/* * kmp_runtime.cpp -- KPTS runtime support library */ //===----------------------------------------------------------------------===// // // 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_atomic.h" #include "kmp_environment.h" #include "kmp_error.h" #include "kmp_i18n.h" #include "kmp_io.h" #include "kmp_itt.h" #include "kmp_settings.h" #include "kmp_stats.h" #include "kmp_str.h" #include "kmp_wait_release.h" #include "kmp_wrapper_getpid.h" #include "kmp_dispatch.h" #if KMP_USE_HIER_SCHED #include "kmp_dispatch_hier.h" #endif #if OMPT_SUPPORT #include "ompt-specific.h" #endif #if OMPD_SUPPORT #include "ompd-specific.h" #endif #if OMP_PROFILING_SUPPORT #include "llvm/Support/TimeProfiler.h" static char *ProfileTraceFile = nullptr; #endif /* these are temporary issues to be dealt with */ #define KMP_USE_PRCTL 0 #if KMP_OS_WINDOWS #include <process.h> #endif #if KMP_OS_WINDOWS // windows does not need include files as it doesn't use shared memory #else #include <sys/mman.h> #include <sys/stat.h> #include <fcntl.h> #define SHM_SIZE 1024 #endif #if defined(KMP_GOMP_COMPAT) char const __kmp_version_alt_comp[] = KMP_VERSION_PREFIX "alternative compiler support: yes"; #endif /* defined(KMP_GOMP_COMPAT) */ char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: 5.0 (201611)"; #ifdef KMP_DEBUG char const __kmp_version_lock[] = KMP_VERSION_PREFIX "lock type: run time selectable"; #endif /* KMP_DEBUG */ #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* ------------------------------------------------------------------------ */ #if KMP_USE_MONITOR kmp_info_t __kmp_monitor; #endif /* Forward declarations */ void __kmp_cleanup(void); static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid, int gtid); static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, kmp_internal_control_t *new_icvs, ident_t *loc); #if KMP_AFFINITY_SUPPORTED static void __kmp_partition_places(kmp_team_t *team, int update_master_only = 0); #endif static void __kmp_do_serial_initialize(void); void __kmp_fork_barrier(int gtid, int tid); void __kmp_join_barrier(int gtid); void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc, kmp_internal_control_t *new_icvs, ident_t *loc); #ifdef USE_LOAD_BALANCE static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc); #endif static int __kmp_expand_threads(int nNeed); #if KMP_OS_WINDOWS static int __kmp_unregister_root_other_thread(int gtid); #endif static void __kmp_reap_thread(kmp_info_t *thread, int is_root); kmp_info_t *__kmp_thread_pool_insert_pt = NULL; void __kmp_resize_dist_barrier(kmp_team_t *team, int old_nthreads, int new_nthreads); void __kmp_add_threads_to_team(kmp_team_t *team, int new_nthreads); /* Calculate the identifier of the current thread */ /* fast (and somewhat portable) way to get unique identifier of executing thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */ int __kmp_get_global_thread_id() { int i; kmp_info_t **other_threads; size_t stack_data; char *stack_addr; size_t stack_size; char *stack_base; KA_TRACE( 1000, ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n", __kmp_nth, __kmp_all_nth)); /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to a parallel region, made it return KMP_GTID_DNE to force serial_initialize by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee __kmp_init_gtid for this to work. */ if (!TCR_4(__kmp_init_gtid)) return KMP_GTID_DNE; #ifdef KMP_TDATA_GTID if (TCR_4(__kmp_gtid_mode) >= 3) { KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n")); return __kmp_gtid; } #endif if (TCR_4(__kmp_gtid_mode) >= 2) { KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n")); return __kmp_gtid_get_specific(); } KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n")); stack_addr = (char *)&stack_data; other_threads = __kmp_threads; /* ATT: The code below is a source of potential bugs due to unsynchronized access to __kmp_threads array. For example: 1. Current thread loads other_threads[i] to thr and checks it, it is non-NULL. 2. Current thread is suspended by OS. 3. Another thread unregisters and finishes (debug versions of free() may fill memory with something like 0xEF). 4. Current thread is resumed. 5. Current thread reads junk from *thr. TODO: Fix it. --ln */ for (i = 0; i < __kmp_threads_capacity; i++) { kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]); if (!thr) continue; stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize); stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase); /* stack grows down -- search through all of the active threads */ if (stack_addr <= stack_base) { size_t stack_diff = stack_base - stack_addr; if (stack_diff <= stack_size) { /* The only way we can be closer than the allocated */ /* stack size is if we are running on this thread. */ KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i); return i; } } } /* get specific to try and determine our gtid */ KA_TRACE(1000, ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n")); i = __kmp_gtid_get_specific(); /*fprintf( stderr, "=== %d\n", i ); */ /* GROO */ /* if we havn't been assigned a gtid, then return code */ if (i < 0) return i; /* dynamically updated stack window for uber threads to avoid get_specific call */ if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) { KMP_FATAL(StackOverflow, i); } stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; if (stack_addr > stack_base) { TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr); TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize, other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base); } else { TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize, stack_base - stack_addr); } /* Reprint stack bounds for ubermaster since they have been refined */ if (__kmp_storage_map) { char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize; __kmp_print_storage_map_gtid(i, stack_beg, stack_end, other_threads[i]->th.th_info.ds.ds_stacksize, "th_%d stack (refinement)", i); } return i; } int __kmp_get_global_thread_id_reg() { int gtid; if (!__kmp_init_serial) { gtid = KMP_GTID_DNE; } else #ifdef KMP_TDATA_GTID if (TCR_4(__kmp_gtid_mode) >= 3) { KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n")); gtid = __kmp_gtid; } else #endif if (TCR_4(__kmp_gtid_mode) >= 2) { KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n")); gtid = __kmp_gtid_get_specific(); } else { KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using internal alg.\n")); gtid = __kmp_get_global_thread_id(); } /* we must be a new uber master sibling thread */ if (gtid == KMP_GTID_DNE) { KA_TRACE(10, ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n")); __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (!__kmp_init_serial) { __kmp_do_serial_initialize(); gtid = __kmp_gtid_get_specific(); } else { gtid = __kmp_register_root(FALSE); } __kmp_release_bootstrap_lock(&__kmp_initz_lock); /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */ } KMP_DEBUG_ASSERT(gtid >= 0); return gtid; } /* caller must hold forkjoin_lock */ void __kmp_check_stack_overlap(kmp_info_t *th) { int f; char *stack_beg = NULL; char *stack_end = NULL; int gtid; KA_TRACE(10, ("__kmp_check_stack_overlap: called\n")); if (__kmp_storage_map) { stack_end = (char *)th->th.th_info.ds.ds_stackbase; stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; gtid = __kmp_gtid_from_thread(th); if (gtid == KMP_GTID_MONITOR) { __kmp_print_storage_map_gtid( gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, "th_%s stack (%s)", "mon", (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); } else { __kmp_print_storage_map_gtid( gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, "th_%d stack (%s)", gtid, (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); } } /* No point in checking ubermaster threads since they use refinement and * cannot overlap */ gtid = __kmp_gtid_from_thread(th); if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) { KA_TRACE(10, ("__kmp_check_stack_overlap: performing extensive checking\n")); if (stack_beg == NULL) { stack_end = (char *)th->th.th_info.ds.ds_stackbase; stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; } for (f = 0; f < __kmp_threads_capacity; f++) { kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]); if (f_th && f_th != th) { char *other_stack_end = (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase); char *other_stack_beg = other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize); if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) || (stack_end > other_stack_beg && stack_end < other_stack_end)) { /* Print the other stack values before the abort */ if (__kmp_storage_map) __kmp_print_storage_map_gtid( -1, other_stack_beg, other_stack_end, (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize), "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th)); __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit), __kmp_msg_null); } } } } KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n")); } /* ------------------------------------------------------------------------ */ void __kmp_infinite_loop(void) { static int done = FALSE; while (!done) { KMP_YIELD(TRUE); } } #define MAX_MESSAGE 512 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size, char const *format, ...) { char buffer[MAX_MESSAGE]; va_list ap; va_start(ap, format); KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1, p2, (unsigned long)size, format); __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); __kmp_vprintf(kmp_err, buffer, ap); #if KMP_PRINT_DATA_PLACEMENT int node; if (gtid >= 0) { if (p1 <= p2 && (char *)p2 - (char *)p1 == size) { if (__kmp_storage_map_verbose) { node = __kmp_get_host_node(p1); if (node < 0) /* doesn't work, so don't try this next time */ __kmp_storage_map_verbose = FALSE; else { char *last; int lastNode; int localProc = __kmp_get_cpu_from_gtid(gtid); const int page_size = KMP_GET_PAGE_SIZE(); p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1)); p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1)); if (localProc >= 0) __kmp_printf_no_lock(" GTID %d localNode %d\n", gtid, localProc >> 1); else __kmp_printf_no_lock(" GTID %d\n", gtid); #if KMP_USE_PRCTL /* The more elaborate format is disabled for now because of the prctl * hanging bug. */ do { last = p1; lastNode = node; /* This loop collates adjacent pages with the same host node. */ do { (char *)p1 += page_size; } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode); __kmp_printf_no_lock(" %p-%p memNode %d\n", last, (char *)p1 - 1, lastNode); } while (p1 <= p2); #else __kmp_printf_no_lock(" %p-%p memNode %d\n", p1, (char *)p1 + (page_size - 1), __kmp_get_host_node(p1)); if (p1 < p2) { __kmp_printf_no_lock(" %p-%p memNode %d\n", p2, (char *)p2 + (page_size - 1), __kmp_get_host_node(p2)); } #endif } } } else __kmp_printf_no_lock(" %s\n", KMP_I18N_STR(StorageMapWarning)); } #endif /* KMP_PRINT_DATA_PLACEMENT */ __kmp_release_bootstrap_lock(&__kmp_stdio_lock); va_end(ap); } void __kmp_warn(char const *format, ...) { char buffer[MAX_MESSAGE]; va_list ap; if (__kmp_generate_warnings == kmp_warnings_off) { return; } va_start(ap, format); KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format); __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); __kmp_vprintf(kmp_err, buffer, ap); __kmp_release_bootstrap_lock(&__kmp_stdio_lock); va_end(ap); } void __kmp_abort_process() { // Later threads may stall here, but that's ok because abort() will kill them. __kmp_acquire_bootstrap_lock(&__kmp_exit_lock); if (__kmp_debug_buf) { __kmp_dump_debug_buffer(); } if (KMP_OS_WINDOWS) { // Let other threads know of abnormal termination and prevent deadlock // if abort happened during library initialization or shutdown __kmp_global.g.g_abort = SIGABRT; /* On Windows* OS by default abort() causes pop-up error box, which stalls nightly testing. Unfortunately, we cannot reliably suppress pop-up error boxes. _set_abort_behavior() works well, but this function is not available in VS7 (this is not problem for DLL, but it is a problem for static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not help, at least in some versions of MS C RTL. It seems following sequence is the only way to simulate abort() and avoid pop-up error box. */ raise(SIGABRT); _exit(3); // Just in case, if signal ignored, exit anyway. } else { __kmp_unregister_library(); abort(); } __kmp_infinite_loop(); __kmp_release_bootstrap_lock(&__kmp_exit_lock); } // __kmp_abort_process void __kmp_abort_thread(void) { // TODO: Eliminate g_abort global variable and this function. // In case of abort just call abort(), it will kill all the threads. __kmp_infinite_loop(); } // __kmp_abort_thread /* Print out the storage map for the major kmp_info_t thread data structures that are allocated together. */ static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) { __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d", gtid); __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team, sizeof(kmp_desc_t), "th_%d.th_info", gtid); __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head, sizeof(kmp_local_t), "th_%d.th_local", gtid); __kmp_print_storage_map_gtid( gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier], sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid); __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier], &thr->th.th_bar[bs_plain_barrier + 1], sizeof(kmp_balign_t), "th_%d.th_bar[plain]", gtid); __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier], &thr->th.th_bar[bs_forkjoin_barrier + 1], sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]", gtid); #if KMP_FAST_REDUCTION_BARRIER __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier], &thr->th.th_bar[bs_reduction_barrier + 1], sizeof(kmp_balign_t), "th_%d.th_bar[reduction]", gtid); #endif // KMP_FAST_REDUCTION_BARRIER } /* Print out the storage map for the major kmp_team_t team data structures that are allocated together. */ static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team, int team_id, int num_thr) { int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2; __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d", header, team_id); __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0], &team->t.t_bar[bs_last_barrier], sizeof(kmp_balign_team_t) * bs_last_barrier, "%s_%d.t_bar", header, team_id); __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier], &team->t.t_bar[bs_plain_barrier + 1], sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]", header, team_id); __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier], &team->t.t_bar[bs_forkjoin_barrier + 1], sizeof(kmp_balign_team_t), "%s_%d.t_bar[forkjoin]", header, team_id); #if KMP_FAST_REDUCTION_BARRIER __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier], &team->t.t_bar[bs_reduction_barrier + 1], sizeof(kmp_balign_team_t), "%s_%d.t_bar[reduction]", header, team_id); #endif // KMP_FAST_REDUCTION_BARRIER __kmp_print_storage_map_gtid( -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr], sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id); __kmp_print_storage_map_gtid( -1, &team->t.t_threads[0], &team->t.t_threads[num_thr], sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id); __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0], &team->t.t_disp_buffer[num_disp_buff], sizeof(dispatch_shared_info_t) * num_disp_buff, "%s_%d.t_disp_buffer", header, team_id); } static void __kmp_init_allocator() { __kmp_init_memkind(); __kmp_init_target_mem(); } static void __kmp_fini_allocator() { __kmp_fini_memkind(); } /* ------------------------------------------------------------------------ */ #if ENABLE_LIBOMPTARGET static void __kmp_init_omptarget() { __kmp_init_target_task(); } #endif /* ------------------------------------------------------------------------ */ #if KMP_DYNAMIC_LIB #if KMP_OS_WINDOWS BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) { //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock ); switch (fdwReason) { case DLL_PROCESS_ATTACH: KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n")); return TRUE; case DLL_PROCESS_DETACH: KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific())); // According to Windows* documentation for DllMain entry point: // for DLL_PROCESS_DETACH, lpReserved is used for telling the difference: // lpReserved == NULL when FreeLibrary() is called, // lpReserved != NULL when the process is terminated. // When FreeLibrary() is called, worker threads remain alive. So the // runtime's state is consistent and executing proper shutdown is OK. // When the process is terminated, worker threads have exited or been // forcefully terminated by the OS and only the shutdown thread remains. // This can leave the runtime in an inconsistent state. // Hence, only attempt proper cleanup when FreeLibrary() is called. // Otherwise, rely on OS to reclaim resources. if (lpReserved == NULL) __kmp_internal_end_library(__kmp_gtid_get_specific()); return TRUE; case DLL_THREAD_ATTACH: KA_TRACE(10, ("DllMain: THREAD_ATTACH\n")); /* if we want to register new siblings all the time here call * __kmp_get_gtid(); */ return TRUE; case DLL_THREAD_DETACH: KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific())); __kmp_internal_end_thread(__kmp_gtid_get_specific()); return TRUE; } return TRUE; } #endif /* KMP_OS_WINDOWS */ #endif /* KMP_DYNAMIC_LIB */ /* __kmp_parallel_deo -- Wait until it's our turn. */ void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { int gtid = *gtid_ref; #ifdef BUILD_PARALLEL_ORDERED kmp_team_t *team = __kmp_team_from_gtid(gtid); #endif /* BUILD_PARALLEL_ORDERED */ if (__kmp_env_consistency_check) { if (__kmp_threads[gtid]->th.th_root->r.r_active) #if KMP_USE_DYNAMIC_LOCK __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0); #else __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL); #endif } #ifdef BUILD_PARALLEL_ORDERED if (!team->t.t_serialized) { KMP_MB(); KMP_WAIT(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ, NULL); KMP_MB(); } #endif /* BUILD_PARALLEL_ORDERED */ } /* __kmp_parallel_dxo -- Signal the next task. */ void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { int gtid = *gtid_ref; #ifdef BUILD_PARALLEL_ORDERED int tid = __kmp_tid_from_gtid(gtid); kmp_team_t *team = __kmp_team_from_gtid(gtid); #endif /* BUILD_PARALLEL_ORDERED */ if (__kmp_env_consistency_check) { if (__kmp_threads[gtid]->th.th_root->r.r_active) __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref); } #ifdef BUILD_PARALLEL_ORDERED if (!team->t.t_serialized) { KMP_MB(); /* Flush all pending memory write invalidates. */ /* use the tid of the next thread in this team */ /* TODO replace with general release procedure */ team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc); KMP_MB(); /* Flush all pending memory write invalidates. */ } #endif /* BUILD_PARALLEL_ORDERED */ } /* ------------------------------------------------------------------------ */ /* The BARRIER for a SINGLE process section is always explicit */ int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) { int status; kmp_info_t *th; kmp_team_t *team; if (!TCR_4(__kmp_init_parallel)) __kmp_parallel_initialize(); __kmp_resume_if_soft_paused(); th = __kmp_threads[gtid]; team = th->th.th_team; status = 0; th->th.th_ident = id_ref; if (team->t.t_serialized) { status = 1; } else { kmp_int32 old_this = th->th.th_local.this_construct; ++th->th.th_local.this_construct; /* try to set team count to thread count--success means thread got the single block */ /* TODO: Should this be acquire or release? */ if (team->t.t_construct == old_this) { status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this, th->th.th_local.this_construct); } #if USE_ITT_BUILD if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 && KMP_MASTER_GTID(gtid) && th->th.th_teams_microtask == NULL && team->t.t_active_level == 1) { // Only report metadata by primary thread of active team at level 1 __kmp_itt_metadata_single(id_ref); } #endif /* USE_ITT_BUILD */ } if (__kmp_env_consistency_check) { if (status && push_ws) { __kmp_push_workshare(gtid, ct_psingle, id_ref); } else { __kmp_check_workshare(gtid, ct_psingle, id_ref); } } #if USE_ITT_BUILD if (status) { __kmp_itt_single_start(gtid); } #endif /* USE_ITT_BUILD */ return status; } void __kmp_exit_single(int gtid) { #if USE_ITT_BUILD __kmp_itt_single_end(gtid); #endif /* USE_ITT_BUILD */ if (__kmp_env_consistency_check) __kmp_pop_workshare(gtid, ct_psingle, NULL); } /* determine if we can go parallel or must use a serialized parallel region and * how many threads we can use * set_nproc is the number of threads requested for the team * returns 0 if we should serialize or only use one thread, * otherwise the number of threads to use * The forkjoin lock is held by the caller. */ static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team, int master_tid, int set_nthreads, int enter_teams) { int capacity; int new_nthreads; KMP_DEBUG_ASSERT(__kmp_init_serial); KMP_DEBUG_ASSERT(root && parent_team); kmp_info_t *this_thr = parent_team->t.t_threads[master_tid]; // If dyn-var is set, dynamically adjust the number of desired threads, // according to the method specified by dynamic_mode. new_nthreads = set_nthreads; if (!get__dynamic_2(parent_team, master_tid)) { ; } #ifdef USE_LOAD_BALANCE else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) { new_nthreads = __kmp_load_balance_nproc(root, set_nthreads); if (new_nthreads == 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid)); return 1; } if (new_nthreads < set_nthreads) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads)); } } #endif /* USE_LOAD_BALANCE */ else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) { new_nthreads = __kmp_avail_proc - __kmp_nth + (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); if (new_nthreads <= 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid)); return 1; } if (new_nthreads < set_nthreads) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads)); } else { new_nthreads = set_nthreads; } } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) { if (set_nthreads > 2) { new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]); new_nthreads = (new_nthreads % set_nthreads) + 1; if (new_nthreads == 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid)); return 1; } if (new_nthreads < set_nthreads) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads)); } } } else { KMP_ASSERT(0); } // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT. if (__kmp_nth + new_nthreads - (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > __kmp_max_nth) { int tl_nthreads = __kmp_max_nth - __kmp_nth + (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); if (tl_nthreads <= 0) { tl_nthreads = 1; } // If dyn-var is false, emit a 1-time warning. if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { __kmp_reserve_warn = 1; __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads), KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); } if (tl_nthreads == 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid)); return 1; } KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads)); new_nthreads = tl_nthreads; } // Respect OMP_THREAD_LIMIT int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads; int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit; if (cg_nthreads + new_nthreads - (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > max_cg_threads) { int tl_nthreads = max_cg_threads - cg_nthreads + (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); if (tl_nthreads <= 0) { tl_nthreads = 1; } // If dyn-var is false, emit a 1-time warning. if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { __kmp_reserve_warn = 1; __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads), KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); } if (tl_nthreads == 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid)); return 1; } KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads)); new_nthreads = tl_nthreads; } // Check if the threads array is large enough, or needs expanding. // See comment in __kmp_register_root() about the adjustment if // __kmp_threads[0] == NULL. capacity = __kmp_threads_capacity; if (TCR_PTR(__kmp_threads[0]) == NULL) { --capacity; } // If it is not for initializing the hidden helper team, we need to take // __kmp_hidden_helper_threads_num out of the capacity because it is included // in __kmp_threads_capacity. if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)) { capacity -= __kmp_hidden_helper_threads_num; } if (__kmp_nth + new_nthreads - (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > capacity) { // Expand the threads array. int slotsRequired = __kmp_nth + new_nthreads - (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) - capacity; int slotsAdded = __kmp_expand_threads(slotsRequired); if (slotsAdded < slotsRequired) { // The threads array was not expanded enough. new_nthreads -= (slotsRequired - slotsAdded); KMP_ASSERT(new_nthreads >= 1); // If dyn-var is false, emit a 1-time warning. if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { __kmp_reserve_warn = 1; if (__kmp_tp_cached) { __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads), KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity), KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null); } else { __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads), KMP_HNT(SystemLimitOnThreads), __kmp_msg_null); } } } } #ifdef KMP_DEBUG if (new_nthreads == 1) { KC_TRACE(10, ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_gtid(), set_nthreads)); } else { KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_gtid(), new_nthreads, set_nthreads)); } #endif // KMP_DEBUG return new_nthreads; } /* Allocate threads from the thread pool and assign them to the new team. We are assured that there are enough threads available, because we checked on that earlier within critical section forkjoin */ static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team, kmp_info_t *master_th, int master_gtid, int fork_teams_workers) { int i; int use_hot_team; KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc)); KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid()); KMP_MB(); /* first, let's setup the primary thread */ master_th->th.th_info.ds.ds_tid = 0; master_th->th.th_team = team; master_th->th.th_team_nproc = team->t.t_nproc; master_th->th.th_team_master = master_th; master_th->th.th_team_serialized = FALSE; master_th->th.th_dispatch = &team->t.t_dispatch[0]; /* make sure we are not the optimized hot team */ #if KMP_NESTED_HOT_TEAMS use_hot_team = 0; kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams; if (hot_teams) { // hot teams array is not allocated if // KMP_HOT_TEAMS_MAX_LEVEL=0 int level = team->t.t_active_level - 1; // index in array of hot teams if (master_th->th.th_teams_microtask) { // are we inside the teams? if (master_th->th.th_teams_size.nteams > 1) { ++level; // level was not increased in teams construct for // team_of_masters } if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && master_th->th.th_teams_level == team->t.t_level) { ++level; // level was not increased in teams construct for // team_of_workers before the parallel } // team->t.t_level will be increased inside parallel } if (level < __kmp_hot_teams_max_level) { if (hot_teams[level].hot_team) { // hot team has already been allocated for given level KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team); use_hot_team = 1; // the team is ready to use } else { use_hot_team = 0; // AC: threads are not allocated yet hot_teams[level].hot_team = team; // remember new hot team hot_teams[level].hot_team_nth = team->t.t_nproc; } } else { use_hot_team = 0; } } #else use_hot_team = team == root->r.r_hot_team; #endif if (!use_hot_team) { /* install the primary thread */ team->t.t_threads[0] = master_th; __kmp_initialize_info(master_th, team, 0, master_gtid); /* now, install the worker threads */ for (i = 1; i < team->t.t_nproc; i++) { /* fork or reallocate a new thread and install it in team */ kmp_info_t *thr = __kmp_allocate_thread(root, team, i); team->t.t_threads[i] = thr; KMP_DEBUG_ASSERT(thr); KMP_DEBUG_ASSERT(thr->th.th_team == team); /* align team and thread arrived states */ KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", __kmp_gtid_from_tid(0, team), team->t.t_id, 0, __kmp_gtid_from_tid(i, team), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived)); thr->th.th_teams_microtask = master_th->th.th_teams_microtask; thr->th.th_teams_level = master_th->th.th_teams_level; thr->th.th_teams_size = master_th->th.th_teams_size; { // Initialize threads' barrier data. int b; kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #if USE_DEBUGGER balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; #endif } } } #if KMP_AFFINITY_SUPPORTED // Do not partition the places list for teams construct workers who // haven't actually been forked to do real work yet. This partitioning // will take place in the parallel region nested within the teams construct. if (!fork_teams_workers) { __kmp_partition_places(team); } #endif if (team->t.t_nproc > 1 && __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { team->t.b->update_num_threads(team->t.t_nproc); __kmp_add_threads_to_team(team, team->t.t_nproc); } } if (__kmp_display_affinity && team->t.t_display_affinity != 1) { for (i = 0; i < team->t.t_nproc; i++) { kmp_info_t *thr = team->t.t_threads[i]; if (thr->th.th_prev_num_threads != team->t.t_nproc || thr->th.th_prev_level != team->t.t_level) { team->t.t_display_affinity = 1; break; } } } KMP_MB(); } #if KMP_ARCH_X86 || KMP_ARCH_X86_64 // Propagate any changes to the floating point control registers out to the team // We try to avoid unnecessary writes to the relevant cache line in the team // structure, so we don't make changes unless they are needed. inline static void propagateFPControl(kmp_team_t *team) { if (__kmp_inherit_fp_control) { kmp_int16 x87_fpu_control_word; kmp_uint32 mxcsr; // Get primary thread's values of FPU control flags (both X87 and vector) __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); __kmp_store_mxcsr(&mxcsr); mxcsr &= KMP_X86_MXCSR_MASK; // There is no point looking at t_fp_control_saved here. // If it is TRUE, we still have to update the values if they are different // from those we now have. If it is FALSE we didn't save anything yet, but // our objective is the same. We have to ensure that the values in the team // are the same as those we have. // So, this code achieves what we need whether or not t_fp_control_saved is // true. By checking whether the value needs updating we avoid unnecessary // writes that would put the cache-line into a written state, causing all // threads in the team to have to read it again. KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word); KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr); // Although we don't use this value, other code in the runtime wants to know // whether it should restore them. So we must ensure it is correct. KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE); } else { // Similarly here. Don't write to this cache-line in the team structure // unless we have to. KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE); } } // Do the opposite, setting the hardware registers to the updated values from // the team. inline static void updateHWFPControl(kmp_team_t *team) { if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) { // Only reset the fp control regs if they have been changed in the team. // the parallel region that we are exiting. kmp_int16 x87_fpu_control_word; kmp_uint32 mxcsr; __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); __kmp_store_mxcsr(&mxcsr); mxcsr &= KMP_X86_MXCSR_MASK; if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) { __kmp_clear_x87_fpu_status_word(); __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word); } if (team->t.t_mxcsr != mxcsr) { __kmp_load_mxcsr(&team->t.t_mxcsr); } } } #else #define propagateFPControl(x) ((void)0) #define updateHWFPControl(x) ((void)0) #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc); // forward declaration /* Run a parallel region that has been serialized, so runs only in a team of the single primary thread. */ void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) { kmp_info_t *this_thr; kmp_team_t *serial_team; KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid)); /* Skip all this code for autopar serialized loops since it results in unacceptable overhead */ if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR)) return; if (!TCR_4(__kmp_init_parallel)) __kmp_parallel_initialize(); __kmp_resume_if_soft_paused(); this_thr = __kmp_threads[global_tid]; serial_team = this_thr->th.th_serial_team; /* utilize the serialized team held by this thread */ KMP_DEBUG_ASSERT(serial_team); KMP_MB(); if (__kmp_tasking_mode != tskm_immediate_exec) { KMP_DEBUG_ASSERT( this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]); KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] == NULL); KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr->th.th_task_team, this_thr->th.th_team)); this_thr->th.th_task_team = NULL; } kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind; if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { proc_bind = proc_bind_false; } else if (proc_bind == proc_bind_default) { // No proc_bind clause was specified, so use the current value // of proc-bind-var for this parallel region. proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind; } // Reset for next parallel region this_thr->th.th_set_proc_bind = proc_bind_default; // Reset num_threads for next parallel region this_thr->th.th_set_nproc = 0; #if OMPT_SUPPORT ompt_data_t ompt_parallel_data = ompt_data_none; void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid); if (ompt_enabled.enabled && this_thr->th.ompt_thread_info.state != ompt_state_overhead) { ompt_task_info_t *parent_task_info; parent_task_info = OMPT_CUR_TASK_INFO(this_thr); parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); if (ompt_enabled.ompt_callback_parallel_begin) { int team_size = 1; ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)( &(parent_task_info->task_data), &(parent_task_info->frame), &ompt_parallel_data, team_size, ompt_parallel_invoker_program | ompt_parallel_team, codeptr); } } #endif // OMPT_SUPPORT if (this_thr->th.th_team != serial_team) { // Nested level will be an index in the nested nthreads array int level = this_thr->th.th_team->t.t_level; if (serial_team->t.t_serialized) { /* this serial team was already used TODO increase performance by making this locks more specific */ kmp_team_t *new_team; __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1, #if OMPT_SUPPORT ompt_parallel_data, #endif proc_bind, &this_thr->th.th_current_task->td_icvs, 0 USE_NESTED_HOT_ARG(NULL)); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); KMP_ASSERT(new_team); /* setup new serialized team and install it */ new_team->t.t_threads[0] = this_thr; new_team->t.t_parent = this_thr->th.th_team; serial_team = new_team; this_thr->th.th_serial_team = serial_team; KF_TRACE( 10, ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n", global_tid, serial_team)); /* TODO the above breaks the requirement that if we run out of resources, then we can still guarantee that serialized teams are ok, since we may need to allocate a new one */ } else { KF_TRACE( 10, ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n", global_tid, serial_team)); } /* we have to initialize this serial team */ KMP_DEBUG_ASSERT(serial_team->t.t_threads); KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr); KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team); serial_team->t.t_ident = loc; serial_team->t.t_serialized = 1; serial_team->t.t_nproc = 1; serial_team->t.t_parent = this_thr->th.th_team; serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched; this_thr->th.th_team = serial_team; serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid; KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d curtask=%p\n", global_tid, this_thr->th.th_current_task)); KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1); this_thr->th.th_current_task->td_flags.executing = 0; __kmp_push_current_task_to_thread(this_thr, serial_team, 0); /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an implicit task for each serialized task represented by team->t.t_serialized? */ copy_icvs(&this_thr->th.th_current_task->td_icvs, &this_thr->th.th_current_task->td_parent->td_icvs); // Thread value exists in the nested nthreads array for the next nested // level if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { this_thr->th.th_current_task->td_icvs.nproc = __kmp_nested_nth.nth[level + 1]; } if (__kmp_nested_proc_bind.used && (level + 1 < __kmp_nested_proc_bind.used)) { this_thr->th.th_current_task->td_icvs.proc_bind = __kmp_nested_proc_bind.bind_types[level + 1]; } #if USE_DEBUGGER serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger. #endif this_thr->th.th_info.ds.ds_tid = 0; /* set thread cache values */ this_thr->th.th_team_nproc = 1; this_thr->th.th_team_master = this_thr; this_thr->th.th_team_serialized = 1; serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1; serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level; serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save propagateFPControl(serial_team); /* check if we need to allocate dispatch buffers stack */ KMP_DEBUG_ASSERT(serial_team->t.t_dispatch); if (!serial_team->t.t_dispatch->th_disp_buffer) { serial_team->t.t_dispatch->th_disp_buffer = (dispatch_private_info_t *)__kmp_allocate( sizeof(dispatch_private_info_t)); } this_thr->th.th_dispatch = serial_team->t.t_dispatch; KMP_MB(); } else { /* this serialized team is already being used, * that's fine, just add another nested level */ KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team); KMP_DEBUG_ASSERT(serial_team->t.t_threads); KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr); ++serial_team->t.t_serialized; this_thr->th.th_team_serialized = serial_team->t.t_serialized; // Nested level will be an index in the nested nthreads array int level = this_thr->th.th_team->t.t_level; // Thread value exists in the nested nthreads array for the next nested // level if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { this_thr->th.th_current_task->td_icvs.nproc = __kmp_nested_nth.nth[level + 1]; } serial_team->t.t_level++; KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team->t.t_level)); /* allocate/push dispatch buffers stack */ KMP_DEBUG_ASSERT(serial_team->t.t_dispatch); { dispatch_private_info_t *disp_buffer = (dispatch_private_info_t *)__kmp_allocate( sizeof(dispatch_private_info_t)); disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer; serial_team->t.t_dispatch->th_disp_buffer = disp_buffer; } this_thr->th.th_dispatch = serial_team->t.t_dispatch; KMP_MB(); } KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq); // Perform the display affinity functionality for // serialized parallel regions if (__kmp_display_affinity) { if (this_thr->th.th_prev_level != serial_team->t.t_level || this_thr->th.th_prev_num_threads != 1) { // NULL means use the affinity-format-var ICV __kmp_aux_display_affinity(global_tid, NULL); this_thr->th.th_prev_level = serial_team->t.t_level; this_thr->th.th_prev_num_threads = 1; } } if (__kmp_env_consistency_check) __kmp_push_parallel(global_tid, NULL); #if OMPT_SUPPORT serial_team->t.ompt_team_info.master_return_address = codeptr; if (ompt_enabled.enabled && this_thr->th.ompt_thread_info.state != ompt_state_overhead) { OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); ompt_lw_taskteam_t lw_taskteam; __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid, &ompt_parallel_data, codeptr); __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1); // don't use lw_taskteam after linking. content was swaped /* OMPT implicit task begin */ if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr), OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid), ompt_task_implicit); // TODO: Can this be ompt_task_initial? OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(global_tid); } /* OMPT state */ this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); } #endif } // Test if this fork is for a team closely nested in a teams construct static inline bool __kmp_is_fork_in_teams(kmp_info_t *master_th, microtask_t microtask, int level, int teams_level, kmp_va_list ap) { return (master_th->th.th_teams_microtask && ap && microtask != (microtask_t)__kmp_teams_master && level == teams_level); } // Test if this fork is for the teams construct, i.e. to form the outer league // of teams static inline bool __kmp_is_entering_teams(int active_level, int level, int teams_level, kmp_va_list ap) { return ((ap == NULL && active_level == 0) || (ap && teams_level > 0 && teams_level == level)); } // AC: This is start of parallel that is nested inside teams construct. // The team is actual (hot), all workers are ready at the fork barrier. // No lock needed to initialize the team a bit, then free workers. static inline int __kmp_fork_in_teams(ident_t *loc, int gtid, kmp_team_t *parent_team, kmp_int32 argc, kmp_info_t *master_th, kmp_root_t *root, enum fork_context_e call_context, microtask_t microtask, launch_t invoker, int master_set_numthreads, int level, #if OMPT_SUPPORT ompt_data_t ompt_parallel_data, void *return_address, #endif kmp_va_list ap) { void **argv; int i; parent_team->t.t_ident = loc; __kmp_alloc_argv_entries(argc, parent_team, TRUE); parent_team->t.t_argc = argc; argv = (void **)parent_team->t.t_argv; for (i = argc - 1; i >= 0; --i) { *argv++ = va_arg(kmp_va_deref(ap), void *); } // Increment our nested depth levels, but not increase the serialization if (parent_team == master_th->th.th_serial_team) { // AC: we are in serialized parallel __kmpc_serialized_parallel(loc, gtid); KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1); if (call_context == fork_context_gnu) { // AC: need to decrement t_serialized for enquiry functions to work // correctly, will restore at join time parent_team->t.t_serialized--; return TRUE; } #if OMPD_SUPPORT parent_team->t.t_pkfn = microtask; #endif #if OMPT_SUPPORT void *dummy; void **exit_frame_p; ompt_data_t *implicit_task_data; ompt_lw_taskteam_t lw_taskteam; if (ompt_enabled.enabled) { __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, &ompt_parallel_data, return_address); exit_frame_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr); __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); // Don't use lw_taskteam after linking. Content was swapped. /* OMPT implicit task begin */ implicit_task_data = OMPT_CUR_TASK_DATA(master_th); if (ompt_enabled.ompt_callback_implicit_task) { OMPT_CUR_TASK_INFO(master_th)->thread_num = __kmp_tid_from_gtid(gtid); ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th), implicit_task_data, 1, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } /* OMPT state */ master_th->th.ompt_thread_info.state = ompt_state_work_parallel; } else { exit_frame_p = &dummy; } #endif // AC: need to decrement t_serialized for enquiry functions to work // correctly, will restore at join time parent_team->t.t_serialized--; { KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv #if OMPT_SUPPORT , exit_frame_p #endif ); } #if OMPT_SUPPORT if (ompt_enabled.enabled) { *exit_frame_p = NULL; OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none; if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, implicit_task_data, 1, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th); __ompt_lw_taskteam_unlink(master_th); if (ompt_enabled.ompt_callback_parallel_end) { ompt_callbacks.ompt_callback(ompt_callback_parallel_end)( &ompt_parallel_data, OMPT_CUR_TASK_DATA(master_th), OMPT_INVOKER(call_context) | ompt_parallel_team, return_address); } master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif return TRUE; } parent_team->t.t_pkfn = microtask; parent_team->t.t_invoke = invoker; KMP_ATOMIC_INC(&root->r.r_in_parallel); parent_team->t.t_active_level++; parent_team->t.t_level++; parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save // If the threads allocated to the team are less than the thread limit, update // the thread limit here. th_teams_size.nth is specific to this team nested // in a teams construct, the team is fully created, and we're about to do // the actual fork. Best to do this here so that the subsequent uses below // and in the join have the correct value. master_th->th.th_teams_size.nth = parent_team->t.t_nproc; #if OMPT_SUPPORT if (ompt_enabled.enabled) { ompt_lw_taskteam_t lw_taskteam; __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, &ompt_parallel_data, return_address); __ompt_lw_taskteam_link(&lw_taskteam, master_th, 1, true); } #endif /* Change number of threads in the team if requested */ if (master_set_numthreads) { // The parallel has num_threads clause if (master_set_numthreads <= master_th->th.th_teams_size.nth) { // AC: only can reduce number of threads dynamically, can't increase kmp_info_t **other_threads = parent_team->t.t_threads; // NOTE: if using distributed barrier, we need to run this code block // even when the team size appears not to have changed from the max. int old_proc = master_th->th.th_teams_size.nth; if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { __kmp_resize_dist_barrier(parent_team, old_proc, master_set_numthreads); __kmp_add_threads_to_team(parent_team, master_set_numthreads); } parent_team->t.t_nproc = master_set_numthreads; for (i = 0; i < master_set_numthreads; ++i) { other_threads[i]->th.th_team_nproc = master_set_numthreads; } } // Keep extra threads hot in the team for possible next parallels master_th->th.th_set_nproc = 0; } #if USE_DEBUGGER if (__kmp_debugging) { // Let debugger override number of threads. int nth = __kmp_omp_num_threads(loc); if (nth > 0) { // 0 means debugger doesn't want to change num threads master_set_numthreads = nth; } } #endif // Figure out the proc_bind policy for the nested parallel within teams kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; // proc_bind_default means don't update kmp_proc_bind_t proc_bind_icv = proc_bind_default; if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { proc_bind = proc_bind_false; } else { // No proc_bind clause specified; use current proc-bind-var if (proc_bind == proc_bind_default) { proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; } /* else: The proc_bind policy was specified explicitly on parallel clause. This overrides proc-bind-var for this parallel region, but does not change proc-bind-var. */ // Figure the value of proc-bind-var for the child threads. if ((level + 1 < __kmp_nested_proc_bind.used) && (__kmp_nested_proc_bind.bind_types[level + 1] != master_th->th.th_current_task->td_icvs.proc_bind)) { proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; } } KMP_CHECK_UPDATE(parent_team->t.t_proc_bind, proc_bind); // Need to change the bind-var ICV to correct value for each implicit task if (proc_bind_icv != proc_bind_default && master_th->th.th_current_task->td_icvs.proc_bind != proc_bind_icv) { kmp_info_t **other_threads = parent_team->t.t_threads; for (i = 0; i < master_th->th.th_team_nproc; ++i) { other_threads[i]->th.th_current_task->td_icvs.proc_bind = proc_bind_icv; } } // Reset for next parallel region master_th->th.th_set_proc_bind = proc_bind_default; #if USE_ITT_BUILD && USE_ITT_NOTIFY if (((__itt_frame_submit_v3_ptr && __itt_get_timestamp_ptr) || KMP_ITT_DEBUG) && __kmp_forkjoin_frames_mode == 3 && parent_team->t.t_active_level == 1 // only report frames at level 1 && master_th->th.th_teams_size.nteams == 1) { kmp_uint64 tmp_time = __itt_get_timestamp(); master_th->th.th_frame_time = tmp_time; parent_team->t.t_region_time = tmp_time; } if (__itt_stack_caller_create_ptr) { KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL); // create new stack stitching id before entering fork barrier parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); } #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ #if KMP_AFFINITY_SUPPORTED __kmp_partition_places(parent_team); #endif KF_TRACE(10, ("__kmp_fork_in_teams: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid)); __kmp_internal_fork(loc, gtid, parent_team); KF_TRACE(10, ("__kmp_fork_in_teams: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid)); if (call_context == fork_context_gnu) return TRUE; /* Invoke microtask for PRIMARY thread */ KA_TRACE(20, ("__kmp_fork_in_teams: T#%d(%d:0) invoke microtask = %p\n", gtid, parent_team->t.t_id, parent_team->t.t_pkfn)); if (!parent_team->t.t_invoke(gtid)) { KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread"); } KA_TRACE(20, ("__kmp_fork_in_teams: T#%d(%d:0) done microtask = %p\n", gtid, parent_team->t.t_id, parent_team->t.t_pkfn)); KMP_MB(); /* Flush all pending memory write invalidates. */ KA_TRACE(20, ("__kmp_fork_in_teams: parallel exit T#%d\n", gtid)); return TRUE; } // Create a serialized parallel region static inline int __kmp_serial_fork_call(ident_t *loc, int gtid, enum fork_context_e call_context, kmp_int32 argc, microtask_t microtask, launch_t invoker, kmp_info_t *master_th, kmp_team_t *parent_team, #if OMPT_SUPPORT ompt_data_t *ompt_parallel_data, void **return_address, ompt_data_t **parent_task_data, #endif kmp_va_list ap) { kmp_team_t *team; int i; void **argv; /* josh todo: hypothetical question: what do we do for OS X*? */ #if KMP_OS_LINUX && \ (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) void *args[argc]; #else void **args = (void **)KMP_ALLOCA(argc * sizeof(void *)); #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \ KMP_ARCH_AARCH64) */ KA_TRACE( 20, ("__kmp_serial_fork_call: T#%d serializing parallel region\n", gtid)); __kmpc_serialized_parallel(loc, gtid); #if OMPD_SUPPORT master_th->th.th_serial_team->t.t_pkfn = microtask; #endif if (call_context == fork_context_intel) { /* TODO this sucks, use the compiler itself to pass args! :) */ master_th->th.th_serial_team->t.t_ident = loc; if (!ap) { // revert change made in __kmpc_serialized_parallel() master_th->th.th_serial_team->t.t_level--; // Get args from parent team for teams construct #if OMPT_SUPPORT void *dummy; void **exit_frame_p; ompt_task_info_t *task_info; ompt_lw_taskteam_t lw_taskteam; if (ompt_enabled.enabled) { __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, ompt_parallel_data, *return_address); __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); // don't use lw_taskteam after linking. content was swaped task_info = OMPT_CUR_TASK_INFO(master_th); exit_frame_p = &(task_info->frame.exit_frame.ptr); if (ompt_enabled.ompt_callback_implicit_task) { OMPT_CUR_TASK_INFO(master_th)->thread_num = __kmp_tid_from_gtid(gtid); ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th), &(task_info->task_data), 1, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } /* OMPT state */ master_th->th.ompt_thread_info.state = ompt_state_work_parallel; } else { exit_frame_p = &dummy; } #endif { KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv #if OMPT_SUPPORT , exit_frame_p #endif ); } #if OMPT_SUPPORT if (ompt_enabled.enabled) { *exit_frame_p = NULL; if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, &(task_info->task_data), 1, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } *ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th); __ompt_lw_taskteam_unlink(master_th); if (ompt_enabled.ompt_callback_parallel_end) { ompt_callbacks.ompt_callback(ompt_callback_parallel_end)( ompt_parallel_data, *parent_task_data, OMPT_INVOKER(call_context) | ompt_parallel_team, *return_address); } master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif } else if (microtask == (microtask_t)__kmp_teams_master) { KMP_DEBUG_ASSERT(master_th->th.th_team == master_th->th.th_serial_team); team = master_th->th.th_team; // team->t.t_pkfn = microtask; team->t.t_invoke = invoker; __kmp_alloc_argv_entries(argc, team, TRUE); team->t.t_argc = argc; argv = (void **)team->t.t_argv; if (ap) { for (i = argc - 1; i >= 0; --i) *argv++ = va_arg(kmp_va_deref(ap), void *); } else { for (i = 0; i < argc; ++i) // Get args from parent team for teams construct argv[i] = parent_team->t.t_argv[i]; } // AC: revert change made in __kmpc_serialized_parallel() // because initial code in teams should have level=0 team->t.t_level--; // AC: call special invoker for outer "parallel" of teams construct invoker(gtid); #if OMPT_SUPPORT if (ompt_enabled.enabled) { ompt_task_info_t *task_info = OMPT_CUR_TASK_INFO(master_th); if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, &(task_info->task_data), 0, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_initial); } if (ompt_enabled.ompt_callback_parallel_end) { ompt_callbacks.ompt_callback(ompt_callback_parallel_end)( ompt_parallel_data, *parent_task_data, OMPT_INVOKER(call_context) | ompt_parallel_league, *return_address); } master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif } else { argv = args; for (i = argc - 1; i >= 0; --i) *argv++ = va_arg(kmp_va_deref(ap), void *); KMP_MB(); #if OMPT_SUPPORT void *dummy; void **exit_frame_p; ompt_task_info_t *task_info; ompt_lw_taskteam_t lw_taskteam; ompt_data_t *implicit_task_data; if (ompt_enabled.enabled) { __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, ompt_parallel_data, *return_address); __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); // don't use lw_taskteam after linking. content was swaped task_info = OMPT_CUR_TASK_INFO(master_th); exit_frame_p = &(task_info->frame.exit_frame.ptr); /* OMPT implicit task begin */ implicit_task_data = OMPT_CUR_TASK_DATA(master_th); if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th), implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); OMPT_CUR_TASK_INFO(master_th)->thread_num = __kmp_tid_from_gtid(gtid); } /* OMPT state */ master_th->th.ompt_thread_info.state = ompt_state_work_parallel; } else { exit_frame_p = &dummy; } #endif { KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); __kmp_invoke_microtask(microtask, gtid, 0, argc, args #if OMPT_SUPPORT , exit_frame_p #endif ); } #if OMPT_SUPPORT if (ompt_enabled.enabled) { *exit_frame_p = NULL; if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, &(task_info->task_data), 1, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } *ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th); __ompt_lw_taskteam_unlink(master_th); if (ompt_enabled.ompt_callback_parallel_end) { ompt_callbacks.ompt_callback(ompt_callback_parallel_end)( ompt_parallel_data, *parent_task_data, OMPT_INVOKER(call_context) | ompt_parallel_team, *return_address); } master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif } } else if (call_context == fork_context_gnu) { #if OMPT_SUPPORT if (ompt_enabled.enabled) { ompt_lw_taskteam_t lwt; __ompt_lw_taskteam_init(&lwt, master_th, gtid, ompt_parallel_data, *return_address); lwt.ompt_task_info.frame.exit_frame = ompt_data_none; __ompt_lw_taskteam_link(&lwt, master_th, 1); } // don't use lw_taskteam after linking. content was swaped #endif // we were called from GNU native code KA_TRACE(20, ("__kmp_serial_fork_call: T#%d serial exit\n", gtid)); return FALSE; } else { KMP_ASSERT2(call_context < fork_context_last, "__kmp_serial_fork_call: unknown fork_context parameter"); } KA_TRACE(20, ("__kmp_serial_fork_call: T#%d serial exit\n", gtid)); KMP_MB(); return FALSE; } /* most of the work for a fork */ /* return true if we really went parallel, false if serialized */ int __kmp_fork_call(ident_t *loc, int gtid, enum fork_context_e call_context, // Intel, GNU, ... kmp_int32 argc, microtask_t microtask, launch_t invoker, kmp_va_list ap) { void **argv; int i; int master_tid; int master_this_cons; kmp_team_t *team; kmp_team_t *parent_team; kmp_info_t *master_th; kmp_root_t *root; int nthreads; int master_active; int master_set_numthreads; int level; int active_level; int teams_level; #if KMP_NESTED_HOT_TEAMS kmp_hot_team_ptr_t **p_hot_teams; #endif { // KMP_TIME_BLOCK KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call); KMP_COUNT_VALUE(OMP_PARALLEL_args, argc); KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid)); if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) { /* Some systems prefer the stack for the root thread(s) to start with */ /* some gap from the parent stack to prevent false sharing. */ void *dummy = KMP_ALLOCA(__kmp_stkpadding); /* These 2 lines below are so this does not get optimized out */ if (__kmp_stkpadding > KMP_MAX_STKPADDING) __kmp_stkpadding += (short)((kmp_int64)dummy); } /* initialize if needed */ KMP_DEBUG_ASSERT( __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown if (!TCR_4(__kmp_init_parallel)) __kmp_parallel_initialize(); __kmp_resume_if_soft_paused(); /* setup current data */ // AC: potentially unsafe, not in sync with library shutdown, // __kmp_threads can be freed master_th = __kmp_threads[gtid]; parent_team = master_th->th.th_team; master_tid = master_th->th.th_info.ds.ds_tid; master_this_cons = master_th->th.th_local.this_construct; root = master_th->th.th_root; master_active = root->r.r_active; master_set_numthreads = master_th->th.th_set_nproc; #if OMPT_SUPPORT ompt_data_t ompt_parallel_data = ompt_data_none; ompt_data_t *parent_task_data; ompt_frame_t *ompt_frame; void *return_address = NULL; if (ompt_enabled.enabled) { __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame, NULL, NULL); return_address = OMPT_LOAD_RETURN_ADDRESS(gtid); } #endif // Assign affinity to root thread if it hasn't happened yet __kmp_assign_root_init_mask(); // Nested level will be an index in the nested nthreads array level = parent_team->t.t_level; // used to launch non-serial teams even if nested is not allowed active_level = parent_team->t.t_active_level; // needed to check nesting inside the teams teams_level = master_th->th.th_teams_level; #if KMP_NESTED_HOT_TEAMS p_hot_teams = &master_th->th.th_hot_teams; if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) { *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate( sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level); (*p_hot_teams)[0].hot_team = root->r.r_hot_team; // it is either actual or not needed (when active_level > 0) (*p_hot_teams)[0].hot_team_nth = 1; } #endif #if OMPT_SUPPORT if (ompt_enabled.enabled) { if (ompt_enabled.ompt_callback_parallel_begin) { int team_size = master_set_numthreads ? master_set_numthreads : get__nproc_2(parent_team, master_tid); int flags = OMPT_INVOKER(call_context) | ((microtask == (microtask_t)__kmp_teams_master) ? ompt_parallel_league : ompt_parallel_team); ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)( parent_task_data, ompt_frame, &ompt_parallel_data, team_size, flags, return_address); } master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif master_th->th.th_ident = loc; // Parallel closely nested in teams construct: if (__kmp_is_fork_in_teams(master_th, microtask, level, teams_level, ap)) { return __kmp_fork_in_teams(loc, gtid, parent_team, argc, master_th, root, call_context, microtask, invoker, master_set_numthreads, level, #if OMPT_SUPPORT ompt_parallel_data, return_address, #endif ap); } // End parallel closely nested in teams construct #if KMP_DEBUG if (__kmp_tasking_mode != tskm_immediate_exec) { KMP_DEBUG_ASSERT(master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]); } #endif // Need this to happen before we determine the number of threads, not while // we are allocating the team //__kmp_push_current_task_to_thread(master_th, parent_team, 0); // Determine the number of threads int enter_teams = __kmp_is_entering_teams(active_level, level, teams_level, ap); if ((!enter_teams && (parent_team->t.t_active_level >= master_th->th.th_current_task->td_icvs.max_active_levels)) || (__kmp_library == library_serial)) { KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team\n", gtid)); nthreads = 1; } else { nthreads = master_set_numthreads ? master_set_numthreads // TODO: get nproc directly from current task : get__nproc_2(parent_team, master_tid); // Check if we need to take forkjoin lock? (no need for serialized // parallel out of teams construct). if (nthreads > 1) { /* determine how many new threads we can use */ __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); /* AC: If we execute teams from parallel region (on host), then teams should be created but each can only have 1 thread if nesting is disabled. If teams called from serial region, then teams and their threads should be created regardless of the nesting setting. */ nthreads = __kmp_reserve_threads(root, parent_team, master_tid, nthreads, enter_teams); if (nthreads == 1) { // Free lock for single thread execution here; for multi-thread // execution it will be freed later after team of threads created // and initialized __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); } } } KMP_DEBUG_ASSERT(nthreads > 0); // If we temporarily changed the set number of threads then restore it now master_th->th.th_set_nproc = 0; if (nthreads == 1) { return __kmp_serial_fork_call(loc, gtid, call_context, argc, microtask, invoker, master_th, parent_team, #if OMPT_SUPPORT &ompt_parallel_data, &return_address, &parent_task_data, #endif ap); } // if (nthreads == 1) // GEH: only modify the executing flag in the case when not serialized // serialized case is handled in kmpc_serialized_parallel KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level, master_th, master_th->th.th_current_task, master_th->th.th_current_task->td_icvs.max_active_levels)); // TODO: GEH - cannot do this assertion because root thread not set up as // executing // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 ); master_th->th.th_current_task->td_flags.executing = 0; if (!master_th->th.th_teams_microtask || level > teams_level) { /* Increment our nested depth level */ KMP_ATOMIC_INC(&root->r.r_in_parallel); } // See if we need to make a copy of the ICVs. int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc; if ((level + 1 < __kmp_nested_nth.used) && (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) { nthreads_icv = __kmp_nested_nth.nth[level + 1]; } else { nthreads_icv = 0; // don't update } // Figure out the proc_bind_policy for the new team. kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; // proc_bind_default means don't update kmp_proc_bind_t proc_bind_icv = proc_bind_default; if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { proc_bind = proc_bind_false; } else { // No proc_bind clause specified; use current proc-bind-var for this // parallel region if (proc_bind == proc_bind_default) { proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; } // Have teams construct take proc_bind value from KMP_TEAMS_PROC_BIND if (master_th->th.th_teams_microtask && microtask == (microtask_t)__kmp_teams_master) { proc_bind = __kmp_teams_proc_bind; } /* else: The proc_bind policy was specified explicitly on parallel clause. This overrides proc-bind-var for this parallel region, but does not change proc-bind-var. */ // Figure the value of proc-bind-var for the child threads. if ((level + 1 < __kmp_nested_proc_bind.used) && (__kmp_nested_proc_bind.bind_types[level + 1] != master_th->th.th_current_task->td_icvs.proc_bind)) { // Do not modify the proc bind icv for the two teams construct forks // They just let the proc bind icv pass through if (!master_th->th.th_teams_microtask || !(microtask == (microtask_t)__kmp_teams_master || ap == NULL)) proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; } } // Reset for next parallel region master_th->th.th_set_proc_bind = proc_bind_default; if ((nthreads_icv > 0) || (proc_bind_icv != proc_bind_default)) { kmp_internal_control_t new_icvs; copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs); new_icvs.next = NULL; if (nthreads_icv > 0) { new_icvs.nproc = nthreads_icv; } if (proc_bind_icv != proc_bind_default) { new_icvs.proc_bind = proc_bind_icv; } /* allocate a new parallel team */ KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n")); team = __kmp_allocate_team(root, nthreads, nthreads, #if OMPT_SUPPORT ompt_parallel_data, #endif proc_bind, &new_icvs, argc USE_NESTED_HOT_ARG(master_th)); if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, &new_icvs); } else { /* allocate a new parallel team */ KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n")); team = __kmp_allocate_team(root, nthreads, nthreads, #if OMPT_SUPPORT ompt_parallel_data, #endif proc_bind, &master_th->th.th_current_task->td_icvs, argc USE_NESTED_HOT_ARG(master_th)); if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) copy_icvs((kmp_internal_control_t *)team->t.b->team_icvs, &master_th->th.th_current_task->td_icvs); } KF_TRACE( 10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team)); /* setup the new team */ KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid); KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons); KMP_CHECK_UPDATE(team->t.t_ident, loc); KMP_CHECK_UPDATE(team->t.t_parent, parent_team); KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask); #if OMPT_SUPPORT KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address, return_address); #endif KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe // TODO: parent_team->t.t_level == INT_MAX ??? if (!master_th->th.th_teams_microtask || level > teams_level) { int new_level = parent_team->t.t_level + 1; KMP_CHECK_UPDATE(team->t.t_level, new_level); new_level = parent_team->t.t_active_level + 1; KMP_CHECK_UPDATE(team->t.t_active_level, new_level); } else { // AC: Do not increase parallel level at start of the teams construct int new_level = parent_team->t.t_level; KMP_CHECK_UPDATE(team->t.t_level, new_level); new_level = parent_team->t.t_active_level; KMP_CHECK_UPDATE(team->t.t_active_level, new_level); } kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid); // set primary thread's schedule as new run-time schedule KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched); KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq); KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator); // Update the floating point rounding in the team if required. propagateFPControl(team); #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_parallel_begin(); #endif if (__kmp_tasking_mode != tskm_immediate_exec) { // Set primary thread's task team to team's task team. Unless this is hot // team, it should be NULL. KMP_DEBUG_ASSERT(master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]); KA_TRACE(20, ("__kmp_fork_call: Primary T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", __kmp_gtid_from_thread(master_th), master_th->th.th_task_team, parent_team, team->t.t_task_team[master_th->th.th_task_state], team)); if (active_level || master_th->th.th_task_team) { // Take a memo of primary thread's task_state KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack); if (master_th->th.th_task_state_top >= master_th->th.th_task_state_stack_sz) { // increase size kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz; kmp_uint8 *old_stack, *new_stack; kmp_uint32 i; new_stack = (kmp_uint8 *)__kmp_allocate(new_size); for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) { new_stack[i] = master_th->th.th_task_state_memo_stack[i]; } for (i = master_th->th.th_task_state_stack_sz; i < new_size; ++i) { // zero-init rest of stack new_stack[i] = 0; } old_stack = master_th->th.th_task_state_memo_stack; master_th->th.th_task_state_memo_stack = new_stack; master_th->th.th_task_state_stack_sz = new_size; __kmp_free(old_stack); } // Store primary thread's task_state on stack master_th->th .th_task_state_memo_stack[master_th->th.th_task_state_top] = master_th->th.th_task_state; master_th->th.th_task_state_top++; #if KMP_NESTED_HOT_TEAMS if (master_th->th.th_hot_teams && active_level < __kmp_hot_teams_max_level && team == master_th->th.th_hot_teams[active_level].hot_team) { // Restore primary thread's nested state if nested hot team master_th->th.th_task_state = master_th->th .th_task_state_memo_stack[master_th->th.th_task_state_top]; } else { #endif master_th->th.th_task_state = 0; #if KMP_NESTED_HOT_TEAMS } #endif } #if !KMP_NESTED_HOT_TEAMS KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) || (team == root->r.r_hot_team)); #endif } KA_TRACE( 20, ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n", gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id, team->t.t_nproc)); KMP_DEBUG_ASSERT(team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))); KMP_MB(); /* now, setup the arguments */ argv = (void **)team->t.t_argv; if (ap) { for (i = argc - 1; i >= 0; --i) { void *new_argv = va_arg(kmp_va_deref(ap), void *); KMP_CHECK_UPDATE(*argv, new_argv); argv++; } } else { for (i = 0; i < argc; ++i) { // Get args from parent team for teams construct KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]); } } /* now actually fork the threads */ KMP_CHECK_UPDATE(team->t.t_master_active, master_active); if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong root->r.r_active = TRUE; __kmp_fork_team_threads(root, team, master_th, gtid, !ap); __kmp_setup_icv_copy(team, nthreads, &master_th->th.th_current_task->td_icvs, loc); #if OMPT_SUPPORT master_th->th.ompt_thread_info.state = ompt_state_work_parallel; #endif __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); #if USE_ITT_BUILD if (team->t.t_active_level == 1 // only report frames at level 1 && !master_th->th.th_teams_microtask) { // not in teams construct #if USE_ITT_NOTIFY if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 1)) { kmp_uint64 tmp_time = 0; if (__itt_get_timestamp_ptr) tmp_time = __itt_get_timestamp(); // Internal fork - report frame begin master_th->th.th_frame_time = tmp_time; if (__kmp_forkjoin_frames_mode == 3) team->t.t_region_time = tmp_time; } else // only one notification scheme (either "submit" or "forking/joined", not both) #endif /* USE_ITT_NOTIFY */ if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) && __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) { // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer. __kmp_itt_region_forking(gtid, team->t.t_nproc, 0); } } #endif /* USE_ITT_BUILD */ /* now go on and do the work */ KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team); KMP_MB(); KF_TRACE(10, ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n", root, team, master_th, gtid)); #if USE_ITT_BUILD if (__itt_stack_caller_create_ptr) { // create new stack stitching id before entering fork barrier if (!enter_teams) { KMP_DEBUG_ASSERT(team->t.t_stack_id == NULL); team->t.t_stack_id = __kmp_itt_stack_caller_create(); } else if (parent_team->t.t_serialized) { // keep stack stitching id in the serialized parent_team; // current team will be used for parallel inside the teams; // if parent_team is active, then it already keeps stack stitching id // for the league of teams KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL); parent_team->t.t_stack_id = __kmp_itt_stack_caller_create(); } } #endif /* USE_ITT_BUILD */ // AC: skip __kmp_internal_fork at teams construct, let only primary // threads execute if (ap) { __kmp_internal_fork(loc, gtid, team); KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid)); } if (call_context == fork_context_gnu) { KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid)); return TRUE; } /* Invoke microtask for PRIMARY thread */ KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid, team->t.t_id, team->t.t_pkfn)); } // END of timer KMP_fork_call block #if KMP_STATS_ENABLED // If beginning a teams construct, then change thread state stats_state_e previous_state = KMP_GET_THREAD_STATE(); if (!ap) { KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION); } #endif if (!team->t.t_invoke(gtid)) { KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread"); } #if KMP_STATS_ENABLED // If was beginning of a teams construct, then reset thread state if (!ap) { KMP_SET_THREAD_STATE(previous_state); } #endif KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid, team->t.t_id, team->t.t_pkfn)); KMP_MB(); /* Flush all pending memory write invalidates. */ KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid)); #if OMPT_SUPPORT if (ompt_enabled.enabled) { master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif return TRUE; } #if OMPT_SUPPORT static inline void __kmp_join_restore_state(kmp_info_t *thread, kmp_team_t *team) { // restore state outside the region thread->th.ompt_thread_info.state = ((team->t.t_serialized) ? ompt_state_work_serial : ompt_state_work_parallel); } static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread, kmp_team_t *team, ompt_data_t *parallel_data, int flags, void *codeptr) { ompt_task_info_t *task_info = __ompt_get_task_info_object(0); if (ompt_enabled.ompt_callback_parallel_end) { ompt_callbacks.ompt_callback(ompt_callback_parallel_end)( parallel_data, &(task_info->task_data), flags, codeptr); } task_info->frame.enter_frame = ompt_data_none; __kmp_join_restore_state(thread, team); } #endif void __kmp_join_call(ident_t *loc, int gtid #if OMPT_SUPPORT , enum fork_context_e fork_context #endif , int exit_teams) { KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call); kmp_team_t *team; kmp_team_t *parent_team; kmp_info_t *master_th; kmp_root_t *root; int master_active; KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid)); /* setup current data */ master_th = __kmp_threads[gtid]; root = master_th->th.th_root; team = master_th->th.th_team; parent_team = team->t.t_parent; master_th->th.th_ident = loc; #if OMPT_SUPPORT void *team_microtask = (void *)team->t.t_pkfn; // For GOMP interface with serialized parallel, need the // __kmpc_end_serialized_parallel to call hooks for OMPT end-implicit-task // and end-parallel events. if (ompt_enabled.enabled && !(team->t.t_serialized && fork_context == fork_context_gnu)) { master_th->th.ompt_thread_info.state = ompt_state_overhead; } #endif #if KMP_DEBUG if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) { KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", __kmp_gtid_from_thread(master_th), team, team->t.t_task_team[master_th->th.th_task_state], master_th->th.th_task_team)); KMP_DEBUG_ASSERT(master_th->th.th_task_team == team->t.t_task_team[master_th->th.th_task_state]); } #endif if (team->t.t_serialized) { if (master_th->th.th_teams_microtask) { // We are in teams construct int level = team->t.t_level; int tlevel = master_th->th.th_teams_level; if (level == tlevel) { // AC: we haven't incremented it earlier at start of teams construct, // so do it here - at the end of teams construct team->t.t_level++; } else if (level == tlevel + 1) { // AC: we are exiting parallel inside teams, need to increment // serialization in order to restore it in the next call to // __kmpc_end_serialized_parallel team->t.t_serialized++; } } __kmpc_end_serialized_parallel(loc, gtid); #if OMPT_SUPPORT if (ompt_enabled.enabled) { if (fork_context == fork_context_gnu) { __ompt_lw_taskteam_unlink(master_th); } __kmp_join_restore_state(master_th, parent_team); } #endif return; } master_active = team->t.t_master_active; if (!exit_teams) { // AC: No barrier for internal teams at exit from teams construct. // But there is barrier for external team (league). __kmp_internal_join(loc, gtid, team); #if USE_ITT_BUILD if (__itt_stack_caller_create_ptr) { KMP_DEBUG_ASSERT(team->t.t_stack_id != NULL); // destroy the stack stitching id after join barrier __kmp_itt_stack_caller_destroy((__itt_caller)team->t.t_stack_id); team->t.t_stack_id = NULL; } #endif } else { master_th->th.th_task_state = 0; // AC: no tasking in teams (out of any parallel) #if USE_ITT_BUILD if (__itt_stack_caller_create_ptr && parent_team->t.t_serialized) { KMP_DEBUG_ASSERT(parent_team->t.t_stack_id != NULL); // destroy the stack stitching id on exit from the teams construct // if parent_team is active, then the id will be destroyed later on // by master of the league of teams __kmp_itt_stack_caller_destroy((__itt_caller)parent_team->t.t_stack_id); parent_team->t.t_stack_id = NULL; } #endif } KMP_MB(); #if OMPT_SUPPORT ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data); void *codeptr = team->t.ompt_team_info.master_return_address; #endif #if USE_ITT_BUILD // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer. if (team->t.t_active_level == 1 && (!master_th->th.th_teams_microtask || /* not in teams construct */ master_th->th.th_teams_size.nteams == 1)) { master_th->th.th_ident = loc; // only one notification scheme (either "submit" or "forking/joined", not // both) if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && __kmp_forkjoin_frames_mode == 3) __kmp_itt_frame_submit(gtid, team->t.t_region_time, master_th->th.th_frame_time, 0, loc, master_th->th.th_team_nproc, 1); else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) && !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames) __kmp_itt_region_joined(gtid); } // active_level == 1 #endif /* USE_ITT_BUILD */ #if KMP_AFFINITY_SUPPORTED if (!exit_teams) { // Restore master thread's partition. master_th->th.th_first_place = team->t.t_first_place; master_th->th.th_last_place = team->t.t_last_place; } #endif // KMP_AFFINITY_SUPPORTED if (master_th->th.th_teams_microtask && !exit_teams && team->t.t_pkfn != (microtask_t)__kmp_teams_master && team->t.t_level == master_th->th.th_teams_level + 1) { // AC: We need to leave the team structure intact at the end of parallel // inside the teams construct, so that at the next parallel same (hot) team // works, only adjust nesting levels #if OMPT_SUPPORT ompt_data_t ompt_parallel_data = ompt_data_none; if (ompt_enabled.enabled) { ompt_task_info_t *task_info = __ompt_get_task_info_object(0); if (ompt_enabled.ompt_callback_implicit_task) { int ompt_team_size = team->t.t_nproc; ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size, OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); } task_info->frame.exit_frame = ompt_data_none; task_info->task_data = ompt_data_none; ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th); __ompt_lw_taskteam_unlink(master_th); } #endif /* Decrement our nested depth level */ team->t.t_level--; team->t.t_active_level--; KMP_ATOMIC_DEC(&root->r.r_in_parallel); // Restore number of threads in the team if needed. This code relies on // the proper adjustment of th_teams_size.nth after the fork in // __kmp_teams_master on each teams primary thread in the case that // __kmp_reserve_threads reduced it. if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) { int old_num = master_th->th.th_team_nproc; int new_num = master_th->th.th_teams_size.nth; kmp_info_t **other_threads = team->t.t_threads; team->t.t_nproc = new_num; for (int i = 0; i < old_num; ++i) { other_threads[i]->th.th_team_nproc = new_num; } // Adjust states of non-used threads of the team for (int i = old_num; i < new_num; ++i) { // Re-initialize thread's barrier data. KMP_DEBUG_ASSERT(other_threads[i]); kmp_balign_t *balign = other_threads[i]->th.th_bar; for (int b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #if USE_DEBUGGER balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; #endif } if (__kmp_tasking_mode != tskm_immediate_exec) { // Synchronize thread's task state other_threads[i]->th.th_task_state = master_th->th.th_task_state; } } } #if OMPT_SUPPORT if (ompt_enabled.enabled) { __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data, OMPT_INVOKER(fork_context) | ompt_parallel_team, codeptr); } #endif return; } /* do cleanup and restore the parent team */ master_th->th.th_info.ds.ds_tid = team->t.t_master_tid; master_th->th.th_local.this_construct = team->t.t_master_this_cons; master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid]; /* jc: The following lock has instructions with REL and ACQ semantics, separating the parallel user code called in this parallel region from the serial user code called after this function returns. */ __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); if (!master_th->th.th_teams_microtask || team->t.t_level > master_th->th.th_teams_level) { /* Decrement our nested depth level */ KMP_ATOMIC_DEC(&root->r.r_in_parallel); } KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0); #if OMPT_SUPPORT if (ompt_enabled.enabled) { ompt_task_info_t *task_info = __ompt_get_task_info_object(0); if (ompt_enabled.ompt_callback_implicit_task) { int flags = (team_microtask == (void *)__kmp_teams_master) ? ompt_task_initial : ompt_task_implicit; int ompt_team_size = (flags == ompt_task_initial) ? 0 : team->t.t_nproc; ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size, OMPT_CUR_TASK_INFO(master_th)->thread_num, flags); } task_info->frame.exit_frame = ompt_data_none; task_info->task_data = ompt_data_none; } #endif KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0, master_th, team)); __kmp_pop_current_task_from_thread(master_th); master_th->th.th_def_allocator = team->t.t_def_allocator; #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_parallel_end(); #endif updateHWFPControl(team); if (root->r.r_active != master_active) root->r.r_active = master_active; __kmp_free_team(root, team USE_NESTED_HOT_ARG( master_th)); // this will free worker threads /* this race was fun to find. make sure the following is in the critical region otherwise assertions may fail occasionally since the old team may be reallocated and the hierarchy appears inconsistent. it is actually safe to run and won't cause any bugs, but will cause those assertion failures. it's only one deref&assign so might as well put this in the critical region */ master_th->th.th_team = parent_team; master_th->th.th_team_nproc = parent_team->t.t_nproc; master_th->th.th_team_master = parent_team->t.t_threads[0]; master_th->th.th_team_serialized = parent_team->t.t_serialized; /* restore serialized team, if need be */ if (parent_team->t.t_serialized && parent_team != master_th->th.th_serial_team && parent_team != root->r.r_root_team) { __kmp_free_team(root, master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL)); master_th->th.th_serial_team = parent_team; } if (__kmp_tasking_mode != tskm_immediate_exec) { if (master_th->th.th_task_state_top > 0) { // Restore task state from memo stack KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack); // Remember primary thread's state if we re-use this nested hot team master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] = master_th->th.th_task_state; --master_th->th.th_task_state_top; // pop // Now restore state at this level master_th->th.th_task_state = master_th->th .th_task_state_memo_stack[master_th->th.th_task_state_top]; } else if (team != root->r.r_hot_team) { // Reset the task state of primary thread if we are not hot team because // in this case all the worker threads will be free, and their task state // will be reset. If not reset the primary's, the task state will be // inconsistent. master_th->th.th_task_state = 0; } // Copy the task team from the parent team to the primary thread master_th->th.th_task_team = parent_team->t.t_task_team[master_th->th.th_task_state]; KA_TRACE(20, ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n", __kmp_gtid_from_thread(master_th), master_th->th.th_task_team, parent_team)); } // TODO: GEH - cannot do this assertion because root thread not set up as // executing // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 ); master_th->th.th_current_task->td_flags.executing = 1; __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); #if KMP_AFFINITY_SUPPORTED if (master_th->th.th_team->t.t_level == 0 && __kmp_affinity.flags.reset) { __kmp_reset_root_init_mask(gtid); } #endif #if OMPT_SUPPORT int flags = OMPT_INVOKER(fork_context) | ((team_microtask == (void *)__kmp_teams_master) ? ompt_parallel_league : ompt_parallel_team); if (ompt_enabled.enabled) { __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, flags, codeptr); } #endif KMP_MB(); KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid)); } /* Check whether we should push an internal control record onto the serial team stack. If so, do it. */ void __kmp_save_internal_controls(kmp_info_t *thread) { if (thread->th.th_team != thread->th.th_serial_team) { return; } if (thread->th.th_team->t.t_serialized > 1) { int push = 0; if (thread->th.th_team->t.t_control_stack_top == NULL) { push = 1; } else { if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level != thread->th.th_team->t.t_serialized) { push = 1; } } if (push) { /* push a record on the serial team's stack */ kmp_internal_control_t *control = (kmp_internal_control_t *)__kmp_allocate( sizeof(kmp_internal_control_t)); copy_icvs(control, &thread->th.th_current_task->td_icvs); control->serial_nesting_level = thread->th.th_team->t.t_serialized; control->next = thread->th.th_team->t.t_control_stack_top; thread->th.th_team->t.t_control_stack_top = control; } } } /* Changes set_nproc */ void __kmp_set_num_threads(int new_nth, int gtid) { kmp_info_t *thread; kmp_root_t *root; KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth)); KMP_DEBUG_ASSERT(__kmp_init_serial); if (new_nth < 1) new_nth = 1; else if (new_nth > __kmp_max_nth) new_nth = __kmp_max_nth; KMP_COUNT_VALUE(OMP_set_numthreads, new_nth); thread = __kmp_threads[gtid]; if (thread->th.th_current_task->td_icvs.nproc == new_nth) return; // nothing to do __kmp_save_internal_controls(thread); set__nproc(thread, new_nth); // If this omp_set_num_threads() call will cause the hot team size to be // reduced (in the absence of a num_threads clause), then reduce it now, // rather than waiting for the next parallel region. root = thread->th.th_root; if (__kmp_init_parallel && (!root->r.r_active) && (root->r.r_hot_team->t.t_nproc > new_nth) #if KMP_NESTED_HOT_TEAMS && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode #endif ) { kmp_team_t *hot_team = root->r.r_hot_team; int f; __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { __kmp_resize_dist_barrier(hot_team, hot_team->t.t_nproc, new_nth); } // Release the extra threads we don't need any more. for (f = new_nth; f < hot_team->t.t_nproc; f++) { KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL); if (__kmp_tasking_mode != tskm_immediate_exec) { // When decreasing team size, threads no longer in the team should unref // task team. hot_team->t.t_threads[f]->th.th_task_team = NULL; } __kmp_free_thread(hot_team->t.t_threads[f]); hot_team->t.t_threads[f] = NULL; } hot_team->t.t_nproc = new_nth; #if KMP_NESTED_HOT_TEAMS if (thread->th.th_hot_teams) { KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team); thread->th.th_hot_teams[0].hot_team_nth = new_nth; } #endif if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { hot_team->t.b->update_num_threads(new_nth); __kmp_add_threads_to_team(hot_team, new_nth); } __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); // Update the t_nproc field in the threads that are still active. for (f = 0; f < new_nth; f++) { KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL); hot_team->t.t_threads[f]->th.th_team_nproc = new_nth; } // Special flag in case omp_set_num_threads() call hot_team->t.t_size_changed = -1; } } /* Changes max_active_levels */ void __kmp_set_max_active_levels(int gtid, int max_active_levels) { kmp_info_t *thread; KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels)); KMP_DEBUG_ASSERT(__kmp_init_serial); // validate max_active_levels if (max_active_levels < 0) { KMP_WARNING(ActiveLevelsNegative, max_active_levels); // We ignore this call if the user has specified a negative value. // The current setting won't be changed. The last valid setting will be // used. A warning will be issued (if warnings are allowed as controlled by // the KMP_WARNINGS env var). KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels)); return; } if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) { // it's OK, the max_active_levels is within the valid range: [ 0; // KMP_MAX_ACTIVE_LEVELS_LIMIT ] // We allow a zero value. (implementation defined behavior) } else { KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels, KMP_MAX_ACTIVE_LEVELS_LIMIT); max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT; // Current upper limit is MAX_INT. (implementation defined behavior) // If the input exceeds the upper limit, we correct the input to be the // upper limit. (implementation defined behavior) // Actually, the flow should never get here until we use MAX_INT limit. } KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels)); thread = __kmp_threads[gtid]; __kmp_save_internal_controls(thread); set__max_active_levels(thread, max_active_levels); } /* Gets max_active_levels */ int __kmp_get_max_active_levels(int gtid) { kmp_info_t *thread; KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid)); KMP_DEBUG_ASSERT(__kmp_init_serial); thread = __kmp_threads[gtid]; KMP_DEBUG_ASSERT(thread->th.th_current_task); KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task, thread->th.th_current_task->td_icvs.max_active_levels)); return thread->th.th_current_task->td_icvs.max_active_levels; } // nteams-var per-device ICV void __kmp_set_num_teams(int num_teams) { if (num_teams > 0) __kmp_nteams = num_teams; } int __kmp_get_max_teams(void) { return __kmp_nteams; } // teams-thread-limit-var per-device ICV void __kmp_set_teams_thread_limit(int limit) { if (limit > 0) __kmp_teams_thread_limit = limit; } int __kmp_get_teams_thread_limit(void) { return __kmp_teams_thread_limit; } KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int)); KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int)); /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */ void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) { kmp_info_t *thread; kmp_sched_t orig_kind; // kmp_team_t *team; KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n", gtid, (int)kind, chunk)); KMP_DEBUG_ASSERT(__kmp_init_serial); // Check if the kind parameter is valid, correct if needed. // Valid parameters should fit in one of two intervals - standard or extended: // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper> // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103 orig_kind = kind; kind = __kmp_sched_without_mods(kind); if (kind <= kmp_sched_lower || kind >= kmp_sched_upper || (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) { // TODO: Hint needs attention in case we change the default schedule. __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind), KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"), __kmp_msg_null); kind = kmp_sched_default; chunk = 0; // ignore chunk value in case of bad kind } thread = __kmp_threads[gtid]; __kmp_save_internal_controls(thread); if (kind < kmp_sched_upper_std) { if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) { // differ static chunked vs. unchunked: chunk should be invalid to // indicate unchunked schedule (which is the default) thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static; } else { thread->th.th_current_task->td_icvs.sched.r_sched_type = __kmp_sch_map[kind - kmp_sched_lower - 1]; } } else { // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std - // kmp_sched_lower - 2 ]; thread->th.th_current_task->td_icvs.sched.r_sched_type = __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std - kmp_sched_lower - 2]; } __kmp_sched_apply_mods_intkind( orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type)); if (kind == kmp_sched_auto || chunk < 1) { // ignore parameter chunk for schedule auto thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK; } else { thread->th.th_current_task->td_icvs.sched.chunk = chunk; } } /* Gets def_sched_var ICV values */ void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) { kmp_info_t *thread; enum sched_type th_type; KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid)); KMP_DEBUG_ASSERT(__kmp_init_serial); thread = __kmp_threads[gtid]; th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type; switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)) { case kmp_sch_static: case kmp_sch_static_greedy: case kmp_sch_static_balanced: *kind = kmp_sched_static; __kmp_sched_apply_mods_stdkind(kind, th_type); *chunk = 0; // chunk was not set, try to show this fact via zero value return; case kmp_sch_static_chunked: *kind = kmp_sched_static; break; case kmp_sch_dynamic_chunked: *kind = kmp_sched_dynamic; break; case kmp_sch_guided_chunked: case kmp_sch_guided_iterative_chunked: case kmp_sch_guided_analytical_chunked: *kind = kmp_sched_guided; break; case kmp_sch_auto: *kind = kmp_sched_auto; break; case kmp_sch_trapezoidal: *kind = kmp_sched_trapezoidal; break; #if KMP_STATIC_STEAL_ENABLED case kmp_sch_static_steal: *kind = kmp_sched_static_steal; break; #endif default: KMP_FATAL(UnknownSchedulingType, th_type); } __kmp_sched_apply_mods_stdkind(kind, th_type); *chunk = thread->th.th_current_task->td_icvs.sched.chunk; } int __kmp_get_ancestor_thread_num(int gtid, int level) { int ii, dd; kmp_team_t *team; kmp_info_t *thr; KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level)); KMP_DEBUG_ASSERT(__kmp_init_serial); // validate level if (level == 0) return 0; if (level < 0) return -1; thr = __kmp_threads[gtid]; team = thr->th.th_team; ii = team->t.t_level; if (level > ii) return -1; if (thr->th.th_teams_microtask) { // AC: we are in teams region where multiple nested teams have same level int tlevel = thr->th.th_teams_level; // the level of the teams construct if (level <= tlevel) { // otherwise usual algorithm works (will not touch the teams) KMP_DEBUG_ASSERT(ii >= tlevel); // AC: As we need to pass by the teams league, we need to artificially // increase ii if (ii == tlevel) { ii += 2; // three teams have same level } else { ii++; // two teams have same level } } } if (ii == level) return __kmp_tid_from_gtid(gtid); dd = team->t.t_serialized; level++; while (ii > level) { for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { } if ((team->t.t_serialized) && (!dd)) { team = team->t.t_parent; continue; } if (ii > level) { team = team->t.t_parent; dd = team->t.t_serialized; ii--; } } return (dd > 1) ? (0) : (team->t.t_master_tid); } int __kmp_get_team_size(int gtid, int level) { int ii, dd; kmp_team_t *team; kmp_info_t *thr; KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level)); KMP_DEBUG_ASSERT(__kmp_init_serial); // validate level if (level == 0) return 1; if (level < 0) return -1; thr = __kmp_threads[gtid]; team = thr->th.th_team; ii = team->t.t_level; if (level > ii) return -1; if (thr->th.th_teams_microtask) { // AC: we are in teams region where multiple nested teams have same level int tlevel = thr->th.th_teams_level; // the level of the teams construct if (level <= tlevel) { // otherwise usual algorithm works (will not touch the teams) KMP_DEBUG_ASSERT(ii >= tlevel); // AC: As we need to pass by the teams league, we need to artificially // increase ii if (ii == tlevel) { ii += 2; // three teams have same level } else { ii++; // two teams have same level } } } while (ii > level) { for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { } if (team->t.t_serialized && (!dd)) { team = team->t.t_parent; continue; } if (ii > level) { team = team->t.t_parent; ii--; } } return team->t.t_nproc; } kmp_r_sched_t __kmp_get_schedule_global() { // This routine created because pairs (__kmp_sched, __kmp_chunk) and // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults // independently. So one can get the updated schedule here. kmp_r_sched_t r_sched; // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static, // __kmp_guided. __kmp_sched should keep original value, so that user can set // KMP_SCHEDULE multiple times, and thus have different run-time schedules in // different roots (even in OMP 2.5) enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched); enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched); if (s == kmp_sch_static) { // replace STATIC with more detailed schedule (balanced or greedy) r_sched.r_sched_type = __kmp_static; } else if (s == kmp_sch_guided_chunked) { // replace GUIDED with more detailed schedule (iterative or analytical) r_sched.r_sched_type = __kmp_guided; } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other r_sched.r_sched_type = __kmp_sched; } SCHEDULE_SET_MODIFIERS(r_sched.r_sched_type, sched_modifiers); if (__kmp_chunk < KMP_DEFAULT_CHUNK) { // __kmp_chunk may be wrong here (if it was not ever set) r_sched.chunk = KMP_DEFAULT_CHUNK; } else { r_sched.chunk = __kmp_chunk; } return r_sched; } /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE) at least argc number of *t_argv entries for the requested team. */ static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) { KMP_DEBUG_ASSERT(team); if (!realloc || argc > team->t.t_max_argc) { KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0)); /* if previously allocated heap space for args, free them */ if (realloc && team->t.t_argv != &team->t.t_inline_argv[0]) __kmp_free((void *)team->t.t_argv); if (argc <= KMP_INLINE_ARGV_ENTRIES) { /* use unused space in the cache line for arguments */ team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES; KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc)); team->t.t_argv = &team->t.t_inline_argv[0]; if (__kmp_storage_map) { __kmp_print_storage_map_gtid( -1, &team->t.t_inline_argv[0], &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES], (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv", team->t.t_id); } } else { /* allocate space for arguments in the heap */ team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1)) ? KMP_MIN_MALLOC_ARGV_ENTRIES : 2 * argc; KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc)); team->t.t_argv = (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc); if (__kmp_storage_map) { __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0], &team->t.t_argv[team->t.t_max_argc], sizeof(void *) * team->t.t_max_argc, "team_%d.t_argv", team->t.t_id); } } } } static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) { int i; int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2; team->t.t_threads = (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth); team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate( sizeof(dispatch_shared_info_t) * num_disp_buff); team->t.t_dispatch = (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth); team->t.t_implicit_task_taskdata = (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth); team->t.t_max_nproc = max_nth; /* setup dispatch buffers */ for (i = 0; i < num_disp_buff; ++i) { team->t.t_disp_buffer[i].buffer_index = i; team->t.t_disp_buffer[i].doacross_buf_idx = i; } } static void __kmp_free_team_arrays(kmp_team_t *team) { /* Note: this does not free the threads in t_threads (__kmp_free_threads) */ int i; for (i = 0; i < team->t.t_max_nproc; ++i) { if (team->t.t_dispatch[i].th_disp_buffer != NULL) { __kmp_free(team->t.t_dispatch[i].th_disp_buffer); team->t.t_dispatch[i].th_disp_buffer = NULL; } } #if KMP_USE_HIER_SCHED __kmp_dispatch_free_hierarchies(team); #endif __kmp_free(team->t.t_threads); __kmp_free(team->t.t_disp_buffer); __kmp_free(team->t.t_dispatch); __kmp_free(team->t.t_implicit_task_taskdata); team->t.t_threads = NULL; team->t.t_disp_buffer = NULL; team->t.t_dispatch = NULL; team->t.t_implicit_task_taskdata = 0; } static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) { kmp_info_t **oldThreads = team->t.t_threads; __kmp_free(team->t.t_disp_buffer); __kmp_free(team->t.t_dispatch); __kmp_free(team->t.t_implicit_task_taskdata); __kmp_allocate_team_arrays(team, max_nth); KMP_MEMCPY(team->t.t_threads, oldThreads, team->t.t_nproc * sizeof(kmp_info_t *)); __kmp_free(oldThreads); } static kmp_internal_control_t __kmp_get_global_icvs(void) { kmp_r_sched_t r_sched = __kmp_get_schedule_global(); // get current state of scheduling globals KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0); kmp_internal_control_t g_icvs = { 0, // int serial_nesting_level; //corresponds to value of th_team_serialized (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic // adjustment of threads (per thread) (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for // whether blocktime is explicitly set __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime #if KMP_USE_MONITOR __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime // intervals #endif __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for // next parallel region (per thread) // (use a max ub on value if __kmp_parallel_initialize not called yet) __kmp_cg_max_nth, // int thread_limit; __kmp_dflt_max_active_levels, // int max_active_levels; //internal control // for max_active_levels r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule // {sched,chunk} pair __kmp_nested_proc_bind.bind_types[0], __kmp_default_device, NULL // struct kmp_internal_control *next; }; return g_icvs; } static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) { kmp_internal_control_t gx_icvs; gx_icvs.serial_nesting_level = 0; // probably =team->t.t_serial like in save_inter_controls copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs); gx_icvs.next = NULL; return gx_icvs; } static void __kmp_initialize_root(kmp_root_t *root) { int f; kmp_team_t *root_team; kmp_team_t *hot_team; int hot_team_max_nth; kmp_r_sched_t r_sched = __kmp_get_schedule_global(); // get current state of scheduling globals kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); KMP_DEBUG_ASSERT(root); KMP_ASSERT(!root->r.r_begin); /* setup the root state structure */ __kmp_init_lock(&root->r.r_begin_lock); root->r.r_begin = FALSE; root->r.r_active = FALSE; root->r.r_in_parallel = 0; root->r.r_blocktime = __kmp_dflt_blocktime; #if KMP_AFFINITY_SUPPORTED root->r.r_affinity_assigned = FALSE; #endif /* setup the root team for this task */ /* allocate the root team structure */ KF_TRACE(10, ("__kmp_initialize_root: before root_team\n")); root_team = __kmp_allocate_team(root, 1, // new_nproc 1, // max_nproc #if OMPT_SUPPORT ompt_data_none, // root parallel id #endif __kmp_nested_proc_bind.bind_types[0], &r_icvs, 0 // argc USE_NESTED_HOT_ARG(NULL) // primary thread is unknown ); #if USE_DEBUGGER // Non-NULL value should be assigned to make the debugger display the root // team. TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0)); #endif KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team)); root->r.r_root_team = root_team; root_team->t.t_control_stack_top = NULL; /* initialize root team */ root_team->t.t_threads[0] = NULL; root_team->t.t_nproc = 1; root_team->t.t_serialized = 1; // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; root_team->t.t_sched.sched = r_sched.sched; KA_TRACE( 20, ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n", root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); /* setup the hot team for this task */ /* allocate the hot team structure */ KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n")); hot_team = __kmp_allocate_team(root, 1, // new_nproc __kmp_dflt_team_nth_ub * 2, // max_nproc #if OMPT_SUPPORT ompt_data_none, // root parallel id #endif __kmp_nested_proc_bind.bind_types[0], &r_icvs, 0 // argc USE_NESTED_HOT_ARG(NULL) // primary thread is unknown ); KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team)); root->r.r_hot_team = hot_team; root_team->t.t_control_stack_top = NULL; /* first-time initialization */ hot_team->t.t_parent = root_team; /* initialize hot team */ hot_team_max_nth = hot_team->t.t_max_nproc; for (f = 0; f < hot_team_max_nth; ++f) { hot_team->t.t_threads[f] = NULL; } hot_team->t.t_nproc = 1; // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; hot_team->t.t_sched.sched = r_sched.sched; hot_team->t.t_size_changed = 0; } #ifdef KMP_DEBUG typedef struct kmp_team_list_item { kmp_team_p const *entry; struct kmp_team_list_item *next; } kmp_team_list_item_t; typedef kmp_team_list_item_t *kmp_team_list_t; static void __kmp_print_structure_team_accum( // Add team to list of teams. kmp_team_list_t list, // List of teams. kmp_team_p const *team // Team to add. ) { // List must terminate with item where both entry and next are NULL. // Team is added to the list only once. // List is sorted in ascending order by team id. // Team id is *not* a key. kmp_team_list_t l; KMP_DEBUG_ASSERT(list != NULL); if (team == NULL) { return; } __kmp_print_structure_team_accum(list, team->t.t_parent); __kmp_print_structure_team_accum(list, team->t.t_next_pool); // Search list for the team. l = list; while (l->next != NULL && l->entry != team) { l = l->next; } if (l->next != NULL) { return; // Team has been added before, exit. } // Team is not found. Search list again for insertion point. l = list; while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) { l = l->next; } // Insert team. { kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC( sizeof(kmp_team_list_item_t)); *item = *l; l->entry = team; l->next = item; } } static void __kmp_print_structure_team(char const *title, kmp_team_p const *team ) { __kmp_printf("%s", title); if (team != NULL) { __kmp_printf("%2x %p\n", team->t.t_id, team); } else { __kmp_printf(" - (nil)\n"); } } static void __kmp_print_structure_thread(char const *title, kmp_info_p const *thread) { __kmp_printf("%s", title); if (thread != NULL) { __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread); } else { __kmp_printf(" - (nil)\n"); } } void __kmp_print_structure(void) { kmp_team_list_t list; // Initialize list of teams. list = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t)); list->entry = NULL; list->next = NULL; __kmp_printf("\n------------------------------\nGlobal Thread " "Table\n------------------------------\n"); { int gtid; for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { __kmp_printf("%2d", gtid); if (__kmp_threads != NULL) { __kmp_printf(" %p", __kmp_threads[gtid]); } if (__kmp_root != NULL) { __kmp_printf(" %p", __kmp_root[gtid]); } __kmp_printf("\n"); } } // Print out __kmp_threads array. __kmp_printf("\n------------------------------\nThreads\n--------------------" "----------\n"); if (__kmp_threads != NULL) { int gtid; for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { kmp_info_t const *thread = __kmp_threads[gtid]; if (thread != NULL) { __kmp_printf("GTID %2d %p:\n", gtid, thread); __kmp_printf(" Our Root: %p\n", thread->th.th_root); __kmp_print_structure_team(" Our Team: ", thread->th.th_team); __kmp_print_structure_team(" Serial Team: ", thread->th.th_serial_team); __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc); __kmp_print_structure_thread(" Primary: ", thread->th.th_team_master); __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized); __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc); __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind); __kmp_print_structure_thread(" Next in pool: ", thread->th.th_next_pool); __kmp_printf("\n"); __kmp_print_structure_team_accum(list, thread->th.th_team); __kmp_print_structure_team_accum(list, thread->th.th_serial_team); } } } else { __kmp_printf("Threads array is not allocated.\n"); } // Print out __kmp_root array. __kmp_printf("\n------------------------------\nUbers\n----------------------" "--------\n"); if (__kmp_root != NULL) { int gtid; for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { kmp_root_t const *root = __kmp_root[gtid]; if (root != NULL) { __kmp_printf("GTID %2d %p:\n", gtid, root); __kmp_print_structure_team(" Root Team: ", root->r.r_root_team); __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team); __kmp_print_structure_thread(" Uber Thread: ", root->r.r_uber_thread); __kmp_printf(" Active?: %2d\n", root->r.r_active); __kmp_printf(" In Parallel: %2d\n", KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel)); __kmp_printf("\n"); __kmp_print_structure_team_accum(list, root->r.r_root_team); __kmp_print_structure_team_accum(list, root->r.r_hot_team); } } } else { __kmp_printf("Ubers array is not allocated.\n"); } __kmp_printf("\n------------------------------\nTeams\n----------------------" "--------\n"); while (list->next != NULL) { kmp_team_p const *team = list->entry; int i; __kmp_printf("Team %2x %p:\n", team->t.t_id, team); __kmp_print_structure_team(" Parent Team: ", team->t.t_parent); __kmp_printf(" Primary TID: %2d\n", team->t.t_master_tid); __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc); __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized); __kmp_printf(" Number threads: %2d\n", team->t.t_nproc); for (i = 0; i < team->t.t_nproc; ++i) { __kmp_printf(" Thread %2d: ", i); __kmp_print_structure_thread("", team->t.t_threads[i]); } __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool); __kmp_printf("\n"); list = list->next; } // Print out __kmp_thread_pool and __kmp_team_pool. __kmp_printf("\n------------------------------\nPools\n----------------------" "--------\n"); __kmp_print_structure_thread("Thread pool: ", CCAST(kmp_info_t *, __kmp_thread_pool)); __kmp_print_structure_team("Team pool: ", CCAST(kmp_team_t *, __kmp_team_pool)); __kmp_printf("\n"); // Free team list. while (list != NULL) { kmp_team_list_item_t *item = list; list = list->next; KMP_INTERNAL_FREE(item); } } #endif //--------------------------------------------------------------------------- // Stuff for per-thread fast random number generator // Table of primes static const unsigned __kmp_primes[] = { 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877, 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231, 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201, 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3, 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7, 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9, 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45, 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7, 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363, 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3, 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f}; //--------------------------------------------------------------------------- // __kmp_get_random: Get a random number using a linear congruential method. unsigned short __kmp_get_random(kmp_info_t *thread) { unsigned x = thread->th.th_x; unsigned short r = (unsigned short)(x >> 16); thread->th.th_x = x * thread->th.th_a + 1; KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n", thread->th.th_info.ds.ds_tid, r)); return r; } //-------------------------------------------------------- // __kmp_init_random: Initialize a random number generator void __kmp_init_random(kmp_info_t *thread) { unsigned seed = thread->th.th_info.ds.ds_tid; thread->th.th_a = __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))]; thread->th.th_x = (seed + 1) * thread->th.th_a + 1; KA_TRACE(30, ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a)); } #if KMP_OS_WINDOWS /* reclaim array entries for root threads that are already dead, returns number * reclaimed */ static int __kmp_reclaim_dead_roots(void) { int i, r = 0; for (i = 0; i < __kmp_threads_capacity; ++i) { if (KMP_UBER_GTID(i) && !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) && !__kmp_root[i] ->r.r_active) { // AC: reclaim only roots died in non-active state r += __kmp_unregister_root_other_thread(i); } } return r; } #endif /* This function attempts to create free entries in __kmp_threads and __kmp_root, and returns the number of free entries generated. For Windows* OS static library, the first mechanism used is to reclaim array entries for root threads that are already dead. On all platforms, expansion is attempted on the arrays __kmp_threads_ and __kmp_root, with appropriate update to __kmp_threads_capacity. Array capacity is increased by doubling with clipping to __kmp_tp_capacity, if threadprivate cache array has been created. Synchronization with __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock. After any dead root reclamation, if the clipping value allows array expansion to result in the generation of a total of nNeed free slots, the function does that expansion. If not, nothing is done beyond the possible initial root thread reclamation. If any argument is negative, the behavior is undefined. */ static int __kmp_expand_threads(int nNeed) { int added = 0; int minimumRequiredCapacity; int newCapacity; kmp_info_t **newThreads; kmp_root_t **newRoot; // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so // resizing __kmp_threads does not need additional protection if foreign // threads are present #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB /* only for Windows static library */ /* reclaim array entries for root threads that are already dead */ added = __kmp_reclaim_dead_roots(); if (nNeed) { nNeed -= added; if (nNeed < 0) nNeed = 0; } #endif if (nNeed <= 0) return added; // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become // > __kmp_max_nth in one of two ways: // // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0] // may not be reused by another thread, so we may need to increase // __kmp_threads_capacity to __kmp_max_nth + 1. // // 2) New foreign root(s) are encountered. We always register new foreign // roots. This may cause a smaller # of threads to be allocated at // subsequent parallel regions, but the worker threads hang around (and // eventually go to sleep) and need slots in the __kmp_threads[] array. // // Anyway, that is the reason for moving the check to see if // __kmp_max_nth was exceeded into __kmp_reserve_threads() // instead of having it performed here. -BB KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity); /* compute expansion headroom to check if we can expand */ if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) { /* possible expansion too small -- give up */ return added; } minimumRequiredCapacity = __kmp_threads_capacity + nNeed; newCapacity = __kmp_threads_capacity; do { newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1) : __kmp_sys_max_nth; } while (newCapacity < minimumRequiredCapacity); newThreads = (kmp_info_t **)__kmp_allocate( (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE); newRoot = (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity); KMP_MEMCPY(newThreads, __kmp_threads, __kmp_threads_capacity * sizeof(kmp_info_t *)); KMP_MEMCPY(newRoot, __kmp_root, __kmp_threads_capacity * sizeof(kmp_root_t *)); // Put old __kmp_threads array on a list. Any ongoing references to the old // list will be valid. This list is cleaned up at library shutdown. kmp_old_threads_list_t *node = (kmp_old_threads_list_t *)__kmp_allocate(sizeof(kmp_old_threads_list_t)); node->threads = __kmp_threads; node->next = __kmp_old_threads_list; __kmp_old_threads_list = node; *(kmp_info_t * *volatile *)&__kmp_threads = newThreads; *(kmp_root_t * *volatile *)&__kmp_root = newRoot; added += newCapacity - __kmp_threads_capacity; *(volatile int *)&__kmp_threads_capacity = newCapacity; if (newCapacity > __kmp_tp_capacity) { __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) { __kmp_threadprivate_resize_cache(newCapacity); } else { // increase __kmp_tp_capacity to correspond with kmp_threads size *(volatile int *)&__kmp_tp_capacity = newCapacity; } __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); } return added; } /* Register the current thread as a root thread and obtain our gtid. We must have the __kmp_initz_lock held at this point. Argument TRUE only if are the thread that calls from __kmp_do_serial_initialize() */ int __kmp_register_root(int initial_thread) { kmp_info_t *root_thread; kmp_root_t *root; int gtid; int capacity; __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); KA_TRACE(20, ("__kmp_register_root: entered\n")); KMP_MB(); /* 2007-03-02: If initial thread did not invoke OpenMP RTL yet, and this thread is not an initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not work as expected -- it may return false (that means there is at least one empty slot in __kmp_threads array), but it is possible the only free slot is #0, which is reserved for initial thread and so cannot be used for this one. Following code workarounds this bug. However, right solution seems to be not reserving slot #0 for initial thread because: (1) there is no magic in slot #0, (2) we cannot detect initial thread reliably (the first thread which does serial initialization may be not a real initial thread). */ capacity = __kmp_threads_capacity; if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) { --capacity; } // If it is not for initializing the hidden helper team, we need to take // __kmp_hidden_helper_threads_num out of the capacity because it is included // in __kmp_threads_capacity. if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)) { capacity -= __kmp_hidden_helper_threads_num; } /* see if there are too many threads */ if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) { if (__kmp_tp_cached) { __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity), KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null); } else { __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads), __kmp_msg_null); } } // When hidden helper task is enabled, __kmp_threads is organized as follows: // 0: initial thread, also a regular OpenMP thread. // [1, __kmp_hidden_helper_threads_num]: slots for hidden helper threads. // [__kmp_hidden_helper_threads_num + 1, __kmp_threads_capacity): slots for // regular OpenMP threads. if (TCR_4(__kmp_init_hidden_helper_threads)) { // Find an available thread slot for hidden helper thread. Slots for hidden // helper threads start from 1 to __kmp_hidden_helper_threads_num. for (gtid = 1; TCR_PTR(__kmp_threads[gtid]) != NULL && gtid <= __kmp_hidden_helper_threads_num; gtid++) ; KMP_ASSERT(gtid <= __kmp_hidden_helper_threads_num); KA_TRACE(1, ("__kmp_register_root: found slot in threads array for " "hidden helper thread: T#%d\n", gtid)); } else { /* find an available thread slot */ // Don't reassign the zero slot since we need that to only be used by // initial thread. Slots for hidden helper threads should also be skipped. if (initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) { gtid = 0; } else { for (gtid = __kmp_hidden_helper_threads_num + 1; TCR_PTR(__kmp_threads[gtid]) != NULL; gtid++) ; } KA_TRACE( 1, ("__kmp_register_root: found slot in threads array: T#%d\n", gtid)); KMP_ASSERT(gtid < __kmp_threads_capacity); } /* update global accounting */ __kmp_all_nth++; TCW_4(__kmp_nth, __kmp_nth + 1); // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low // numbers of procs, and method #2 (keyed API call) for higher numbers. if (__kmp_adjust_gtid_mode) { if (__kmp_all_nth >= __kmp_tls_gtid_min) { if (TCR_4(__kmp_gtid_mode) != 2) { TCW_4(__kmp_gtid_mode, 2); } } else { if (TCR_4(__kmp_gtid_mode) != 1) { TCW_4(__kmp_gtid_mode, 1); } } } #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime to zero if necessary */ /* Middle initialization might not have occurred yet */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { if (__kmp_nth > __kmp_avail_proc) { __kmp_zero_bt = TRUE; } } #endif /* KMP_ADJUST_BLOCKTIME */ /* setup this new hierarchy */ if (!(root = __kmp_root[gtid])) { root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t)); KMP_DEBUG_ASSERT(!root->r.r_root_team); } #if KMP_STATS_ENABLED // Initialize stats as soon as possible (right after gtid assignment). __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid); __kmp_stats_thread_ptr->startLife(); KMP_SET_THREAD_STATE(SERIAL_REGION); KMP_INIT_PARTITIONED_TIMERS(OMP_serial); #endif __kmp_initialize_root(root); /* setup new root thread structure */ if (root->r.r_uber_thread) { root_thread = root->r.r_uber_thread; } else { root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t)); if (__kmp_storage_map) { __kmp_print_thread_storage_map(root_thread, gtid); } root_thread->th.th_info.ds.ds_gtid = gtid; #if OMPT_SUPPORT root_thread->th.ompt_thread_info.thread_data = ompt_data_none; #endif root_thread->th.th_root = root; if (__kmp_env_consistency_check) { root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid); } #if USE_FAST_MEMORY __kmp_initialize_fast_memory(root_thread); #endif /* USE_FAST_MEMORY */ #if KMP_USE_BGET KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL); __kmp_initialize_bget(root_thread); #endif __kmp_init_random(root_thread); // Initialize random number generator } /* setup the serial team held in reserve by the root thread */ if (!root_thread->th.th_serial_team) { kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); KF_TRACE(10, ("__kmp_register_root: before serial_team\n")); root_thread->th.th_serial_team = __kmp_allocate_team( root, 1, 1, #if OMPT_SUPPORT ompt_data_none, // root parallel id #endif proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL)); } KMP_ASSERT(root_thread->th.th_serial_team); KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n", root_thread->th.th_serial_team)); /* drop root_thread into place */ TCW_SYNC_PTR(__kmp_threads[gtid], root_thread); root->r.r_root_team->t.t_threads[0] = root_thread; root->r.r_hot_team->t.t_threads[0] = root_thread; root_thread->th.th_serial_team->t.t_threads[0] = root_thread; // AC: the team created in reserve, not for execution (it is unused for now). root_thread->th.th_serial_team->t.t_serialized = 0; root->r.r_uber_thread = root_thread; /* initialize the thread, get it ready to go */ __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid); TCW_4(__kmp_init_gtid, TRUE); /* prepare the primary thread for get_gtid() */ __kmp_gtid_set_specific(gtid); #if USE_ITT_BUILD __kmp_itt_thread_name(gtid); #endif /* USE_ITT_BUILD */ #ifdef KMP_TDATA_GTID __kmp_gtid = gtid; #endif __kmp_create_worker(gtid, root_thread, __kmp_stksize); KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid); KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team), root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); { // Initialize barrier data. int b; for (b = 0; b < bs_last_barrier; ++b) { root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE; #if USE_DEBUGGER root_thread->th.th_bar[b].bb.b_worker_arrived = 0; #endif } } KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == KMP_INIT_BARRIER_STATE); #if KMP_AFFINITY_SUPPORTED root_thread->th.th_current_place = KMP_PLACE_UNDEFINED; root_thread->th.th_new_place = KMP_PLACE_UNDEFINED; root_thread->th.th_first_place = KMP_PLACE_UNDEFINED; root_thread->th.th_last_place = KMP_PLACE_UNDEFINED; #endif /* KMP_AFFINITY_SUPPORTED */ root_thread->th.th_def_allocator = __kmp_def_allocator; root_thread->th.th_prev_level = 0; root_thread->th.th_prev_num_threads = 1; kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t)); tmp->cg_root = root_thread; tmp->cg_thread_limit = __kmp_cg_max_nth; tmp->cg_nthreads = 1; KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with" " cg_nthreads init to 1\n", root_thread, tmp)); tmp->up = NULL; root_thread->th.th_cg_roots = tmp; __kmp_root_counter++; #if OMPT_SUPPORT if (!initial_thread && ompt_enabled.enabled) { kmp_info_t *root_thread = ompt_get_thread(); ompt_set_thread_state(root_thread, ompt_state_overhead); if (ompt_enabled.ompt_callback_thread_begin) { ompt_callbacks.ompt_callback(ompt_callback_thread_begin)( ompt_thread_initial, __ompt_get_thread_data_internal()); } ompt_data_t *task_data; ompt_data_t *parallel_data; __ompt_get_task_info_internal(0, NULL, &task_data, NULL, ¶llel_data, NULL); if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial); } ompt_set_thread_state(root_thread, ompt_state_work_serial); } #endif #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_thread_begin(); #endif KMP_MB(); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); return gtid; } #if KMP_NESTED_HOT_TEAMS static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level, const int max_level) { int i, n, nth; kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams; if (!hot_teams || !hot_teams[level].hot_team) { return 0; } KMP_DEBUG_ASSERT(level < max_level); kmp_team_t *team = hot_teams[level].hot_team; nth = hot_teams[level].hot_team_nth; n = nth - 1; // primary thread is not freed if (level < max_level - 1) { for (i = 0; i < nth; ++i) { kmp_info_t *th = team->t.t_threads[i]; n += __kmp_free_hot_teams(root, th, level + 1, max_level); if (i > 0 && th->th.th_hot_teams) { __kmp_free(th->th.th_hot_teams); th->th.th_hot_teams = NULL; } } } __kmp_free_team(root, team, NULL); return n; } #endif // Resets a root thread and clear its root and hot teams. // Returns the number of __kmp_threads entries directly and indirectly freed. static int __kmp_reset_root(int gtid, kmp_root_t *root) { kmp_team_t *root_team = root->r.r_root_team; kmp_team_t *hot_team = root->r.r_hot_team; int n = hot_team->t.t_nproc; int i; KMP_DEBUG_ASSERT(!root->r.r_active); root->r.r_root_team = NULL; root->r.r_hot_team = NULL; // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team // before call to __kmp_free_team(). __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL)); #if KMP_NESTED_HOT_TEAMS if (__kmp_hot_teams_max_level > 0) { // need to free nested hot teams and their threads if any for (i = 0; i < hot_team->t.t_nproc; ++i) { kmp_info_t *th = hot_team->t.t_threads[i]; if (__kmp_hot_teams_max_level > 1) { n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level); } if (th->th.th_hot_teams) { __kmp_free(th->th.th_hot_teams); th->th.th_hot_teams = NULL; } } } #endif __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL)); // Before we can reap the thread, we need to make certain that all other // threads in the teams that had this root as ancestor have stopped trying to // steal tasks. if (__kmp_tasking_mode != tskm_immediate_exec) { __kmp_wait_to_unref_task_teams(); } #if KMP_OS_WINDOWS /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */ KA_TRACE( 10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread)); __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread); #endif /* KMP_OS_WINDOWS */ #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_thread_end(); #endif #if OMPT_SUPPORT ompt_data_t *task_data; ompt_data_t *parallel_data; __ompt_get_task_info_internal(0, NULL, &task_data, NULL, ¶llel_data, NULL); if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial); } if (ompt_enabled.ompt_callback_thread_end) { ompt_callbacks.ompt_callback(ompt_callback_thread_end)( &(root->r.r_uber_thread->th.ompt_thread_info.thread_data)); } #endif TCW_4(__kmp_nth, __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth. i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--; KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p" " to %d\n", root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots, root->r.r_uber_thread->th.th_cg_roots->cg_nthreads)); if (i == 1) { // need to free contention group structure KMP_DEBUG_ASSERT(root->r.r_uber_thread == root->r.r_uber_thread->th.th_cg_roots->cg_root); KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL); __kmp_free(root->r.r_uber_thread->th.th_cg_roots); root->r.r_uber_thread->th.th_cg_roots = NULL; } __kmp_reap_thread(root->r.r_uber_thread, 1); // We canot put root thread to __kmp_thread_pool, so we have to reap it // instead of freeing. root->r.r_uber_thread = NULL; /* mark root as no longer in use */ root->r.r_begin = FALSE; return n; } void __kmp_unregister_root_current_thread(int gtid) { KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid)); /* this lock should be ok, since unregister_root_current_thread is never called during an abort, only during a normal close. furthermore, if you have the forkjoin lock, you should never try to get the initz lock */ __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid)); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); return; } kmp_root_t *root = __kmp_root[gtid]; KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]); KMP_ASSERT(KMP_UBER_GTID(gtid)); KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root); KMP_ASSERT(root->r.r_active == FALSE); KMP_MB(); kmp_info_t *thread = __kmp_threads[gtid]; kmp_team_t *team = thread->th.th_team; kmp_task_team_t *task_team = thread->th.th_task_team; // we need to wait for the proxy tasks before finishing the thread if (task_team != NULL && (task_team->tt.tt_found_proxy_tasks || task_team->tt.tt_hidden_helper_task_encountered)) { #if OMPT_SUPPORT // the runtime is shutting down so we won't report any events thread->th.ompt_thread_info.state = ompt_state_undefined; #endif __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL)); } __kmp_reset_root(gtid, root); KMP_MB(); KC_TRACE(10, ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid)); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); } #if KMP_OS_WINDOWS /* __kmp_forkjoin_lock must be already held Unregisters a root thread that is not the current thread. Returns the number of __kmp_threads entries freed as a result. */ static int __kmp_unregister_root_other_thread(int gtid) { kmp_root_t *root = __kmp_root[gtid]; int r; KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid)); KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]); KMP_ASSERT(KMP_UBER_GTID(gtid)); KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root); KMP_ASSERT(root->r.r_active == FALSE); r = __kmp_reset_root(gtid, root); KC_TRACE(10, ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid)); return r; } #endif #if KMP_DEBUG void __kmp_task_info() { kmp_int32 gtid = __kmp_entry_gtid(); kmp_int32 tid = __kmp_tid_from_gtid(gtid); kmp_info_t *this_thr = __kmp_threads[gtid]; kmp_team_t *steam = this_thr->th.th_serial_team; kmp_team_t *team = this_thr->th.th_team; __kmp_printf( "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p " "ptask=%p\n", gtid, tid, this_thr, team, steam, this_thr->th.th_current_task, team->t.t_implicit_task_taskdata[tid].td_parent); } #endif // KMP_DEBUG /* TODO optimize with one big memclr, take out what isn't needed, split responsibility to workers as much as possible, and delay initialization of features as much as possible */ static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team, int tid, int gtid) { /* this_thr->th.th_info.ds.ds_gtid is setup in kmp_allocate_thread/create_worker. this_thr->th.th_serial_team is setup in __kmp_allocate_thread */ KMP_DEBUG_ASSERT(this_thr != NULL); KMP_DEBUG_ASSERT(this_thr->th.th_serial_team); KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(team->t.t_threads); KMP_DEBUG_ASSERT(team->t.t_dispatch); kmp_info_t *master = team->t.t_threads[0]; KMP_DEBUG_ASSERT(master); KMP_DEBUG_ASSERT(master->th.th_root); KMP_MB(); TCW_SYNC_PTR(this_thr->th.th_team, team); this_thr->th.th_info.ds.ds_tid = tid; this_thr->th.th_set_nproc = 0; if (__kmp_tasking_mode != tskm_immediate_exec) // When tasking is possible, threads are not safe to reap until they are // done tasking; this will be set when tasking code is exited in wait this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP; else // no tasking --> always safe to reap this_thr->th.th_reap_state = KMP_SAFE_TO_REAP; this_thr->th.th_set_proc_bind = proc_bind_default; #if KMP_AFFINITY_SUPPORTED this_thr->th.th_new_place = this_thr->th.th_current_place; #endif this_thr->th.th_root = master->th.th_root; /* setup the thread's cache of the team structure */ this_thr->th.th_team_nproc = team->t.t_nproc; this_thr->th.th_team_master = master; this_thr->th.th_team_serialized = team->t.t_serialized; KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata); KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n", tid, gtid, this_thr, this_thr->th.th_current_task)); __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr, team, tid, TRUE); KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n", tid, gtid, this_thr, this_thr->th.th_current_task)); // TODO: Initialize ICVs from parent; GEH - isn't that already done in // __kmp_initialize_team()? /* TODO no worksharing in speculative threads */ this_thr->th.th_dispatch = &team->t.t_dispatch[tid]; this_thr->th.th_local.this_construct = 0; if (!this_thr->th.th_pri_common) { this_thr->th.th_pri_common = (struct common_table *)__kmp_allocate(sizeof(struct common_table)); if (__kmp_storage_map) { __kmp_print_storage_map_gtid( gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1, sizeof(struct common_table), "th_%d.th_pri_common\n", gtid); } this_thr->th.th_pri_head = NULL; } if (this_thr != master && // Primary thread's CG root is initialized elsewhere this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set // Make new thread's CG root same as primary thread's KMP_DEBUG_ASSERT(master->th.th_cg_roots); kmp_cg_root_t *tmp = this_thr->th.th_cg_roots; if (tmp) { // worker changes CG, need to check if old CG should be freed int i = tmp->cg_nthreads--; KA_TRACE(100, ("__kmp_initialize_info: Thread %p decrement cg_nthreads" " on node %p of thread %p to %d\n", this_thr, tmp, tmp->cg_root, tmp->cg_nthreads)); if (i == 1) { __kmp_free(tmp); // last thread left CG --> free it } } this_thr->th.th_cg_roots = master->th.th_cg_roots; // Increment new thread's CG root's counter to add the new thread this_thr->th.th_cg_roots->cg_nthreads++; KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on" " node %p of thread %p to %d\n", this_thr, this_thr->th.th_cg_roots, this_thr->th.th_cg_roots->cg_root, this_thr->th.th_cg_roots->cg_nthreads)); this_thr->th.th_current_task->td_icvs.thread_limit = this_thr->th.th_cg_roots->cg_thread_limit; } /* Initialize dynamic dispatch */ { volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch; // Use team max_nproc since this will never change for the team. size_t disp_size = sizeof(dispatch_private_info_t) * (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers); KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid, team->t.t_max_nproc)); KMP_ASSERT(dispatch); KMP_DEBUG_ASSERT(team->t.t_dispatch); KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]); dispatch->th_disp_index = 0; dispatch->th_doacross_buf_idx = 0; if (!dispatch->th_disp_buffer) { dispatch->th_disp_buffer = (dispatch_private_info_t *)__kmp_allocate(disp_size); if (__kmp_storage_map) { __kmp_print_storage_map_gtid( gtid, &dispatch->th_disp_buffer[0], &dispatch->th_disp_buffer[team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers], disp_size, "th_%d.th_dispatch.th_disp_buffer " "(team_%d.t_dispatch[%d].th_disp_buffer)", gtid, team->t.t_id, gtid); } } else { memset(&dispatch->th_disp_buffer[0], '\0', disp_size); } dispatch->th_dispatch_pr_current = 0; dispatch->th_dispatch_sh_current = 0; dispatch->th_deo_fcn = 0; /* ORDERED */ dispatch->th_dxo_fcn = 0; /* END ORDERED */ } this_thr->th.th_next_pool = NULL; if (!this_thr->th.th_task_state_memo_stack) { size_t i; this_thr->th.th_task_state_memo_stack = (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8)); this_thr->th.th_task_state_top = 0; this_thr->th.th_task_state_stack_sz = 4; for (i = 0; i < this_thr->th.th_task_state_stack_sz; ++i) // zero init the stack this_thr->th.th_task_state_memo_stack[i] = 0; } KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here); KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0); KMP_MB(); } /* allocate a new thread for the requesting team. this is only called from within a forkjoin critical section. we will first try to get an available thread from the thread pool. if none is available, we will fork a new one assuming we are able to create a new one. this should be assured, as the caller should check on this first. */ kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, int new_tid) { kmp_team_t *serial_team; kmp_info_t *new_thr; int new_gtid; KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid())); KMP_DEBUG_ASSERT(root && team); #if !KMP_NESTED_HOT_TEAMS KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid())); #endif KMP_MB(); /* first, try to get one from the thread pool */ if (__kmp_thread_pool) { new_thr = CCAST(kmp_info_t *, __kmp_thread_pool); __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool; if (new_thr == __kmp_thread_pool_insert_pt) { __kmp_thread_pool_insert_pt = NULL; } TCW_4(new_thr->th.th_in_pool, FALSE); __kmp_suspend_initialize_thread(new_thr); __kmp_lock_suspend_mx(new_thr); if (new_thr->th.th_active_in_pool == TRUE) { KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE); KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); new_thr->th.th_active_in_pool = FALSE; } __kmp_unlock_suspend_mx(new_thr); KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n", __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid)); KMP_ASSERT(!new_thr->th.th_team); KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity); /* setup the thread structure */ __kmp_initialize_info(new_thr, team, new_tid, new_thr->th.th_info.ds.ds_gtid); KMP_DEBUG_ASSERT(new_thr->th.th_serial_team); TCW_4(__kmp_nth, __kmp_nth + 1); new_thr->th.th_task_state = 0; new_thr->th.th_task_state_top = 0; new_thr->th.th_task_state_stack_sz = 4; if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { // Make sure pool thread has transitioned to waiting on own thread struct KMP_DEBUG_ASSERT(new_thr->th.th_used_in_team.load() == 0); // Thread activated in __kmp_allocate_team when increasing team size } #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime back to zero if necessary */ /* Middle initialization might not have occurred yet */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { if (__kmp_nth > __kmp_avail_proc) { __kmp_zero_bt = TRUE; } } #endif /* KMP_ADJUST_BLOCKTIME */ #if KMP_DEBUG // If thread entered pool via __kmp_free_thread, wait_flag should != // KMP_BARRIER_PARENT_FLAG. int b; kmp_balign_t *balign = new_thr->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #endif KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n", __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid)); KMP_MB(); return new_thr; } /* no, well fork a new one */ KMP_ASSERT(__kmp_nth == __kmp_all_nth); KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity); #if KMP_USE_MONITOR // If this is the first worker thread the RTL is creating, then also // launch the monitor thread. We try to do this as early as possible. if (!TCR_4(__kmp_init_monitor)) { __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); if (!TCR_4(__kmp_init_monitor)) { KF_TRACE(10, ("before __kmp_create_monitor\n")); TCW_4(__kmp_init_monitor, 1); __kmp_create_monitor(&__kmp_monitor); KF_TRACE(10, ("after __kmp_create_monitor\n")); #if KMP_OS_WINDOWS // AC: wait until monitor has started. This is a fix for CQ232808. // The reason is that if the library is loaded/unloaded in a loop with // small (parallel) work in between, then there is high probability that // monitor thread started after the library shutdown. At shutdown it is // too late to cope with the problem, because when the primary thread is // in DllMain (process detach) the monitor has no chances to start (it is // blocked), and primary thread has no means to inform the monitor that // the library has gone, because all the memory which the monitor can // access is going to be released/reset. while (TCR_4(__kmp_init_monitor) < 2) { KMP_YIELD(TRUE); } KF_TRACE(10, ("after monitor thread has started\n")); #endif } __kmp_release_bootstrap_lock(&__kmp_monitor_lock); } #endif KMP_MB(); { int new_start_gtid = TCR_4(__kmp_init_hidden_helper_threads) ? 1 : __kmp_hidden_helper_threads_num + 1; for (new_gtid = new_start_gtid; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) { KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity); } if (TCR_4(__kmp_init_hidden_helper_threads)) { KMP_DEBUG_ASSERT(new_gtid <= __kmp_hidden_helper_threads_num); } } /* allocate space for it. */ new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t)); TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr); #if USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG // suppress race conditions detection on synchronization flags in debug mode // this helps to analyze library internals eliminating false positives __itt_suppress_mark_range( __itt_suppress_range, __itt_suppress_threading_errors, &new_thr->th.th_sleep_loc, sizeof(new_thr->th.th_sleep_loc)); __itt_suppress_mark_range( __itt_suppress_range, __itt_suppress_threading_errors, &new_thr->th.th_reap_state, sizeof(new_thr->th.th_reap_state)); #if KMP_OS_WINDOWS __itt_suppress_mark_range( __itt_suppress_range, __itt_suppress_threading_errors, &new_thr->th.th_suspend_init, sizeof(new_thr->th.th_suspend_init)); #else __itt_suppress_mark_range(__itt_suppress_range, __itt_suppress_threading_errors, &new_thr->th.th_suspend_init_count, sizeof(new_thr->th.th_suspend_init_count)); #endif // TODO: check if we need to also suppress b_arrived flags __itt_suppress_mark_range(__itt_suppress_range, __itt_suppress_threading_errors, CCAST(kmp_uint64 *, &new_thr->th.th_bar[0].bb.b_go), sizeof(new_thr->th.th_bar[0].bb.b_go)); __itt_suppress_mark_range(__itt_suppress_range, __itt_suppress_threading_errors, CCAST(kmp_uint64 *, &new_thr->th.th_bar[1].bb.b_go), sizeof(new_thr->th.th_bar[1].bb.b_go)); __itt_suppress_mark_range(__itt_suppress_range, __itt_suppress_threading_errors, CCAST(kmp_uint64 *, &new_thr->th.th_bar[2].bb.b_go), sizeof(new_thr->th.th_bar[2].bb.b_go)); #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */ if (__kmp_storage_map) { __kmp_print_thread_storage_map(new_thr, new_gtid); } // add the reserve serialized team, initialized from the team's primary thread { kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team); KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n")); new_thr->th.th_serial_team = serial_team = (kmp_team_t *)__kmp_allocate_team(root, 1, 1, #if OMPT_SUPPORT ompt_data_none, // root parallel id #endif proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL)); } KMP_ASSERT(serial_team); serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for // execution (it is unused for now). serial_team->t.t_threads[0] = new_thr; KF_TRACE(10, ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n", new_thr)); /* setup the thread structures */ __kmp_initialize_info(new_thr, team, new_tid, new_gtid); #if USE_FAST_MEMORY __kmp_initialize_fast_memory(new_thr); #endif /* USE_FAST_MEMORY */ #if KMP_USE_BGET KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL); __kmp_initialize_bget(new_thr); #endif __kmp_init_random(new_thr); // Initialize random number generator /* Initialize these only once when thread is grabbed for a team allocation */ KA_TRACE(20, ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n", __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); int b; kmp_balign_t *balign = new_thr->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_go = KMP_INIT_BARRIER_STATE; balign[b].bb.team = NULL; balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING; balign[b].bb.use_oncore_barrier = 0; } TCW_PTR(new_thr->th.th_sleep_loc, NULL); new_thr->th.th_sleep_loc_type = flag_unset; new_thr->th.th_spin_here = FALSE; new_thr->th.th_next_waiting = 0; #if KMP_OS_UNIX new_thr->th.th_blocking = false; #endif #if KMP_AFFINITY_SUPPORTED new_thr->th.th_current_place = KMP_PLACE_UNDEFINED; new_thr->th.th_new_place = KMP_PLACE_UNDEFINED; new_thr->th.th_first_place = KMP_PLACE_UNDEFINED; new_thr->th.th_last_place = KMP_PLACE_UNDEFINED; #endif new_thr->th.th_def_allocator = __kmp_def_allocator; new_thr->th.th_prev_level = 0; new_thr->th.th_prev_num_threads = 1; TCW_4(new_thr->th.th_in_pool, FALSE); new_thr->th.th_active_in_pool = FALSE; TCW_4(new_thr->th.th_active, TRUE); /* adjust the global counters */ __kmp_all_nth++; __kmp_nth++; // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low // numbers of procs, and method #2 (keyed API call) for higher numbers. if (__kmp_adjust_gtid_mode) { if (__kmp_all_nth >= __kmp_tls_gtid_min) { if (TCR_4(__kmp_gtid_mode) != 2) { TCW_4(__kmp_gtid_mode, 2); } } else { if (TCR_4(__kmp_gtid_mode) != 1) { TCW_4(__kmp_gtid_mode, 1); } } } #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime back to zero if necessary */ /* Middle initialization might not have occurred yet */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { if (__kmp_nth > __kmp_avail_proc) { __kmp_zero_bt = TRUE; } } #endif /* KMP_ADJUST_BLOCKTIME */ /* actually fork it and create the new worker thread */ KF_TRACE( 10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr)); __kmp_create_worker(new_gtid, new_thr, __kmp_stksize); KF_TRACE(10, ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr)); KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(), new_gtid)); KMP_MB(); return new_thr; } /* Reinitialize team for reuse. The hot team code calls this case at every fork barrier, so EPCC barrier test are extremely sensitive to changes in it, esp. writes to the team struct, which cause a cache invalidation in all threads. IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */ static void __kmp_reinitialize_team(kmp_team_t *team, kmp_internal_control_t *new_icvs, ident_t *loc) { KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n", team->t.t_threads[0], team)); KMP_DEBUG_ASSERT(team && new_icvs); KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc); KMP_CHECK_UPDATE(team->t.t_ident, loc); KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID()); // Copy ICVs to the primary thread's implicit taskdata __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE); copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs); KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n", team->t.t_threads[0], team)); } /* Initialize the team data structure. This assumes the t_threads and t_max_nproc are already set. Also, we don't touch the arguments */ static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, kmp_internal_control_t *new_icvs, ident_t *loc) { KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team)); /* verify */ KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc); KMP_DEBUG_ASSERT(team->t.t_threads); KMP_MB(); team->t.t_master_tid = 0; /* not needed */ /* team->t.t_master_bar; not needed */ team->t.t_serialized = new_nproc > 1 ? 0 : 1; team->t.t_nproc = new_nproc; /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */ team->t.t_next_pool = NULL; /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess * up hot team */ TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */ team->t.t_invoke = NULL; /* not needed */ // TODO???: team->t.t_max_active_levels = new_max_active_levels; team->t.t_sched.sched = new_icvs->sched.sched; #if KMP_ARCH_X86 || KMP_ARCH_X86_64 team->t.t_fp_control_saved = FALSE; /* not needed */ team->t.t_x87_fpu_control_word = 0; /* not needed */ team->t.t_mxcsr = 0; /* not needed */ #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ team->t.t_construct = 0; team->t.t_ordered.dt.t_value = 0; team->t.t_master_active = FALSE; #ifdef KMP_DEBUG team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */ #endif #if KMP_OS_WINDOWS team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */ #endif team->t.t_control_stack_top = NULL; __kmp_reinitialize_team(team, new_icvs, loc); KMP_MB(); KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team)); } #if KMP_AFFINITY_SUPPORTED // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism. // It calculates the worker + primary thread's partition based upon the parent // thread's partition, and binds each worker to a thread in their partition. // The primary thread's partition should already include its current binding. static void __kmp_partition_places(kmp_team_t *team, int update_master_only) { // Do not partition places for the hidden helper team if (KMP_HIDDEN_HELPER_TEAM(team)) return; // Copy the primary thread's place partition to the team struct kmp_info_t *master_th = team->t.t_threads[0]; KMP_DEBUG_ASSERT(master_th != NULL); kmp_proc_bind_t proc_bind = team->t.t_proc_bind; int first_place = master_th->th.th_first_place; int last_place = master_th->th.th_last_place; int masters_place = master_th->th.th_current_place; int num_masks = __kmp_affinity.num_masks; team->t.t_first_place = first_place; team->t.t_last_place = last_place; KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]), team->t.t_id, masters_place, first_place, last_place)); switch (proc_bind) { case proc_bind_default: // Serial teams might have the proc_bind policy set to proc_bind_default. // Not an issue -- we don't rebind primary thread for any proc_bind policy. KMP_DEBUG_ASSERT(team->t.t_nproc == 1); break; case proc_bind_primary: { int f; int n_th = team->t.t_nproc; for (f = 1; f < n_th; f++) { kmp_info_t *th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th != NULL); th->th.th_first_place = first_place; th->th.th_last_place = last_place; th->th.th_new_place = masters_place; if (__kmp_display_affinity && masters_place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } KA_TRACE(100, ("__kmp_partition_places: primary: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, masters_place, first_place, last_place)); } } break; case proc_bind_close: { int f; int n_th = team->t.t_nproc; int n_places; if (first_place <= last_place) { n_places = last_place - first_place + 1; } else { n_places = num_masks - first_place + last_place + 1; } if (n_th <= n_places) { int place = masters_place; for (f = 1; f < n_th; f++) { kmp_info_t *th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th != NULL); if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } th->th.th_first_place = first_place; th->th.th_last_place = last_place; th->th.th_new_place = place; if (__kmp_display_affinity && place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, place, first_place, last_place)); } } else { int S, rem, gap, s_count; S = n_th / n_places; s_count = 0; rem = n_th - (S * n_places); gap = rem > 0 ? n_places / rem : n_places; int place = masters_place; int gap_ct = gap; for (f = 0; f < n_th; f++) { kmp_info_t *th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th != NULL); th->th.th_first_place = first_place; th->th.th_last_place = last_place; th->th.th_new_place = place; if (__kmp_display_affinity && place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } s_count++; if ((s_count == S) && rem && (gap_ct == gap)) { // do nothing, add an extra thread to place on next iteration } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { // we added an extra thread to this place; move to next place if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } s_count = 0; gap_ct = 1; rem--; } else if (s_count == S) { // place full; don't add extra if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } gap_ct++; s_count = 0; } KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place, first_place, last_place)); } KMP_DEBUG_ASSERT(place == masters_place); } } break; case proc_bind_spread: { int f; int n_th = team->t.t_nproc; int n_places; int thidx; if (first_place <= last_place) { n_places = last_place - first_place + 1; } else { n_places = num_masks - first_place + last_place + 1; } if (n_th <= n_places) { int place = -1; if (n_places != num_masks) { int S = n_places / n_th; int s_count, rem, gap, gap_ct; place = masters_place; rem = n_places - n_th * S; gap = rem ? n_th / rem : 1; gap_ct = gap; thidx = n_th; if (update_master_only == 1) thidx = 1; for (f = 0; f < thidx; f++) { kmp_info_t *th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th != NULL); th->th.th_first_place = place; th->th.th_new_place = place; if (__kmp_display_affinity && place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } s_count = 1; while (s_count < S) { if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } s_count++; } if (rem && (gap_ct == gap)) { if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } rem--; gap_ct = 0; } th->th.th_last_place = place; gap_ct++; if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], num_masks: %u\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place, th->th.th_last_place, num_masks)); } } else { /* Having uniform space of available computation places I can create T partitions of round(P/T) size and put threads into the first place of each partition. */ double current = static_cast<double>(masters_place); double spacing = (static_cast<double>(n_places + 1) / static_cast<double>(n_th)); int first, last; kmp_info_t *th; thidx = n_th + 1; if (update_master_only == 1) thidx = 1; for (f = 0; f < thidx; f++) { first = static_cast<int>(current); last = static_cast<int>(current + spacing) - 1; KMP_DEBUG_ASSERT(last >= first); if (first >= n_places) { if (masters_place) { first -= n_places; last -= n_places; if (first == (masters_place + 1)) { KMP_DEBUG_ASSERT(f == n_th); first--; } if (last == masters_place) { KMP_DEBUG_ASSERT(f == (n_th - 1)); last--; } } else { KMP_DEBUG_ASSERT(f == n_th); first = 0; last = 0; } } if (last >= n_places) { last = (n_places - 1); } place = first; current += spacing; if (f < n_th) { KMP_DEBUG_ASSERT(0 <= first); KMP_DEBUG_ASSERT(n_places > first); KMP_DEBUG_ASSERT(0 <= last); KMP_DEBUG_ASSERT(n_places > last); KMP_DEBUG_ASSERT(last_place >= first_place); th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th); th->th.th_first_place = first; th->th.th_new_place = place; th->th.th_last_place = last; if (__kmp_display_affinity && place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place, th->th.th_last_place, spacing)); } } } KMP_DEBUG_ASSERT(update_master_only || place == masters_place); } else { int S, rem, gap, s_count; S = n_th / n_places; s_count = 0; rem = n_th - (S * n_places); gap = rem > 0 ? n_places / rem : n_places; int place = masters_place; int gap_ct = gap; thidx = n_th; if (update_master_only == 1) thidx = 1; for (f = 0; f < thidx; f++) { kmp_info_t *th = team->t.t_threads[f]; KMP_DEBUG_ASSERT(th != NULL); th->th.th_first_place = place; th->th.th_last_place = place; th->th.th_new_place = place; if (__kmp_display_affinity && place != th->th.th_current_place && team->t.t_display_affinity != 1) { team->t.t_display_affinity = 1; } s_count++; if ((s_count == S) && rem && (gap_ct == gap)) { // do nothing, add an extra thread to place on next iteration } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { // we added an extra thread to this place; move on to next place if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } s_count = 0; gap_ct = 1; rem--; } else if (s_count == S) { // place is full; don't add extra thread if (place == last_place) { place = first_place; } else if (place == (num_masks - 1)) { place = 0; } else { place++; } gap_ct++; s_count = 0; } KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place, th->th.th_last_place)); } KMP_DEBUG_ASSERT(update_master_only || place == masters_place); } } break; default: break; } KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id)); } #endif // KMP_AFFINITY_SUPPORTED /* allocate a new team data structure to use. take one off of the free pool if available */ kmp_team_t * __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, #if OMPT_SUPPORT ompt_data_t ompt_parallel_data, #endif kmp_proc_bind_t new_proc_bind, kmp_internal_control_t *new_icvs, int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) { KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team); int f; kmp_team_t *team; int use_hot_team = !root->r.r_active; int level = 0; int do_place_partition = 1; KA_TRACE(20, ("__kmp_allocate_team: called\n")); KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0); KMP_DEBUG_ASSERT(max_nproc >= new_nproc); KMP_MB(); #if KMP_NESTED_HOT_TEAMS kmp_hot_team_ptr_t *hot_teams; if (master) { team = master->th.th_team; level = team->t.t_active_level; if (master->th.th_teams_microtask) { // in teams construct? if (master->th.th_teams_size.nteams > 1 && ( // #teams > 1 team->t.t_pkfn == (microtask_t)__kmp_teams_master || // inner fork of the teams master->th.th_teams_level < team->t.t_level)) { // or nested parallel inside the teams ++level; // not increment if #teams==1, or for outer fork of the teams; // increment otherwise } // Do not perform the place partition if inner fork of the teams // Wait until nested parallel region encountered inside teams construct if ((master->th.th_teams_size.nteams == 1 && master->th.th_teams_level >= team->t.t_level) || (team->t.t_pkfn == (microtask_t)__kmp_teams_master)) do_place_partition = 0; } hot_teams = master->th.th_hot_teams; if (level < __kmp_hot_teams_max_level && hot_teams && hot_teams[level].hot_team) { // hot team has already been allocated for given level use_hot_team = 1; } else { use_hot_team = 0; } } else { // check we won't access uninitialized hot_teams, just in case KMP_DEBUG_ASSERT(new_nproc == 1); } #endif // Optimization to use a "hot" team if (use_hot_team && new_nproc > 1) { KMP_DEBUG_ASSERT(new_nproc <= max_nproc); #if KMP_NESTED_HOT_TEAMS team = hot_teams[level].hot_team; #else team = root->r.r_hot_team; #endif #if KMP_DEBUG if (__kmp_tasking_mode != tskm_immediate_exec) { KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[0], team->t.t_task_team[1])); } #endif if (team->t.t_nproc != new_nproc && __kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { // Distributed barrier may need a resize int old_nthr = team->t.t_nproc; __kmp_resize_dist_barrier(team, old_nthr, new_nproc); } // If not doing the place partition, then reset the team's proc bind // to indicate that partitioning of all threads still needs to take place if (do_place_partition == 0) team->t.t_proc_bind = proc_bind_default; // Has the number of threads changed? /* Let's assume the most common case is that the number of threads is unchanged, and put that case first. */ if (team->t.t_nproc == new_nproc) { // Check changes in number of threads KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n")); // This case can mean that omp_set_num_threads() was called and the hot // team size was already reduced, so we check the special flag if (team->t.t_size_changed == -1) { team->t.t_size_changed = 1; } else { KMP_CHECK_UPDATE(team->t.t_size_changed, 0); } // TODO???: team->t.t_max_active_levels = new_max_active_levels; kmp_r_sched_t new_sched = new_icvs->sched; // set primary thread's schedule as new run-time schedule KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched); __kmp_reinitialize_team(team, new_icvs, root->r.r_uber_thread->th.th_ident); KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0, team->t.t_threads[0], team)); __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); #if KMP_AFFINITY_SUPPORTED if ((team->t.t_size_changed == 0) && (team->t.t_proc_bind == new_proc_bind)) { if (new_proc_bind == proc_bind_spread) { if (do_place_partition) { // add flag to update only master for spread __kmp_partition_places(team, 1); } } KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind, team->t.t_first_place, team->t.t_last_place)); } else { if (do_place_partition) { KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); __kmp_partition_places(team); } } #else KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); #endif /* KMP_AFFINITY_SUPPORTED */ } else if (team->t.t_nproc > new_nproc) { KA_TRACE(20, ("__kmp_allocate_team: decreasing hot team thread count to %d\n", new_nproc)); team->t.t_size_changed = 1; if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { // Barrier size already reduced earlier in this function // Activate team threads via th_used_in_team __kmp_add_threads_to_team(team, new_nproc); } #if KMP_NESTED_HOT_TEAMS if (__kmp_hot_teams_mode == 0) { // AC: saved number of threads should correspond to team's value in this // mode, can be bigger in mode 1, when hot team has threads in reserve KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc); hot_teams[level].hot_team_nth = new_nproc; #endif // KMP_NESTED_HOT_TEAMS /* release the extra threads we don't need any more */ for (f = new_nproc; f < team->t.t_nproc; f++) { KMP_DEBUG_ASSERT(team->t.t_threads[f]); if (__kmp_tasking_mode != tskm_immediate_exec) { // When decreasing team size, threads no longer in the team should // unref task team. team->t.t_threads[f]->th.th_task_team = NULL; } __kmp_free_thread(team->t.t_threads[f]); team->t.t_threads[f] = NULL; } #if KMP_NESTED_HOT_TEAMS } // (__kmp_hot_teams_mode == 0) else { // When keeping extra threads in team, switch threads to wait on own // b_go flag for (f = new_nproc; f < team->t.t_nproc; ++f) { KMP_DEBUG_ASSERT(team->t.t_threads[f]); kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar; for (int b = 0; b < bs_last_barrier; ++b) { if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) { balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG; } KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0); } } } #endif // KMP_NESTED_HOT_TEAMS team->t.t_nproc = new_nproc; // TODO???: team->t.t_max_active_levels = new_max_active_levels; KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched); __kmp_reinitialize_team(team, new_icvs, root->r.r_uber_thread->th.th_ident); // Update remaining threads for (f = 0; f < new_nproc; ++f) { team->t.t_threads[f]->th.th_team_nproc = new_nproc; } // restore the current task state of the primary thread: should be the // implicit task KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0, team->t.t_threads[0], team)); __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); #ifdef KMP_DEBUG for (f = 0; f < team->t.t_nproc; f++) { KMP_DEBUG_ASSERT(team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc); } #endif if (do_place_partition) { KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); #if KMP_AFFINITY_SUPPORTED __kmp_partition_places(team); #endif } } else { // team->t.t_nproc < new_nproc KA_TRACE(20, ("__kmp_allocate_team: increasing hot team thread count to %d\n", new_nproc)); int old_nproc = team->t.t_nproc; // save old value and use to update only team->t.t_size_changed = 1; #if KMP_NESTED_HOT_TEAMS int avail_threads = hot_teams[level].hot_team_nth; if (new_nproc < avail_threads) avail_threads = new_nproc; kmp_info_t **other_threads = team->t.t_threads; for (f = team->t.t_nproc; f < avail_threads; ++f) { // Adjust barrier data of reserved threads (if any) of the team // Other data will be set in __kmp_initialize_info() below. int b; kmp_balign_t *balign = other_threads[f]->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #if USE_DEBUGGER balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; #endif } } if (hot_teams[level].hot_team_nth >= new_nproc) { // we have all needed threads in reserve, no need to allocate any // this only possible in mode 1, cannot have reserved threads in mode 0 KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1); team->t.t_nproc = new_nproc; // just get reserved threads involved } else { // We may have some threads in reserve, but not enough; // get reserved threads involved if any. team->t.t_nproc = hot_teams[level].hot_team_nth; hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size #endif // KMP_NESTED_HOT_TEAMS if (team->t.t_max_nproc < new_nproc) { /* reallocate larger arrays */ __kmp_reallocate_team_arrays(team, new_nproc); __kmp_reinitialize_team(team, new_icvs, NULL); } #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED /* Temporarily set full mask for primary thread before creation of workers. The reason is that workers inherit the affinity from the primary thread, so if a lot of workers are created on the single core quickly, they don't get a chance to set their own affinity for a long time. */ kmp_affinity_raii_t new_temp_affinity{__kmp_affin_fullMask}; #endif /* allocate new threads for the hot team */ for (f = team->t.t_nproc; f < new_nproc; f++) { kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f); KMP_DEBUG_ASSERT(new_worker); team->t.t_threads[f] = new_worker; KA_TRACE(20, ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier].b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived)); { // Initialize barrier data for new threads. int b; kmp_balign_t *balign = new_worker->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #if USE_DEBUGGER balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; #endif } } } #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED /* Restore initial primary thread's affinity mask */ new_temp_affinity.restore(); #endif #if KMP_NESTED_HOT_TEAMS } // end of check of t_nproc vs. new_nproc vs. hot_team_nth #endif // KMP_NESTED_HOT_TEAMS if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) { // Barrier size already increased earlier in this function // Activate team threads via th_used_in_team __kmp_add_threads_to_team(team, new_nproc); } /* make sure everyone is syncronized */ // new threads below __kmp_initialize_team(team, new_nproc, new_icvs, root->r.r_uber_thread->th.th_ident); /* reinitialize the threads */ KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc); for (f = 0; f < team->t.t_nproc; ++f) __kmp_initialize_info(team->t.t_threads[f], team, f, __kmp_gtid_from_tid(f, team)); // set th_task_state for new threads in hot team with older thread's state kmp_uint8 old_state = team->t.t_threads[old_nproc - 1]->th.th_task_state; for (f = old_nproc; f < team->t.t_nproc; ++f) team->t.t_threads[f]->th.th_task_state = old_state; #ifdef KMP_DEBUG for (f = 0; f < team->t.t_nproc; ++f) { KMP_DEBUG_ASSERT(team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc); } #endif if (do_place_partition) { KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); #if KMP_AFFINITY_SUPPORTED __kmp_partition_places(team); #endif } } // Check changes in number of threads kmp_info_t *master = team->t.t_threads[0]; if (master->th.th_teams_microtask) { for (f = 1; f < new_nproc; ++f) { // propagate teams construct specific info to workers kmp_info_t *thr = team->t.t_threads[f]; thr->th.th_teams_microtask = master->th.th_teams_microtask; thr->th.th_teams_level = master->th.th_teams_level; thr->th.th_teams_size = master->th.th_teams_size; } } #if KMP_NESTED_HOT_TEAMS if (level) { // Sync barrier state for nested hot teams, not needed for outermost hot // team. for (f = 1; f < new_nproc; ++f) { kmp_info_t *thr = team->t.t_threads[f]; int b; kmp_balign_t *balign = thr->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); #if USE_DEBUGGER balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; #endif } } } #endif // KMP_NESTED_HOT_TEAMS /* reallocate space for arguments if necessary */ __kmp_alloc_argv_entries(argc, team, TRUE); KMP_CHECK_UPDATE(team->t.t_argc, argc); // The hot team re-uses the previous task team, // if untouched during the previous release->gather phase. KF_TRACE(10, (" hot_team = %p\n", team)); #if KMP_DEBUG if (__kmp_tasking_mode != tskm_immediate_exec) { KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0], team->t.t_task_team[1])); } #endif #if OMPT_SUPPORT __ompt_team_assign_id(team, ompt_parallel_data); #endif KMP_MB(); return team; } /* next, let's try to take one from the team pool */ KMP_MB(); for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) { /* TODO: consider resizing undersized teams instead of reaping them, now that we have a resizing mechanism */ if (team->t.t_max_nproc >= max_nproc) { /* take this team from the team pool */ __kmp_team_pool = team->t.t_next_pool; if (max_nproc > 1 && __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { if (!team->t.b) { // Allocate barrier structure team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); } } /* setup the team for fresh use */ __kmp_initialize_team(team, new_nproc, new_icvs, NULL); KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1])); team->t.t_task_team[0] = NULL; team->t.t_task_team[1] = NULL; /* reallocate space for arguments if necessary */ __kmp_alloc_argv_entries(argc, team, TRUE); KMP_CHECK_UPDATE(team->t.t_argc, argc); KA_TRACE( 20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n", team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); { // Initialize barrier data. int b; for (b = 0; b < bs_last_barrier; ++b) { team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE; #if USE_DEBUGGER team->t.t_bar[b].b_master_arrived = 0; team->t.t_bar[b].b_team_arrived = 0; #endif } } team->t.t_proc_bind = new_proc_bind; KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n", team->t.t_id)); #if OMPT_SUPPORT __ompt_team_assign_id(team, ompt_parallel_data); #endif KMP_MB(); return team; } /* reap team if it is too small, then loop back and check the next one */ // not sure if this is wise, but, will be redone during the hot-teams // rewrite. /* TODO: Use technique to find the right size hot-team, don't reap them */ team = __kmp_reap_team(team); __kmp_team_pool = team; } /* nothing available in the pool, no matter, make a new team! */ KMP_MB(); team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t)); /* and set it up */ team->t.t_max_nproc = max_nproc; if (max_nproc > 1 && __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { // Allocate barrier structure team->t.b = distributedBarrier::allocate(__kmp_dflt_team_nth_ub); } /* NOTE well, for some reason allocating one big buffer and dividing it up seems to really hurt performance a lot on the P4, so, let's not use this */ __kmp_allocate_team_arrays(team, max_nproc); KA_TRACE(20, ("__kmp_allocate_team: making a new team\n")); __kmp_initialize_team(team, new_nproc, new_icvs, NULL); KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1])); team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes // memory, no need to duplicate team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes // memory, no need to duplicate if (__kmp_storage_map) { __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc); } /* allocate space for arguments */ __kmp_alloc_argv_entries(argc, team, FALSE); team->t.t_argc = argc; KA_TRACE(20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n", team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); { // Initialize barrier data. int b; for (b = 0; b < bs_last_barrier; ++b) { team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE; #if USE_DEBUGGER team->t.t_bar[b].b_master_arrived = 0; team->t.t_bar[b].b_team_arrived = 0; #endif } } team->t.t_proc_bind = new_proc_bind; #if OMPT_SUPPORT __ompt_team_assign_id(team, ompt_parallel_data); team->t.ompt_serialized_team_info = NULL; #endif KMP_MB(); KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n", team->t.t_id)); return team; } /* TODO implement hot-teams at all levels */ /* TODO implement lazy thread release on demand (disband request) */ /* free the team. return it to the team pool. release all the threads * associated with it */ void __kmp_free_team(kmp_root_t *root, kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) { int f; KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(), team->t.t_id)); /* verify state */ KMP_DEBUG_ASSERT(root); KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc); KMP_DEBUG_ASSERT(team->t.t_threads); int use_hot_team = team == root->r.r_hot_team; #if KMP_NESTED_HOT_TEAMS int level; if (master) { level = team->t.t_active_level - 1; if (master->th.th_teams_microtask) { // in teams construct? if (master->th.th_teams_size.nteams > 1) { ++level; // level was not increased in teams construct for // team_of_masters } if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && master->th.th_teams_level == team->t.t_level) { ++level; // level was not increased in teams construct for // team_of_workers before the parallel } // team->t.t_level will be increased inside parallel } #if KMP_DEBUG kmp_hot_team_ptr_t *hot_teams = master->th.th_hot_teams; #endif if (level < __kmp_hot_teams_max_level) { KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team); use_hot_team = 1; } } #endif // KMP_NESTED_HOT_TEAMS /* team is done working */ TCW_SYNC_PTR(team->t.t_pkfn, NULL); // Important for Debugging Support Library. #if KMP_OS_WINDOWS team->t.t_copyin_counter = 0; // init counter for possible reuse #endif // Do not reset pointer to parent team to NULL for hot teams. /* if we are non-hot team, release our threads */ if (!use_hot_team) { if (__kmp_tasking_mode != tskm_immediate_exec) { // Wait for threads to reach reapable state for (f = 1; f < team->t.t_nproc; ++f) { KMP_DEBUG_ASSERT(team->t.t_threads[f]); kmp_info_t *th = team->t.t_threads[f]; volatile kmp_uint32 *state = &th->th.th_reap_state; while (*state != KMP_SAFE_TO_REAP) { #if KMP_OS_WINDOWS // On Windows a thread can be killed at any time, check this DWORD ecode; if (!__kmp_is_thread_alive(th, &ecode)) { *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread break; } #endif // first check if thread is sleeping kmp_flag_64<> fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th); if (fl.is_sleeping()) fl.resume(__kmp_gtid_from_thread(th)); KMP_CPU_PAUSE(); } } // Delete task teams int tt_idx; for (tt_idx = 0; tt_idx < 2; ++tt_idx) { kmp_task_team_t *task_team = team->t.t_task_team[tt_idx]; if (task_team != NULL) { for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams KMP_DEBUG_ASSERT(team->t.t_threads[f]); team->t.t_threads[f]->th.th_task_team = NULL; } KA_TRACE( 20, ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n", __kmp_get_gtid(), task_team, team->t.t_id)); #if KMP_NESTED_HOT_TEAMS __kmp_free_task_team(master, task_team); #endif team->t.t_task_team[tt_idx] = NULL; } } } // Reset pointer to parent team only for non-hot teams. team->t.t_parent = NULL; team->t.t_level = 0; team->t.t_active_level = 0; /* free the worker threads */ for (f = 1; f < team->t.t_nproc; ++f) { KMP_DEBUG_ASSERT(team->t.t_threads[f]); if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { KMP_COMPARE_AND_STORE_ACQ32(&(team->t.t_threads[f]->th.th_used_in_team), 1, 2); } __kmp_free_thread(team->t.t_threads[f]); } if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { if (team->t.b) { // wake up thread at old location team->t.b->go_release(); if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) { for (f = 1; f < team->t.t_nproc; ++f) { if (team->t.b->sleep[f].sleep) { __kmp_atomic_resume_64( team->t.t_threads[f]->th.th_info.ds.ds_gtid, (kmp_atomic_flag_64<> *)NULL); } } } // Wait for threads to be removed from team for (int f = 1; f < team->t.t_nproc; ++f) { while (team->t.t_threads[f]->th.th_used_in_team.load() != 0) KMP_CPU_PAUSE(); } } } for (f = 1; f < team->t.t_nproc; ++f) { team->t.t_threads[f] = NULL; } if (team->t.t_max_nproc > 1 && __kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { distributedBarrier::deallocate(team->t.b); team->t.b = NULL; } /* put the team back in the team pool */ /* TODO limit size of team pool, call reap_team if pool too large */ team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool); __kmp_team_pool = (volatile kmp_team_t *)team; } else { // Check if team was created for primary threads in teams construct // See if first worker is a CG root KMP_DEBUG_ASSERT(team->t.t_threads[1] && team->t.t_threads[1]->th.th_cg_roots); if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) { // Clean up the CG root nodes on workers so that this team can be re-used for (f = 1; f < team->t.t_nproc; ++f) { kmp_info_t *thr = team->t.t_threads[f]; KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots && thr->th.th_cg_roots->cg_root == thr); // Pop current CG root off list kmp_cg_root_t *tmp = thr->th.th_cg_roots; thr->th.th_cg_roots = tmp->up; KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving" " up to node %p. cg_nthreads was %d\n", thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads)); int i = tmp->cg_nthreads--; if (i == 1) { __kmp_free(tmp); // free CG if we are the last thread in it } // Restore current task's thread_limit from CG root if (thr->th.th_cg_roots) thr->th.th_current_task->td_icvs.thread_limit = thr->th.th_cg_roots->cg_thread_limit; } } } KMP_MB(); } /* reap the team. destroy it, reclaim all its resources and free its memory */ kmp_team_t *__kmp_reap_team(kmp_team_t *team) { kmp_team_t *next_pool = team->t.t_next_pool; KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(team->t.t_dispatch); KMP_DEBUG_ASSERT(team->t.t_disp_buffer); KMP_DEBUG_ASSERT(team->t.t_threads); KMP_DEBUG_ASSERT(team->t.t_argv); /* TODO clean the threads that are a part of this? */ /* free stuff */ __kmp_free_team_arrays(team); if (team->t.t_argv != &team->t.t_inline_argv[0]) __kmp_free((void *)team->t.t_argv); __kmp_free(team); KMP_MB(); return next_pool; } // Free the thread. Don't reap it, just place it on the pool of available // threads. // // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid // binding for the affinity mechanism to be useful. // // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid. // However, we want to avoid a potential performance problem by always // scanning through the list to find the correct point at which to insert // the thread (potential N**2 behavior). To do this we keep track of the // last place a thread struct was inserted (__kmp_thread_pool_insert_pt). // With single-level parallelism, threads will always be added to the tail // of the list, kept track of by __kmp_thread_pool_insert_pt. With nested // parallelism, all bets are off and we may need to scan through the entire // free list. // // This change also has a potentially large performance benefit, for some // applications. Previously, as threads were freed from the hot team, they // would be placed back on the free list in inverse order. If the hot team // grew back to it's original size, then the freed thread would be placed // back on the hot team in reverse order. This could cause bad cache // locality problems on programs where the size of the hot team regularly // grew and shrunk. // // Now, for single-level parallelism, the OMP tid is always == gtid. void __kmp_free_thread(kmp_info_t *this_th) { int gtid; kmp_info_t **scan; KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n", __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid)); KMP_DEBUG_ASSERT(this_th); // When moving thread to pool, switch thread to wait on own b_go flag, and // uninitialized (NULL team). int b; kmp_balign_t *balign = this_th->th.th_bar; for (b = 0; b < bs_last_barrier; ++b) { if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG; balign[b].bb.team = NULL; balign[b].bb.leaf_kids = 0; } this_th->th.th_task_state = 0; this_th->th.th_reap_state = KMP_SAFE_TO_REAP; /* put thread back on the free pool */ TCW_PTR(this_th->th.th_team, NULL); TCW_PTR(this_th->th.th_root, NULL); TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */ while (this_th->th.th_cg_roots) { this_th->th.th_cg_roots->cg_nthreads--; KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node" " %p of thread %p to %d\n", this_th, this_th->th.th_cg_roots, this_th->th.th_cg_roots->cg_root, this_th->th.th_cg_roots->cg_nthreads)); kmp_cg_root_t *tmp = this_th->th.th_cg_roots; if (tmp->cg_root == this_th) { // Thread is a cg_root KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0); KA_TRACE( 5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp)); this_th->th.th_cg_roots = tmp->up; __kmp_free(tmp); } else { // Worker thread if (tmp->cg_nthreads == 0) { // last thread leaves contention group __kmp_free(tmp); } this_th->th.th_cg_roots = NULL; break; } } /* If the implicit task assigned to this thread can be used by other threads * -> multiple threads can share the data and try to free the task at * __kmp_reap_thread at exit. This duplicate use of the task data can happen * with higher probability when hot team is disabled but can occurs even when * the hot team is enabled */ __kmp_free_implicit_task(this_th); this_th->th.th_current_task = NULL; // If the __kmp_thread_pool_insert_pt is already past the new insert // point, then we need to re-scan the entire list. gtid = this_th->th.th_info.ds.ds_gtid; if (__kmp_thread_pool_insert_pt != NULL) { KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL); if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) { __kmp_thread_pool_insert_pt = NULL; } } // Scan down the list to find the place to insert the thread. // scan is the address of a link in the list, possibly the address of // __kmp_thread_pool itself. // // In the absence of nested parallelism, the for loop will have 0 iterations. if (__kmp_thread_pool_insert_pt != NULL) { scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool); } else { scan = CCAST(kmp_info_t **, &__kmp_thread_pool); } for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid); scan = &((*scan)->th.th_next_pool)) ; // Insert the new element on the list, and set __kmp_thread_pool_insert_pt // to its address. TCW_PTR(this_th->th.th_next_pool, *scan); __kmp_thread_pool_insert_pt = *scan = this_th; KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)); TCW_4(this_th->th.th_in_pool, TRUE); __kmp_suspend_initialize_thread(this_th); __kmp_lock_suspend_mx(this_th); if (this_th->th.th_active == TRUE) { KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); this_th->th.th_active_in_pool = TRUE; } #if KMP_DEBUG else { KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE); } #endif __kmp_unlock_suspend_mx(this_th); TCW_4(__kmp_nth, __kmp_nth - 1); #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime back to user setting or default if necessary */ /* Middle initialization might never have occurred */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); if (__kmp_nth <= __kmp_avail_proc) { __kmp_zero_bt = FALSE; } } #endif /* KMP_ADJUST_BLOCKTIME */ KMP_MB(); } /* ------------------------------------------------------------------------ */ void *__kmp_launch_thread(kmp_info_t *this_thr) { #if OMP_PROFILING_SUPPORT ProfileTraceFile = getenv("LIBOMPTARGET_PROFILE"); // TODO: add a configuration option for time granularity if (ProfileTraceFile) llvm::timeTraceProfilerInitialize(500 /* us */, "libomptarget"); #endif int gtid = this_thr->th.th_info.ds.ds_gtid; /* void *stack_data;*/ kmp_team_t **volatile pteam; KMP_MB(); KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid)); if (__kmp_env_consistency_check) { this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak? } #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_thread_begin(); #endif #if OMPT_SUPPORT ompt_data_t *thread_data = nullptr; if (ompt_enabled.enabled) { thread_data = &(this_thr->th.ompt_thread_info.thread_data); *thread_data = ompt_data_none; this_thr->th.ompt_thread_info.state = ompt_state_overhead; this_thr->th.ompt_thread_info.wait_id = 0; this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0); this_thr->th.ompt_thread_info.parallel_flags = 0; if (ompt_enabled.ompt_callback_thread_begin) { ompt_callbacks.ompt_callback(ompt_callback_thread_begin)( ompt_thread_worker, thread_data); } this_thr->th.ompt_thread_info.state = ompt_state_idle; } #endif /* This is the place where threads wait for work */ while (!TCR_4(__kmp_global.g.g_done)) { KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]); KMP_MB(); /* wait for work to do */ KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid)); /* No tid yet since not part of a team */ __kmp_fork_barrier(gtid, KMP_GTID_DNE); #if OMPT_SUPPORT if (ompt_enabled.enabled) { this_thr->th.ompt_thread_info.state = ompt_state_overhead; } #endif pteam = &this_thr->th.th_team; /* have we been allocated? */ if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) { /* we were just woken up, so run our new task */ if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) { int rc; KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n", gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam)->t.t_pkfn)); updateHWFPControl(*pteam); #if OMPT_SUPPORT if (ompt_enabled.enabled) { this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; } #endif rc = (*pteam)->t.t_invoke(gtid); KMP_ASSERT(rc); KMP_MB(); KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n", gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam)->t.t_pkfn)); } #if OMPT_SUPPORT if (ompt_enabled.enabled) { /* no frame set while outside task */ __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none; this_thr->th.ompt_thread_info.state = ompt_state_overhead; } #endif /* join barrier after parallel region */ __kmp_join_barrier(gtid); } } #if OMPD_SUPPORT if (ompd_state & OMPD_ENABLE_BP) ompd_bp_thread_end(); #endif #if OMPT_SUPPORT if (ompt_enabled.ompt_callback_thread_end) { ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data); } #endif this_thr->th.th_task_team = NULL; /* run the destructors for the threadprivate data for this thread */ __kmp_common_destroy_gtid(gtid); KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid)); KMP_MB(); #if OMP_PROFILING_SUPPORT llvm::timeTraceProfilerFinishThread(); #endif return this_thr; } /* ------------------------------------------------------------------------ */ void __kmp_internal_end_dest(void *specific_gtid) { // Make sure no significant bits are lost int gtid; __kmp_type_convert((kmp_intptr_t)specific_gtid - 1, >id); KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid)); /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage * this is because 0 is reserved for the nothing-stored case */ __kmp_internal_end_thread(gtid); } #if KMP_OS_UNIX && KMP_DYNAMIC_LIB __attribute__((destructor)) void __kmp_internal_end_dtor(void) { __kmp_internal_end_atexit(); } #endif /* [Windows] josh: when the atexit handler is called, there may still be more than one thread alive */ void __kmp_internal_end_atexit(void) { KA_TRACE(30, ("__kmp_internal_end_atexit\n")); /* [Windows] josh: ideally, we want to completely shutdown the library in this atexit handler, but stat code that depends on thread specific data for gtid fails because that data becomes unavailable at some point during the shutdown, so we call __kmp_internal_end_thread instead. We should eventually remove the dependency on __kmp_get_specific_gtid in the stat code and use __kmp_internal_end_library to cleanly shutdown the library. // TODO: Can some of this comment about GVS be removed? I suspect that the offending stat code is executed when the calling thread tries to clean up a dead root thread's data structures, resulting in GVS code trying to close the GVS structures for that thread, but since the stat code uses __kmp_get_specific_gtid to get the gtid with the assumption that the calling thread is cleaning up itself instead of another thread, it get confused. This happens because allowing a thread to unregister and cleanup another thread is a recent modification for addressing an issue. Based on the current design (20050722), a thread may end up trying to unregister another thread only if thread death does not trigger the calling of __kmp_internal_end_thread. For Linux* OS, there is the thread specific data destructor function to detect thread death. For Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there is nothing. Thus, the workaround is applicable only for Windows static stat library. */ __kmp_internal_end_library(-1); #if KMP_OS_WINDOWS __kmp_close_console(); #endif } static void __kmp_reap_thread(kmp_info_t *thread, int is_root) { // It is assumed __kmp_forkjoin_lock is acquired. int gtid; KMP_DEBUG_ASSERT(thread != NULL); gtid = thread->th.th_info.ds.ds_gtid; if (!is_root) { if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) { /* Assume the threads are at the fork barrier here */ KA_TRACE( 20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n", gtid)); if (__kmp_barrier_gather_pattern[bs_forkjoin_barrier] == bp_dist_bar) { while ( !KMP_COMPARE_AND_STORE_ACQ32(&(thread->th.th_used_in_team), 0, 3)) KMP_CPU_PAUSE(); __kmp_resume_32(gtid, (kmp_flag_32<false, false> *)NULL); } else { /* Need release fence here to prevent seg faults for tree forkjoin barrier (GEH) */ kmp_flag_64<> flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread); __kmp_release_64(&flag); } } // Terminate OS thread. __kmp_reap_worker(thread); // The thread was killed asynchronously. If it was actively // spinning in the thread pool, decrement the global count. // // There is a small timing hole here - if the worker thread was just waking // up after sleeping in the pool, had reset it's th_active_in_pool flag but // not decremented the global counter __kmp_thread_pool_active_nth yet, then // the global counter might not get updated. // // Currently, this can only happen as the library is unloaded, // so there are no harmful side effects. if (thread->th.th_active_in_pool) { thread->th.th_active_in_pool = FALSE; KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0); } } __kmp_free_implicit_task(thread); // Free the fast memory for tasking #if USE_FAST_MEMORY __kmp_free_fast_memory(thread); #endif /* USE_FAST_MEMORY */ __kmp_suspend_uninitialize_thread(thread); KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread); TCW_SYNC_PTR(__kmp_threads[gtid], NULL); --__kmp_all_nth; // __kmp_nth was decremented when thread is added to the pool. #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime back to user setting or default if necessary */ /* Middle initialization might never have occurred */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); if (__kmp_nth <= __kmp_avail_proc) { __kmp_zero_bt = FALSE; } } #endif /* KMP_ADJUST_BLOCKTIME */ /* free the memory being used */ if (__kmp_env_consistency_check) { if (thread->th.th_cons) { __kmp_free_cons_stack(thread->th.th_cons); thread->th.th_cons = NULL; } } if (thread->th.th_pri_common != NULL) { __kmp_free(thread->th.th_pri_common); thread->th.th_pri_common = NULL; } if (thread->th.th_task_state_memo_stack != NULL) { __kmp_free(thread->th.th_task_state_memo_stack); thread->th.th_task_state_memo_stack = NULL; } #if KMP_USE_BGET if (thread->th.th_local.bget_data != NULL) { __kmp_finalize_bget(thread); } #endif #if KMP_AFFINITY_SUPPORTED if (thread->th.th_affin_mask != NULL) { KMP_CPU_FREE(thread->th.th_affin_mask); thread->th.th_affin_mask = NULL; } #endif /* KMP_AFFINITY_SUPPORTED */ #if KMP_USE_HIER_SCHED if (thread->th.th_hier_bar_data != NULL) { __kmp_free(thread->th.th_hier_bar_data); thread->th.th_hier_bar_data = NULL; } #endif __kmp_reap_team(thread->th.th_serial_team); thread->th.th_serial_team = NULL; __kmp_free(thread); KMP_MB(); } // __kmp_reap_thread static void __kmp_itthash_clean(kmp_info_t *th) { #if USE_ITT_NOTIFY if (__kmp_itt_region_domains.count > 0) { for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { kmp_itthash_entry_t *bucket = __kmp_itt_region_domains.buckets[i]; while (bucket) { kmp_itthash_entry_t *next = bucket->next_in_bucket; __kmp_thread_free(th, bucket); bucket = next; } } } if (__kmp_itt_barrier_domains.count > 0) { for (int i = 0; i < KMP_MAX_FRAME_DOMAINS; ++i) { kmp_itthash_entry_t *bucket = __kmp_itt_barrier_domains.buckets[i]; while (bucket) { kmp_itthash_entry_t *next = bucket->next_in_bucket; __kmp_thread_free(th, bucket); bucket = next; } } } #endif } static void __kmp_internal_end(void) { int i; /* First, unregister the library */ __kmp_unregister_library(); #if KMP_OS_WINDOWS /* In Win static library, we can't tell when a root actually dies, so we reclaim the data structures for any root threads that have died but not unregistered themselves, in order to shut down cleanly. In Win dynamic library we also can't tell when a thread dies. */ __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of // dead roots #endif for (i = 0; i < __kmp_threads_capacity; i++) if (__kmp_root[i]) if (__kmp_root[i]->r.r_active) break; KMP_MB(); /* Flush all pending memory write invalidates. */ TCW_SYNC_4(__kmp_global.g.g_done, TRUE); if (i < __kmp_threads_capacity) { #if KMP_USE_MONITOR // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor?? KMP_MB(); /* Flush all pending memory write invalidates. */ // Need to check that monitor was initialized before reaping it. If we are // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then // __kmp_monitor will appear to contain valid data, but it is only valid in // the parent process, not the child. // New behavior (201008): instead of keying off of the flag // __kmp_init_parallel, the monitor thread creation is keyed off // of the new flag __kmp_init_monitor. __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); if (TCR_4(__kmp_init_monitor)) { __kmp_reap_monitor(&__kmp_monitor); TCW_4(__kmp_init_monitor, 0); } __kmp_release_bootstrap_lock(&__kmp_monitor_lock); KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n")); #endif // KMP_USE_MONITOR } else { /* TODO move this to cleanup code */ #ifdef KMP_DEBUG /* make sure that everything has properly ended */ for (i = 0; i < __kmp_threads_capacity; i++) { if (__kmp_root[i]) { // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC: // there can be uber threads alive here KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active? } } #endif KMP_MB(); // Reap the worker threads. // This is valid for now, but be careful if threads are reaped sooner. while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool. // Get the next thread from the pool. kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool); __kmp_thread_pool = thread->th.th_next_pool; // Reap it. KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP); thread->th.th_next_pool = NULL; thread->th.th_in_pool = FALSE; __kmp_reap_thread(thread, 0); } __kmp_thread_pool_insert_pt = NULL; // Reap teams. while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool. // Get the next team from the pool. kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool); __kmp_team_pool = team->t.t_next_pool; // Reap it. team->t.t_next_pool = NULL; __kmp_reap_team(team); } __kmp_reap_task_teams(); #if KMP_OS_UNIX // Threads that are not reaped should not access any resources since they // are going to be deallocated soon, so the shutdown sequence should wait // until all threads either exit the final spin-waiting loop or begin // sleeping after the given blocktime. for (i = 0; i < __kmp_threads_capacity; i++) { kmp_info_t *thr = __kmp_threads[i]; while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking)) KMP_CPU_PAUSE(); } #endif for (i = 0; i < __kmp_threads_capacity; ++i) { // TBD: Add some checking... // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL ); } /* Make sure all threadprivate destructors get run by joining with all worker threads before resetting this flag */ TCW_SYNC_4(__kmp_init_common, FALSE); KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n")); KMP_MB(); #if KMP_USE_MONITOR // See note above: One of the possible fixes for CQ138434 / CQ140126 // // FIXME: push both code fragments down and CSE them? // push them into __kmp_cleanup() ? __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); if (TCR_4(__kmp_init_monitor)) { __kmp_reap_monitor(&__kmp_monitor); TCW_4(__kmp_init_monitor, 0); } __kmp_release_bootstrap_lock(&__kmp_monitor_lock); KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n")); #endif } /* else !__kmp_global.t_active */ TCW_4(__kmp_init_gtid, FALSE); KMP_MB(); /* Flush all pending memory write invalidates. */ __kmp_cleanup(); #if OMPT_SUPPORT ompt_fini(); #endif } void __kmp_internal_end_library(int gtid_req) { /* if we have already cleaned up, don't try again, it wouldn't be pretty */ /* this shouldn't be a race condition because __kmp_internal_end() is the only place to clear __kmp_serial_init */ /* we'll check this later too, after we get the lock */ // 2009-09-06: We do not set g_abort without setting g_done. This check looks // redundant, because the next check will work in any case. if (__kmp_global.g.g_abort) { KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n")); /* TODO abort? */ return; } if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { KA_TRACE(10, ("__kmp_internal_end_library: already finished\n")); return; } // If hidden helper team has been initialized, we need to deinit it if (TCR_4(__kmp_init_hidden_helper) && !TCR_4(__kmp_hidden_helper_team_done)) { TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE); // First release the main thread to let it continue its work __kmp_hidden_helper_main_thread_release(); // Wait until the hidden helper team has been destroyed __kmp_hidden_helper_threads_deinitz_wait(); } KMP_MB(); /* Flush all pending memory write invalidates. */ /* find out who we are and what we should do */ { int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); KA_TRACE( 10, ("__kmp_internal_end_library: enter T#%d (%d)\n", gtid, gtid_req)); if (gtid == KMP_GTID_SHUTDOWN) { KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n")); return; } else if (gtid == KMP_GTID_MONITOR) { KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not " "registered, or system shutdown\n")); return; } else if (gtid == KMP_GTID_DNE) { KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n")); /* we don't know who we are, but we may still shutdown the library */ } else if (KMP_UBER_GTID(gtid)) { /* unregister ourselves as an uber thread. gtid is no longer valid */ if (__kmp_root[gtid]->r.r_active) { __kmp_global.g.g_abort = -1; TCW_SYNC_4(__kmp_global.g.g_done, TRUE); __kmp_unregister_library(); KA_TRACE(10, ("__kmp_internal_end_library: root still active, abort T#%d\n", gtid)); return; } else { __kmp_itthash_clean(__kmp_threads[gtid]); KA_TRACE( 10, ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid)); __kmp_unregister_root_current_thread(gtid); } } else { /* worker threads may call this function through the atexit handler, if they * call exit() */ /* For now, skip the usual subsequent processing and just dump the debug buffer. TODO: do a thorough shutdown instead */ #ifdef DUMP_DEBUG_ON_EXIT if (__kmp_debug_buf) __kmp_dump_debug_buffer(); #endif // added unregister library call here when we switch to shm linux // if we don't, it will leave lots of files in /dev/shm // cleanup shared memory file before exiting. __kmp_unregister_library(); return; } } /* synchronize the termination process */ __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); /* have we already finished */ if (__kmp_global.g.g_abort) { KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n")); /* TODO abort? */ __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } /* We need this lock to enforce mutex between this reading of __kmp_threads_capacity and the writing by __kmp_register_root. Alternatively, we can use a counter of roots that is atomically updated by __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and __kmp_internal_end_*. */ __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); /* now we can safely conduct the actual termination */ __kmp_internal_end(); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); __kmp_release_bootstrap_lock(&__kmp_initz_lock); KA_TRACE(10, ("__kmp_internal_end_library: exit\n")); #ifdef DUMP_DEBUG_ON_EXIT if (__kmp_debug_buf) __kmp_dump_debug_buffer(); #endif #if KMP_OS_WINDOWS __kmp_close_console(); #endif __kmp_fini_allocator(); } // __kmp_internal_end_library void __kmp_internal_end_thread(int gtid_req) { int i; /* if we have already cleaned up, don't try again, it wouldn't be pretty */ /* this shouldn't be a race condition because __kmp_internal_end() is the * only place to clear __kmp_serial_init */ /* we'll check this later too, after we get the lock */ // 2009-09-06: We do not set g_abort without setting g_done. This check looks // redundant, because the next check will work in any case. if (__kmp_global.g.g_abort) { KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n")); /* TODO abort? */ return; } if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n")); return; } // If hidden helper team has been initialized, we need to deinit it if (TCR_4(__kmp_init_hidden_helper) && !TCR_4(__kmp_hidden_helper_team_done)) { TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE); // First release the main thread to let it continue its work __kmp_hidden_helper_main_thread_release(); // Wait until the hidden helper team has been destroyed __kmp_hidden_helper_threads_deinitz_wait(); } KMP_MB(); /* Flush all pending memory write invalidates. */ /* find out who we are and what we should do */ { int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); KA_TRACE(10, ("__kmp_internal_end_thread: enter T#%d (%d)\n", gtid, gtid_req)); if (gtid == KMP_GTID_SHUTDOWN) { KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n")); return; } else if (gtid == KMP_GTID_MONITOR) { KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not " "registered, or system shutdown\n")); return; } else if (gtid == KMP_GTID_DNE) { KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n")); return; /* we don't know who we are */ } else if (KMP_UBER_GTID(gtid)) { /* unregister ourselves as an uber thread. gtid is no longer valid */ if (__kmp_root[gtid]->r.r_active) { __kmp_global.g.g_abort = -1; TCW_SYNC_4(__kmp_global.g.g_done, TRUE); KA_TRACE(10, ("__kmp_internal_end_thread: root still active, abort T#%d\n", gtid)); return; } else { KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n", gtid)); __kmp_unregister_root_current_thread(gtid); } } else { /* just a worker thread, let's leave */ KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid)); if (gtid >= 0) { __kmp_threads[gtid]->th.th_task_team = NULL; } KA_TRACE(10, ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n", gtid)); return; } } #if KMP_DYNAMIC_LIB if (__kmp_pause_status != kmp_hard_paused) // AC: lets not shutdown the dynamic library at the exit of uber thread, // because we will better shutdown later in the library destructor. { KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req)); return; } #endif /* synchronize the termination process */ __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); /* have we already finished */ if (__kmp_global.g.g_abort) { KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n")); /* TODO abort? */ __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } /* We need this lock to enforce mutex between this reading of __kmp_threads_capacity and the writing by __kmp_register_root. Alternatively, we can use a counter of roots that is atomically updated by __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and __kmp_internal_end_*. */ /* should we finish the run-time? are all siblings done? */ __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); for (i = 0; i < __kmp_threads_capacity; ++i) { if (KMP_UBER_GTID(i)) { KA_TRACE( 10, ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i)); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } } /* now we can safely conduct the actual termination */ __kmp_internal_end(); __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); __kmp_release_bootstrap_lock(&__kmp_initz_lock); KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req)); #ifdef DUMP_DEBUG_ON_EXIT if (__kmp_debug_buf) __kmp_dump_debug_buffer(); #endif } // __kmp_internal_end_thread // ----------------------------------------------------------------------------- // Library registration stuff. static long __kmp_registration_flag = 0; // Random value used to indicate library initialization. static char *__kmp_registration_str = NULL; // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>. static inline char *__kmp_reg_status_name() { /* On RHEL 3u5 if linked statically, getpid() returns different values in each thread. If registration and unregistration go in different threads (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env env var can not be found, because the name will contain different pid. */ // macOS* complains about name being too long with additional getuid() #if KMP_OS_UNIX && !KMP_OS_DARWIN && KMP_DYNAMIC_LIB return __kmp_str_format("__KMP_REGISTERED_LIB_%d_%d", (int)getpid(), (int)getuid()); #else return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid()); #endif } // __kmp_reg_status_get #if defined(KMP_USE_SHM) // If /dev/shm is not accessible, we will create a temporary file under /tmp. char *temp_reg_status_file_name = nullptr; #endif void __kmp_register_library_startup(void) { char *name = __kmp_reg_status_name(); // Name of the environment variable. int done = 0; union { double dtime; long ltime; } time; #if KMP_ARCH_X86 || KMP_ARCH_X86_64 __kmp_initialize_system_tick(); #endif __kmp_read_system_time(&time.dtime); __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL); __kmp_registration_str = __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag, __kmp_registration_flag, KMP_LIBRARY_FILE); KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name, __kmp_registration_str)); while (!done) { char *value = NULL; // Actual value of the environment variable. #if defined(KMP_USE_SHM) char *shm_name = __kmp_str_format("/%s", name); int shm_preexist = 0; char *data1; int fd1 = shm_open(shm_name, O_CREAT | O_EXCL | O_RDWR, 0666); if ((fd1 == -1) && (errno == EEXIST)) { // file didn't open because it already exists. // try opening existing file fd1 = shm_open(shm_name, O_RDWR, 0666); if (fd1 == -1) { // file didn't open // error out here __kmp_fatal(KMP_MSG(FunctionError, "Can't open SHM"), KMP_ERR(0), __kmp_msg_null); } else { // able to open existing file shm_preexist = 1; } } else if (fd1 == -1) { // SHM didn't open; it was due to error other than already exists. Try to // create a temp file under /tmp. // TODO: /tmp might not always be the temporary directory. For now we will // not consider TMPDIR. If /tmp is not accessible, we simply error out. char *temp_file_name = __kmp_str_format("/tmp/%sXXXXXX", name); fd1 = mkstemp(temp_file_name); if (fd1 == -1) { // error out here. __kmp_fatal(KMP_MSG(FunctionError, "Can't open TEMP"), KMP_ERR(errno), __kmp_msg_null); } temp_reg_status_file_name = temp_file_name; } if (shm_preexist == 0) { // we created SHM now set size if (ftruncate(fd1, SHM_SIZE) == -1) { // error occured setting size; __kmp_fatal(KMP_MSG(FunctionError, "Can't set size of SHM"), KMP_ERR(errno), __kmp_msg_null); } } data1 = (char *)mmap(0, SHM_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, fd1, 0); if (data1 == MAP_FAILED) { // failed to map shared memory __kmp_fatal(KMP_MSG(FunctionError, "Can't map SHM"), KMP_ERR(errno), __kmp_msg_null); } if (shm_preexist == 0) { // set data to SHM, set value KMP_STRCPY_S(data1, SHM_SIZE, __kmp_registration_str); } // Read value from either what we just wrote or existing file. value = __kmp_str_format("%s", data1); // read value from SHM munmap(data1, SHM_SIZE); close(fd1); #else // Windows and unix with static library // Set environment variable, but do not overwrite if it is exist. __kmp_env_set(name, __kmp_registration_str, 0); // read value to see if it got set value = __kmp_env_get(name); #endif if (value != NULL && strcmp(value, __kmp_registration_str) == 0) { done = 1; // Ok, environment variable set successfully, exit the loop. } else { // Oops. Write failed. Another copy of OpenMP RTL is in memory. // Check whether it alive or dead. int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead. char *tail = value; char *flag_addr_str = NULL; char *flag_val_str = NULL; char const *file_name = NULL; __kmp_str_split(tail, '-', &flag_addr_str, &tail); __kmp_str_split(tail, '-', &flag_val_str, &tail); file_name = tail; if (tail != NULL) { unsigned long *flag_addr = 0; unsigned long flag_val = 0; KMP_SSCANF(flag_addr_str, "%p", RCAST(void **, &flag_addr)); KMP_SSCANF(flag_val_str, "%lx", &flag_val); if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) { // First, check whether environment-encoded address is mapped into // addr space. // If so, dereference it to see if it still has the right value. if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) { neighbor = 1; } else { // If not, then we know the other copy of the library is no longer // running. neighbor = 2; } } } switch (neighbor) { case 0: // Cannot parse environment variable -- neighbor status unknown. // Assume it is the incompatible format of future version of the // library. Assume the other library is alive. // WARN( ... ); // TODO: Issue a warning. file_name = "unknown library"; KMP_FALLTHROUGH(); // Attention! Falling to the next case. That's intentional. case 1: { // Neighbor is alive. // Check it is allowed. char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK"); if (!__kmp_str_match_true(duplicate_ok)) { // That's not allowed. Issue fatal error. __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name), KMP_HNT(DuplicateLibrary), __kmp_msg_null); } KMP_INTERNAL_FREE(duplicate_ok); __kmp_duplicate_library_ok = 1; done = 1; // Exit the loop. } break; case 2: { // Neighbor is dead. #if defined(KMP_USE_SHM) // close shared memory. shm_unlink(shm_name); // this removes file in /dev/shm #else // Clear the variable and try to register library again. __kmp_env_unset(name); #endif } break; default: { KMP_DEBUG_ASSERT(0); } break; } } KMP_INTERNAL_FREE((void *)value); #if defined(KMP_USE_SHM) KMP_INTERNAL_FREE((void *)shm_name); #endif } // while KMP_INTERNAL_FREE((void *)name); } // func __kmp_register_library_startup void __kmp_unregister_library(void) { char *name = __kmp_reg_status_name(); char *value = NULL; #if defined(KMP_USE_SHM) bool use_shm = true; char *shm_name = __kmp_str_format("/%s", name); int fd1 = shm_open(shm_name, O_RDONLY, 0666); if (fd1 == -1) { // File did not open. Try the temporary file. use_shm = false; KMP_DEBUG_ASSERT(temp_reg_status_file_name); fd1 = open(temp_reg_status_file_name, O_RDONLY); if (fd1 == -1) { // give it up now. return; } } char *data1 = (char *)mmap(0, SHM_SIZE, PROT_READ, MAP_SHARED, fd1, 0); if (data1 != MAP_FAILED) { value = __kmp_str_format("%s", data1); // read value from SHM munmap(data1, SHM_SIZE); } close(fd1); #else value = __kmp_env_get(name); #endif KMP_DEBUG_ASSERT(__kmp_registration_flag != 0); KMP_DEBUG_ASSERT(__kmp_registration_str != NULL); if (value != NULL && strcmp(value, __kmp_registration_str) == 0) { // Ok, this is our variable. Delete it. #if defined(KMP_USE_SHM) if (use_shm) { shm_unlink(shm_name); // this removes file in /dev/shm } else { KMP_DEBUG_ASSERT(temp_reg_status_file_name); unlink(temp_reg_status_file_name); // this removes the temp file } #else __kmp_env_unset(name); #endif } #if defined(KMP_USE_SHM) KMP_INTERNAL_FREE(shm_name); if (!use_shm) { KMP_DEBUG_ASSERT(temp_reg_status_file_name); KMP_INTERNAL_FREE(temp_reg_status_file_name); } #endif KMP_INTERNAL_FREE(__kmp_registration_str); KMP_INTERNAL_FREE(value); KMP_INTERNAL_FREE(name); __kmp_registration_flag = 0; __kmp_registration_str = NULL; } // __kmp_unregister_library // End of Library registration stuff. // ----------------------------------------------------------------------------- #if KMP_MIC_SUPPORTED static void __kmp_check_mic_type() { kmp_cpuid_t cpuid_state = {0}; kmp_cpuid_t *cs_p = &cpuid_state; __kmp_x86_cpuid(1, 0, cs_p); // We don't support mic1 at the moment if ((cs_p->eax & 0xff0) == 0xB10) { __kmp_mic_type = mic2; } else if ((cs_p->eax & 0xf0ff0) == 0x50670) { __kmp_mic_type = mic3; } else { __kmp_mic_type = non_mic; } } #endif /* KMP_MIC_SUPPORTED */ #if KMP_HAVE_UMWAIT static void __kmp_user_level_mwait_init() { struct kmp_cpuid buf; __kmp_x86_cpuid(7, 0, &buf); __kmp_waitpkg_enabled = ((buf.ecx >> 5) & 1); __kmp_umwait_enabled = __kmp_waitpkg_enabled && __kmp_user_level_mwait; __kmp_tpause_enabled = __kmp_waitpkg_enabled && (__kmp_tpause_state > 0); KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n", __kmp_umwait_enabled)); } #elif KMP_HAVE_MWAIT #ifndef AT_INTELPHIUSERMWAIT // Spurious, non-existent value that should always fail to return anything. // Will be replaced with the correct value when we know that. #define AT_INTELPHIUSERMWAIT 10000 #endif // getauxval() function is available in RHEL7 and SLES12. If a system with an // earlier OS is used to build the RTL, we'll use the following internal // function when the entry is not found. unsigned long getauxval(unsigned long) KMP_WEAK_ATTRIBUTE_EXTERNAL; unsigned long getauxval(unsigned long) { return 0; } static void __kmp_user_level_mwait_init() { // When getauxval() and correct value of AT_INTELPHIUSERMWAIT are available // use them to find if the user-level mwait is enabled. Otherwise, forcibly // set __kmp_mwait_enabled=TRUE on Intel MIC if the environment variable // KMP_USER_LEVEL_MWAIT was set to TRUE. if (__kmp_mic_type == mic3) { unsigned long res = getauxval(AT_INTELPHIUSERMWAIT); if ((res & 0x1) || __kmp_user_level_mwait) { __kmp_mwait_enabled = TRUE; if (__kmp_user_level_mwait) { KMP_INFORM(EnvMwaitWarn); } } else { __kmp_mwait_enabled = FALSE; } } KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, " "__kmp_mwait_enabled = %d\n", __kmp_mic_type, __kmp_mwait_enabled)); } #endif /* KMP_HAVE_UMWAIT */ static void __kmp_do_serial_initialize(void) { int i, gtid; size_t size; KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n")); KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4); KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4); KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8); KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8); KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *)); #if OMPT_SUPPORT ompt_pre_init(); #endif #if OMPD_SUPPORT __kmp_env_dump(); ompd_init(); #endif __kmp_validate_locks(); #if ENABLE_LIBOMPTARGET /* Initialize functions from libomptarget */ __kmp_init_omptarget(); #endif /* Initialize internal memory allocator */ __kmp_init_allocator(); /* Register the library startup via an environment variable or via mapped shared memory file and check to see whether another copy of the library is already registered. Since forked child process is often terminated, we postpone the registration till middle initialization in the child */ if (__kmp_need_register_serial) __kmp_register_library_startup(); /* TODO reinitialization of library */ if (TCR_4(__kmp_global.g.g_done)) { KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n")); } __kmp_global.g.g_abort = 0; TCW_SYNC_4(__kmp_global.g.g_done, FALSE); /* initialize the locks */ #if KMP_USE_ADAPTIVE_LOCKS #if KMP_DEBUG_ADAPTIVE_LOCKS __kmp_init_speculative_stats(); #endif #endif #if KMP_STATS_ENABLED __kmp_stats_init(); #endif __kmp_init_lock(&__kmp_global_lock); __kmp_init_queuing_lock(&__kmp_dispatch_lock); __kmp_init_lock(&__kmp_debug_lock); __kmp_init_atomic_lock(&__kmp_atomic_lock); __kmp_init_atomic_lock(&__kmp_atomic_lock_1i); __kmp_init_atomic_lock(&__kmp_atomic_lock_2i); __kmp_init_atomic_lock(&__kmp_atomic_lock_4i); __kmp_init_atomic_lock(&__kmp_atomic_lock_4r); __kmp_init_atomic_lock(&__kmp_atomic_lock_8i); __kmp_init_atomic_lock(&__kmp_atomic_lock_8r); __kmp_init_atomic_lock(&__kmp_atomic_lock_8c); __kmp_init_atomic_lock(&__kmp_atomic_lock_10r); __kmp_init_atomic_lock(&__kmp_atomic_lock_16r); __kmp_init_atomic_lock(&__kmp_atomic_lock_16c); __kmp_init_atomic_lock(&__kmp_atomic_lock_20c); __kmp_init_atomic_lock(&__kmp_atomic_lock_32c); __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock); __kmp_init_bootstrap_lock(&__kmp_exit_lock); #if KMP_USE_MONITOR __kmp_init_bootstrap_lock(&__kmp_monitor_lock); #endif __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock); /* conduct initialization and initial setup of configuration */ __kmp_runtime_initialize(); #if KMP_MIC_SUPPORTED __kmp_check_mic_type(); #endif // Some global variable initialization moved here from kmp_env_initialize() #ifdef KMP_DEBUG kmp_diag = 0; #endif __kmp_abort_delay = 0; // From __kmp_init_dflt_team_nth() /* assume the entire machine will be used */ __kmp_dflt_team_nth_ub = __kmp_xproc; if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) { __kmp_dflt_team_nth_ub = KMP_MIN_NTH; } if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) { __kmp_dflt_team_nth_ub = __kmp_sys_max_nth; } __kmp_max_nth = __kmp_sys_max_nth; __kmp_cg_max_nth = __kmp_sys_max_nth; __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default if (__kmp_teams_max_nth > __kmp_sys_max_nth) { __kmp_teams_max_nth = __kmp_sys_max_nth; } // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME" // part __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME; #if KMP_USE_MONITOR __kmp_monitor_wakeups = KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); __kmp_bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); #endif // From "KMP_LIBRARY" part of __kmp_env_initialize() __kmp_library = library_throughput; // From KMP_SCHEDULE initialization __kmp_static = kmp_sch_static_balanced; // AC: do not use analytical here, because it is non-monotonous //__kmp_guided = kmp_sch_guided_iterative_chunked; //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no // need to repeat assignment // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch // bit control and barrier method control parts #if KMP_FAST_REDUCTION_BARRIER #define kmp_reduction_barrier_gather_bb ((int)1) #define kmp_reduction_barrier_release_bb ((int)1) #define kmp_reduction_barrier_gather_pat __kmp_barrier_gather_pat_dflt #define kmp_reduction_barrier_release_pat __kmp_barrier_release_pat_dflt #endif // KMP_FAST_REDUCTION_BARRIER for (i = bs_plain_barrier; i < bs_last_barrier; i++) { __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt; __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt; __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt; __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt; #if KMP_FAST_REDUCTION_BARRIER if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only ( // lin_64 ): hyper,1 __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb; __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb; __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat; __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat; } #endif // KMP_FAST_REDUCTION_BARRIER } #if KMP_FAST_REDUCTION_BARRIER #undef kmp_reduction_barrier_release_pat #undef kmp_reduction_barrier_gather_pat #undef kmp_reduction_barrier_release_bb #undef kmp_reduction_barrier_gather_bb #endif // KMP_FAST_REDUCTION_BARRIER #if KMP_MIC_SUPPORTED if (__kmp_mic_type == mic2) { // KNC // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] = 1; // forkjoin release __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; } #if KMP_FAST_REDUCTION_BARRIER if (__kmp_mic_type == mic2) { // KNC __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar; __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar; } #endif // KMP_FAST_REDUCTION_BARRIER #endif // KMP_MIC_SUPPORTED // From KMP_CHECKS initialization #ifdef KMP_DEBUG __kmp_env_checks = TRUE; /* development versions have the extra checks */ #else __kmp_env_checks = FALSE; /* port versions do not have the extra checks */ #endif // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization __kmp_foreign_tp = TRUE; __kmp_global.g.g_dynamic = FALSE; __kmp_global.g.g_dynamic_mode = dynamic_default; __kmp_init_nesting_mode(); __kmp_env_initialize(NULL); #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT __kmp_user_level_mwait_init(); #endif // Print all messages in message catalog for testing purposes. #ifdef KMP_DEBUG char const *val = __kmp_env_get("KMP_DUMP_CATALOG"); if (__kmp_str_match_true(val)) { kmp_str_buf_t buffer; __kmp_str_buf_init(&buffer); __kmp_i18n_dump_catalog(&buffer); __kmp_printf("%s", buffer.str); __kmp_str_buf_free(&buffer); } __kmp_env_free(&val); #endif __kmp_threads_capacity = __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub); // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part __kmp_tp_capacity = __kmp_default_tp_capacity( __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified); // If the library is shut down properly, both pools must be NULL. Just in // case, set them to NULL -- some memory may leak, but subsequent code will // work even if pools are not freed. KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL); KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL); KMP_DEBUG_ASSERT(__kmp_team_pool == NULL); __kmp_thread_pool = NULL; __kmp_thread_pool_insert_pt = NULL; __kmp_team_pool = NULL; /* Allocate all of the variable sized records */ /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are * expandable */ /* Since allocation is cache-aligned, just add extra padding at the end */ size = (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity + CACHE_LINE; __kmp_threads = (kmp_info_t **)__kmp_allocate(size); __kmp_root = (kmp_root_t **)((char *)__kmp_threads + sizeof(kmp_info_t *) * __kmp_threads_capacity); /* init thread counts */ KMP_DEBUG_ASSERT(__kmp_all_nth == 0); // Asserts fail if the library is reinitializing and KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination. __kmp_all_nth = 0; __kmp_nth = 0; /* setup the uber master thread and hierarchy */ gtid = __kmp_register_root(TRUE); KA_TRACE(10, ("__kmp_do_serial_initialize T#%d\n", gtid)); KMP_ASSERT(KMP_UBER_GTID(gtid)); KMP_ASSERT(KMP_INITIAL_GTID(gtid)); KMP_MB(); /* Flush all pending memory write invalidates. */ __kmp_common_initialize(); #if KMP_OS_UNIX /* invoke the child fork handler */ __kmp_register_atfork(); #endif #if !KMP_DYNAMIC_LIB || \ ((KMP_COMPILER_ICC || KMP_COMPILER_ICX) && KMP_OS_DARWIN) { /* Invoke the exit handler when the program finishes, only for static library and macOS* dynamic. For other dynamic libraries, we already have _fini and DllMain. */ int rc = atexit(__kmp_internal_end_atexit); if (rc != 0) { __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc), __kmp_msg_null); } } #endif #if KMP_HANDLE_SIGNALS #if KMP_OS_UNIX /* NOTE: make sure that this is called before the user installs their own signal handlers so that the user handlers are called first. this way they can return false, not call our handler, avoid terminating the library, and continue execution where they left off. */ __kmp_install_signals(FALSE); #endif /* KMP_OS_UNIX */ #if KMP_OS_WINDOWS __kmp_install_signals(TRUE); #endif /* KMP_OS_WINDOWS */ #endif /* we have finished the serial initialization */ __kmp_init_counter++; __kmp_init_serial = TRUE; if (__kmp_version) { __kmp_print_version_1(); } if (__kmp_settings) { __kmp_env_print(); } if (__kmp_display_env || __kmp_display_env_verbose) { __kmp_env_print_2(); } #if OMPT_SUPPORT ompt_post_init(); #endif KMP_MB(); KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n")); } void __kmp_serial_initialize(void) { if (__kmp_init_serial) { return; } __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (__kmp_init_serial) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } __kmp_do_serial_initialize(); __kmp_release_bootstrap_lock(&__kmp_initz_lock); } static void __kmp_do_middle_initialize(void) { int i, j; int prev_dflt_team_nth; if (!__kmp_init_serial) { __kmp_do_serial_initialize(); } KA_TRACE(10, ("__kmp_middle_initialize: enter\n")); if (UNLIKELY(!__kmp_need_register_serial)) { // We are in a forked child process. The registration was skipped during // serial initialization in __kmp_atfork_child handler. Do it here. __kmp_register_library_startup(); } // Save the previous value for the __kmp_dflt_team_nth so that // we can avoid some reinitialization if it hasn't changed. prev_dflt_team_nth = __kmp_dflt_team_nth; #if KMP_AFFINITY_SUPPORTED // __kmp_affinity_initialize() will try to set __kmp_ncores to the // number of cores on the machine. __kmp_affinity_initialize(__kmp_affinity); #endif /* KMP_AFFINITY_SUPPORTED */ KMP_ASSERT(__kmp_xproc > 0); if (__kmp_avail_proc == 0) { __kmp_avail_proc = __kmp_xproc; } // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3), // correct them now j = 0; while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) { __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub = __kmp_avail_proc; j++; } if (__kmp_dflt_team_nth == 0) { #ifdef KMP_DFLT_NTH_CORES // Default #threads = #cores __kmp_dflt_team_nth = __kmp_ncores; KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth)); #else // Default #threads = #available OS procs __kmp_dflt_team_nth = __kmp_avail_proc; KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth)); #endif /* KMP_DFLT_NTH_CORES */ } if (__kmp_dflt_team_nth < KMP_MIN_NTH) { __kmp_dflt_team_nth = KMP_MIN_NTH; } if (__kmp_dflt_team_nth > __kmp_sys_max_nth) { __kmp_dflt_team_nth = __kmp_sys_max_nth; } if (__kmp_nesting_mode > 0) __kmp_set_nesting_mode_threads(); // There's no harm in continuing if the following check fails, // but it indicates an error in the previous logic. KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub); if (__kmp_dflt_team_nth != prev_dflt_team_nth) { // Run through the __kmp_threads array and set the num threads icv for each // root thread that is currently registered with the RTL (which has not // already explicitly set its nthreads-var with a call to // omp_set_num_threads()). for (i = 0; i < __kmp_threads_capacity; i++) { kmp_info_t *thread = __kmp_threads[i]; if (thread == NULL) continue; if (thread->th.th_current_task->td_icvs.nproc != 0) continue; set__nproc(__kmp_threads[i], __kmp_dflt_team_nth); } } KA_TRACE( 20, ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n", __kmp_dflt_team_nth)); #ifdef KMP_ADJUST_BLOCKTIME /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */ if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); if (__kmp_nth > __kmp_avail_proc) { __kmp_zero_bt = TRUE; } } #endif /* KMP_ADJUST_BLOCKTIME */ /* we have finished middle initialization */ TCW_SYNC_4(__kmp_init_middle, TRUE); KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n")); } void __kmp_middle_initialize(void) { if (__kmp_init_middle) { return; } __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (__kmp_init_middle) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } __kmp_do_middle_initialize(); __kmp_release_bootstrap_lock(&__kmp_initz_lock); } void __kmp_parallel_initialize(void) { int gtid = __kmp_entry_gtid(); // this might be a new root /* synchronize parallel initialization (for sibling) */ if (TCR_4(__kmp_init_parallel)) return; __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (TCR_4(__kmp_init_parallel)) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } /* TODO reinitialization after we have already shut down */ if (TCR_4(__kmp_global.g.g_done)) { KA_TRACE( 10, ("__kmp_parallel_initialize: attempt to init while shutting down\n")); __kmp_infinite_loop(); } /* jc: The lock __kmp_initz_lock is already held, so calling __kmp_serial_initialize would cause a deadlock. So we call __kmp_do_serial_initialize directly. */ if (!__kmp_init_middle) { __kmp_do_middle_initialize(); } __kmp_assign_root_init_mask(); __kmp_resume_if_hard_paused(); /* begin initialization */ KA_TRACE(10, ("__kmp_parallel_initialize: enter\n")); KMP_ASSERT(KMP_UBER_GTID(gtid)); #if KMP_ARCH_X86 || KMP_ARCH_X86_64 // Save the FP control regs. // Worker threads will set theirs to these values at thread startup. __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); __kmp_store_mxcsr(&__kmp_init_mxcsr); __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK; #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ #if KMP_OS_UNIX #if KMP_HANDLE_SIGNALS /* must be after __kmp_serial_initialize */ __kmp_install_signals(TRUE); #endif #endif __kmp_suspend_initialize(); #if defined(USE_LOAD_BALANCE) if (__kmp_global.g.g_dynamic_mode == dynamic_default) { __kmp_global.g.g_dynamic_mode = dynamic_load_balance; } #else if (__kmp_global.g.g_dynamic_mode == dynamic_default) { __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; } #endif if (__kmp_version) { __kmp_print_version_2(); } /* we have finished parallel initialization */ TCW_SYNC_4(__kmp_init_parallel, TRUE); KMP_MB(); KA_TRACE(10, ("__kmp_parallel_initialize: exit\n")); __kmp_release_bootstrap_lock(&__kmp_initz_lock); } void __kmp_hidden_helper_initialize() { if (TCR_4(__kmp_init_hidden_helper)) return; // __kmp_parallel_initialize is required before we initialize hidden helper if (!TCR_4(__kmp_init_parallel)) __kmp_parallel_initialize(); // Double check. Note that this double check should not be placed before // __kmp_parallel_initialize as it will cause dead lock. __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (TCR_4(__kmp_init_hidden_helper)) { __kmp_release_bootstrap_lock(&__kmp_initz_lock); return; } #if KMP_AFFINITY_SUPPORTED // Initialize hidden helper affinity settings. // The above __kmp_parallel_initialize() will initialize // regular affinity (and topology) if not already done. if (!__kmp_hh_affinity.flags.initialized) __kmp_affinity_initialize(__kmp_hh_affinity); #endif // Set the count of hidden helper tasks to be executed to zero KMP_ATOMIC_ST_REL(&__kmp_unexecuted_hidden_helper_tasks, 0); // Set the global variable indicating that we're initializing hidden helper // team/threads TCW_SYNC_4(__kmp_init_hidden_helper_threads, TRUE); // Platform independent initialization __kmp_do_initialize_hidden_helper_threads(); // Wait here for the finish of initialization of hidden helper teams __kmp_hidden_helper_threads_initz_wait(); // We have finished hidden helper initialization TCW_SYNC_4(__kmp_init_hidden_helper, TRUE); __kmp_release_bootstrap_lock(&__kmp_initz_lock); } /* ------------------------------------------------------------------------ */ void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr, kmp_team_t *team) { kmp_disp_t *dispatch; KMP_MB(); /* none of the threads have encountered any constructs, yet. */ this_thr->th.th_local.this_construct = 0; #if KMP_CACHE_MANAGE KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived); #endif /* KMP_CACHE_MANAGE */ dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch); KMP_DEBUG_ASSERT(dispatch); KMP_DEBUG_ASSERT(team->t.t_dispatch); // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[ // this_thr->th.th_info.ds.ds_tid ] ); dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */ dispatch->th_doacross_buf_idx = 0; // reset doacross dispatch buffer counter if (__kmp_env_consistency_check) __kmp_push_parallel(gtid, team->t.t_ident); KMP_MB(); /* Flush all pending memory write invalidates. */ } void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr, kmp_team_t *team) { if (__kmp_env_consistency_check) __kmp_pop_parallel(gtid, team->t.t_ident); __kmp_finish_implicit_task(this_thr); } int __kmp_invoke_task_func(int gtid) { int rc; int tid = __kmp_tid_from_gtid(gtid); kmp_info_t *this_thr = __kmp_threads[gtid]; kmp_team_t *team = this_thr->th.th_team; __kmp_run_before_invoked_task(gtid, tid, this_thr, team); #if USE_ITT_BUILD if (__itt_stack_caller_create_ptr) { // inform ittnotify about entering user's code if (team->t.t_stack_id != NULL) { __kmp_itt_stack_callee_enter((__itt_caller)team->t.t_stack_id); } else { KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL); __kmp_itt_stack_callee_enter( (__itt_caller)team->t.t_parent->t.t_stack_id); } } #endif /* USE_ITT_BUILD */ #if INCLUDE_SSC_MARKS SSC_MARK_INVOKING(); #endif #if OMPT_SUPPORT void *dummy; void **exit_frame_p; ompt_data_t *my_task_data; ompt_data_t *my_parallel_data; int ompt_team_size; if (ompt_enabled.enabled) { exit_frame_p = &(team->t.t_implicit_task_taskdata[tid] .ompt_task_info.frame.exit_frame.ptr); } else { exit_frame_p = &dummy; } my_task_data = &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data); my_parallel_data = &(team->t.ompt_team_info.parallel_data); if (ompt_enabled.ompt_callback_implicit_task) { ompt_team_size = team->t.t_nproc; ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size, __kmp_tid_from_gtid(gtid), ompt_task_implicit); OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid); } #endif #if KMP_STATS_ENABLED stats_state_e previous_state = KMP_GET_THREAD_STATE(); if (previous_state == stats_state_e::TEAMS_REGION) { KMP_PUSH_PARTITIONED_TIMER(OMP_teams); } else { KMP_PUSH_PARTITIONED_TIMER(OMP_parallel); } KMP_SET_THREAD_STATE(IMPLICIT_TASK); #endif rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid, tid, (int)team->t.t_argc, (void **)team->t.t_argv #if OMPT_SUPPORT , exit_frame_p #endif ); #if OMPT_SUPPORT *exit_frame_p = NULL; this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_team; #endif #if KMP_STATS_ENABLED if (previous_state == stats_state_e::TEAMS_REGION) { KMP_SET_THREAD_STATE(previous_state); } KMP_POP_PARTITIONED_TIMER(); #endif #if USE_ITT_BUILD if (__itt_stack_caller_create_ptr) { // inform ittnotify about leaving user's code if (team->t.t_stack_id != NULL) { __kmp_itt_stack_callee_leave((__itt_caller)team->t.t_stack_id); } else { KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL); __kmp_itt_stack_callee_leave( (__itt_caller)team->t.t_parent->t.t_stack_id); } } #endif /* USE_ITT_BUILD */ __kmp_run_after_invoked_task(gtid, tid, this_thr, team); return rc; } void __kmp_teams_master(int gtid) { // This routine is called by all primary threads in teams construct kmp_info_t *thr = __kmp_threads[gtid]; kmp_team_t *team = thr->th.th_team; ident_t *loc = team->t.t_ident; thr->th.th_set_nproc = thr->th.th_teams_size.nth; KMP_DEBUG_ASSERT(thr->th.th_teams_microtask); KMP_DEBUG_ASSERT(thr->th.th_set_nproc); KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid, __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask)); // This thread is a new CG root. Set up the proper variables. kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t)); tmp->cg_root = thr; // Make thr the CG root // Init to thread limit stored when league primary threads were forked tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit; tmp->cg_nthreads = 1; // Init counter to one active thread, this one KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init" " cg_nthreads to 1\n", thr, tmp)); tmp->up = thr->th.th_cg_roots; thr->th.th_cg_roots = tmp; // Launch league of teams now, but not let workers execute // (they hang on fork barrier until next parallel) #if INCLUDE_SSC_MARKS SSC_MARK_FORKING(); #endif __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc, (microtask_t)thr->th.th_teams_microtask, // "wrapped" task VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL); #if INCLUDE_SSC_MARKS SSC_MARK_JOINING(); #endif // If the team size was reduced from the limit, set it to the new size if (thr->th.th_team_nproc < thr->th.th_teams_size.nth) thr->th.th_teams_size.nth = thr->th.th_team_nproc; // AC: last parameter "1" eliminates join barrier which won't work because // worker threads are in a fork barrier waiting for more parallel regions __kmp_join_call(loc, gtid #if OMPT_SUPPORT , fork_context_intel #endif , 1); } int __kmp_invoke_teams_master(int gtid) { kmp_info_t *this_thr = __kmp_threads[gtid]; kmp_team_t *team = this_thr->th.th_team; #if KMP_DEBUG if (!__kmp_threads[gtid]->th.th_team->t.t_serialized) KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master); #endif __kmp_run_before_invoked_task(gtid, 0, this_thr, team); #if OMPT_SUPPORT int tid = __kmp_tid_from_gtid(gtid); ompt_data_t *task_data = &team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data; ompt_data_t *parallel_data = &team->t.ompt_team_info.parallel_data; if (ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_begin, parallel_data, task_data, team->t.t_nproc, tid, ompt_task_initial); OMPT_CUR_TASK_INFO(this_thr)->thread_num = tid; } #endif __kmp_teams_master(gtid); #if OMPT_SUPPORT this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_league; #endif __kmp_run_after_invoked_task(gtid, 0, this_thr, team); return 1; } /* this sets the requested number of threads for the next parallel region encountered by this team. since this should be enclosed in the forkjoin critical section it should avoid race conditions with asymmetrical nested parallelism */ void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) { kmp_info_t *thr = __kmp_threads[gtid]; if (num_threads > 0) thr->th.th_set_nproc = num_threads; } static void __kmp_push_thread_limit(kmp_info_t *thr, int num_teams, int num_threads) { KMP_DEBUG_ASSERT(thr); // Remember the number of threads for inner parallel regions if (!TCR_4(__kmp_init_middle)) __kmp_middle_initialize(); // get internal globals calculated __kmp_assign_root_init_mask(); KMP_DEBUG_ASSERT(__kmp_avail_proc); KMP_DEBUG_ASSERT(__kmp_dflt_team_nth); if (num_threads == 0) { if (__kmp_teams_thread_limit > 0) { num_threads = __kmp_teams_thread_limit; } else { num_threads = __kmp_avail_proc / num_teams; } // adjust num_threads w/o warning as it is not user setting // num_threads = min(num_threads, nthreads-var, thread-limit-var) // no thread_limit clause specified - do not change thread-limit-var ICV if (num_threads > __kmp_dflt_team_nth) { num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV } if (num_threads > thr->th.th_current_task->td_icvs.thread_limit) { num_threads = thr->th.th_current_task->td_icvs.thread_limit; } // prevent team size to exceed thread-limit-var if (num_teams * num_threads > __kmp_teams_max_nth) { num_threads = __kmp_teams_max_nth / num_teams; } if (num_threads == 0) { num_threads = 1; } } else { if (num_threads < 0) { __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_threads, 1), __kmp_msg_null); num_threads = 1; } // This thread will be the primary thread of the league primary threads // Store new thread limit; old limit is saved in th_cg_roots list thr->th.th_current_task->td_icvs.thread_limit = num_threads; // num_threads = min(num_threads, nthreads-var) if (num_threads > __kmp_dflt_team_nth) { num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV } if (num_teams * num_threads > __kmp_teams_max_nth) { int new_threads = __kmp_teams_max_nth / num_teams; if (new_threads == 0) { new_threads = 1; } if (new_threads != num_threads) { if (!__kmp_reserve_warn) { // user asked for too many threads __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_threads, new_threads), KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); } } num_threads = new_threads; } } thr->th.th_teams_size.nth = num_threads; } /* this sets the requested number of teams for the teams region and/or the number of threads for the next parallel region encountered */ void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams, int num_threads) { kmp_info_t *thr = __kmp_threads[gtid]; if (num_teams < 0) { // OpenMP specification requires requested values to be positive, // but people can send us any value, so we'd better check __kmp_msg(kmp_ms_warning, KMP_MSG(NumTeamsNotPositive, num_teams, 1), __kmp_msg_null); num_teams = 1; } if (num_teams == 0) { if (__kmp_nteams > 0) { num_teams = __kmp_nteams; } else { num_teams = 1; // default number of teams is 1. } } if (num_teams > __kmp_teams_max_nth) { // if too many teams requested? if (!__kmp_reserve_warn) { __kmp_reserve_warn = 1; __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth), KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); } num_teams = __kmp_teams_max_nth; } // Set number of teams (number of threads in the outer "parallel" of the // teams) thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; __kmp_push_thread_limit(thr, num_teams, num_threads); } /* This sets the requested number of teams for the teams region and/or the number of threads for the next parallel region encountered */ void __kmp_push_num_teams_51(ident_t *id, int gtid, int num_teams_lb, int num_teams_ub, int num_threads) { kmp_info_t *thr = __kmp_threads[gtid]; KMP_DEBUG_ASSERT(num_teams_lb >= 0 && num_teams_ub >= 0); KMP_DEBUG_ASSERT(num_teams_ub >= num_teams_lb); KMP_DEBUG_ASSERT(num_threads >= 0); if (num_teams_lb > num_teams_ub) { __kmp_fatal(KMP_MSG(FailedToCreateTeam, num_teams_lb, num_teams_ub), KMP_HNT(SetNewBound, __kmp_teams_max_nth), __kmp_msg_null); } int num_teams = 1; // defalt number of teams is 1. if (num_teams_lb == 0 && num_teams_ub > 0) num_teams_lb = num_teams_ub; if (num_teams_lb == 0 && num_teams_ub == 0) { // no num_teams clause num_teams = (__kmp_nteams > 0) ? __kmp_nteams : num_teams; if (num_teams > __kmp_teams_max_nth) { if (!__kmp_reserve_warn) { __kmp_reserve_warn = 1; __kmp_msg(kmp_ms_warning, KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth), KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); } num_teams = __kmp_teams_max_nth; } } else if (num_teams_lb == num_teams_ub) { // requires exact number of teams num_teams = num_teams_ub; } else { // num_teams_lb <= num_teams <= num_teams_ub if (num_threads <= 0) { if (num_teams_ub > __kmp_teams_max_nth) { num_teams = num_teams_lb; } else { num_teams = num_teams_ub; } } else { num_teams = (num_threads > __kmp_teams_max_nth) ? num_teams : __kmp_teams_max_nth / num_threads; if (num_teams < num_teams_lb) { num_teams = num_teams_lb; } else if (num_teams > num_teams_ub) { num_teams = num_teams_ub; } } } // Set number of teams (number of threads in the outer "parallel" of the // teams) thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; __kmp_push_thread_limit(thr, num_teams, num_threads); } // Set the proc_bind var to use in the following parallel region. void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) { kmp_info_t *thr = __kmp_threads[gtid]; thr->th.th_set_proc_bind = proc_bind; } /* Launch the worker threads into the microtask. */ void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) { kmp_info_t *this_thr = __kmp_threads[gtid]; #ifdef KMP_DEBUG int f; #endif /* KMP_DEBUG */ KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(this_thr->th.th_team == team); KMP_ASSERT(KMP_MASTER_GTID(gtid)); KMP_MB(); /* Flush all pending memory write invalidates. */ team->t.t_construct = 0; /* no single directives seen yet */ team->t.t_ordered.dt.t_value = 0; /* thread 0 enters the ordered section first */ /* Reset the identifiers on the dispatch buffer */ KMP_DEBUG_ASSERT(team->t.t_disp_buffer); if (team->t.t_max_nproc > 1) { int i; for (i = 0; i < __kmp_dispatch_num_buffers; ++i) { team->t.t_disp_buffer[i].buffer_index = i; team->t.t_disp_buffer[i].doacross_buf_idx = i; } } else { team->t.t_disp_buffer[0].buffer_index = 0; team->t.t_disp_buffer[0].doacross_buf_idx = 0; } KMP_MB(); /* Flush all pending memory write invalidates. */ KMP_ASSERT(this_thr->th.th_team == team); #ifdef KMP_DEBUG for (f = 0; f < team->t.t_nproc; f++) { KMP_DEBUG_ASSERT(team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc); } #endif /* KMP_DEBUG */ /* release the worker threads so they may begin working */ __kmp_fork_barrier(gtid, 0); } void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) { kmp_info_t *this_thr = __kmp_threads[gtid]; KMP_DEBUG_ASSERT(team); KMP_DEBUG_ASSERT(this_thr->th.th_team == team); KMP_ASSERT(KMP_MASTER_GTID(gtid)); KMP_MB(); /* Flush all pending memory write invalidates. */ /* Join barrier after fork */ #ifdef KMP_DEBUG if (__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) { __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid, __kmp_threads[gtid]); __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, " "team->t.t_nproc=%d\n", gtid, __kmp_threads[gtid]->th.th_team_nproc, team, team->t.t_nproc); __kmp_print_structure(); } KMP_DEBUG_ASSERT(__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc); #endif /* KMP_DEBUG */ __kmp_join_barrier(gtid); /* wait for everyone */ #if OMPT_SUPPORT if (ompt_enabled.enabled && this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) { int ds_tid = this_thr->th.th_info.ds.ds_tid; ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr); this_thr->th.ompt_thread_info.state = ompt_state_overhead; #if OMPT_OPTIONAL void *codeptr = NULL; if (KMP_MASTER_TID(ds_tid) && (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) || ompt_callbacks.ompt_callback(ompt_callback_sync_region))) codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address; if (ompt_enabled.ompt_callback_sync_region_wait) { ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)( ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data, codeptr); } if (ompt_enabled.ompt_callback_sync_region) { ompt_callbacks.ompt_callback(ompt_callback_sync_region)( ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data, codeptr); } #endif if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) { ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial? } } #endif KMP_MB(); /* Flush all pending memory write invalidates. */ KMP_ASSERT(this_thr->th.th_team == team); } /* ------------------------------------------------------------------------ */ #ifdef USE_LOAD_BALANCE // Return the worker threads actively spinning in the hot team, if we // are at the outermost level of parallelism. Otherwise, return 0. static int __kmp_active_hot_team_nproc(kmp_root_t *root) { int i; int retval; kmp_team_t *hot_team; if (root->r.r_active) { return 0; } hot_team = root->r.r_hot_team; if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { return hot_team->t.t_nproc - 1; // Don't count primary thread } // Skip the primary thread - it is accounted for elsewhere. retval = 0; for (i = 1; i < hot_team->t.t_nproc; i++) { if (hot_team->t.t_threads[i]->th.th_active) { retval++; } } return retval; } // Perform an automatic adjustment to the number of // threads used by the next parallel region. static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) { int retval; int pool_active; int hot_team_active; int team_curr_active; int system_active; KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root, set_nproc)); KMP_DEBUG_ASSERT(root); KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0] ->th.th_current_task->td_icvs.dynamic == TRUE); KMP_DEBUG_ASSERT(set_nproc > 1); if (set_nproc == 1) { KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n")); return 1; } // Threads that are active in the thread pool, active in the hot team for this // particular root (if we are at the outer par level), and the currently // executing thread (to become the primary thread) are available to add to the // new team, but are currently contributing to the system load, and must be // accounted for. pool_active = __kmp_thread_pool_active_nth; hot_team_active = __kmp_active_hot_team_nproc(root); team_curr_active = pool_active + hot_team_active + 1; // Check the system load. system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active); KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active)); if (system_active < 0) { // There was an error reading the necessary info from /proc, so use the // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode // = dynamic_thread_limit, we shouldn't wind up getting back here. __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit"); // Make this call behave like the thread limit algorithm. retval = __kmp_avail_proc - __kmp_nth + (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); if (retval > set_nproc) { retval = set_nproc; } if (retval < KMP_MIN_NTH) { retval = KMP_MIN_NTH; } KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n", retval)); return retval; } // There is a slight delay in the load balance algorithm in detecting new // running procs. The real system load at this instant should be at least as // large as the #active omp thread that are available to add to the team. if (system_active < team_curr_active) { system_active = team_curr_active; } retval = __kmp_avail_proc - system_active + team_curr_active; if (retval > set_nproc) { retval = set_nproc; } if (retval < KMP_MIN_NTH) { retval = KMP_MIN_NTH; } KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval)); return retval; } // __kmp_load_balance_nproc() #endif /* USE_LOAD_BALANCE */ /* ------------------------------------------------------------------------ */ /* NOTE: this is called with the __kmp_init_lock held */ void __kmp_cleanup(void) { int f; KA_TRACE(10, ("__kmp_cleanup: enter\n")); if (TCR_4(__kmp_init_parallel)) { #if KMP_HANDLE_SIGNALS __kmp_remove_signals(); #endif TCW_4(__kmp_init_parallel, FALSE); } if (TCR_4(__kmp_init_middle)) { #if KMP_AFFINITY_SUPPORTED __kmp_affinity_uninitialize(); #endif /* KMP_AFFINITY_SUPPORTED */ __kmp_cleanup_hierarchy(); TCW_4(__kmp_init_middle, FALSE); } KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n")); if (__kmp_init_serial) { __kmp_runtime_destroy(); __kmp_init_serial = FALSE; } __kmp_cleanup_threadprivate_caches(); for (f = 0; f < __kmp_threads_capacity; f++) { if (__kmp_root[f] != NULL) { __kmp_free(__kmp_root[f]); __kmp_root[f] = NULL; } } __kmp_free(__kmp_threads); // __kmp_threads and __kmp_root were allocated at once, as single block, so // there is no need in freeing __kmp_root. __kmp_threads = NULL; __kmp_root = NULL; __kmp_threads_capacity = 0; // Free old __kmp_threads arrays if they exist. kmp_old_threads_list_t *ptr = __kmp_old_threads_list; while (ptr) { kmp_old_threads_list_t *next = ptr->next; __kmp_free(ptr->threads); __kmp_free(ptr); ptr = next; } #if KMP_USE_DYNAMIC_LOCK __kmp_cleanup_indirect_user_locks(); #else __kmp_cleanup_user_locks(); #endif #if OMPD_SUPPORT if (ompd_state) { __kmp_free(ompd_env_block); ompd_env_block = NULL; ompd_env_block_size = 0; } #endif #if KMP_AFFINITY_SUPPORTED KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file)); __kmp_cpuinfo_file = NULL; #endif /* KMP_AFFINITY_SUPPORTED */ #if KMP_USE_ADAPTIVE_LOCKS #if KMP_DEBUG_ADAPTIVE_LOCKS __kmp_print_speculative_stats(); #endif #endif KMP_INTERNAL_FREE(__kmp_nested_nth.nth); __kmp_nested_nth.nth = NULL; __kmp_nested_nth.size = 0; __kmp_nested_nth.used = 0; KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types); __kmp_nested_proc_bind.bind_types = NULL; __kmp_nested_proc_bind.size = 0; __kmp_nested_proc_bind.used = 0; if (__kmp_affinity_format) { KMP_INTERNAL_FREE(__kmp_affinity_format); __kmp_affinity_format = NULL; } __kmp_i18n_catclose(); #if KMP_USE_HIER_SCHED __kmp_hier_scheds.deallocate(); #endif #if KMP_STATS_ENABLED __kmp_stats_fini(); #endif KA_TRACE(10, ("__kmp_cleanup: exit\n")); } /* ------------------------------------------------------------------------ */ int __kmp_ignore_mppbeg(void) { char *env; if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) { if (__kmp_str_match_false(env)) return FALSE; } // By default __kmpc_begin() is no-op. return TRUE; } int __kmp_ignore_mppend(void) { char *env; if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) { if (__kmp_str_match_false(env)) return FALSE; } // By default __kmpc_end() is no-op. return TRUE; } void __kmp_internal_begin(void) { int gtid; kmp_root_t *root; /* this is a very important step as it will register new sibling threads and assign these new uber threads a new gtid */ gtid = __kmp_entry_gtid(); root = __kmp_threads[gtid]->th.th_root; KMP_ASSERT(KMP_UBER_GTID(gtid)); if (root->r.r_begin) return; __kmp_acquire_lock(&root->r.r_begin_lock, gtid); if (root->r.r_begin) { __kmp_release_lock(&root->r.r_begin_lock, gtid); return; } root->r.r_begin = TRUE; __kmp_release_lock(&root->r.r_begin_lock, gtid); } /* ------------------------------------------------------------------------ */ void __kmp_user_set_library(enum library_type arg) { int gtid; kmp_root_t *root; kmp_info_t *thread; /* first, make sure we are initialized so we can get our gtid */ gtid = __kmp_entry_gtid(); thread = __kmp_threads[gtid]; root = thread->th.th_root; KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg, library_serial)); if (root->r.r_in_parallel) { /* Must be called in serial section of top-level thread */ KMP_WARNING(SetLibraryIncorrectCall); return; } switch (arg) { case library_serial: thread->th.th_set_nproc = 0; set__nproc(thread, 1); break; case library_turnaround: thread->th.th_set_nproc = 0; set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub); break; case library_throughput: thread->th.th_set_nproc = 0; set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub); break; default: KMP_FATAL(UnknownLibraryType, arg); } __kmp_aux_set_library(arg); } void __kmp_aux_set_stacksize(size_t arg) { if (!__kmp_init_serial) __kmp_serial_initialize(); #if KMP_OS_DARWIN if (arg & (0x1000 - 1)) { arg &= ~(0x1000 - 1); if (arg + 0x1000) /* check for overflow if we round up */ arg += 0x1000; } #endif __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); /* only change the default stacksize before the first parallel region */ if (!TCR_4(__kmp_init_parallel)) { size_t value = arg; /* argument is in bytes */ if (value < __kmp_sys_min_stksize) value = __kmp_sys_min_stksize; else if (value > KMP_MAX_STKSIZE) value = KMP_MAX_STKSIZE; __kmp_stksize = value; __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */ } __kmp_release_bootstrap_lock(&__kmp_initz_lock); } /* set the behaviour of the runtime library */ /* TODO this can cause some odd behaviour with sibling parallelism... */ void __kmp_aux_set_library(enum library_type arg) { __kmp_library = arg; switch (__kmp_library) { case library_serial: { KMP_INFORM(LibraryIsSerial); } break; case library_turnaround: if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set) __kmp_use_yield = 2; // only yield when oversubscribed break; case library_throughput: if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME; break; default: KMP_FATAL(UnknownLibraryType, arg); } } /* Getting team information common for all team API */ // Returns NULL if not in teams construct static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) { kmp_info_t *thr = __kmp_entry_thread(); teams_serialized = 0; if (thr->th.th_teams_microtask) { kmp_team_t *team = thr->th.th_team; int tlevel = thr->th.th_teams_level; // the level of the teams construct int ii = team->t.t_level; teams_serialized = team->t.t_serialized; int level = tlevel + 1; KMP_DEBUG_ASSERT(ii >= tlevel); while (ii > level) { for (teams_serialized = team->t.t_serialized; (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) { } if (team->t.t_serialized && (!teams_serialized)) { team = team->t.t_parent; continue; } if (ii > level) { team = team->t.t_parent; ii--; } } return team; } return NULL; } int __kmp_aux_get_team_num() { int serialized; kmp_team_t *team = __kmp_aux_get_team_info(serialized); if (team) { if (serialized > 1) { return 0; // teams region is serialized ( 1 team of 1 thread ). } else { return team->t.t_master_tid; } } return 0; } int __kmp_aux_get_num_teams() { int serialized; kmp_team_t *team = __kmp_aux_get_team_info(serialized); if (team) { if (serialized > 1) { return 1; } else { return team->t.t_parent->t.t_nproc; } } return 1; } /* ------------------------------------------------------------------------ */ /* * Affinity Format Parser * * Field is in form of: %[[[0].]size]type * % and type are required (%% means print a literal '%') * type is either single char or long name surrounded by {}, * e.g., N or {num_threads} * 0 => leading zeros * . => right justified when size is specified * by default output is left justified * size is the *minimum* field length * All other characters are printed as is * * Available field types: * L {thread_level} - omp_get_level() * n {thread_num} - omp_get_thread_num() * h {host} - name of host machine * P {process_id} - process id (integer) * T {thread_identifier} - native thread identifier (integer) * N {num_threads} - omp_get_num_threads() * A {ancestor_tnum} - omp_get_ancestor_thread_num(omp_get_level()-1) * a {thread_affinity} - comma separated list of integers or integer ranges * (values of affinity mask) * * Implementation-specific field types can be added * If a type is unknown, print "undefined" */ // Structure holding the short name, long name, and corresponding data type // for snprintf. A table of these will represent the entire valid keyword // field types. typedef struct kmp_affinity_format_field_t { char short_name; // from spec e.g., L -> thread level const char *long_name; // from spec thread_level -> thread level char field_format; // data type for snprintf (typically 'd' or 's' // for integer or string) } kmp_affinity_format_field_t; static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = { #if KMP_AFFINITY_SUPPORTED {'A', "thread_affinity", 's'}, #endif {'t', "team_num", 'd'}, {'T', "num_teams", 'd'}, {'L', "nesting_level", 'd'}, {'n', "thread_num", 'd'}, {'N', "num_threads", 'd'}, {'a', "ancestor_tnum", 'd'}, {'H', "host", 's'}, {'P', "process_id", 'd'}, {'i', "native_thread_id", 'd'}}; // Return the number of characters it takes to hold field static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th, const char **ptr, kmp_str_buf_t *field_buffer) { int rc, format_index, field_value; const char *width_left, *width_right; bool pad_zeros, right_justify, parse_long_name, found_valid_name; static const int FORMAT_SIZE = 20; char format[FORMAT_SIZE] = {0}; char absolute_short_name = 0; KMP_DEBUG_ASSERT(gtid >= 0); KMP_DEBUG_ASSERT(th); KMP_DEBUG_ASSERT(**ptr == '%'); KMP_DEBUG_ASSERT(field_buffer); __kmp_str_buf_clear(field_buffer); // Skip the initial % (*ptr)++; // Check for %% first if (**ptr == '%') { __kmp_str_buf_cat(field_buffer, "%", 1); (*ptr)++; // skip over the second % return 1; } // Parse field modifiers if they are present pad_zeros = false; if (**ptr == '0') { pad_zeros = true; (*ptr)++; // skip over 0 } right_justify = false; if (**ptr == '.') { right_justify = true; (*ptr)++; // skip over . } // Parse width of field: [width_left, width_right) width_left = width_right = NULL; if (**ptr >= '0' && **ptr <= '9') { width_left = *ptr; SKIP_DIGITS(*ptr); width_right = *ptr; } // Create the format for KMP_SNPRINTF based on flags parsed above format_index = 0; format[format_index++] = '%'; if (!right_justify) format[format_index++] = '-'; if (pad_zeros) format[format_index++] = '0'; if (width_left && width_right) { int i = 0; // Only allow 8 digit number widths. // This also prevents overflowing format variable while (i < 8 && width_left < width_right) { format[format_index++] = *width_left; width_left++; i++; } } // Parse a name (long or short) // Canonicalize the name into absolute_short_name found_valid_name = false; parse_long_name = (**ptr == '{'); if (parse_long_name) (*ptr)++; // skip initial left brace for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) / sizeof(__kmp_affinity_format_table[0]); ++i) { char short_name = __kmp_affinity_format_table[i].short_name; const char *long_name = __kmp_affinity_format_table[i].long_name; char field_format = __kmp_affinity_format_table[i].field_format; if (parse_long_name) { size_t length = KMP_STRLEN(long_name); if (strncmp(*ptr, long_name, length) == 0) { found_valid_name = true; (*ptr) += length; // skip the long name } } else if (**ptr == short_name) { found_valid_name = true; (*ptr)++; // skip the short name } if (found_valid_name) { format[format_index++] = field_format; format[format_index++] = '\0'; absolute_short_name = short_name; break; } } if (parse_long_name) { if (**ptr != '}') { absolute_short_name = 0; } else { (*ptr)++; // skip over the right brace } } // Attempt to fill the buffer with the requested // value using snprintf within __kmp_str_buf_print() switch (absolute_short_name) { case 't': rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num()); break; case 'T': rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams()); break; case 'L': rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level); break; case 'n': rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid)); break; case 'H': { static const int BUFFER_SIZE = 256; char buf[BUFFER_SIZE]; __kmp_expand_host_name(buf, BUFFER_SIZE); rc = __kmp_str_buf_print(field_buffer, format, buf); } break; case 'P': rc = __kmp_str_buf_print(field_buffer, format, getpid()); break; case 'i': rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid()); break; case 'N': rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc); break; case 'a': field_value = __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1); rc = __kmp_str_buf_print(field_buffer, format, field_value); break; #if KMP_AFFINITY_SUPPORTED case 'A': { kmp_str_buf_t buf; __kmp_str_buf_init(&buf); __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask); rc = __kmp_str_buf_print(field_buffer, format, buf.str); __kmp_str_buf_free(&buf); } break; #endif default: // According to spec, If an implementation does not have info for field // type, then "undefined" is printed rc = __kmp_str_buf_print(field_buffer, "%s", "undefined"); // Skip the field if (parse_long_name) { SKIP_TOKEN(*ptr); if (**ptr == '}') (*ptr)++; } else { (*ptr)++; } } KMP_ASSERT(format_index <= FORMAT_SIZE); return rc; } /* * Return number of characters needed to hold the affinity string * (not including null byte character) * The resultant string is printed to buffer, which the caller can then * handle afterwards */ size_t __kmp_aux_capture_affinity(int gtid, const char *format, kmp_str_buf_t *buffer) { const char *parse_ptr; size_t retval; const kmp_info_t *th; kmp_str_buf_t field; KMP_DEBUG_ASSERT(buffer); KMP_DEBUG_ASSERT(gtid >= 0); __kmp_str_buf_init(&field); __kmp_str_buf_clear(buffer); th = __kmp_threads[gtid]; retval = 0; // If format is NULL or zero-length string, then we use // affinity-format-var ICV parse_ptr = format; if (parse_ptr == NULL || *parse_ptr == '\0') { parse_ptr = __kmp_affinity_format; } KMP_DEBUG_ASSERT(parse_ptr); while (*parse_ptr != '\0') { // Parse a field if (*parse_ptr == '%') { // Put field in the buffer int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field); __kmp_str_buf_catbuf(buffer, &field); retval += rc; } else { // Put literal character in buffer __kmp_str_buf_cat(buffer, parse_ptr, 1); retval++; parse_ptr++; } } __kmp_str_buf_free(&field); return retval; } // Displays the affinity string to stdout void __kmp_aux_display_affinity(int gtid, const char *format) { kmp_str_buf_t buf; __kmp_str_buf_init(&buf); __kmp_aux_capture_affinity(gtid, format, &buf); __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str); __kmp_str_buf_free(&buf); } /* ------------------------------------------------------------------------ */ void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) { int blocktime = arg; /* argument is in milliseconds */ #if KMP_USE_MONITOR int bt_intervals; #endif kmp_int8 bt_set; __kmp_save_internal_controls(thread); /* Normalize and set blocktime for the teams */ if (blocktime < KMP_MIN_BLOCKTIME) blocktime = KMP_MIN_BLOCKTIME; else if (blocktime > KMP_MAX_BLOCKTIME) blocktime = KMP_MAX_BLOCKTIME; set__blocktime_team(thread->th.th_team, tid, blocktime); set__blocktime_team(thread->th.th_serial_team, 0, blocktime); #if KMP_USE_MONITOR /* Calculate and set blocktime intervals for the teams */ bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups); set__bt_intervals_team(thread->th.th_team, tid, bt_intervals); set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals); #endif /* Set whether blocktime has been set to "TRUE" */ bt_set = TRUE; set__bt_set_team(thread->th.th_team, tid, bt_set); set__bt_set_team(thread->th.th_serial_team, 0, bt_set); #if KMP_USE_MONITOR KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", __kmp_gtid_from_tid(tid, thread->th.th_team), thread->th.th_team->t.t_id, tid, blocktime, bt_intervals, __kmp_monitor_wakeups)); #else KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n", __kmp_gtid_from_tid(tid, thread->th.th_team), thread->th.th_team->t.t_id, tid, blocktime)); #endif } void __kmp_aux_set_defaults(char const *str, size_t len) { if (!__kmp_init_serial) { __kmp_serial_initialize(); } __kmp_env_initialize(str); if (__kmp_settings || __kmp_display_env || __kmp_display_env_verbose) { __kmp_env_print(); } } // __kmp_aux_set_defaults /* ------------------------------------------------------------------------ */ /* internal fast reduction routines */ PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method( ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck) { // Default reduction method: critical construct ( lck != NULL, like in current // PAROPT ) // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method // can be selected by RTL // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method // can be selected by RTL // Finally, it's up to OpenMP RTL to make a decision on which method to select // among generated by PAROPT. PACKED_REDUCTION_METHOD_T retval; int team_size; KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 ) #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \ (loc && \ ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))) #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func)) retval = critical_reduce_block; // another choice of getting a team size (with 1 dynamic deference) is slower team_size = __kmp_get_team_num_threads(global_tid); if (team_size == 1) { retval = empty_reduce_block; } else { int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || \ KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD int teamsize_cutoff = 4; #if KMP_MIC_SUPPORTED if (__kmp_mic_type != non_mic) { teamsize_cutoff = 8; } #endif int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; if (tree_available) { if (team_size <= teamsize_cutoff) { if (atomic_available) { retval = atomic_reduce_block; } } else { retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER; } } else if (atomic_available) { retval = atomic_reduce_block; } #else #error "Unknown or unsupported OS" #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_HURD // basic tuning if (atomic_available) { if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ??? retval = atomic_reduce_block; } } // otherwise: use critical section #elif KMP_OS_DARWIN int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; if (atomic_available && (num_vars <= 3)) { retval = atomic_reduce_block; } else if (tree_available) { if ((reduce_size > (9 * sizeof(kmp_real64))) && (reduce_size < (2000 * sizeof(kmp_real64)))) { retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER; } } // otherwise: use critical section #else #error "Unknown or unsupported OS" #endif #else #error "Unknown or unsupported architecture" #endif } // KMP_FORCE_REDUCTION // If the team is serialized (team_size == 1), ignore the forced reduction // method and stay with the unsynchronized method (empty_reduce_block) if (__kmp_force_reduction_method != reduction_method_not_defined && team_size != 1) { PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block; int atomic_available, tree_available; switch ((forced_retval = __kmp_force_reduction_method)) { case critical_reduce_block: KMP_ASSERT(lck); // lck should be != 0 break; case atomic_reduce_block: atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; if (!atomic_available) { KMP_WARNING(RedMethodNotSupported, "atomic"); forced_retval = critical_reduce_block; } break; case tree_reduce_block: tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; if (!tree_available) { KMP_WARNING(RedMethodNotSupported, "tree"); forced_retval = critical_reduce_block; } else { #if KMP_FAST_REDUCTION_BARRIER forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER; #endif } break; default: KMP_ASSERT(0); // "unsupported method specified" } retval = forced_retval; } KA_TRACE(10, ("reduction method selected=%08x\n", retval)); #undef FAST_REDUCTION_TREE_METHOD_GENERATED #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED return (retval); } // this function is for testing set/get/determine reduce method kmp_int32 __kmp_get_reduce_method(void) { return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8); } // Soft pause sets up threads to ignore blocktime and just go to sleep. // Spin-wait code checks __kmp_pause_status and reacts accordingly. void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; } // Hard pause shuts down the runtime completely. Resume happens naturally when // OpenMP is used subsequently. void __kmp_hard_pause() { __kmp_pause_status = kmp_hard_paused; __kmp_internal_end_thread(-1); } // Soft resume sets __kmp_pause_status, and wakes up all threads. void __kmp_resume_if_soft_paused() { if (__kmp_pause_status == kmp_soft_paused) { __kmp_pause_status = kmp_not_paused; for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) { kmp_info_t *thread = __kmp_threads[gtid]; if (thread) { // Wake it if sleeping kmp_flag_64<> fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread); if (fl.is_sleeping()) fl.resume(gtid); else if (__kmp_try_suspend_mx(thread)) { // got suspend lock __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep } else { // thread holds the lock and may sleep soon do { // until either the thread sleeps, or we can get the lock if (fl.is_sleeping()) { fl.resume(gtid); break; } else if (__kmp_try_suspend_mx(thread)) { __kmp_unlock_suspend_mx(thread); break; } } while (1); } } } } } // This function is called via __kmpc_pause_resource. Returns 0 if successful. // TODO: add warning messages int __kmp_pause_resource(kmp_pause_status_t level) { if (level == kmp_not_paused) { // requesting resume if (__kmp_pause_status == kmp_not_paused) { // error message about runtime not being paused, so can't resume return 1; } else { KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused || __kmp_pause_status == kmp_hard_paused); __kmp_pause_status = kmp_not_paused; return 0; } } else if (level == kmp_soft_paused) { // requesting soft pause if (__kmp_pause_status != kmp_not_paused) { // error message about already being paused return 1; } else { __kmp_soft_pause(); return 0; } } else if (level == kmp_hard_paused) { // requesting hard pause if (__kmp_pause_status != kmp_not_paused) { // error message about already being paused return 1; } else { __kmp_hard_pause(); return 0; } } else { // error message about invalid level return 1; } } void __kmp_omp_display_env(int verbose) { __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); if (__kmp_init_serial == 0) __kmp_do_serial_initialize(); __kmp_display_env_impl(!verbose, verbose); __kmp_release_bootstrap_lock(&__kmp_initz_lock); } // The team size is changing, so distributed barrier must be modified void __kmp_resize_dist_barrier(kmp_team_t *team, int old_nthreads, int new_nthreads) { KMP_DEBUG_ASSERT(__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar); kmp_info_t **other_threads = team->t.t_threads; // We want all the workers to stop waiting on the barrier while we adjust the // size of the team. for (int f = 1; f < old_nthreads; ++f) { KMP_DEBUG_ASSERT(other_threads[f] != NULL); // Ignore threads that are already inactive or not present in the team if (team->t.t_threads[f]->th.th_used_in_team.load() == 0) { // teams construct causes thread_limit to get passed in, and some of // those could be inactive; just ignore them continue; } // If thread is transitioning still to in_use state, wait for it if (team->t.t_threads[f]->th.th_used_in_team.load() == 3) { while (team->t.t_threads[f]->th.th_used_in_team.load() == 3) KMP_CPU_PAUSE(); } // The thread should be in_use now KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 1); // Transition to unused state team->t.t_threads[f]->th.th_used_in_team.store(2); KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 2); } // Release all the workers team->t.b->go_release(); KMP_MFENCE(); // Workers should see transition status 2 and move to 0; but may need to be // woken up first int count = old_nthreads - 1; while (count > 0) { count = old_nthreads - 1; for (int f = 1; f < old_nthreads; ++f) { if (other_threads[f]->th.th_used_in_team.load() != 0) { if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) { // Wake up the workers kmp_atomic_flag_64<> *flag = (kmp_atomic_flag_64<> *)CCAST( void *, other_threads[f]->th.th_sleep_loc); __kmp_atomic_resume_64(other_threads[f]->th.th_info.ds.ds_gtid, flag); } } else { KMP_DEBUG_ASSERT(team->t.t_threads[f]->th.th_used_in_team.load() == 0); count--; } } } // Now update the barrier size team->t.b->update_num_threads(new_nthreads); team->t.b->go_reset(); } void __kmp_add_threads_to_team(kmp_team_t *team, int new_nthreads) { // Add the threads back to the team KMP_DEBUG_ASSERT(team); // Threads were paused and pointed at th_used_in_team temporarily during a // resize of the team. We're going to set th_used_in_team to 3 to indicate to // the thread that it should transition itself back into the team. Then, if // blocktime isn't infinite, the thread could be sleeping, so we send a resume // to wake it up. for (int f = 1; f < new_nthreads; ++f) { KMP_DEBUG_ASSERT(team->t.t_threads[f]); KMP_COMPARE_AND_STORE_ACQ32(&(team->t.t_threads[f]->th.th_used_in_team), 0, 3); if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) { // Wake up sleeping threads __kmp_resume_32(team->t.t_threads[f]->th.th_info.ds.ds_gtid, (kmp_flag_32<false, false> *)NULL); } } // The threads should be transitioning to the team; when they are done, they // should have set th_used_in_team to 1. This loop forces master to wait until // all threads have moved into the team and are waiting in the barrier. int count = new_nthreads - 1; while (count > 0) { count = new_nthreads - 1; for (int f = 1; f < new_nthreads; ++f) { if (team->t.t_threads[f]->th.th_used_in_team.load() == 1) { count--; } } } } // Globals and functions for hidden helper task kmp_info_t **__kmp_hidden_helper_threads; kmp_info_t *__kmp_hidden_helper_main_thread; std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; #if KMP_OS_LINUX kmp_int32 __kmp_hidden_helper_threads_num = 8; kmp_int32 __kmp_enable_hidden_helper = TRUE; #else kmp_int32 __kmp_hidden_helper_threads_num = 0; kmp_int32 __kmp_enable_hidden_helper = FALSE; #endif namespace { std::atomic<kmp_int32> __kmp_hit_hidden_helper_threads_num; void __kmp_hidden_helper_wrapper_fn(int *gtid, int *, ...) { // This is an explicit synchronization on all hidden helper threads in case // that when a regular thread pushes a hidden helper task to one hidden // helper thread, the thread has not been awaken once since they're released // by the main thread after creating the team. KMP_ATOMIC_INC(&__kmp_hit_hidden_helper_threads_num); while (KMP_ATOMIC_LD_ACQ(&__kmp_hit_hidden_helper_threads_num) != __kmp_hidden_helper_threads_num) ; // If main thread, then wait for signal if (__kmpc_master(nullptr, *gtid)) { // First, unset the initial state and release the initial thread TCW_4(__kmp_init_hidden_helper_threads, FALSE); __kmp_hidden_helper_initz_release(); __kmp_hidden_helper_main_thread_wait(); // Now wake up all worker threads for (int i = 1; i < __kmp_hit_hidden_helper_threads_num; ++i) { __kmp_hidden_helper_worker_thread_signal(); } } } } // namespace void __kmp_hidden_helper_threads_initz_routine() { // Create a new root for hidden helper team/threads const int gtid = __kmp_register_root(TRUE); __kmp_hidden_helper_main_thread = __kmp_threads[gtid]; __kmp_hidden_helper_threads = &__kmp_threads[gtid]; __kmp_hidden_helper_main_thread->th.th_set_nproc = __kmp_hidden_helper_threads_num; KMP_ATOMIC_ST_REL(&__kmp_hit_hidden_helper_threads_num, 0); __kmpc_fork_call(nullptr, 0, __kmp_hidden_helper_wrapper_fn); // Set the initialization flag to FALSE TCW_SYNC_4(__kmp_init_hidden_helper, FALSE); __kmp_hidden_helper_threads_deinitz_release(); } /* Nesting Mode: Set via KMP_NESTING_MODE, which takes an integer. Note: we skip duplicate topology levels, and skip levels with only one entity. KMP_NESTING_MODE=0 is the default, and doesn't use nesting mode. KMP_NESTING_MODE=1 sets as many nesting levels as there are distinct levels in the topology, and initializes the number of threads at each of those levels to the number of entities at each level, respectively, below the entity at the parent level. KMP_NESTING_MODE=N, where N>1, attempts to create up to N nesting levels, but starts with nesting OFF -- max-active-levels-var is 1 -- and requires the user to turn nesting on explicitly. This is an even more experimental option to this experimental feature, and may change or go away in the future. */ // Allocate space to store nesting levels void __kmp_init_nesting_mode() { int levels = KMP_HW_LAST; __kmp_nesting_mode_nlevels = levels; __kmp_nesting_nth_level = (int *)KMP_INTERNAL_MALLOC(levels * sizeof(int)); for (int i = 0; i < levels; ++i) __kmp_nesting_nth_level[i] = 0; if (__kmp_nested_nth.size < levels) { __kmp_nested_nth.nth = (int *)KMP_INTERNAL_REALLOC(__kmp_nested_nth.nth, levels * sizeof(int)); __kmp_nested_nth.size = levels; } } // Set # threads for top levels of nesting; must be called after topology set void __kmp_set_nesting_mode_threads() { kmp_info_t *thread = __kmp_threads[__kmp_entry_gtid()]; if (__kmp_nesting_mode == 1) __kmp_nesting_mode_nlevels = KMP_MAX_ACTIVE_LEVELS_LIMIT; else if (__kmp_nesting_mode > 1) __kmp_nesting_mode_nlevels = __kmp_nesting_mode; if (__kmp_topology) { // use topology info int loc, hw_level; for (loc = 0, hw_level = 0; hw_level < __kmp_topology->get_depth() && loc < __kmp_nesting_mode_nlevels; loc++, hw_level++) { __kmp_nesting_nth_level[loc] = __kmp_topology->get_ratio(hw_level); if (__kmp_nesting_nth_level[loc] == 1) loc--; } // Make sure all cores are used if (__kmp_nesting_mode > 1 && loc > 1) { int core_level = __kmp_topology->get_level(KMP_HW_CORE); int num_cores = __kmp_topology->get_count(core_level); int upper_levels = 1; for (int level = 0; level < loc - 1; ++level) upper_levels *= __kmp_nesting_nth_level[level]; if (upper_levels * __kmp_nesting_nth_level[loc - 1] < num_cores) __kmp_nesting_nth_level[loc - 1] = num_cores / __kmp_nesting_nth_level[loc - 2]; } __kmp_nesting_mode_nlevels = loc; __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; } else { // no topology info available; provide a reasonable guesstimation if (__kmp_avail_proc >= 4) { __kmp_nesting_nth_level[0] = __kmp_avail_proc / 2; __kmp_nesting_nth_level[1] = 2; __kmp_nesting_mode_nlevels = 2; } else { __kmp_nesting_nth_level[0] = __kmp_avail_proc; __kmp_nesting_mode_nlevels = 1; } __kmp_nested_nth.used = __kmp_nesting_mode_nlevels; } for (int i = 0; i < __kmp_nesting_mode_nlevels; ++i) { __kmp_nested_nth.nth[i] = __kmp_nesting_nth_level[i]; } set__nproc(thread, __kmp_nesting_nth_level[0]); if (__kmp_nesting_mode > 1 && __kmp_nesting_mode_nlevels > __kmp_nesting_mode) __kmp_nesting_mode_nlevels = __kmp_nesting_mode; if (get__max_active_levels(thread) > 1) { // if max levels was set, set nesting mode levels to same __kmp_nesting_mode_nlevels = get__max_active_levels(thread); } if (__kmp_nesting_mode == 1) // turn on nesting for this case only set__max_active_levels(thread, __kmp_nesting_mode_nlevels); } // Empty symbols to export (see exports_so.txt) when feature is disabled extern "C" { #if !KMP_STATS_ENABLED void __kmp_reset_stats() {} #endif #if !USE_DEBUGGER int __kmp_omp_debug_struct_info = FALSE; int __kmp_debugging = FALSE; #endif #if !USE_ITT_BUILD || !USE_ITT_NOTIFY void __kmp_itt_fini_ittlib() {} void __kmp_itt_init_ittlib() {} #endif } // end of file