloc_topology_t &tp = __kmp_hwloc_topology; int NC = __kmp_hwloc_count_children_by_type(tp, obj, HWLOC_OBJ_CORE, &core); for (int core_id = 0; core_id < NC; ++core_id, core = core->next_cousin) { hwloc_obj_t pu = NULL; KMP_DEBUG_ASSERT(core != NULL); int num_active_threads = 0; int NT = __kmp_hwloc_count_children_by_type(tp, core, HWLOC_OBJ_PU, &pu); // int NT = core->arity; pu = core->first_child; // faster? for (int pu_id = 0; pu_id < NT; ++pu_id, pu = pu->next_cousin) { KMP_DEBUG_ASSERT(pu != NULL); if (!KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask)) continue; // skip inactive (inaccessible) unit Address addr(depth + 2); KA_TRACE(20, ("Hwloc inserting %d (%d) %d (%d) %d (%d) into address2os\n", obj->os_index, obj->logical_index, core->os_index, core->logical_index, pu->os_index, pu->logical_index)); for (int i = 0; i < depth; ++i) addr.labels[i] = labels[i]; // package, etc. addr.labels[depth] = core_id; // core addr.labels[depth + 1] = pu_id; // pu addrPair[nActiveThreads] = AddrUnsPair(addr, pu->os_index); __kmp_pu_os_idx[nActiveThreads] = pu->os_index; nActiveThreads++; ++num_active_threads; // count active threads per core } if (num_active_threads) { // were there any active threads on the core? ++__kmp_ncores; // count total active cores ++num_active_cores; // count active cores per socket if (num_active_threads > __kmp_nThreadsPerCore) __kmp_nThreadsPerCore = num_active_threads; // calc maximum } } return 0; } // Check if NUMA node detected below the package, // and if tile object is detected and return its depth static int __kmp_hwloc_check_numa() { hwloc_topology_t &tp = __kmp_hwloc_topology; hwloc_obj_t hT, hC, hL, hN, hS; // hwloc objects (pointers to) int depth, l2cache_depth, package_depth; // Get some PU hT = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PU, 0); if (hT == NULL) // something has gone wrong return 1; // check NUMA node below PACKAGE hN = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_NUMANODE, hT); hS = hwloc_get_ancestor_obj_by_type(tp, HWLOC_OBJ_PACKAGE, hT); KMP_DEBUG_ASSERT(hS != NULL); if (hN != NULL && hN->depth > hS->depth) { __kmp_numa_detected = TRUE; // socket includes node(s) if (__kmp_affinity_gran == affinity_gran_node) { __kmp_affinity_gran = affinity_gran_numa; } } package_depth = hwloc_get_type_depth(tp, HWLOC_OBJ_PACKAGE); l2cache_depth = hwloc_get_cache_type_depth(tp, 2, HWLOC_OBJ_CACHE_UNIFIED); // check tile, get object by depth because of multiple caches possible depth = (l2cache_depth < package_depth) ? package_depth : l2cache_depth; hL = hwloc_get_ancestor_obj_by_depth(tp, depth, hT); hC = NULL; // not used, but reset it here just in case if (hL != NULL && __kmp_hwloc_count_children_by_type(tp, hL, HWLOC_OBJ_CORE, &hC) > 1) __kmp_tile_depth = depth; // tile consists of multiple cores return 0; } static int __kmp_affinity_create_hwloc_map(AddrUnsPair **address2os, kmp_i18n_id_t *const msg_id) { hwloc_topology_t &tp = __kmp_hwloc_topology; // shortcut of a long name *address2os = NULL; *msg_id = kmp_i18n_null; // Save the affinity mask for the current thread. kmp_affin_mask_t *oldMask; KMP_CPU_ALLOC(oldMask); __kmp_get_system_affinity(oldMask, TRUE); __kmp_hwloc_check_numa(); if (!KMP_AFFINITY_CAPABLE()) { // Hack to try and infer the machine topology using only the data // available from cpuid on the current thread, and __kmp_xproc. KMP_ASSERT(__kmp_affinity_type == affinity_none); nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj( hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0), HWLOC_OBJ_CORE); __kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj( hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0), HWLOC_OBJ_PU); __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore; nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg; if (__kmp_affinity_verbose) { KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY"); KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc); if (__kmp_affinity_uniform_topology()) { KMP_INFORM(Uniform, "KMP_AFFINITY"); } else { KMP_INFORM(NonUniform, "KMP_AFFINITY"); } KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg, __kmp_nThreadsPerCore, __kmp_ncores); } KMP_CPU_FREE(oldMask); return 0; } int depth = 3; int levels[5] = {0, 1, 2, 3, 4}; // package, [node,] [tile,] core, thread int labels[3] = {0}; // package [,node] [,tile] - head of lables array if (__kmp_numa_detected) ++depth; if (__kmp_tile_depth) ++depth; // Allocate the data structure to be returned. AddrUnsPair *retval = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc); KMP_DEBUG_ASSERT(__kmp_pu_os_idx == NULL); __kmp_pu_os_idx = (int *)__kmp_allocate(sizeof(int) * __kmp_avail_proc); // When affinity is off, this routine will still be called to set // __kmp_ncores, as well as __kmp_nThreadsPerCore, // nCoresPerPkg, & nPackages. Make sure all these vars are set // correctly, and return if affinity is not enabled. hwloc_obj_t socket, node, tile; int nActiveThreads = 0; int socket_id = 0; // re-calculate globals to count only accessible resources __kmp_ncores = nPackages = nCoresPerPkg = __kmp_nThreadsPerCore = 0; nNodePerPkg = nTilePerPkg = nTilePerNode = nCorePerNode = nCorePerTile = 0; for (socket = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0); socket != NULL; socket = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PACKAGE, socket), socket_id++) { labels[0] = socket_id; if (__kmp_numa_detected) { int NN; int n_active_nodes = 0; node = NULL; NN = __kmp_hwloc_count_children_by_type(tp, socket, HWLOC_OBJ_NUMANODE, &node); for (int node_id = 0; node_id < NN; ++node_id, node = node->next_cousin) { labels[1] = node_id; if (__kmp_tile_depth) { // NUMA + tiles int NT; int n_active_tiles = 0; tile = NULL; NT = __kmp_hwloc_count_children_by_depth(tp, node, __kmp_tile_depth, &tile); for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) { labels[2] = tl_id; int n_active_cores = 0; __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores, tile, 3, labels); if (n_active_cores) { // were there any active cores on the socket? ++n_active_tiles; // count active tiles per node if (n_active_cores > nCorePerTile) nCorePerTile = n_active_cores; // calc maximum } } if (n_active_tiles) { // were there any active tiles on the socket? ++n_active_nodes; // count active nodes per package if (n_active_tiles > nTilePerNode) nTilePerNode = n_active_tiles; // calc maximum } } else { // NUMA, no tiles int n_active_cores = 0; __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores, node, 2, labels); if (n_active_cores) { // were there any active cores on the socket? ++n_active_nodes; // count active nodes per package if (n_active_cores > nCorePerNode) nCorePerNode = n_active_cores; // calc maximum } } } if (n_active_nodes) { // were there any active nodes on the socket? ++nPackages; // count total active packages if (n_active_nodes > nNodePerPkg) nNodePerPkg = n_active_nodes; // calc maximum } } else { if (__kmp_tile_depth) { // no NUMA, tiles int NT; int n_active_tiles = 0; tile = NULL; NT = __kmp_hwloc_count_children_by_depth(tp, socket, __kmp_tile_depth, &tile); for (int tl_id = 0; tl_id < NT; ++tl_id, tile = tile->next_cousin) { labels[1] = tl_id; int n_active_cores = 0; __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores, tile, 2, labels); if (n_active_cores) { // were there any active cores on the socket? ++n_active_tiles; // count active tiles per package if (n_active_cores > nCorePerTile) nCorePerTile = n_active_cores; // calc maximum } } if (n_active_tiles) { // were there any active tiles on the socket? ++nPackages; // count total active packages if (n_active_tiles > nTilePerPkg) nTilePerPkg = n_active_tiles; // calc maximum } } else { // no NUMA, no tiles int n_active_cores = 0; __kmp_hwloc_process_obj_core_pu(retval, nActiveThreads, n_active_cores, socket, 1, labels); if (n_active_cores) { // were there any active cores on the socket? ++nPackages; // count total active packages if (n_active_cores > nCoresPerPkg) nCoresPerPkg = n_active_cores; // calc maximum } } } } // If there's only one thread context to bind to, return now. KMP_DEBUG_ASSERT(nActiveThreads == __kmp_avail_proc); KMP_ASSERT(nActiveThreads > 0); if (nActiveThreads == 1) { __kmp_ncores = nPackages = 1; __kmp_nThreadsPerCore = nCoresPerPkg = 1; if (__kmp_affinity_verbose) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask); KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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); // Check to see if the machine topology is uniform int nPUs = nPackages * __kmp_nThreadsPerCore; if (__kmp_numa_detected) { if (__kmp_tile_depth) { // NUMA + tiles nPUs *= (nNodePerPkg * nTilePerNode * nCorePerTile); } else { // NUMA, no tiles nPUs *= (nNodePerPkg * nCorePerNode); } } else { if (__kmp_tile_depth) { // no NUMA, tiles nPUs *= (nTilePerPkg * nCorePerTile); } else { // no NUMA, no tiles nPUs *= nCoresPerPkg; } } unsigned uniform = (nPUs == nActiveThreads); // Print the machine topology summary. if (__kmp_affinity_verbose) { char mask[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask); } KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc); if (uniform) { KMP_INFORM(Uniform, "KMP_AFFINITY"); } else { KMP_INFORM(NonUniform, "KMP_AFFINITY"); } if (__kmp_numa_detected) { if (__kmp_tile_depth) { // NUMA + tiles KMP_INFORM(TopologyExtraNoTi, "KMP_AFFINITY", nPackages, nNodePerPkg, nTilePerNode, nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores); } else { // NUMA, no tiles KMP_INFORM(TopologyExtraNode, "KMP_AFFINITY", nPackages, nNodePerPkg, nCorePerNode, __kmp_nThreadsPerCore, __kmp_ncores); nPUs *= (nNodePerPkg * nCorePerNode); } } else { if (__kmp_tile_depth) { // no NUMA, tiles KMP_INFORM(TopologyExtraTile, "KMP_AFFINITY", nPackages, nTilePerPkg, nCorePerTile, __kmp_nThreadsPerCore, __kmp_ncores); } else { // no NUMA, no tiles kmp_str_buf_t buf; __kmp_str_buf_init(&buf); __kmp_str_buf_print(&buf, "%d", nPackages); KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg, __kmp_nThreadsPerCore, __kmp_ncores); __kmp_str_buf_free(&buf); } } } if (__kmp_affinity_type == affinity_none) { __kmp_free(retval); KMP_CPU_FREE(oldMask); return 0; } int depth_full = depth; // number of levels before compressing // Find any levels with radiix 1, and remove them from the map // (except for the package level). depth = __kmp_affinity_remove_radix_one_levels(retval, nActiveThreads, depth, levels); KMP_DEBUG_ASSERT(__kmp_affinity_gran != affinity_gran_default); if (__kmp_affinity_gran_levels < 0) { // Set the granularity level based on what levels are modeled // in the machine topology map. __kmp_affinity_gran_levels = 0; // lowest level (e.g. fine) if (__kmp_affinity_gran > affinity_gran_thread) { for (int i = 1; i <= depth_full; ++i) { if (__kmp_affinity_gran <= i) // only count deeper levels break; if (levels[depth_full - i] > 0) __kmp_affinity_gran_levels++; } } if (__kmp_affinity_gran > affinity_gran_package) __kmp_affinity_gran_levels++; // e.g. granularity = group } if (__kmp_affinity_verbose) __kmp_affinity_print_hwloc_tp(retval, nActiveThreads, depth, levels); KMP_CPU_FREE(oldMask); *address2os = retval; return depth; } #endif // KMP_USE_HWLOC // If we don't know how to retrieve the machine's processor topology, or // 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; *msg_id = kmp_i18n_null; // 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; } // 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) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, __kmp_affin_fullMask); KMP_INFORM(AffCapableUseFlat, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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; } // Contruct the data structure to be returned. *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(**address2os) * __kmp_avail_proc); 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); } 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; } #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; *msg_id = kmp_i18n_null; // If we aren't affinity capable, then return now. // The flat mapping will be used. if (!KMP_AFFINITY_CAPABLE()) { // FIXME set *msg_id return -1; } // Contruct 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); 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; } #endif /* KMP_GROUP_AFFINITY */ #if KMP_ARCH_X86 || KMP_ARCH_X86_64 static int __kmp_cpuid_mask_width(int count) { int r = 0; while ((1 << r) < count) ++r; return r; } class apicThreadInfo { public: unsigned osId; // param to __kmp_affinity_bind_thread unsigned apicId; // from cpuid after binding unsigned maxCoresPerPkg; // "" unsigned maxThreadsPerPkg; // "" unsigned pkgId; // inferred from above values unsigned coreId; // "" unsigned threadId; // "" }; static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a, const void *b) { const apicThreadInfo *aa = (const apicThreadInfo *)a; const apicThreadInfo *bb = (const apicThreadInfo *)b; if (aa->pkgId < bb->pkgId) return -1; if (aa->pkgId > bb->pkgId) return 1; if (aa->coreId < bb->coreId) return -1; if (aa->coreId > bb->coreId) return 1; if (aa->threadId < bb->threadId) return -1; if (aa->threadId > bb->threadId) return 1; return 0; } // 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) { kmp_cpuid buf; *address2os = NULL; *msg_id = kmp_i18n_null; // 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; } // 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); // Get an upper bound on the number of threads per package using cpuid(1). // On some OS/chps combinations where HT is supported by the chip but is // disabled, this value will be 2 on a single core chip. Usually, it will be // 2 if HT is enabled and 1 if HT is disabled. __kmp_x86_cpuid(1, 0, &buf); int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff; if (maxThreadsPerPkg == 0) { maxThreadsPerPkg = 1; } // The num cores per pkg comes from cpuid(4). 1 must be added to the encoded // value. // // The author of cpu_count.cpp treated this only an upper bound on the // number of cores, but I haven't seen any cases where it was greater than // the actual number of cores, so we will treat it as exact in this block of // code. // // First, we need to check if cpuid(4) is supported on this chip. To see if // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or // greater. __kmp_x86_cpuid(0, 0, &buf); if (buf.eax >= 4) { __kmp_x86_cpuid(4, 0, &buf); nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1; } else { nCoresPerPkg = 1; } // There is no way to reliably tell if HT is enabled without issuing the // cpuid instruction from every thread, can correlating the cpuid info, so // if the machine is not affinity capable, we assume that HT is off. We have // seen quite a few machines where maxThreadsPerPkg is 2, yet the machine // does not support HT. // // - Older OSes are usually found on machines with older chips, which do not // support HT. // - The performance penalty for mistakenly identifying a machine as HT when // it isn't (which results in blocktime being incorrecly 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; } // 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); // 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 // of this field determines the width of the core# + thread# fields in the // Apic Id. It is also an upper bound on the number of threads per // package, but it has been verified that situations happen were it is not // exact. In particular, on certain OS/chip combinations where Intel(R) // Hyper-Threading Technology is supported by the chip but has been // disabled, the value of this field will be 2 (for a single core chip). // On other OS/chip combinations supporting Intel(R) Hyper-Threading // Technology, the value of this field will be 1 when Intel(R) // Hyper-Threading Technology is disabled and 2 when it is enabled. // - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value // of this field (+1) determines the width of the core# field in the Apic // Id. The comments in "cpucount.cpp" say that this value is an upper // bound, but the IA-32 architecture manual says that it is exactly the // number of cores per package, and I haven't seen any case where it // wasn't. // // From this information, deduce the package Id, core Id, and thread Id, // and set the corresponding fields in the apicThreadInfo struct. unsigned i; apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate( __kmp_avail_proc * sizeof(apicThreadInfo)); unsigned nApics = 0; 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_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; } 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 // value. // // First, we need to check if cpuid(4) is supported on this chip. To see if // cpuid(n) is supported, issue cpuid(0) and check if eax has the value n // or greater. __kmp_x86_cpuid(0, 0, &buf); if (buf.eax >= 4) { __kmp_x86_cpuid(4, 0, &buf); threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1; } else { threadInfo[nApics].maxCoresPerPkg = 1; } // Infer the pkgId / coreId / threadId using only the info obtained locally. int widthCT = __kmp_cpuid_mask_width(threadInfo[nApics].maxThreadsPerPkg); 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; } 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) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask); KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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; } // 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 // [0 .. coresPerPkg-1] and threadId's are usually assigned // [0..threadsPerCore-1], we don't want to make any such assumptions. // // For that matter, we don't know what coresPerPkg and threadsPerCore (or the // total # packages) are at this point - we want to determine that now. We // only have an upper bound on the first two figures. // // We also perform a consistency check at this point: the values returned by // the cpuid instruction for any thread bound to a given package had better // return the same info for maxThreadsPerPkg and maxCoresPerPkg. nPackages = 1; nCoresPerPkg = 1; __kmp_nThreadsPerCore = 1; unsigned nCores = 1; unsigned pkgCt = 1; // to determine radii unsigned lastPkgId = threadInfo[0].pkgId; unsigned coreCt = 1; unsigned lastCoreId = threadInfo[0].coreId; unsigned threadCt = 1; unsigned lastThreadId = threadInfo[0].threadId; // intra-pkg consist checks unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg; unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg; for (i = 1; i < nApics; i++) { if (threadInfo[i].pkgId != lastPkgId) { nCores++; pkgCt++; lastPkgId = threadInfo[i].pkgId; if ((int)coreCt > nCoresPerPkg) nCoresPerPkg = coreCt; coreCt = 1; lastCoreId = threadInfo[i].coreId; if ((int)threadCt > __kmp_nThreadsPerCore) __kmp_nThreadsPerCore = threadCt; threadCt = 1; lastThreadId = threadInfo[i].threadId; // This is a different package, so go on to the next iteration without // doing any consistency checks. Reset the consistency check vars, though. prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg; prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg; continue; } if (threadInfo[i].coreId != lastCoreId) { nCores++; coreCt++; lastCoreId = threadInfo[i].coreId; if ((int)threadCt > __kmp_nThreadsPerCore) __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; } // 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; } } 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) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask); KMP_INFORM(AffUseGlobCpuid, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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 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); KMP_ASSERT(depth > 0); *address2os = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics); for (i = 0; i < nApics; ++i) { Address addr(depth); unsigned os = threadInfo[i].osId; int d = 0; if (pkgLevel >= 0) { addr.labels[d++] = threadInfo[i].pkgId; } if (coreLevel >= 0) { addr.labels[d++] = threadInfo[i].coreId; } if (threadLevel >= 0) { addr.labels[d++] = threadInfo[i].threadId; } (*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. __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++; } } if (__kmp_affinity_verbose) { __kmp_affinity_print_topology(*address2os, nApics, depth, pkgLevel, coreLevel, threadLevel); } __kmp_free(threadInfo); KMP_CPU_FREE(oldMask); return depth; } // Intel(R) microarchitecture code name Nehalem, Dunnington and later // architectures support a newer interface for specifying the x2APIC Ids, // based on cpuid leaf 11. static int __kmp_affinity_create_x2apicid_map(AddrUnsPair **address2os, kmp_i18n_id_t *const msg_id) { kmp_cpuid buf; *address2os = NULL; *msg_id = kmp_i18n_null; // Check to see if cpuid leaf 11 is supported. __kmp_x86_cpuid(0, 0, &buf); if (buf.eax < 11) { *msg_id = kmp_i18n_str_NoLeaf11Support; return -1; } __kmp_x86_cpuid(11, 0, &buf); if (buf.ebx == 0) { *msg_id = kmp_i18n_str_NoLeaf11Support; return -1; } // Find the number of levels in the machine topology. While we're at it, get // the default values for __kmp_nThreadsPerCore & nCoresPerPkg. We will try to // get more accurate values later by explicitly counting them, but get // reasonable defaults now, in case we return early. int level; int threadLevel = -1; int coreLevel = -1; int pkgLevel = -1; __kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1; for (level = 0;; level++) { if (level > 31) { // FIXME: Hack for DPD200163180 // // If level is big then something went wrong -> exiting // // There could actually be 32 valid levels in the machine topology, but so // far, the only machine we have seen which does not exit this loop before // iteration 32 has fubar x2APIC settings. // // For now, just reject this case based upon loop trip count. *msg_id = kmp_i18n_str_InvalidCpuidInfo; return -1; } __kmp_x86_cpuid(11, level, &buf); if (buf.ebx == 0) { if (pkgLevel < 0) { // Will infer nPackages from __kmp_xproc pkgLevel = level; level++; } break; } int kind = (buf.ecx >> 8) & 0xff; if (kind == 1) { // SMT level threadLevel = level; coreLevel = -1; pkgLevel = -1; __kmp_nThreadsPerCore = buf.ebx & 0xffff; if (__kmp_nThreadsPerCore == 0) { *msg_id = kmp_i18n_str_InvalidCpuidInfo; return -1; } } else if (kind == 2) { // core level coreLevel = level; pkgLevel = -1; nCoresPerPkg = buf.ebx & 0xffff; if (nCoresPerPkg == 0) { *msg_id = kmp_i18n_str_InvalidCpuidInfo; return -1; } } else { if (level <= 0) { *msg_id = kmp_i18n_str_InvalidCpuidInfo; return -1; } if (pkgLevel >= 0) { continue; } pkgLevel = level; nPackages = buf.ebx & 0xffff; if (nPackages == 0) { *msg_id = kmp_i18n_str_InvalidCpuidInfo; return -1; } } } int depth = level; // In the above loop, "level" was counted from the finest level (usually // thread) to the coarsest. The caller expects that we will place the labels // in (*address2os)[].first.labels[] in the inverse order, so we need to // invert the vars saying which level means what. if (threadLevel >= 0) { threadLevel = depth - threadLevel - 1; } if (coreLevel >= 0) { coreLevel = depth - coreLevel - 1; } KMP_DEBUG_ASSERT(pkgLevel >= 0); pkgLevel = depth - pkgLevel - 1; // 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); __kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore; nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg; if (__kmp_affinity_verbose) { KMP_INFORM(AffNotCapableUseLocCpuidL11, "KMP_AFFINITY"); KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc); if (__kmp_affinity_uniform_topology()) { KMP_INFORM(Uniform, "KMP_AFFINITY"); } else { KMP_INFORM(NonUniform, "KMP_AFFINITY"); } KMP_INFORM(Topology, "KMP_AFFINITY", nPackages, nCoresPerPkg, __kmp_nThreadsPerCore, __kmp_ncores); } 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. AddrUnsPair *retval = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * __kmp_avail_proc); // 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; KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) { // 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_affinity_dispatch->bind_thread(proc); // Extract labels for each level in the machine topology map from Apic ID. Address addr(depth); int prev_shift = 0; for (level = 0; level < depth; level++) { __kmp_x86_cpuid(11, level, &buf); unsigned apicId = buf.edx; if (buf.ebx == 0) { if (level != depth - 1) { KMP_CPU_FREE(oldMask); *msg_id = kmp_i18n_str_InconsistentCpuidInfo; return -1; } addr.labels[depth - level - 1] = apicId >> prev_shift; level++; break; } int shift = buf.eax & 0x1f; int mask = (1 << shift) - 1; addr.labels[depth - level - 1] = (apicId & mask) >> prev_shift; prev_shift = shift; } if (level != depth) { KMP_CPU_FREE(oldMask); *msg_id = kmp_i18n_str_InconsistentCpuidInfo; return -1; } retval[nApics] = AddrUnsPair(addr, proc); 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, return now. KMP_ASSERT(nApics > 0); if (nApics == 1) { __kmp_ncores = nPackages = 1; __kmp_nThreadsPerCore = nCoresPerPkg = 1; if (__kmp_affinity_verbose) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, oldMask); KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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[pkgLevel]; 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); // Find the radix at each of the levels. unsigned *totals = (unsigned *)__kmp_allocate(depth * sizeof(unsigned)); unsigned *counts = (unsigned *)__kmp_allocate(depth * sizeof(unsigned)); unsigned *maxCt = (unsigned *)__kmp_allocate(depth * sizeof(unsigned)); unsigned *last = (unsigned *)__kmp_allocate(depth * sizeof(unsigned)); for (level = 0; level < depth; level++) { totals[level] = 1; maxCt[level] = 1; counts[level] = 1; last[level] = retval[0].first.labels[level]; } // From here on, the iteration variable "level" runs from the finest level to // the coarsest, i.e. we iterate forward through // (*address2os)[].first.labels[] - in the previous loops, we iterated // backwards. for (proc = 1; (int)proc < nApics; proc++) { int level; for (level = 0; level < depth; level++) { if (retval[proc].first.labels[level] != last[level]) { int j; for (j = level + 1; j < depth; j++) { totals[j]++; counts[j] = 1; // The line below causes printing incorrect topology information in // case the max value for some level (maxCt[level]) is encountered // earlier than some less value while going through the array. For // example, let pkg0 has 4 cores and pkg1 has 2 cores. Then // maxCt[1] == 2 // whereas it must be 4. // TODO!!! Check if it can be commented safely // maxCt[j] = 1; last[j] = retval[proc].first.labels[j]; } totals[level]++; counts[level]++; if (counts[level] > maxCt[level]) { maxCt[level] = counts[level]; } last[level] = retval[proc].first.labels[level]; break; } else if (level == depth - 1) { __kmp_free(last); __kmp_free(maxCt); __kmp_free(counts); __kmp_free(totals); __kmp_free(retval); KMP_CPU_FREE(oldMask); *msg_id = kmp_i18n_str_x2ApicIDsNotUnique; return -1; } } } // 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. if (threadLevel >= 0) { __kmp_nThreadsPerCore = maxCt[threadLevel]; } else { __kmp_nThreadsPerCore = 1; } nPackages = totals[pkgLevel]; if (coreLevel >= 0) { __kmp_ncores = totals[coreLevel]; nCoresPerPkg = maxCt[coreLevel]; } else { __kmp_ncores = nPackages; nCoresPerPkg = 1; } // Check to see if the machine topology is uniform unsigned prod = maxCt[0]; for (level = 1; level < depth; level++) { prod *= maxCt[level]; } bool uniform = (prod == totals[level - 1]); // Print the machine topology summary. if (__kmp_affinity_verbose) { char mask[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(mask, KMP_AFFIN_MASK_PRINT_LEN, oldMask); KMP_INFORM(AffUseGlobCpuidL11, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", mask); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", mask); } KMP_INFORM(AvailableOSProc, "KMP_AFFINITY", __kmp_avail_proc); if (uniform) { KMP_INFORM(Uniform, "KMP_AFFINITY"); } else { KMP_INFORM(NonUniform, "KMP_AFFINITY"); } kmp_str_buf_t buf; __kmp_str_buf_init(&buf); __kmp_str_buf_print(&buf, "%d", totals[0]); for (level = 1; level <= pkgLevel; level++) { __kmp_str_buf_print(&buf, " x %d", maxCt[level]); } KMP_INFORM(TopologyExtra, "KMP_AFFINITY", buf.str, nCoresPerPkg, __kmp_nThreadsPerCore, __kmp_ncores); __kmp_str_buf_free(&buf); } 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(last); __kmp_free(maxCt); __kmp_free(counts); __kmp_free(totals); __kmp_free(retval); KMP_CPU_FREE(oldMask); return 0; } // Find any levels with radiix 1, and remove them from the map // (except for the package level). int new_depth = 0; for (level = 0; level < depth; level++) { if ((maxCt[level] == 1) && (level != pkgLevel)) { continue; } new_depth++; } // If we are removing any levels, allocate a new vector to return, // and copy the relevant information to it. if (new_depth != depth) { AddrUnsPair *new_retval = (AddrUnsPair *)__kmp_allocate(sizeof(AddrUnsPair) * nApics); for (proc = 0; (int)proc < nApics; proc++) { Address addr(new_depth); new_retval[proc] = AddrUnsPair(addr, retval[proc].second); } int new_level = 0; int newPkgLevel = -1; int newCoreLevel = -1; int newThreadLevel = -1; for (level = 0; level < depth; level++) { if ((maxCt[level] == 1) && (level != pkgLevel)) { // Remove this level. Never remove the package level continue; } if (level == pkgLevel) { newPkgLevel = new_level; } if (level == coreLevel) { newCoreLevel = new_level; } if (level == threadLevel) { newThreadLevel = new_level; } for (proc = 0; (int)proc < nApics; proc++) { new_retval[proc].first.labels[new_level] = retval[proc].first.labels[level]; } new_level++; } __kmp_free(retval); retval = new_retval; depth = new_depth; pkgLevel = newPkgLevel; coreLevel = newCoreLevel; threadLevel = newThreadLevel; } 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 (__kmp_affinity_gran > affinity_gran_package) { __kmp_affinity_gran_levels++; } } if (__kmp_affinity_verbose) { __kmp_affinity_print_topology(retval, nApics, depth, pkgLevel, coreLevel, threadLevel); } __kmp_free(last); __kmp_free(maxCt); __kmp_free(counts); __kmp_free(totals); KMP_CPU_FREE(oldMask); *address2os = retval; return depth; } #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ #define osIdIndex 0 #define threadIdIndex 1 #define coreIdIndex 2 #define pkgIdIndex 3 #define nodeIdIndex 4 typedef unsigned *ProcCpuInfo; static unsigned maxIndex = pkgIdIndex; static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a, const void *b) { unsigned i; const unsigned *aa = *(unsigned *const *)a; const unsigned *bb = *(unsigned *const *)b; for (i = maxIndex;; i--) { if (aa[i] < bb[i]) return -1; if (aa[i] > bb[i]) return 1; if (i == osIdIndex) break; } return 0; } #if KMP_USE_HIER_SCHED // Set the array sizes for the hierarchy layers static void __kmp_dispatch_set_hierarchy_values() { // Set the maximum number of L1's to number of cores // Set the maximum number of L2's to to either number of cores / 2 for // Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing // Or the number of cores for Intel(R) Xeon(R) processors // Set the maximum number of NUMA nodes and L3's to number of packages __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] = nPackages * nCoresPerPkg * __kmp_nThreadsPerCore; __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores; #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS) if (__kmp_mic_type >= mic3) __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2; else #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS) __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores; __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages; __kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages; __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1; // Set the number of threads per unit // Number of hardware threads per L1/L2/L3/NUMA/LOOP __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1; __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_nThreadsPerCore; #if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS) if (__kmp_mic_type >= mic3) __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] = 2 * __kmp_nThreadsPerCore; else #endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS) __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_nThreadsPerCore; __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] = nCoresPerPkg * __kmp_nThreadsPerCore; __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] = nCoresPerPkg * __kmp_nThreadsPerCore; __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] = nPackages * nCoresPerPkg * __kmp_nThreadsPerCore; } // Return the index into the hierarchy for this tid and layer type (L1, L2, etc) // i.e., this thread's L1 or this thread's L2, etc. int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) { int index = type + 1; int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1]; KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST); if (type == kmp_hier_layer_e::LAYER_THREAD) return tid; else if (type == kmp_hier_layer_e::LAYER_LOOP) return 0; KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0); if (tid >= num_hw_threads) tid = tid % num_hw_threads; return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index]; } // Return the number of t1's per t2 int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) { int i1 = t1 + 1; int i2 = t2 + 1; KMP_DEBUG_ASSERT(i1 <= i2); 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 // 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; *msg_id = kmp_i18n_null; // Scan of the file, and count the number of "processor" (osId) fields, // and find the highest value of for a node_ 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 break; } char s1[] = "processor"; if (strncmp(buf, s1, sizeof(s1) - 1) == 0) { num_records++; continue; } // FIXME - this will match "node_ " unsigned level; if (KMP_SSCANF(buf, "node_%u id", &level) == 1) { if (nodeIdIndex + level >= maxIndex) { maxIndex = nodeIdIndex + level; } 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; } if (num_records > (unsigned)__kmp_xproc) { *line = 0; *msg_id = kmp_i18n_str_TooManyProcRecords; return -1; } // Set the file pointer back to the begginning, 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; } // 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++) { threadInfo[i] = (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned)); } #define CLEANUP_THREAD_INFO \ for (i = 0; i <= num_records; i++) { \ __kmp_free(threadInfo[i]); \ } \ __kmp_free(threadInfo); // A value of UINT_MAX means that we didn't find the field unsigned __index; #define INIT_PROC_INFO(p) \ for (__index = 0; __index <= maxIndex; __index++) { \ (p)[__index] = UINT_MAX; \ } for (i = 0; i <= num_records; i++) { INIT_PROC_INFO(threadInfo[i]); } unsigned num_avail = 0; *line = 0; while (!feof(f)) { // Create an inner scoping level, so that all the goto targets at the end of // the loop appear in an outer scoping level. This avoids warnings about // jumping past an initialization to a target in the same block. { buf[sizeof(buf) - 1] = 1; bool long_line = false; if (!fgets(buf, sizeof(buf), f)) { // Read errors presumably because of EOF // If there is valid data in threadInfo[num_avail], then fake // a blank line in ensure that the last address gets parsed. bool valid = false; for (i = 0; i <= maxIndex; i++) { if (threadInfo[num_avail][i] != UINT_MAX) { valid = true; } } if (!valid) { break; } buf[0] = 0; } else if (!buf[sizeof(buf) - 1]) { // 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; \ } } (*line)++; char s1[] = "processor"; if (strncmp(buf, s1, sizeof(s1) - 1) == 0) { CHECK_LINE; char *p = strchr(buf + sizeof(s1) - 1, ':'); unsigned val; if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1)) goto no_val; if (threadInfo[num_avail][osIdIndex] != UINT_MAX) #if KMP_ARCH_AARCH64 // Handle the old AArch64 /proc/cpuinfo layout differently, // it contains all of the 'processor' entries listed in a // single 'Processor' section, therefore the normal looking // for duplicates in that section will always fail. num_avail++; #else goto dup_field; #endif threadInfo[num_avail][osIdIndex] = val; #if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64) char path[256]; KMP_SNPRINTF( path, sizeof(path), "/sys/devices/system/cpu/cpu%u/topology/physical_package_id", threadInfo[num_avail][osIdIndex]); __kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]); KMP_SNPRINTF(path, sizeof(path), "/sys/devices/system/cpu/cpu%u/topology/core_id", threadInfo[num_avail][osIdIndex]); __kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]); continue; #else } char s2[] = "physical id"; if (strncmp(buf, s2, sizeof(s2) - 1) == 0) { CHECK_LINE; char *p = strchr(buf + sizeof(s2) - 1, ':'); unsigned val; if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1)) goto no_val; if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX) goto dup_field; threadInfo[num_avail][pkgIdIndex] = val; continue; } char s3[] = "core id"; if (strncmp(buf, s3, sizeof(s3) - 1) == 0) { CHECK_LINE; char *p = strchr(buf + sizeof(s3) - 1, ':'); unsigned val; if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1)) goto no_val; if (threadInfo[num_avail][coreIdIndex] != UINT_MAX) goto dup_field; threadInfo[num_avail][coreIdIndex] = val; continue; #endif // KMP_OS_LINUX && USE_SYSFS_INFO } char s4[] = "thread id"; if (strncmp(buf, s4, sizeof(s4) - 1) == 0) { CHECK_LINE; char *p = strchr(buf + sizeof(s4) - 1, ':'); unsigned val; if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1)) goto no_val; if (threadInfo[num_avail][threadIdIndex] != UINT_MAX) goto dup_field; threadInfo[num_avail][threadIdIndex] = val; continue; } unsigned level; if (KMP_SSCANF(buf, "node_%u id", &level) == 1) { CHECK_LINE; char *p = strchr(buf + sizeof(s4) - 1, ':'); unsigned val; if ((p == NULL) || (KMP_SSCANF(p + 1, "%u\n", &val) != 1)) goto no_val; KMP_ASSERT(nodeIdIndex + level <= maxIndex); if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX) goto dup_field; threadInfo[num_avail][nodeIdIndex + level] = val; continue; } // We didn't recognize the leading token on the line. There are lots of // leading tokens that we don't recognize - if the line isn't empty, go on // to the next line. if ((*buf != 0) && (*buf != '\n')) { // If the line is longer than the buffer, read characters // until we find a newline. if (long_line) { int ch; while (((ch = fgetc(f)) != EOF) && (ch != '\n')) ; } 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; } // 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; } if (threadInfo[0][pkgIdIndex] == UINT_MAX) { CLEANUP_THREAD_INFO; *msg_id = kmp_i18n_str_MissingPhysicalIDField; return -1; } // 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; dup_field: CLEANUP_THREAD_INFO; *msg_id = kmp_i18n_str_DuplicateFieldCpuinfo; return -1; } *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. 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 { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, __kmp_affin_fullMask); KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", buf); } 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 // [0 .. coresPerPkg-1] and threadId's are usually assigned // [0..threadsPerCore-1], we don't want to make any such assumptions. // // For that matter, we don't know what coresPerPkg and threadsPerCore (or the // total # packages) are at this point - we want to determine that now. We // only have an upper bound on the first two figures. unsigned *counts = (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned)); unsigned *maxCt = (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned)); unsigned *totals = (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned)); unsigned *lastId = (unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned)); bool assign_thread_ids = false; unsigned threadIdCt; unsigned index; restart_radix_check: threadIdCt = 0; // Initialize the counter arrays with data from threadInfo[0]. if (assign_thread_ids) { if (threadInfo[0][threadIdIndex] == UINT_MAX) { threadInfo[0][threadIdIndex] = threadIdCt++; } else if (threadIdCt <= threadInfo[0][threadIdIndex]) { threadIdCt = threadInfo[0][threadIdIndex] + 1; } } for (index = 0; index <= maxIndex; index++) { counts[index] = 1; maxCt[index] = 1; totals[index] = 1; lastId[index] = threadInfo[0][index]; ; } // Run through the rest of the OS procs. for (i = 1; i < num_avail; i++) { // Find the most significant index whose id differs from the id for the // previous OS proc. for (index = maxIndex; index >= threadIdIndex; index--) { if (assign_thread_ids && (index == threadIdIndex)) { // Auto-assign the thread id field if it wasn't specified. if (threadInfo[i][threadIdIndex] == UINT_MAX) { threadInfo[i][threadIdIndex] = threadIdCt++; } // Apparently the thread id field was specified for some entries and not // others. Start the thread id counter off at the next higher thread id. else if (threadIdCt <= threadInfo[i][threadIdIndex]) { threadIdCt = threadInfo[i][threadIdIndex] + 1; } } if (threadInfo[i][index] != lastId[index]) { // Run through all indices which are less significant, and reset the // counts to 1. At all levels up to and including index, we need to // increment the totals and record the last id. unsigned index2; for (index2 = threadIdIndex; index2 < index; index2++) { totals[index2]++; if (counts[index2] > maxCt[index2]) { maxCt[index2] = counts[index2]; } counts[index2] = 1; lastId[index2] = threadInfo[i][index2]; } counts[index]++; totals[index]++; lastId[index] = threadInfo[i][index]; if (assign_thread_ids && (index > threadIdIndex)) { #if KMP_MIC && REDUCE_TEAM_SIZE // The default team size is the total #threads in the machine // minus 1 thread for every core that has 3 or more threads. teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1); #endif // KMP_MIC && REDUCE_TEAM_SIZE // Restart the thread counter, as we are on a new core. threadIdCt = 0; // Auto-assign the thread id field if it wasn't specified. if (threadInfo[i][threadIdIndex] == UINT_MAX) { threadInfo[i][threadIdIndex] = threadIdCt++; } // Aparrently the thread id field was specified for some entries and // not others. Start the thread id counter off at the next higher // thread id. else if (threadIdCt <= threadInfo[i][threadIdIndex]) { threadIdCt = threadInfo[i][threadIdIndex] + 1; } } break; } } if (index < threadIdIndex) { // If thread ids were specified, it is an error if they are not unique. // Also, check that we waven't already restarted the loop (to be safe - // 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; } // 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; } } #if KMP_MIC && REDUCE_TEAM_SIZE // The default team size is the total #threads in the machine // minus 1 thread for every core that has 3 or more threads. teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1); #endif // KMP_MIC && REDUCE_TEAM_SIZE for (index = threadIdIndex; index <= maxIndex; index++) { if (counts[index] > maxCt[index]) { maxCt[index] = counts[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 { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, __kmp_affin_fullMask); KMP_INFORM(AffCapableUseCpuinfo, "KMP_AFFINITY"); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_AFFINITY", buf); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_AFFINITY", 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_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); } #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; int depth = 0; for (index = threadIdIndex; index <= maxIndex; index++) { if (inMap[index]) { depth++; } } 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; for (i = 0; i < num_avail; ++i) { Address addr(depth); unsigned os = threadInfo[i][osIdIndex]; int src_index; int dst_index = 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; } else if (src_index == coreIdIndex) { coreLevel = dst_index; } else if (src_index == threadIdIndex) { threadLevel = dst_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; } // 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) { // 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; maxOsId = 0; for (i = numAddrs - 1;; --i) { unsigned osId = address2os[i].second; if (osId > maxOsId) { maxOsId = osId; } if (i == 0) break; } kmp_affin_mask_t *osId2Mask; KMP_CPU_ALLOC_ARRAY(osId2Mask, (maxOsId + 1)); // Sort the address2os table according to physical order. Doing so will put // all threads on the same core/package/node in consecutive locations. qsort(address2os, numAddrs, sizeof(*address2os), __kmp_affinity_cmp_Address_labels); 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 // 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); kmp_affin_mask_t *sum; KMP_CPU_ALLOC_ON_STACK(sum); KMP_CPU_ZERO(sum); KMP_CPU_SET(address2os[0].second, 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); 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; 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); } unique++; // Start a new mask. leader = i; leaderAddr = &(address2os[i].first); KMP_CPU_ZERO(sum); KMP_CPU_SET(address2os[i].second, 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; 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); } unique++; KMP_CPU_FREE_FROM_STACK(sum); *maxIndex = maxOsId; *numUnique = unique; return osId2Mask; } // Stuff for the affinity proclist parsers. It's easier to declare these vars // as file-static than to try and pass them through the calling sequence of // the recursive-descent OMP_PLACES parser. static kmp_affin_mask_t *newMasks; static int numNewMasks; static int nextNewMask; #define ADD_MASK(_mask) \ { \ if (nextNewMask >= numNewMasks) { \ int i; \ numNewMasks *= 2; \ kmp_affin_mask_t *temp; \ KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \ for (i = 0; i < numNewMasks / 2; i++) { \ kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \ kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \ KMP_CPU_COPY(dest, src); \ } \ KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \ newMasks = temp; \ } \ KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \ nextNewMask++; \ } #define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \ { \ if (((_osId) > _maxOsId) || \ (!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \ if (__kmp_affinity_verbose || \ (__kmp_affinity_warnings && \ (__kmp_affinity_type != affinity_none))) { \ KMP_WARNING(AffIgnoreInvalidProcID, _osId); \ } \ } else { \ ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \ } \ } // Re-parse the proclist (for the explicit affinity type), and form the list // of affinity newMasks indexed by gtid. static void __kmp_affinity_process_proclist(kmp_affin_mask_t **out_masks, unsigned int *out_numMasks, const char *proclist, kmp_affin_mask_t *osId2Mask, int maxOsId) { int i; const char *scan = proclist; const char *next = proclist; // We use malloc() for the temporary mask vector, so that we can use // realloc() to extend it. numNewMasks = 2; KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks); nextNewMask = 0; kmp_affin_mask_t *sumMask; KMP_CPU_ALLOC(sumMask); int setSize = 0; for (;;) { int start, end, stride; SKIP_WS(scan); next = scan; if (*next == '\0') { break; } if (*next == '{') { int num; setSize = 0; next++; // skip '{' SKIP_WS(next); scan = next; // Read the first integer in the set. KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist"); SKIP_DIGITS(next); num = __kmp_str_to_int(scan, *next); KMP_ASSERT2(num >= 0, "bad explicit proc list"); // Copy the mask for that osId to the sum (union) mask. if ((num > maxOsId) || (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, num); } KMP_CPU_ZERO(sumMask); } else { KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num)); setSize = 1; } for (;;) { // Check for end of set. SKIP_WS(next); if (*next == '}') { next++; // skip '}' break; } // Skip optional comma. if (*next == ',') { next++; } SKIP_WS(next); // Read the next integer in the set. scan = next; KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list"); SKIP_DIGITS(next); num = __kmp_str_to_int(scan, *next); KMP_ASSERT2(num >= 0, "bad explicit proc list"); // Add the mask for that osId to the sum mask. if ((num > maxOsId) || (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, num); } } else { KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num)); setSize++; } } if (setSize > 0) { ADD_MASK(sumMask); } SKIP_WS(next); if (*next == ',') { next++; } scan = next; continue; } // Read the first integer. KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list"); SKIP_DIGITS(next); start = __kmp_str_to_int(scan, *next); KMP_ASSERT2(start >= 0, "bad explicit proc list"); SKIP_WS(next); // If this isn't a range, then add a mask to the list and go on. if (*next != '-') { ADD_MASK_OSID(start, osId2Mask, maxOsId); // Skip optional comma. if (*next == ',') { next++; } scan = next; continue; } // This is a range. Skip over the '-' and read in the 2nd int. next++; // skip '-' SKIP_WS(next); scan = next; KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list"); SKIP_DIGITS(next); end = __kmp_str_to_int(scan, *next); KMP_ASSERT2(end >= 0, "bad explicit proc list"); // Check for a stride parameter stride = 1; SKIP_WS(next); if (*next == ':') { // A stride is specified. Skip over the ':" and read the 3rd int. int sign = +1; next++; // skip ':' SKIP_WS(next); scan = next; if (*next == '-') { sign = -1; next++; SKIP_WS(next); scan = next; } KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list"); SKIP_DIGITS(next); stride = __kmp_str_to_int(scan, *next); KMP_ASSERT2(stride >= 0, "bad explicit proc list"); stride *= sign; } // Do some range checks. KMP_ASSERT2(stride != 0, "bad explicit proc list"); if (stride > 0) { KMP_ASSERT2(start <= end, "bad explicit proc list"); } else { KMP_ASSERT2(start >= end, "bad explicit proc list"); } KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list"); // Add the mask for each OS proc # to the list. if (stride > 0) { do { ADD_MASK_OSID(start, osId2Mask, maxOsId); start += stride; } while (start <= end); } else { do { ADD_MASK_OSID(start, osId2Mask, maxOsId); start += stride; } while (start >= end); } // Skip optional comma. SKIP_WS(next); if (*next == ',') { next++; } scan = next; } *out_numMasks = nextNewMask; if (nextNewMask == 0) { *out_masks = NULL; KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks); return; } KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask); for (i = 0; i < nextNewMask; i++) { kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i); KMP_CPU_COPY(dest, src); } KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks); KMP_CPU_FREE(sumMask); } /*----------------------------------------------------------------------------- Re-parse the OMP_PLACES proc id list, forming the newMasks for the different places. Again, Here is the grammar: place_list := place place_list := place , place_list place := num place := place : num place := place : num : signed place := { subplacelist } place := ! place // (lowest priority) subplace_list := subplace subplace_list := subplace , subplace_list subplace := num subplace := num : num subplace := num : num : signed signed := num signed := + signed signed := - signed -----------------------------------------------------------------------------*/ static void __kmp_process_subplace_list(const char **scan, kmp_affin_mask_t *osId2Mask, int maxOsId, kmp_affin_mask_t *tempMask, int *setSize) { const char *next; for (;;) { int start, count, stride, i; // Read in the starting proc id SKIP_WS(*scan); KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list"); next = *scan; SKIP_DIGITS(next); start = __kmp_str_to_int(*scan, *next); KMP_ASSERT(start >= 0); *scan = next; // valid follow sets are ',' ':' and '}' SKIP_WS(*scan); if (**scan == '}' || **scan == ',') { if ((start > maxOsId) || (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, start); } } else { KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start)); (*setSize)++; } if (**scan == '}') { break; } (*scan)++; // skip ',' continue; } KMP_ASSERT2(**scan == ':', "bad explicit places list"); (*scan)++; // skip ':' // Read count parameter SKIP_WS(*scan); KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list"); next = *scan; SKIP_DIGITS(next); count = __kmp_str_to_int(*scan, *next); KMP_ASSERT(count >= 0); *scan = next; // valid follow sets are ',' ':' and '}' SKIP_WS(*scan); if (**scan == '}' || **scan == ',') { for (i = 0; i < count; i++) { if ((start > maxOsId) || (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, start); } break; // don't proliferate warnings for large count } else { KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start)); start++; (*setSize)++; } } if (**scan == '}') { break; } (*scan)++; // skip ',' continue; } KMP_ASSERT2(**scan == ':', "bad explicit places list"); (*scan)++; // skip ':' // Read stride parameter int sign = +1; for (;;) { SKIP_WS(*scan); if (**scan == '+') { (*scan)++; // skip '+' continue; } if (**scan == '-') { sign *= -1; (*scan)++; // skip '-' continue; } break; } SKIP_WS(*scan); KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list"); next = *scan; SKIP_DIGITS(next); stride = __kmp_str_to_int(*scan, *next); KMP_ASSERT(stride >= 0); *scan = next; stride *= sign; // valid follow sets are ',' and '}' SKIP_WS(*scan); if (**scan == '}' || **scan == ',') { for (i = 0; i < count; i++) { if ((start > maxOsId) || (!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, start); } break; // don't proliferate warnings for large count } else { KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start)); start += stride; (*setSize)++; } } if (**scan == '}') { break; } (*scan)++; // skip ',' continue; } KMP_ASSERT2(0, "bad explicit places list"); } } static void __kmp_process_place(const char **scan, kmp_affin_mask_t *osId2Mask, int maxOsId, kmp_affin_mask_t *tempMask, int *setSize) { const char *next; // valid follow sets are '{' '!' and num SKIP_WS(*scan); if (**scan == '{') { (*scan)++; // skip '{' __kmp_process_subplace_list(scan, osId2Mask, maxOsId, tempMask, setSize); KMP_ASSERT2(**scan == '}', "bad explicit places list"); (*scan)++; // skip '}' } else if (**scan == '!') { (*scan)++; // skip '!' __kmp_process_place(scan, osId2Mask, maxOsId, tempMask, setSize); KMP_CPU_COMPLEMENT(maxOsId, tempMask); } else if ((**scan >= '0') && (**scan <= '9')) { next = *scan; SKIP_DIGITS(next); int num = __kmp_str_to_int(*scan, *next); KMP_ASSERT(num >= 0); if ((num > maxOsId) || (!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(AffIgnoreInvalidProcID, num); } } else { KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num)); (*setSize)++; } *scan = next; // skip num } else { KMP_ASSERT2(0, "bad explicit places list"); } } // static void void __kmp_affinity_process_placelist(kmp_affin_mask_t **out_masks, unsigned int *out_numMasks, const char *placelist, kmp_affin_mask_t *osId2Mask, int maxOsId) { int i, j, count, stride, sign; const char *scan = placelist; const char *next = placelist; numNewMasks = 2; KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks); nextNewMask = 0; // tempMask is modified based on the previous or initial // place to form the current place // previousMask contains the previous place kmp_affin_mask_t *tempMask; kmp_affin_mask_t *previousMask; KMP_CPU_ALLOC(tempMask); KMP_CPU_ZERO(tempMask); KMP_CPU_ALLOC(previousMask); KMP_CPU_ZERO(previousMask); int setSize = 0; for (;;) { __kmp_process_place(&scan, osId2Mask, maxOsId, tempMask, &setSize); // valid follow sets are ',' ':' and EOL SKIP_WS(scan); if (*scan == '\0' || *scan == ',') { if (setSize > 0) { ADD_MASK(tempMask); } KMP_CPU_ZERO(tempMask); setSize = 0; if (*scan == '\0') { break; } scan++; // skip ',' continue; } KMP_ASSERT2(*scan == ':', "bad explicit places list"); scan++; // skip ':' // Read count parameter SKIP_WS(scan); KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list"); next = scan; SKIP_DIGITS(next); count = __kmp_str_to_int(scan, *next); KMP_ASSERT(count >= 0); scan = next; // valid follow sets are ',' ':' and EOL SKIP_WS(scan); if (*scan == '\0' || *scan == ',') { stride = +1; } else { KMP_ASSERT2(*scan == ':', "bad explicit places list"); scan++; // skip ':' // Read stride parameter sign = +1; for (;;) { SKIP_WS(scan); if (*scan == '+') { scan++; // skip '+' continue; } if (*scan == '-') { sign *= -1; scan++; // skip '-' continue; } break; } SKIP_WS(scan); KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list"); next = scan; SKIP_DIGITS(next); stride = __kmp_str_to_int(scan, *next); KMP_DEBUG_ASSERT(stride >= 0); scan = next; stride *= sign; } // Add places determined by initial_place : count : stride for (i = 0; i < count; i++) { if (setSize == 0) { break; } // Add the current place, then build the next place (tempMask) from that KMP_CPU_COPY(previousMask, tempMask); ADD_MASK(previousMask); KMP_CPU_ZERO(tempMask); setSize = 0; KMP_CPU_SET_ITERATE(j, previousMask) { if (!KMP_CPU_ISSET(j, previousMask)) { continue; } if ((j + stride > maxOsId) || (j + stride < 0) || (!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) || (!KMP_CPU_ISSET(j + stride, KMP_CPU_INDEX(osId2Mask, j + stride)))) { if ((__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) && i < count - 1) { KMP_WARNING(AffIgnoreInvalidProcID, j + stride); } continue; } KMP_CPU_SET(j + stride, tempMask); setSize++; } } KMP_CPU_ZERO(tempMask); setSize = 0; // valid follow sets are ',' and EOL SKIP_WS(scan); if (*scan == '\0') { break; } if (*scan == ',') { scan++; // skip ',' continue; } KMP_ASSERT2(0, "bad explicit places list"); } *out_numMasks = nextNewMask; if (nextNewMask == 0) { *out_masks = NULL; KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks); return; } KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask); KMP_CPU_FREE(tempMask); KMP_CPU_FREE(previousMask); for (i = 0; i < nextNewMask; i++) { kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), 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_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 > 3) { 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 (__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_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 += nCoresPerPkg * __kmp_nThreadsPerCore; if (__kmp_pu_os_idx != NULL) { // walk through skipped socket 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 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 * nCoresPerPkg * __kmp_nThreadsPerCore); KMP_DEBUG_ASSERT(n_new == __kmp_hws_socket.num * __kmp_hws_core.num * __kmp_hws_proc.num); nPackages = __kmp_hws_socket.num; // correct nPackages nCoresPerPkg = __kmp_hws_core.num; // correct nCoresPerPkg __kmp_nThreadsPerCore = __kmp_hws_proc.num; // correct __kmp_nThreadsPerCore __kmp_avail_proc = n_new; // correct avail_proc __kmp_ncores = nPackages * __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) { char m[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(m, KMP_AFFIN_MASK_PRINT_LEN, __kmp_affin_fullMask); if (__kmp_affinity_respect_mask) { KMP_INFORM(InitOSProcSetRespect, "KMP_HW_SUBSET", m); } else { KMP_INFORM(InitOSProcSetNotRespect, "KMP_HW_SUBSET", m); } 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) { int core_level = 0; for (int i = 0; i < nprocs; i++) { for (int j = bottom_level; j > 0; j--) { if (address2os[i].first.labels[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, 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++; } } 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; } // 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, 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; } 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 // initialization thread's affinity mask. Otherwise, it is all processors that // we know about on the machine. if (__kmp_affin_fullMask == NULL) { KMP_CPU_ALLOC(__kmp_affin_fullMask); } if (KMP_AFFINITY_CAPABLE()) { if (__kmp_affinity_respect_mask) { __kmp_get_system_affinity(__kmp_affin_fullMask, TRUE); // Count the number of available processors. unsigned i; __kmp_avail_proc = 0; KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) { if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) { continue; } __kmp_avail_proc++; } if (__kmp_avail_proc > __kmp_xproc) { if (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none))) { KMP_WARNING(ErrorInitializeAffinity); } __kmp_affinity_type = affinity_none; KMP_AFFINITY_DISABLE(); return; } } else { __kmp_affinity_entire_machine_mask(__kmp_affin_fullMask); __kmp_avail_proc = __kmp_xproc; } } 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; } 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, an the nowarnings flag was not set. const char *file_name = NULL; int line = 0; #if KMP_USE_HWLOC if (depth < 0 && __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; } else if (depth < 0 && __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)); } 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; } } } #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"); } } FILE *f = fopen("/proc/cpuinfo", "r"); if (f == NULL) { msg_id = kmp_i18n_str_CantOpenCpuinfo; } else { file_name = "/proc/cpuinfo"; depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f); fclose(f); if (depth == 0) { KMP_EXIT_AFF_NONE; } } } #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"); } 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; } KMP_ASSERT(depth > 0); KMP_ASSERT(address2os != NULL); } } #if KMP_USE_HWLOC else if (__kmp_affinity_top_method == affinity_top_method_hwloc) { KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC); if (__kmp_affinity_verbose) { KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY"); } depth = __kmp_affinity_create_hwloc_map(&address2os, &msg_id); if (depth == 0) { KMP_EXIT_AFF_NONE; } } #endif // KMP_USE_HWLOC // If the user has specified that a paricular topology discovery method is to be // used, then we abort if that method fails. The exception is group affinity, // which might have been implicitly set. #if KMP_ARCH_X86 || KMP_ARCH_X86_64 else if (__kmp_affinity_top_method == affinity_top_method_x2apicid) { if (__kmp_affinity_verbose) { 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) { 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) { 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; if (__kmp_cpuinfo_file != NULL) { filename = __kmp_cpuinfo_file; } else { filename = "/proc/cpuinfo"; } if (__kmp_affinity_verbose) { KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename); } FILE *f = fopen(filename, "r"); if (f == NULL) { int code = errno; if (__kmp_cpuinfo_file != NULL) { __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), KMP_HNT(NameComesFrom_CPUINFO_FILE), __kmp_msg_null); } else { __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), __kmp_msg_null); } } int line = 0; depth = __kmp_affinity_create_cpuinfo_map(&address2os, &line, &msg_id, f); fclose(f); if (depth < 0) { KMP_ASSERT(msg_id != kmp_i18n_null); 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) { 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; } // should not fail KMP_ASSERT(depth > 0); KMP_ASSERT(address2os != NULL); } #if KMP_USE_HIER_SCHED __kmp_dispatch_set_hierarchy_values(); #endif if (address2os == NULL) { if (KMP_AFFINITY_CAPABLE() && (__kmp_affinity_verbose || (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none)))) { KMP_WARNING(ErrorInitializeAffinity); } __kmp_affinity_type = affinity_none; __kmp_create_affinity_none_places(); KMP_AFFINITY_DISABLE(); return; } if (__kmp_affinity_gran == affinity_gran_tile #if KMP_USE_HWLOC && __kmp_tile_depth == 0 #endif ) { // tiles requested but not detected, warn user on this KMP_WARNING(AffTilesNoTiles, "KMP_AFFINITY"); } __kmp_apply_thread_places(&address2os, 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); 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. case affinity_logical: __kmp_affinity_compact = 0; if (__kmp_affinity_offset) { __kmp_affinity_offset = __kmp_nThreadsPerCore * __kmp_affinity_offset % __kmp_avail_proc; } goto sortAddresses; 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; case affinity_scatter: if (__kmp_affinity_compact >= depth) { __kmp_affinity_compact = 0; } else { __kmp_affinity_compact = depth - 1 - __kmp_affinity_compact; } goto sortAddresses; case affinity_compact: if (__kmp_affinity_compact >= depth) { __kmp_affinity_compact = depth - 1; } goto sortAddresses; 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()) { // 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 maxprocpercore = __kmp_affinity_max_proc_per_core( address2os, __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); 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: // 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 // __kmp_affinity_compact, then fill out __kmp_affinity_masks. qsort(address2os, __kmp_avail_proc, sizeof(*address2os), __kmp_affinity_cmp_Address_child_num); { int i; unsigned j; for (i = 0, j = 0; i < __kmp_avail_proc; i++) { if ((!__kmp_affinity_dups) && (!address2os[i].first.leader)) { continue; } unsigned osId = address2os[i].second; 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); } 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 assumming 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 // affinity_none, call the real initialization routine, then restore // __kmp_affinity_type to affinity_disabled. int disabled = (__kmp_affinity_type == affinity_disabled); if (!KMP_AFFINITY_CAPABLE()) { KMP_ASSERT(disabled); } if (disabled) { __kmp_affinity_type = affinity_none; } __kmp_aux_affinity_initialize(); if (disabled) { __kmp_affinity_type = affinity_disabled; } } void __kmp_affinity_uninitialize(void) { if (__kmp_affinity_masks != NULL) { KMP_CPU_FREE_ARRAY(__kmp_affinity_masks, __kmp_affinity_num_masks); __kmp_affinity_masks = NULL; } if (__kmp_affin_fullMask != NULL) { KMP_CPU_FREE(__kmp_affin_fullMask); __kmp_affin_fullMask = NULL; } __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 KMPAffinity::destroy_api(); } void __kmp_affinity_set_init_mask(int gtid, int isa_root) { if (!KMP_AFFINITY_CAPABLE()) { return; } kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]); if (th->th.th_affin_mask == NULL) { KMP_CPU_ALLOC(th->th.th_affin_mask); } else { KMP_CPU_ZERO(th->th.th_affin_mask); } // Copy the thread mask to the kmp_info_t strucuture. If // __kmp_affinity_type == affinity_none, copy the "full" mask, i.e. one that // has all of the OS proc ids set, or if __kmp_affinity_respect_mask is set, // then the full mask is the same as the mask of the initialization thread. kmp_affin_mask_t *mask; int i; if (KMP_AFFINITY_NON_PROC_BIND) { if ((__kmp_affinity_type == affinity_none) || (__kmp_affinity_type == affinity_balanced)) { #if KMP_GROUP_AFFINITY if (__kmp_num_proc_groups > 1) { return; } #endif KMP_ASSERT(__kmp_affin_fullMask != NULL); i = 0; mask = __kmp_affin_fullMask; } else { KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0); i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks; mask = KMP_CPU_INDEX(__kmp_affinity_masks, i); } } else { if ((!isa_root) || (__kmp_nested_proc_bind.bind_types[0] == proc_bind_false)) { #if KMP_GROUP_AFFINITY if (__kmp_num_proc_groups > 1) { return; } #endif KMP_ASSERT(__kmp_affin_fullMask != NULL); i = KMP_PLACE_ALL; mask = __kmp_affin_fullMask; } else { // int i = some hash function or just a counter that doesn't // always start at 0. Use gtid for now. KMP_DEBUG_ASSERT(__kmp_affinity_num_masks > 0); i = (gtid + __kmp_affinity_offset) % __kmp_affinity_num_masks; mask = KMP_CPU_INDEX(__kmp_affinity_masks, i); } } th->th.th_current_place = i; if (isa_root) { th->th.th_new_place = i; th->th.th_first_place = 0; th->th.th_last_place = __kmp_affinity_num_masks - 1; } else if (KMP_AFFINITY_NON_PROC_BIND) { // When using a Non-OMP_PROC_BIND affinity method, // set all threads' place-partition-var to the entire place list th->th.th_first_place = 0; th->th.th_last_place = __kmp_affinity_num_masks - 1; } if (i == KMP_PLACE_ALL) { KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to all places\n", gtid)); } else { KA_TRACE(100, ("__kmp_affinity_set_init_mask: binding T#%d to place %d\n", gtid, i)); } KMP_CPU_COPY(th->th.th_affin_mask, mask); if (__kmp_affinity_verbose /* to avoid duplicate printing (will be correctly printed on barrier) */ && (__kmp_affinity_type == affinity_none || (i != KMP_PLACE_ALL && __kmp_affinity_type != affinity_balanced))) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, th->th.th_affin_mask); KMP_INFORM(BoundToOSProcSet, "KMP_AFFINITY", (kmp_int32)getpid(), __kmp_gettid(), gtid, buf); } #if KMP_OS_WINDOWS // On Windows* OS, the process affinity mask might have changed. If the user // didn't request affinity and this call fails, just continue silently. // See CQ171393. if (__kmp_affinity_type == affinity_none) { __kmp_set_system_affinity(th->th.th_affin_mask, FALSE); } else #endif __kmp_set_system_affinity(th->th.th_affin_mask, TRUE); } void __kmp_affinity_set_place(int gtid) { if (!KMP_AFFINITY_CAPABLE()) { return; } kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]); KA_TRACE(100, ("__kmp_affinity_set_place: binding T#%d to place %d (current " "place = %d)\n", gtid, th->th.th_new_place, th->th.th_current_place)); // Check that the new place is within this thread's partition. KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL); KMP_ASSERT(th->th.th_new_place >= 0); KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity_num_masks); if (th->th.th_first_place <= th->th.th_last_place) { KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) && (th->th.th_new_place <= th->th.th_last_place)); } else { KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) || (th->th.th_new_place >= th->th.th_last_place)); } // Copy the thread mask to the kmp_info_t strucuture, // and set this thread's affinity. kmp_affin_mask_t *mask = KMP_CPU_INDEX(__kmp_affinity_masks, th->th.th_new_place); KMP_CPU_COPY(th->th.th_affin_mask, mask); th->th.th_current_place = th->th.th_new_place; if (__kmp_affinity_verbose) { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, th->th.th_affin_mask); KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(), __kmp_gettid(), gtid, buf); } __kmp_set_system_affinity(th->th.th_affin_mask, TRUE); } int __kmp_aux_set_affinity(void **mask) { int gtid; kmp_info_t *th; int retval; if (!KMP_AFFINITY_CAPABLE()) { return -1; } gtid = __kmp_entry_gtid(); KA_TRACE(1000, (""); { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf( "kmp_set_affinity: setting affinity mask for thread %d = %s\n", gtid, buf); }); if (__kmp_env_consistency_check) { if ((mask == NULL) || (*mask == NULL)) { KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity"); } else { unsigned proc; int num_procs = 0; KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) { if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) { KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity"); } if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) { continue; } num_procs++; } if (num_procs == 0) { KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity"); } #if KMP_GROUP_AFFINITY if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) { KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity"); } #endif /* KMP_GROUP_AFFINITY */ } } th = __kmp_threads[gtid]; KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL); retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE); if (retval == 0) { KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask)); } th->th.th_current_place = KMP_PLACE_UNDEFINED; th->th.th_new_place = KMP_PLACE_UNDEFINED; th->th.th_first_place = 0; th->th.th_last_place = __kmp_affinity_num_masks - 1; // Turn off 4.0 affinity for the current tread at this parallel level. th->th.th_current_task->td_icvs.proc_bind = proc_bind_false; return retval; } int __kmp_aux_get_affinity(void **mask) { int gtid; int retval; kmp_info_t *th; if (!KMP_AFFINITY_CAPABLE()) { return -1; } gtid = __kmp_entry_gtid(); th = __kmp_threads[gtid]; KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL); KA_TRACE(1000, (""); { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, th->th.th_affin_mask); __kmp_printf("kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid, buf); }); if (__kmp_env_consistency_check) { if ((mask == NULL) || (*mask == NULL)) { KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity"); } } #if !KMP_OS_WINDOWS retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE); KA_TRACE(1000, (""); { char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, (kmp_affin_mask_t *)(*mask)); __kmp_printf("kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid, buf); }); return retval; #else KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask); return 0; #endif /* KMP_OS_WINDOWS */ } int __kmp_aux_get_affinity_max_proc() { if (!KMP_AFFINITY_CAPABLE()) { return 0; } #if KMP_GROUP_AFFINITY if (__kmp_num_proc_groups > 1) { return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT); } #endif return __kmp_xproc; } int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) { if (!KMP_AFFINITY_CAPABLE()) { return -1; } KA_TRACE(1000, (""); { int gtid = __kmp_entry_gtid(); char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in " "affinity mask for thread %d = %s\n", proc, gtid, buf); }); if (__kmp_env_consistency_check) { if ((mask == NULL) || (*mask == NULL)) { KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc"); } } if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) { return -1; } if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) { return -2; } KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask)); return 0; } int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) { if (!KMP_AFFINITY_CAPABLE()) { return -1; } KA_TRACE(1000, (""); { int gtid = __kmp_entry_gtid(); char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in " "affinity mask for thread %d = %s\n", proc, gtid, buf); }); if (__kmp_env_consistency_check) { if ((mask == NULL) || (*mask == NULL)) { KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc"); } } if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) { return -1; } if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) { return -2; } KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask)); return 0; } int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) { if (!KMP_AFFINITY_CAPABLE()) { return -1; } KA_TRACE(1000, (""); { int gtid = __kmp_entry_gtid(); char buf[KMP_AFFIN_MASK_PRINT_LEN]; __kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, (kmp_affin_mask_t *)(*mask)); __kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in " "affinity mask for thread %d = %s\n", proc, gtid, buf); }); if (__kmp_env_consistency_check) { if ((mask == NULL) || (*mask == NULL)) { KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc"); } } if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) { return -1; } if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) { return 0; } return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask)); } // 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: break; case affinity_gran_core: if (__kmp_nThreadsPerCore > 1) { fine_gran = false; } break; case affinity_gran_package: if (nCoresPerPkg > 1) { fine_gran = false; } break; default: fine_gran = false; } if (__kmp_affinity_uniform_topology()) { 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; 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; 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, __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); // For performance gain consider the special case nthreads == // __kmp_avail_proc if (nthreads == __kmp_avail_proc) { if (fine_gran) { int osID = address2os[tid].second; KMP_CPU_SET(osID, mask); } else { int core = __kmp_affinity_find_core(address2os, 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) { KMP_CPU_SET(osID, mask); } } } } else if (nthreads <= ncores) { int core = 0; for (int i = 0; i < ncores; i++) { // Check if this core from procarr[] is in the mask int in_mask = 0; for (int j = 0; j < nth_per_core; j++) { if (procarr[i * nth_per_core + j] != -1) { in_mask = 1; break; } } if (in_mask) { if (tid == core) { for (int j = 0; j < nth_per_core; j++) { int osID = procarr[i * nth_per_core + j]; if (osID != -1) { KMP_CPU_SET(osID, mask); // For fine granularity it is enough to set the first available // osID for this core if (fine_gran) { break; } } } break; } else { core++; } } } } else { // nthreads > ncores // Array to save the number of processors at each core int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores); // Array to save the number of cores with "x" available processors; int *ncores_with_x_procs = (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1)); // Array to save the number of cores with # procs from x to nth_per_core int *ncores_with_x_to_max_procs = (int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1)); for (int i = 0; i <= nth_per_core; i++) { ncores_with_x_procs[i] = 0; ncores_with_x_to_max_procs[i] = 0; } for (int i = 0; i < ncores; i++) { int cnt = 0; for (int j = 0; j < nth_per_core; j++) { if (procarr[i * nth_per_core + j] != -1) { cnt++; } } nproc_at_core[i] = cnt; ncores_with_x_procs[cnt]++; } for (int i = 0; i <= nth_per_core; i++) { for (int j = i; j <= nth_per_core; j++) { ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j]; } } // Max number of processors int nproc = nth_per_core * ncores; // An array to keep number of threads per each context int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc); for (int i = 0; i < nproc; i++) { newarr[i] = 0; } int nth = nthreads; int flag = 0; while (nth > 0) { for (int j = 1; j <= nth_per_core; j++) { int cnt = ncores_with_x_to_max_procs[j]; for (int i = 0; i < ncores; i++) { // Skip the core with 0 processors if (nproc_at_core[i] == 0) { continue; } for (int k = 0; k < nth_per_core; k++) { if (procarr[i * nth_per_core + k] != -1) { if (newarr[i * nth_per_core + k] == 0) { newarr[i * nth_per_core + k] = 1; cnt--; nth--; break; } else { if (flag != 0) { newarr[i * nth_per_core + k]++; cnt--; nth--; break; } } } } if (cnt == 0 || nth == 0) { break; } } if (nth == 0) { break; } } flag = 1; } int sum = 0; for (int i = 0; i < nproc; i++) { sum += newarr[i]; if (sum > tid) { if (fine_gran) { int osID = procarr[i]; KMP_CPU_SET(osID, mask); } else { int coreID = i / nth_per_core; for (int ii = 0; ii < nth_per_core; ii++) { int osID = procarr[coreID * nth_per_core + ii]; if (osID != -1) { KMP_CPU_SET(osID, mask); } } } break; } } __kmp_free(newarr); } 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); } } #if KMP_OS_LINUX // We don't need this entry for Windows because // there is GetProcessAffinityMask() api // // The intended usage is indicated by these steps: // 1) The user gets the current affinity mask // 2) Then sets the affinity by calling this function // 3) Error check the return value // 4) Use non-OpenMP parallelization // 5) Reset the affinity to what was stored in step 1) #ifdef __cplusplus extern "C" #endif int kmp_set_thread_affinity_mask_initial() // the function returns 0 on success, // -1 if we cannot bind thread // >0 (errno) if an error happened during binding { int gtid = __kmp_get_gtid(); if (gtid < 0) { // Do not touch non-omp threads KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: " "non-omp thread, returning\n")); return -1; } if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) { KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: " "affinity not initialized, returning\n")); return -1; } KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: " "set full mask for thread %d\n", gtid)); KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL); return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE); } #endif #endif // KMP_AFFINITY_SUPPORTED