shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity
llvm-project/openmp/runtime/src/kmp_barrier.cpp
Jonathan Peyton 99f5969565 [OpenMP] Let primary thread gather topology info for each worker thread 
This change has the primary thread create each thread's initial mask
and topology information so it is available immediately after
forking. The setting of mask/topology information is decoupled from the
actual binding. Also add this setting of topology information inside the
__kmp_partition_places mechanism for OMP_PLACES+OMP_PROC_BIND.

Without this, there could be a timing window after the primary
thread signals the workers to fork where worker threads have not yet
established their affinity mask or topology information.

Each worker thread will then bind to the location the primary thread
sets.

Differential Revision: https://reviews.llvm.org/D156727
2023-08-22 15:56:51 -05:00

2671 lines
106 KiB
C++
Raw Blame History

/*
* kmp_barrier.cpp
*/
//===----------------------------------------------------------------------===//
//
// 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_wait_release.h"
#include "kmp_barrier.h"
#include "kmp_itt.h"
#include "kmp_os.h"
#include "kmp_stats.h"
#include "ompt-specific.h"
// for distributed barrier
#include "kmp_affinity.h"
#if KMP_MIC
#include <immintrin.h>
#define USE_NGO_STORES 1
#endif // KMP_MIC
#if KMP_MIC && USE_NGO_STORES
// ICV copying
#define ngo_load(src) __m512d Vt = _mm512_load_pd((void *)(src))
#define ngo_store_icvs(dst, src) _mm512_storenrngo_pd((void *)(dst), Vt)
#define ngo_store_go(dst, src) _mm512_storenrngo_pd((void *)(dst), Vt)
#define ngo_sync() __asm__ volatile("lock; addl $0,0(%%rsp)" ::: "memory")
#else
#define ngo_load(src) ((void)0)
#define ngo_store_icvs(dst, src) copy_icvs((dst), (src))
#define ngo_store_go(dst, src) KMP_MEMCPY((dst), (src), CACHE_LINE)
#define ngo_sync() ((void)0)
#endif /* KMP_MIC && USE_NGO_STORES */
void __kmp_print_structure(void); // Forward declaration
// ---------------------------- Barrier Algorithms ----------------------------
// Distributed barrier
// Compute how many threads to have polling each cache-line.
// We want to limit the number of writes to IDEAL_GO_RESOLUTION.
void distributedBarrier::computeVarsForN(size_t n) {
int nsockets = 1;
if (__kmp_topology) {
int socket_level = __kmp_topology->get_level(KMP_HW_SOCKET);
int core_level = __kmp_topology->get_level(KMP_HW_CORE);
int ncores_per_socket =
__kmp_topology->calculate_ratio(core_level, socket_level);
nsockets = __kmp_topology->get_count(socket_level);
if (nsockets <= 0)
nsockets = 1;
if (ncores_per_socket <= 0)
ncores_per_socket = 1;
threads_per_go = ncores_per_socket >> 1;
if (!fix_threads_per_go) {
// Minimize num_gos
if (threads_per_go > 4) {
if (KMP_OPTIMIZE_FOR_REDUCTIONS) {
threads_per_go = threads_per_go >> 1;
}
if (threads_per_go > 4 && nsockets == 1)
threads_per_go = threads_per_go >> 1;
}
}
if (threads_per_go == 0)
threads_per_go = 1;
fix_threads_per_go = true;
num_gos = n / threads_per_go;
if (n % threads_per_go)
num_gos++;
if (nsockets == 1 || num_gos == 1)
num_groups = 1;
else {
num_groups = num_gos / nsockets;
if (num_gos % nsockets)
num_groups++;
}
if (num_groups <= 0)
num_groups = 1;
gos_per_group = num_gos / num_groups;
if (num_gos % num_groups)
gos_per_group++;
threads_per_group = threads_per_go * gos_per_group;
} else {
num_gos = n / threads_per_go;
if (n % threads_per_go)
num_gos++;
if (num_gos == 1)
num_groups = 1;
else {
num_groups = num_gos / 2;
if (num_gos % 2)
num_groups++;
}
gos_per_group = num_gos / num_groups;
if (num_gos % num_groups)
gos_per_group++;
threads_per_group = threads_per_go * gos_per_group;
}
}
void distributedBarrier::computeGo(size_t n) {
// Minimize num_gos
for (num_gos = 1;; num_gos++)
if (IDEAL_CONTENTION * num_gos >= n)
break;
threads_per_go = n / num_gos;
if (n % num_gos)
threads_per_go++;
while (num_gos > MAX_GOS) {
threads_per_go++;
num_gos = n / threads_per_go;
if (n % threads_per_go)
num_gos++;
}
computeVarsForN(n);
}
// This function is to resize the barrier arrays when the new number of threads
// exceeds max_threads, which is the current size of all the arrays
void distributedBarrier::resize(size_t nthr) {
KMP_DEBUG_ASSERT(nthr > max_threads);
// expand to requested size * 2
max_threads = nthr * 2;
// allocate arrays to new max threads
for (int i = 0; i < MAX_ITERS; ++i) {
if (flags[i])
flags[i] = (flags_s *)KMP_INTERNAL_REALLOC(flags[i],
max_threads * sizeof(flags_s));
else
flags[i] = (flags_s *)KMP_INTERNAL_MALLOC(max_threads * sizeof(flags_s));
}
if (go)
go = (go_s *)KMP_INTERNAL_REALLOC(go, max_threads * sizeof(go_s));
else
go = (go_s *)KMP_INTERNAL_MALLOC(max_threads * sizeof(go_s));
if (iter)
iter = (iter_s *)KMP_INTERNAL_REALLOC(iter, max_threads * sizeof(iter_s));
else
iter = (iter_s *)KMP_INTERNAL_MALLOC(max_threads * sizeof(iter_s));
if (sleep)
sleep =
(sleep_s *)KMP_INTERNAL_REALLOC(sleep, max_threads * sizeof(sleep_s));
else
sleep = (sleep_s *)KMP_INTERNAL_MALLOC(max_threads * sizeof(sleep_s));
}
// This function is to set all the go flags that threads might be waiting
// on, and when blocktime is not infinite, it should be followed by a wake-up
// call to each thread
kmp_uint64 distributedBarrier::go_release() {
kmp_uint64 next_go = iter[0].iter + distributedBarrier::MAX_ITERS;
for (size_t j = 0; j < num_gos; j++) {
go[j].go.store(next_go);
}
return next_go;
}
void distributedBarrier::go_reset() {
for (size_t j = 0; j < max_threads; ++j) {
for (size_t i = 0; i < distributedBarrier::MAX_ITERS; ++i) {
flags[i][j].stillNeed = 1;
}
go[j].go.store(0);
iter[j].iter = 0;
}
}
// This function inits/re-inits the distributed barrier for a particular number
// of threads. If a resize of arrays is needed, it calls the resize function.
void distributedBarrier::init(size_t nthr) {
size_t old_max = max_threads;
if (nthr > max_threads) { // need more space in arrays
resize(nthr);
}
for (size_t i = 0; i < max_threads; i++) {
for (size_t j = 0; j < distributedBarrier::MAX_ITERS; j++) {
flags[j][i].stillNeed = 1;
}
go[i].go.store(0);
iter[i].iter = 0;
if (i >= old_max)
sleep[i].sleep = false;
}
// Recalculate num_gos, etc. based on new nthr
computeVarsForN(nthr);
num_threads = nthr;
if (team_icvs == NULL)
team_icvs = __kmp_allocate(sizeof(kmp_internal_control_t));
}
// This function is used only when KMP_BLOCKTIME is not infinite.
// static
void __kmp_dist_barrier_wakeup(enum barrier_type bt, kmp_team_t *team,
size_t start, size_t stop, size_t inc,
size_t tid) {
KMP_DEBUG_ASSERT(__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
kmp_info_t **other_threads = team->t.t_threads;
for (size_t thr = start; thr < stop; thr += inc) {
KMP_DEBUG_ASSERT(other_threads[thr]);
int gtid = other_threads[thr]->th.th_info.ds.ds_gtid;
// Wake up worker regardless of if it appears to be sleeping or not
__kmp_atomic_resume_64(gtid, (kmp_atomic_flag_64<> *)NULL);
}
}
static void __kmp_dist_barrier_gather(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_dist_gather);
kmp_team_t *team;
distributedBarrier *b;
kmp_info_t **other_threads;
kmp_uint64 my_current_iter, my_next_iter;
kmp_uint32 nproc;
bool group_leader;
team = this_thr->th.th_team;
nproc = this_thr->th.th_team_nproc;
other_threads = team->t.t_threads;
b = team->t.b;
my_current_iter = b->iter[tid].iter;
my_next_iter = (my_current_iter + 1) % distributedBarrier::MAX_ITERS;
group_leader = ((tid % b->threads_per_group) == 0);
KA_TRACE(20,
("__kmp_dist_barrier_gather: T#%d(%d:%d) enter; barrier type %d\n",
gtid, team->t.t_id, tid, bt));
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - save arrive time to the thread
if (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time =
__itt_get_timestamp();
}
#endif
if (group_leader) {
// Start from the thread after the group leader
size_t group_start = tid + 1;
size_t group_end = tid + b->threads_per_group;
size_t threads_pending = 0;
if (group_end > nproc)
group_end = nproc;
do { // wait for threads in my group
threads_pending = 0;
// Check all the flags every time to avoid branch misspredict
for (size_t thr = group_start; thr < group_end; thr++) {
// Each thread uses a different cache line
threads_pending += b->flags[my_current_iter][thr].stillNeed;
}
// Execute tasks here
if (__kmp_tasking_mode != tskm_immediate_exec) {
kmp_task_team_t *task_team = this_thr->th.th_task_team;
if (task_team != NULL) {
if (TCR_SYNC_4(task_team->tt.tt_active)) {
if (KMP_TASKING_ENABLED(task_team)) {
int tasks_completed = FALSE;
__kmp_atomic_execute_tasks_64(
this_thr, gtid, (kmp_atomic_flag_64<> *)NULL, FALSE,
&tasks_completed USE_ITT_BUILD_ARG(itt_sync_obj), 0);
} else
this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
}
} else {
this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
} // if
}
if (TCR_4(__kmp_global.g.g_done)) {
if (__kmp_global.g.g_abort)
__kmp_abort_thread();
break;
} else if (__kmp_tasking_mode != tskm_immediate_exec &&
this_thr->th.th_reap_state == KMP_SAFE_TO_REAP) {
this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
}
} while (threads_pending > 0);
if (reduce) { // Perform reduction if needed
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
// Group leader reduces all threads in group
for (size_t thr = group_start; thr < group_end; thr++) {
(*reduce)(this_thr->th.th_local.reduce_data,
other_threads[thr]->th.th_local.reduce_data);
}
OMPT_REDUCTION_END;
}
// Set flag for next iteration
b->flags[my_next_iter][tid].stillNeed = 1;
// Each thread uses a different cache line; resets stillNeed to 0 to
// indicate it has reached the barrier
b->flags[my_current_iter][tid].stillNeed = 0;
do { // wait for all group leaders
threads_pending = 0;
for (size_t thr = 0; thr < nproc; thr += b->threads_per_group) {
threads_pending += b->flags[my_current_iter][thr].stillNeed;
}
// Execute tasks here
if (__kmp_tasking_mode != tskm_immediate_exec) {
kmp_task_team_t *task_team = this_thr->th.th_task_team;
if (task_team != NULL) {
if (TCR_SYNC_4(task_team->tt.tt_active)) {
if (KMP_TASKING_ENABLED(task_team)) {
int tasks_completed = FALSE;
__kmp_atomic_execute_tasks_64(
this_thr, gtid, (kmp_atomic_flag_64<> *)NULL, FALSE,
&tasks_completed USE_ITT_BUILD_ARG(itt_sync_obj), 0);
} else
this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
}
} else {
this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
} // if
}
if (TCR_4(__kmp_global.g.g_done)) {
if (__kmp_global.g.g_abort)
__kmp_abort_thread();
break;
} else if (__kmp_tasking_mode != tskm_immediate_exec &&
this_thr->th.th_reap_state == KMP_SAFE_TO_REAP) {
this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
}
} while (threads_pending > 0);
if (reduce) { // Perform reduction if needed
if (KMP_MASTER_TID(tid)) { // Master reduces over group leaders
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
for (size_t thr = b->threads_per_group; thr < nproc;
thr += b->threads_per_group) {
(*reduce)(this_thr->th.th_local.reduce_data,
other_threads[thr]->th.th_local.reduce_data);
}
OMPT_REDUCTION_END;
}
}
} else {
// Set flag for next iteration
b->flags[my_next_iter][tid].stillNeed = 1;
// Each thread uses a different cache line; resets stillNeed to 0 to
// indicate it has reached the barrier
b->flags[my_current_iter][tid].stillNeed = 0;
}
KMP_MFENCE();
KA_TRACE(20,
("__kmp_dist_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
static void __kmp_dist_barrier_release(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_dist_release);
kmp_team_t *team;
distributedBarrier *b;
kmp_bstate_t *thr_bar;
kmp_uint64 my_current_iter, next_go;
size_t my_go_index;
bool group_leader;
KA_TRACE(20, ("__kmp_dist_barrier_release: T#%d(%d) enter; barrier type %d\n",
gtid, tid, bt));
thr_bar = &this_thr->th.th_bar[bt].bb;
if (!KMP_MASTER_TID(tid)) {
// workers and non-master group leaders need to check their presence in team
do {
if (this_thr->th.th_used_in_team.load() != 1 &&
this_thr->th.th_used_in_team.load() != 3) {
// Thread is not in use in a team. Wait on location in tid's thread
// struct. The 0 value tells anyone looking that this thread is spinning
// or sleeping until this location becomes 3 again; 3 is the transition
// state to get to 1 which is waiting on go and being in the team
kmp_flag_32<false, false> my_flag(&(this_thr->th.th_used_in_team), 3);
if (KMP_COMPARE_AND_STORE_ACQ32(&(this_thr->th.th_used_in_team), 2,
0) ||
this_thr->th.th_used_in_team.load() == 0) {
my_flag.wait(this_thr, true USE_ITT_BUILD_ARG(itt_sync_obj));
}
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
// In fork barrier where we could not get the object reliably
itt_sync_obj =
__kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
// Cancel wait on previous parallel region...
__kmp_itt_task_starting(itt_sync_obj);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
if (itt_sync_obj != NULL)
// Call prepare as early as possible for "new" barrier
__kmp_itt_task_finished(itt_sync_obj);
} else
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
}
if (this_thr->th.th_used_in_team.load() != 1 &&
this_thr->th.th_used_in_team.load() != 3) // spurious wake-up?
continue;
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
// At this point, the thread thinks it is in use in a team, or in
// transition to be used in a team, but it might have reached this barrier
// before it was marked unused by the team. Unused threads are awoken and
// shifted to wait on local thread struct elsewhere. It also might reach
// this point by being picked up for use by a different team. Either way,
// we need to update the tid.
tid = __kmp_tid_from_gtid(gtid);
team = this_thr->th.th_team;
KMP_DEBUG_ASSERT(tid >= 0);
KMP_DEBUG_ASSERT(team);
b = team->t.b;
my_current_iter = b->iter[tid].iter;
next_go = my_current_iter + distributedBarrier::MAX_ITERS;
my_go_index = tid / b->threads_per_go;
if (this_thr->th.th_used_in_team.load() == 3) {
KMP_COMPARE_AND_STORE_ACQ32(&(this_thr->th.th_used_in_team), 3, 1);
}
// Check if go flag is set
if (b->go[my_go_index].go.load() != next_go) {
// Wait on go flag on team
kmp_atomic_flag_64<false, true> my_flag(
&(b->go[my_go_index].go), next_go, &(b->sleep[tid].sleep));
my_flag.wait(this_thr, true USE_ITT_BUILD_ARG(itt_sync_obj));
KMP_DEBUG_ASSERT(my_current_iter == b->iter[tid].iter ||
b->iter[tid].iter == 0);
KMP_DEBUG_ASSERT(b->sleep[tid].sleep == false);
}
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
// At this point, the thread's go location was set. This means the primary
// thread is safely in the barrier, and so this thread's data is
// up-to-date, but we should check again that this thread is really in
// use in the team, as it could have been woken up for the purpose of
// changing team size, or reaping threads at shutdown.
if (this_thr->th.th_used_in_team.load() == 1)
break;
} while (1);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
group_leader = ((tid % b->threads_per_group) == 0);
if (group_leader) {
// Tell all the threads in my group they can go!
for (size_t go_idx = my_go_index + 1;
go_idx < my_go_index + b->gos_per_group; go_idx++) {
b->go[go_idx].go.store(next_go);
}
// Fence added so that workers can see changes to go. sfence inadequate.
KMP_MFENCE();
}
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs) { // copy ICVs to final dest
__kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team,
tid, FALSE);
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
(kmp_internal_control_t *)team->t.b->team_icvs);
copy_icvs(&thr_bar->th_fixed_icvs,
&team->t.t_implicit_task_taskdata[tid].td_icvs);
}
#endif
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME && group_leader) {
// This thread is now awake and participating in the barrier;
// wake up the other threads in the group
size_t nproc = this_thr->th.th_team_nproc;
size_t group_end = tid + b->threads_per_group;
if (nproc < group_end)
group_end = nproc;
__kmp_dist_barrier_wakeup(bt, team, tid + 1, group_end, 1, tid);
}
} else { // Primary thread
team = this_thr->th.th_team;
b = team->t.b;
my_current_iter = b->iter[tid].iter;
next_go = my_current_iter + distributedBarrier::MAX_ITERS;
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs) {
// primary thread has ICVs in final destination; copy
copy_icvs(&thr_bar->th_fixed_icvs,
&team->t.t_implicit_task_taskdata[tid].td_icvs);
}
#endif
// Tell all the group leaders they can go!
for (size_t go_idx = 0; go_idx < b->num_gos; go_idx += b->gos_per_group) {
b->go[go_idx].go.store(next_go);
}
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
// Wake-up the group leaders
size_t nproc = this_thr->th.th_team_nproc;
__kmp_dist_barrier_wakeup(bt, team, tid + b->threads_per_group, nproc,
b->threads_per_group, tid);
}
// Tell all the threads in my group they can go!
for (size_t go_idx = 1; go_idx < b->gos_per_group; go_idx++) {
b->go[go_idx].go.store(next_go);
}
// Fence added so that workers can see changes to go. sfence inadequate.
KMP_MFENCE();
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
// Wake-up the other threads in my group
size_t nproc = this_thr->th.th_team_nproc;
size_t group_end = tid + b->threads_per_group;
if (nproc < group_end)
group_end = nproc;
__kmp_dist_barrier_wakeup(bt, team, tid + 1, group_end, 1, tid);
}
}
// Update to next iteration
KMP_ASSERT(my_current_iter == b->iter[tid].iter);
b->iter[tid].iter = (b->iter[tid].iter + 1) % distributedBarrier::MAX_ITERS;
KA_TRACE(
20, ("__kmp_dist_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
// Linear Barrier
template <bool cancellable = false>
static bool __kmp_linear_barrier_gather_template(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_linear_gather);
kmp_team_t *team = this_thr->th.th_team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_info_t **other_threads = team->t.t_threads;
KA_TRACE(
20,
("__kmp_linear_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - save arrive time to the thread
if (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time =
__itt_get_timestamp();
}
#endif
// We now perform a linear reduction to signal that all of the threads have
// arrived.
if (!KMP_MASTER_TID(tid)) {
KA_TRACE(20,
("__kmp_linear_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d)"
"arrived(%p): %llu => %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(0, team),
team->t.t_id, 0, &thr_bar->b_arrived, thr_bar->b_arrived,
thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
// Mark arrival to primary thread
/* After performing this write, a worker thread may not assume that the team
is valid any more - it could be deallocated by the primary thread at any
time. */
kmp_flag_64<> flag(&thr_bar->b_arrived, other_threads[0]);
flag.release();
} else {
kmp_balign_team_t *team_bar = &team->t.t_bar[bt];
int nproc = this_thr->th.th_team_nproc;
int i;
// Don't have to worry about sleep bit here or atomic since team setting
kmp_uint64 new_state = team_bar->b_arrived + KMP_BARRIER_STATE_BUMP;
// Collect all the worker team member threads.
for (i = 1; i < nproc; ++i) {
#if KMP_CACHE_MANAGE
// Prefetch next thread's arrived count
if (i + 1 < nproc)
KMP_CACHE_PREFETCH(&other_threads[i + 1]->th.th_bar[bt].bb.b_arrived);
#endif /* KMP_CACHE_MANAGE */
KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%d) "
"arrived(%p) == %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(i, team),
team->t.t_id, i,
&other_threads[i]->th.th_bar[bt].bb.b_arrived, new_state));
// Wait for worker thread to arrive
if (cancellable) {
kmp_flag_64<true, false> flag(
&other_threads[i]->th.th_bar[bt].bb.b_arrived, new_state);
if (flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj)))
return true;
} else {
kmp_flag_64<> flag(&other_threads[i]->th.th_bar[bt].bb.b_arrived,
new_state);
flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
}
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - write min of the thread time and the other thread
// time to the thread.
if (__kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_min_time = KMP_MIN(
this_thr->th.th_bar_min_time, other_threads[i]->th.th_bar_min_time);
}
#endif
if (reduce) {
KA_TRACE(100,
("__kmp_linear_barrier_gather: T#%d(%d:%d) += T#%d(%d:%d)\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(i, team),
team->t.t_id, i));
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
(*reduce)(this_thr->th.th_local.reduce_data,
other_threads[i]->th.th_local.reduce_data);
OMPT_REDUCTION_END;
}
}
// Don't have to worry about sleep bit here or atomic since team setting
team_bar->b_arrived = new_state;
KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) set team %d "
"arrived(%p) = %llu\n",
gtid, team->t.t_id, tid, team->t.t_id, &team_bar->b_arrived,
new_state));
}
KA_TRACE(
20,
("__kmp_linear_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
return false;
}
template <bool cancellable = false>
static bool __kmp_linear_barrier_release_template(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_linear_release);
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_team_t *team;
if (KMP_MASTER_TID(tid)) {
unsigned int i;
kmp_uint32 nproc = this_thr->th.th_team_nproc;
kmp_info_t **other_threads;
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
other_threads = team->t.t_threads;
KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d(%d:%d) primary enter for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
if (nproc > 1) {
#if KMP_BARRIER_ICV_PUSH
{
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_icv_copy);
if (propagate_icvs) {
ngo_load(&team->t.t_implicit_task_taskdata[0].td_icvs);
for (i = 1; i < nproc; ++i) {
__kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[i],
team, i, FALSE);
ngo_store_icvs(&team->t.t_implicit_task_taskdata[i].td_icvs,
&team->t.t_implicit_task_taskdata[0].td_icvs);
}
ngo_sync();
}
}
#endif // KMP_BARRIER_ICV_PUSH
// Now, release all of the worker threads
for (i = 1; i < nproc; ++i) {
#if KMP_CACHE_MANAGE
// Prefetch next thread's go flag
if (i + 1 < nproc)
KMP_CACHE_PREFETCH(&other_threads[i + 1]->th.th_bar[bt].bb.b_go);
#endif /* KMP_CACHE_MANAGE */
KA_TRACE(
20,
("__kmp_linear_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%d) "
"go(%p): %u => %u\n",
gtid, team->t.t_id, tid, other_threads[i]->th.th_info.ds.ds_gtid,
team->t.t_id, i, &other_threads[i]->th.th_bar[bt].bb.b_go,
other_threads[i]->th.th_bar[bt].bb.b_go,
other_threads[i]->th.th_bar[bt].bb.b_go + KMP_BARRIER_STATE_BUMP));
kmp_flag_64<> flag(&other_threads[i]->th.th_bar[bt].bb.b_go,
other_threads[i]);
flag.release();
}
}
} else { // Wait for the PRIMARY thread to release us
KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d wait go(%p) == %u\n",
gtid, &thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
if (cancellable) {
kmp_flag_64<true, false> flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
if (flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj)))
return true;
} else {
kmp_flag_64<> flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
}
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
// In a fork barrier; cannot get the object reliably (or ITTNOTIFY is
// disabled)
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
// Cancel wait on previous parallel region...
__kmp_itt_task_starting(itt_sync_obj);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return false;
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
if (itt_sync_obj != NULL)
// Call prepare as early as possible for "new" barrier
__kmp_itt_task_finished(itt_sync_obj);
} else
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
// Early exit for reaping threads releasing forkjoin barrier
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return false;
// The worker thread may now assume that the team is valid.
#ifdef KMP_DEBUG
tid = __kmp_tid_from_gtid(gtid);
team = __kmp_threads[gtid]->th.th_team;
#endif
KMP_DEBUG_ASSERT(team != NULL);
TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
KA_TRACE(20,
("__kmp_linear_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
KMP_MB(); // Flush all pending memory write invalidates.
}
KA_TRACE(
20,
("__kmp_linear_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
return false;
}
static void __kmp_linear_barrier_gather(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
__kmp_linear_barrier_gather_template<false>(
bt, this_thr, gtid, tid, reduce USE_ITT_BUILD_ARG(itt_sync_obj));
}
static bool __kmp_linear_barrier_gather_cancellable(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
return __kmp_linear_barrier_gather_template<true>(
bt, this_thr, gtid, tid, reduce USE_ITT_BUILD_ARG(itt_sync_obj));
}
static void __kmp_linear_barrier_release(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
__kmp_linear_barrier_release_template<false>(
bt, this_thr, gtid, tid, propagate_icvs USE_ITT_BUILD_ARG(itt_sync_obj));
}
static bool __kmp_linear_barrier_release_cancellable(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
return __kmp_linear_barrier_release_template<true>(
bt, this_thr, gtid, tid, propagate_icvs USE_ITT_BUILD_ARG(itt_sync_obj));
}
// Tree barrier
static void __kmp_tree_barrier_gather(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_tree_gather);
kmp_team_t *team = this_thr->th.th_team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_info_t **other_threads = team->t.t_threads;
kmp_uint32 nproc = this_thr->th.th_team_nproc;
kmp_uint32 branch_bits = __kmp_barrier_gather_branch_bits[bt];
kmp_uint32 branch_factor = 1 << branch_bits;
kmp_uint32 child;
kmp_uint32 child_tid;
kmp_uint64 new_state = 0;
KA_TRACE(
20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - save arrive time to the thread
if (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time =
__itt_get_timestamp();
}
#endif
// Perform tree gather to wait until all threads have arrived; reduce any
// required data as we go
child_tid = (tid << branch_bits) + 1;
if (child_tid < nproc) {
// Parent threads wait for all their children to arrive
new_state = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
child = 1;
do {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
#if KMP_CACHE_MANAGE
// Prefetch next thread's arrived count
if (child + 1 <= branch_factor && child_tid + 1 < nproc)
KMP_CACHE_PREFETCH(
&other_threads[child_tid + 1]->th.th_bar[bt].bb.b_arrived);
#endif /* KMP_CACHE_MANAGE */
KA_TRACE(20,
("__kmp_tree_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%u) "
"arrived(%p) == %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid, &child_bar->b_arrived, new_state));
// Wait for child to arrive
kmp_flag_64<> flag(&child_bar->b_arrived, new_state);
flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - write min of the thread time and a child time to
// the thread.
if (__kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_min_time = KMP_MIN(this_thr->th.th_bar_min_time,
child_thr->th.th_bar_min_time);
}
#endif
if (reduce) {
KA_TRACE(100,
("__kmp_tree_barrier_gather: T#%d(%d:%d) += T#%d(%d:%u)\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid));
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
(*reduce)(this_thr->th.th_local.reduce_data,
child_thr->th.th_local.reduce_data);
OMPT_REDUCTION_END;
}
child++;
child_tid++;
} while (child <= branch_factor && child_tid < nproc);
}
if (!KMP_MASTER_TID(tid)) { // Worker threads
kmp_int32 parent_tid = (tid - 1) >> branch_bits;
KA_TRACE(20,
("__kmp_tree_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d) "
"arrived(%p): %llu => %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(parent_tid, team),
team->t.t_id, parent_tid, &thr_bar->b_arrived, thr_bar->b_arrived,
thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
// Mark arrival to parent thread
/* After performing this write, a worker thread may not assume that the team
is valid any more - it could be deallocated by the primary thread at any
time. */
kmp_flag_64<> flag(&thr_bar->b_arrived, other_threads[parent_tid]);
flag.release();
} else {
// Need to update the team arrived pointer if we are the primary thread
if (nproc > 1) // New value was already computed above
team->t.t_bar[bt].b_arrived = new_state;
else
team->t.t_bar[bt].b_arrived += KMP_BARRIER_STATE_BUMP;
KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) set team %d "
"arrived(%p) = %llu\n",
gtid, team->t.t_id, tid, team->t.t_id,
&team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
}
KA_TRACE(20,
("__kmp_tree_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
static void __kmp_tree_barrier_release(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_tree_release);
kmp_team_t *team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_uint32 nproc;
kmp_uint32 branch_bits = __kmp_barrier_release_branch_bits[bt];
kmp_uint32 branch_factor = 1 << branch_bits;
kmp_uint32 child;
kmp_uint32 child_tid;
// Perform a tree release for all of the threads that have been gathered
if (!KMP_MASTER_TID(
tid)) { // Handle fork barrier workers who aren't part of a team yet
KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d wait go(%p) == %u\n", gtid,
&thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
// Wait for parent thread to release us
kmp_flag_64<> flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
// In fork barrier where we could not get the object reliably (or
// ITTNOTIFY is disabled)
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
// Cancel wait on previous parallel region...
__kmp_itt_task_starting(itt_sync_obj);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
if (itt_sync_obj != NULL)
// Call prepare as early as possible for "new" barrier
__kmp_itt_task_finished(itt_sync_obj);
} else
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
// Early exit for reaping threads releasing forkjoin barrier
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
// The worker thread may now assume that the team is valid.
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
tid = __kmp_tid_from_gtid(gtid);
TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
KA_TRACE(20,
("__kmp_tree_barrier_release: T#%d(%d:%d) set go(%p) = %u\n", gtid,
team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
KMP_MB(); // Flush all pending memory write invalidates.
} else {
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d(%d:%d) primary enter for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
nproc = this_thr->th.th_team_nproc;
child_tid = (tid << branch_bits) + 1;
if (child_tid < nproc) {
kmp_info_t **other_threads = team->t.t_threads;
child = 1;
// Parent threads release all their children
do {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
#if KMP_CACHE_MANAGE
// Prefetch next thread's go count
if (child + 1 <= branch_factor && child_tid + 1 < nproc)
KMP_CACHE_PREFETCH(
&other_threads[child_tid + 1]->th.th_bar[bt].bb.b_go);
#endif /* KMP_CACHE_MANAGE */
#if KMP_BARRIER_ICV_PUSH
{
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_icv_copy);
if (propagate_icvs) {
__kmp_init_implicit_task(team->t.t_ident,
team->t.t_threads[child_tid], team,
child_tid, FALSE);
copy_icvs(&team->t.t_implicit_task_taskdata[child_tid].td_icvs,
&team->t.t_implicit_task_taskdata[0].td_icvs);
}
}
#endif // KMP_BARRIER_ICV_PUSH
KA_TRACE(20,
("__kmp_tree_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%u)"
"go(%p): %u => %u\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
child_bar->b_go + KMP_BARRIER_STATE_BUMP));
// Release child from barrier
kmp_flag_64<> flag(&child_bar->b_go, child_thr);
flag.release();
child++;
child_tid++;
} while (child <= branch_factor && child_tid < nproc);
}
KA_TRACE(
20, ("__kmp_tree_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
// Hyper Barrier
static void __kmp_hyper_barrier_gather(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_hyper_gather);
kmp_team_t *team = this_thr->th.th_team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_info_t **other_threads = team->t.t_threads;
kmp_uint64 new_state = KMP_BARRIER_UNUSED_STATE;
kmp_uint32 num_threads = this_thr->th.th_team_nproc;
kmp_uint32 branch_bits = __kmp_barrier_gather_branch_bits[bt];
kmp_uint32 branch_factor = 1 << branch_bits;
kmp_uint32 offset;
kmp_uint32 level;
KA_TRACE(
20,
("__kmp_hyper_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - save arrive time to the thread
if (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time =
__itt_get_timestamp();
}
#endif
/* Perform a hypercube-embedded tree gather to wait until all of the threads
have arrived, and reduce any required data as we go. */
kmp_flag_64<> p_flag(&thr_bar->b_arrived);
for (level = 0, offset = 1; offset < num_threads;
level += branch_bits, offset <<= branch_bits) {
kmp_uint32 child;
kmp_uint32 child_tid;
if (((tid >> level) & (branch_factor - 1)) != 0) {
kmp_int32 parent_tid = tid & ~((1 << (level + branch_bits)) - 1);
KMP_MB(); // Synchronize parent and child threads.
KA_TRACE(20,
("__kmp_hyper_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d) "
"arrived(%p): %llu => %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(parent_tid, team),
team->t.t_id, parent_tid, &thr_bar->b_arrived,
thr_bar->b_arrived,
thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
// Mark arrival to parent thread
/* After performing this write (in the last iteration of the enclosing for
loop), a worker thread may not assume that the team is valid any more
- it could be deallocated by the primary thread at any time. */
p_flag.set_waiter(other_threads[parent_tid]);
p_flag.release();
break;
}
// Parent threads wait for children to arrive
if (new_state == KMP_BARRIER_UNUSED_STATE)
new_state = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
for (child = 1, child_tid = tid + (1 << level);
child < branch_factor && child_tid < num_threads;
child++, child_tid += (1 << level)) {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
#if KMP_CACHE_MANAGE
kmp_uint32 next_child_tid = child_tid + (1 << level);
// Prefetch next thread's arrived count
if (child + 1 < branch_factor && next_child_tid < num_threads)
KMP_CACHE_PREFETCH(
&other_threads[next_child_tid]->th.th_bar[bt].bb.b_arrived);
#endif /* KMP_CACHE_MANAGE */
KA_TRACE(20,
("__kmp_hyper_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%u) "
"arrived(%p) == %llu\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid, &child_bar->b_arrived, new_state));
// Wait for child to arrive
kmp_flag_64<> c_flag(&child_bar->b_arrived, new_state);
c_flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
KMP_MB(); // Synchronize parent and child threads.
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - write min of the thread time and a child time to
// the thread.
if (__kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_min_time = KMP_MIN(this_thr->th.th_bar_min_time,
child_thr->th.th_bar_min_time);
}
#endif
if (reduce) {
KA_TRACE(100,
("__kmp_hyper_barrier_gather: T#%d(%d:%d) += T#%d(%d:%u)\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid));
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
(*reduce)(this_thr->th.th_local.reduce_data,
child_thr->th.th_local.reduce_data);
OMPT_REDUCTION_END;
}
}
}
if (KMP_MASTER_TID(tid)) {
// Need to update the team arrived pointer if we are the primary thread
if (new_state == KMP_BARRIER_UNUSED_STATE)
team->t.t_bar[bt].b_arrived += KMP_BARRIER_STATE_BUMP;
else
team->t.t_bar[bt].b_arrived = new_state;
KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) set team %d "
"arrived(%p) = %llu\n",
gtid, team->t.t_id, tid, team->t.t_id,
&team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
}
KA_TRACE(
20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
// The reverse versions seem to beat the forward versions overall
#define KMP_REVERSE_HYPER_BAR
static void __kmp_hyper_barrier_release(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_hyper_release);
kmp_team_t *team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_info_t **other_threads;
kmp_uint32 num_threads;
kmp_uint32 branch_bits = __kmp_barrier_release_branch_bits[bt];
kmp_uint32 branch_factor = 1 << branch_bits;
kmp_uint32 child;
kmp_uint32 child_tid;
kmp_uint32 offset;
kmp_uint32 level;
/* Perform a hypercube-embedded tree release for all of the threads that have
been gathered. If KMP_REVERSE_HYPER_BAR is defined (default) the threads
are released in the reverse order of the corresponding gather, otherwise
threads are released in the same order. */
if (KMP_MASTER_TID(tid)) { // primary thread
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d(%d:%d) primary enter for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs) { // primary already has ICVs in final destination; copy
copy_icvs(&thr_bar->th_fixed_icvs,
&team->t.t_implicit_task_taskdata[tid].td_icvs);
}
#endif
} else { // Handle fork barrier workers who aren't part of a team yet
KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d wait go(%p) == %u\n", gtid,
&thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
// Wait for parent thread to release us
kmp_flag_64<> flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
// In fork barrier where we could not get the object reliably
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
// Cancel wait on previous parallel region...
__kmp_itt_task_starting(itt_sync_obj);
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
if (itt_sync_obj != NULL)
// Call prepare as early as possible for "new" barrier
__kmp_itt_task_finished(itt_sync_obj);
} else
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
// Early exit for reaping threads releasing forkjoin barrier
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
// The worker thread may now assume that the team is valid.
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
tid = __kmp_tid_from_gtid(gtid);
TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
KA_TRACE(20,
("__kmp_hyper_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
KMP_MB(); // Flush all pending memory write invalidates.
}
num_threads = this_thr->th.th_team_nproc;
other_threads = team->t.t_threads;
#ifdef KMP_REVERSE_HYPER_BAR
// Count up to correct level for parent
for (level = 0, offset = 1;
offset < num_threads && (((tid >> level) & (branch_factor - 1)) == 0);
level += branch_bits, offset <<= branch_bits)
;
// Now go down from there
for (level -= branch_bits, offset >>= branch_bits; offset != 0;
level -= branch_bits, offset >>= branch_bits)
#else
// Go down the tree, level by level
for (level = 0, offset = 1; offset < num_threads;
level += branch_bits, offset <<= branch_bits)
#endif // KMP_REVERSE_HYPER_BAR
{
#ifdef KMP_REVERSE_HYPER_BAR
/* Now go in reverse order through the children, highest to lowest.
Initial setting of child is conservative here. */
child = num_threads >> ((level == 0) ? level : level - 1);
for (child = (child < branch_factor - 1) ? child : branch_factor - 1,
child_tid = tid + (child << level);
child >= 1; child--, child_tid -= (1 << level))
#else
if (((tid >> level) & (branch_factor - 1)) != 0)
// No need to go lower than this, since this is the level parent would be
// notified
break;
// Iterate through children on this level of the tree
for (child = 1, child_tid = tid + (1 << level);
child < branch_factor && child_tid < num_threads;
child++, child_tid += (1 << level))
#endif // KMP_REVERSE_HYPER_BAR
{
if (child_tid >= num_threads)
continue; // Child doesn't exist so keep going
else {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
#if KMP_CACHE_MANAGE
kmp_uint32 next_child_tid = child_tid - (1 << level);
// Prefetch next thread's go count
#ifdef KMP_REVERSE_HYPER_BAR
if (child - 1 >= 1 && next_child_tid < num_threads)
#else
if (child + 1 < branch_factor && next_child_tid < num_threads)
#endif // KMP_REVERSE_HYPER_BAR
KMP_CACHE_PREFETCH(
&other_threads[next_child_tid]->th.th_bar[bt].bb.b_go);
#endif /* KMP_CACHE_MANAGE */
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs) // push my fixed ICVs to my child
copy_icvs(&child_bar->th_fixed_icvs, &thr_bar->th_fixed_icvs);
#endif // KMP_BARRIER_ICV_PUSH
KA_TRACE(
20,
("__kmp_hyper_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%u)"
"go(%p): %u => %u\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
child_bar->b_go + KMP_BARRIER_STATE_BUMP));
// Release child from barrier
kmp_flag_64<> flag(&child_bar->b_go, child_thr);
flag.release();
}
}
}
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs &&
!KMP_MASTER_TID(tid)) { // copy ICVs locally to final dest
__kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team, tid,
FALSE);
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
&thr_bar->th_fixed_icvs);
}
#endif
KA_TRACE(
20,
("__kmp_hyper_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
// Hierarchical Barrier
// Initialize thread barrier data
/* Initializes/re-initializes the hierarchical barrier data stored on a thread.
Performs the minimum amount of initialization required based on how the team
has changed. Returns true if leaf children will require both on-core and
traditional wake-up mechanisms. For example, if the team size increases,
threads already in the team will respond to on-core wakeup on their parent
thread, but threads newly added to the team will only be listening on the
their local b_go. */
static bool __kmp_init_hierarchical_barrier_thread(enum barrier_type bt,
kmp_bstate_t *thr_bar,
kmp_uint32 nproc, int gtid,
int tid, kmp_team_t *team) {
// Checks to determine if (re-)initialization is needed
bool uninitialized = thr_bar->team == NULL;
bool team_changed = team != thr_bar->team;
bool team_sz_changed = nproc != thr_bar->nproc;
bool tid_changed = tid != thr_bar->old_tid;
bool retval = false;
if (uninitialized || team_sz_changed) {
__kmp_get_hierarchy(nproc, thr_bar);
}
if (uninitialized || team_sz_changed || tid_changed) {
thr_bar->my_level = thr_bar->depth - 1; // default for primary thread
thr_bar->parent_tid = -1; // default for primary thread
if (!KMP_MASTER_TID(tid)) {
// if not primary thread, find parent thread in hierarchy
kmp_uint32 d = 0;
while (d < thr_bar->depth) { // find parent based on level of thread in
// hierarchy, and note level
kmp_uint32 rem;
if (d == thr_bar->depth - 2) { // reached level right below the primary
thr_bar->parent_tid = 0;
thr_bar->my_level = d;
break;
} else if ((rem = tid % thr_bar->skip_per_level[d + 1]) != 0) {
// TODO: can we make the above op faster?
// thread is not a subtree root at next level, so this is max
thr_bar->parent_tid = tid - rem;
thr_bar->my_level = d;
break;
}
++d;
}
}
__kmp_type_convert(7 - ((tid - thr_bar->parent_tid) /
(thr_bar->skip_per_level[thr_bar->my_level])),
&(thr_bar->offset));
thr_bar->old_tid = tid;
thr_bar->wait_flag = KMP_BARRIER_NOT_WAITING;
thr_bar->team = team;
thr_bar->parent_bar =
&team->t.t_threads[thr_bar->parent_tid]->th.th_bar[bt].bb;
}
if (uninitialized || team_changed || tid_changed) {
thr_bar->team = team;
thr_bar->parent_bar =
&team->t.t_threads[thr_bar->parent_tid]->th.th_bar[bt].bb;
retval = true;
}
if (uninitialized || team_sz_changed || tid_changed) {
thr_bar->nproc = nproc;
thr_bar->leaf_kids = thr_bar->base_leaf_kids;
if (thr_bar->my_level == 0)
thr_bar->leaf_kids = 0;
if (thr_bar->leaf_kids && (kmp_uint32)tid + thr_bar->leaf_kids + 1 > nproc)
__kmp_type_convert(nproc - tid - 1, &(thr_bar->leaf_kids));
thr_bar->leaf_state = 0;
for (int i = 0; i < thr_bar->leaf_kids; ++i)
((char *)&(thr_bar->leaf_state))[7 - i] = 1;
}
return retval;
}
static void __kmp_hierarchical_barrier_gather(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
void (*reduce)(void *, void *) USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_hier_gather);
kmp_team_t *team = this_thr->th.th_team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_uint32 nproc = this_thr->th.th_team_nproc;
kmp_info_t **other_threads = team->t.t_threads;
kmp_uint64 new_state = 0;
int level = team->t.t_level;
if (other_threads[0]
->th.th_teams_microtask) // are we inside the teams construct?
if (this_thr->th.th_teams_size.nteams > 1)
++level; // level was not increased in teams construct for team_of_masters
if (level == 1)
thr_bar->use_oncore_barrier = 1;
else
thr_bar->use_oncore_barrier = 0; // Do not use oncore barrier when nested
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) enter for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier imbalance - save arrive time to the thread
if (__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
this_thr->th.th_bar_arrive_time = __itt_get_timestamp();
}
#endif
(void)__kmp_init_hierarchical_barrier_thread(bt, thr_bar, nproc, gtid, tid,
team);
if (thr_bar->my_level) { // not a leaf (my_level==0 means leaf)
kmp_int32 child_tid;
new_state =
(kmp_uint64)team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
thr_bar->use_oncore_barrier) {
if (thr_bar->leaf_kids) {
// First, wait for leaf children to check-in on my b_arrived flag
kmp_uint64 leaf_state =
KMP_MASTER_TID(tid)
? thr_bar->b_arrived | thr_bar->leaf_state
: team->t.t_bar[bt].b_arrived | thr_bar->leaf_state;
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) waiting "
"for leaf kids\n",
gtid, team->t.t_id, tid));
kmp_flag_64<> flag(&thr_bar->b_arrived, leaf_state);
flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
if (reduce) {
OMPT_REDUCTION_DECL(this_thr, gtid);
OMPT_REDUCTION_BEGIN;
for (child_tid = tid + 1; child_tid <= tid + thr_bar->leaf_kids;
++child_tid) {
KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += "
"T#%d(%d:%d)\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid));
(*reduce)(this_thr->th.th_local.reduce_data,
other_threads[child_tid]->th.th_local.reduce_data);
}
OMPT_REDUCTION_END;
}
// clear leaf_state bits
KMP_TEST_THEN_AND64(&thr_bar->b_arrived, ~(thr_bar->leaf_state));
}
// Next, wait for higher level children on each child's b_arrived flag
for (kmp_uint32 d = 1; d < thr_bar->my_level;
++d) { // gather lowest level threads first, but skip 0
kmp_uint32 last = tid + thr_bar->skip_per_level[d + 1],
skip = thr_bar->skip_per_level[d];
if (last > nproc)
last = nproc;
for (child_tid = tid + skip; child_tid < (int)last; child_tid += skip) {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) wait "
"T#%d(%d:%d) "
"arrived(%p) == %llu\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid, &child_bar->b_arrived, new_state));
kmp_flag_64<> flag(&child_bar->b_arrived, new_state);
flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
if (reduce) {
KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += "
"T#%d(%d:%d)\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid));
(*reduce)(this_thr->th.th_local.reduce_data,
child_thr->th.th_local.reduce_data);
}
}
}
} else { // Blocktime is not infinite
for (kmp_uint32 d = 0; d < thr_bar->my_level;
++d) { // Gather lowest level threads first
kmp_uint32 last = tid + thr_bar->skip_per_level[d + 1],
skip = thr_bar->skip_per_level[d];
if (last > nproc)
last = nproc;
for (child_tid = tid + skip; child_tid < (int)last; child_tid += skip) {
kmp_info_t *child_thr = other_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) wait "
"T#%d(%d:%d) "
"arrived(%p) == %llu\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid, &child_bar->b_arrived, new_state));
kmp_flag_64<> flag(&child_bar->b_arrived, new_state);
flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
if (reduce) {
KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += "
"T#%d(%d:%d)\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid));
(*reduce)(this_thr->th.th_local.reduce_data,
child_thr->th.th_local.reduce_data);
}
}
}
}
}
// All subordinates are gathered; now release parent if not primary thread
if (!KMP_MASTER_TID(tid)) { // worker threads release parent in hierarchy
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) releasing"
" T#%d(%d:%d) arrived(%p): %llu => %llu\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(thr_bar->parent_tid, team), team->t.t_id,
thr_bar->parent_tid, &thr_bar->b_arrived, thr_bar->b_arrived,
thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
/* Mark arrival to parent: After performing this write, a worker thread may
not assume that the team is valid any more - it could be deallocated by
the primary thread at any time. */
if (thr_bar->my_level || __kmp_dflt_blocktime != KMP_MAX_BLOCKTIME ||
!thr_bar->use_oncore_barrier) { // Parent is waiting on my b_arrived
// flag; release it
kmp_flag_64<> flag(&thr_bar->b_arrived,
other_threads[thr_bar->parent_tid]);
flag.release();
} else {
// Leaf does special release on "offset" bits of parent's b_arrived flag
thr_bar->b_arrived = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
kmp_flag_oncore flag(&thr_bar->parent_bar->b_arrived,
thr_bar->offset + 1);
flag.set_waiter(other_threads[thr_bar->parent_tid]);
flag.release();
}
} else { // Primary thread needs to update the team's b_arrived value
team->t.t_bar[bt].b_arrived = new_state;
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) set team %d "
"arrived(%p) = %llu\n",
gtid, team->t.t_id, tid, team->t.t_id,
&team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
}
// Is the team access below unsafe or just technically invalid?
KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) exit for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
static void __kmp_hierarchical_barrier_release(
enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
int propagate_icvs USE_ITT_BUILD_ARG(void *itt_sync_obj)) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_hier_release);
kmp_team_t *team;
kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
kmp_uint32 nproc;
bool team_change = false; // indicates on-core barrier shouldn't be used
if (KMP_MASTER_TID(tid)) {
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) primary "
"entered barrier type %d\n",
gtid, team->t.t_id, tid, bt));
} else { // Worker threads
// Wait for parent thread to release me
if (!thr_bar->use_oncore_barrier ||
__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME || thr_bar->my_level != 0 ||
thr_bar->team == NULL) {
// Use traditional method of waiting on my own b_go flag
thr_bar->wait_flag = KMP_BARRIER_OWN_FLAG;
kmp_flag_64<> flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
TCW_8(thr_bar->b_go,
KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
} else { // Thread barrier data is initialized, this is a leaf, blocktime is
// infinite, not nested
// Wait on my "offset" bits on parent's b_go flag
thr_bar->wait_flag = KMP_BARRIER_PARENT_FLAG;
kmp_flag_oncore flag(&thr_bar->parent_bar->b_go, KMP_BARRIER_STATE_BUMP,
thr_bar->offset + 1, bt,
this_thr USE_ITT_BUILD_ARG(itt_sync_obj));
flag.wait(this_thr, TRUE);
if (thr_bar->wait_flag ==
KMP_BARRIER_SWITCHING) { // Thread was switched to own b_go
TCW_8(thr_bar->b_go,
KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
} else { // Reset my bits on parent's b_go flag
(RCAST(volatile char *,
&(thr_bar->parent_bar->b_go)))[thr_bar->offset + 1] = 0;
}
}
thr_bar->wait_flag = KMP_BARRIER_NOT_WAITING;
// Early exit for reaping threads releasing forkjoin barrier
if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
return;
// The worker thread may now assume that the team is valid.
team = __kmp_threads[gtid]->th.th_team;
KMP_DEBUG_ASSERT(team != NULL);
tid = __kmp_tid_from_gtid(gtid);
KA_TRACE(
20,
("__kmp_hierarchical_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
KMP_MB(); // Flush all pending memory write invalidates.
}
nproc = this_thr->th.th_team_nproc;
int level = team->t.t_level;
if (team->t.t_threads[0]
->th.th_teams_microtask) { // are we inside the teams construct?
if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
this_thr->th.th_teams_level == level)
++level; // level was not increased in teams construct for team_of_workers
if (this_thr->th.th_teams_size.nteams > 1)
++level; // level was not increased in teams construct for team_of_masters
}
if (level == 1)
thr_bar->use_oncore_barrier = 1;
else
thr_bar->use_oncore_barrier = 0; // Do not use oncore barrier when nested
// If the team size has increased, we still communicate with old leaves via
// oncore barrier.
unsigned short int old_leaf_kids = thr_bar->leaf_kids;
kmp_uint64 old_leaf_state = thr_bar->leaf_state;
team_change = __kmp_init_hierarchical_barrier_thread(bt, thr_bar, nproc, gtid,
tid, team);
// But if the entire team changes, we won't use oncore barrier at all
if (team_change)
old_leaf_kids = 0;
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs) {
__kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team, tid,
FALSE);
if (KMP_MASTER_TID(
tid)) { // primary already has copy in final destination; copy
copy_icvs(&thr_bar->th_fixed_icvs,
&team->t.t_implicit_task_taskdata[tid].td_icvs);
} else if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
thr_bar->use_oncore_barrier) { // optimization for inf blocktime
if (!thr_bar->my_level) // I'm a leaf in the hierarchy (my_level==0)
// leaves (on-core children) pull parent's fixed ICVs directly to local
// ICV store
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
&thr_bar->parent_bar->th_fixed_icvs);
// non-leaves will get ICVs piggybacked with b_go via NGO store
} else { // blocktime is not infinite; pull ICVs from parent's fixed ICVs
if (thr_bar->my_level) // not a leaf; copy ICVs to my fixed ICVs child can
// access
copy_icvs(&thr_bar->th_fixed_icvs, &thr_bar->parent_bar->th_fixed_icvs);
else // leaves copy parent's fixed ICVs directly to local ICV store
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
&thr_bar->parent_bar->th_fixed_icvs);
}
}
#endif // KMP_BARRIER_ICV_PUSH
// Now, release my children
if (thr_bar->my_level) { // not a leaf
kmp_int32 child_tid;
kmp_uint32 last;
if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
thr_bar->use_oncore_barrier) {
if (KMP_MASTER_TID(tid)) { // do a flat release
// Set local b_go to bump children via NGO store of the cache line
// containing IVCs and b_go.
thr_bar->b_go = KMP_BARRIER_STATE_BUMP;
// Use ngo stores if available; b_go piggybacks in the last 8 bytes of
// the cache line
ngo_load(&thr_bar->th_fixed_icvs);
// This loops over all the threads skipping only the leaf nodes in the
// hierarchy
for (child_tid = thr_bar->skip_per_level[1]; child_tid < (int)nproc;
child_tid += thr_bar->skip_per_level[1]) {
kmp_bstate_t *child_bar =
&team->t.t_threads[child_tid]->th.th_bar[bt].bb;
KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) "
"releasing T#%d(%d:%d)"
" go(%p): %u => %u\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid, &child_bar->b_go, child_bar->b_go,
child_bar->b_go + KMP_BARRIER_STATE_BUMP));
// Use ngo store (if available) to both store ICVs and release child
// via child's b_go
ngo_store_go(&child_bar->th_fixed_icvs, &thr_bar->th_fixed_icvs);
}
ngo_sync();
}
TCW_8(thr_bar->b_go,
KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
// Now, release leaf children
if (thr_bar->leaf_kids) { // if there are any
// We test team_change on the off-chance that the level 1 team changed.
if (team_change ||
old_leaf_kids < thr_bar->leaf_kids) { // some old, some new
if (old_leaf_kids) { // release old leaf kids
thr_bar->b_go |= old_leaf_state;
}
// Release new leaf kids
last = tid + thr_bar->skip_per_level[1];
if (last > nproc)
last = nproc;
for (child_tid = tid + 1 + old_leaf_kids; child_tid < (int)last;
++child_tid) { // skip_per_level[0]=1
kmp_info_t *child_thr = team->t.t_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
KA_TRACE(
20,
("__kmp_hierarchical_barrier_release: T#%d(%d:%d) releasing"
" T#%d(%d:%d) go(%p): %u => %u\n",
gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
child_bar->b_go + KMP_BARRIER_STATE_BUMP));
// Release child using child's b_go flag
kmp_flag_64<> flag(&child_bar->b_go, child_thr);
flag.release();
}
} else { // Release all children at once with leaf_state bits on my own
// b_go flag
thr_bar->b_go |= thr_bar->leaf_state;
}
}
} else { // Blocktime is not infinite; do a simple hierarchical release
for (int d = thr_bar->my_level - 1; d >= 0;
--d) { // Release highest level threads first
last = tid + thr_bar->skip_per_level[d + 1];
kmp_uint32 skip = thr_bar->skip_per_level[d];
if (last > nproc)
last = nproc;
for (child_tid = tid + skip; child_tid < (int)last; child_tid += skip) {
kmp_info_t *child_thr = team->t.t_threads[child_tid];
kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) "
"releasing T#%d(%d:%d) go(%p): %u => %u\n",
gtid, team->t.t_id, tid,
__kmp_gtid_from_tid(child_tid, team), team->t.t_id,
child_tid, &child_bar->b_go, child_bar->b_go,
child_bar->b_go + KMP_BARRIER_STATE_BUMP));
// Release child using child's b_go flag
kmp_flag_64<> flag(&child_bar->b_go, child_thr);
flag.release();
}
}
}
#if KMP_BARRIER_ICV_PUSH
if (propagate_icvs && !KMP_MASTER_TID(tid))
// non-leaves copy ICVs from fixed ICVs to local dest
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
&thr_bar->th_fixed_icvs);
#endif // KMP_BARRIER_ICV_PUSH
}
KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) exit for "
"barrier type %d\n",
gtid, team->t.t_id, tid, bt));
}
// End of Barrier Algorithms
// type traits for cancellable value
// if cancellable is true, then is_cancellable is a normal boolean variable
// if cancellable is false, then is_cancellable is a compile time constant
template <bool cancellable> struct is_cancellable {};
template <> struct is_cancellable<true> {
bool value;
is_cancellable() : value(false) {}
is_cancellable(bool b) : value(b) {}
is_cancellable &operator=(bool b) {
value = b;
return *this;
}
operator bool() const { return value; }
};
template <> struct is_cancellable<false> {
is_cancellable &operator=(bool b) { return *this; }
constexpr operator bool() const { return false; }
};
// Internal function to do a barrier.
/* If is_split is true, do a split barrier, otherwise, do a plain barrier
If reduce is non-NULL, do a split reduction barrier, otherwise, do a split
barrier
When cancellable = false,
Returns 0 if primary thread, 1 if worker thread.
When cancellable = true
Returns 0 if not cancelled, 1 if cancelled. */
template <bool cancellable = false>
static int __kmp_barrier_template(enum barrier_type bt, int gtid, int is_split,
size_t reduce_size, void *reduce_data,
void (*reduce)(void *, void *)) {
KMP_TIME_PARTITIONED_BLOCK(OMP_plain_barrier);
KMP_SET_THREAD_STATE_BLOCK(PLAIN_BARRIER);
int tid = __kmp_tid_from_gtid(gtid);
kmp_info_t *this_thr = __kmp_threads[gtid];
kmp_team_t *team = this_thr->th.th_team;
int status = 0;
is_cancellable<cancellable> cancelled;
#if OMPT_SUPPORT && OMPT_OPTIONAL
ompt_data_t *my_task_data;
ompt_data_t *my_parallel_data;
void *return_address;
ompt_sync_region_t barrier_kind;
#endif
KA_TRACE(15, ("__kmp_barrier: T#%d(%d:%d) has arrived\n", gtid,
__kmp_team_from_gtid(gtid)->t.t_id, __kmp_tid_from_gtid(gtid)));
#if OMPT_SUPPORT
if (ompt_enabled.enabled) {
#if OMPT_OPTIONAL
my_task_data = OMPT_CUR_TASK_DATA(this_thr);
my_parallel_data = OMPT_CUR_TEAM_DATA(this_thr);
return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
barrier_kind = __ompt_get_barrier_kind(bt, this_thr);
if (ompt_enabled.ompt_callback_sync_region) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
barrier_kind, ompt_scope_begin, my_parallel_data, my_task_data,
return_address);
}
if (ompt_enabled.ompt_callback_sync_region_wait) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
barrier_kind, ompt_scope_begin, my_parallel_data, my_task_data,
return_address);
}
#endif
// It is OK to report the barrier state after the barrier begin callback.
// According to the OMPT specification, a compliant implementation may
// even delay reporting this state until the barrier begins to wait.
this_thr->th.ompt_thread_info.state = ompt_state_wait_barrier;
}
#endif
if (!team->t.t_serialized) {
#if USE_ITT_BUILD
// This value will be used in itt notify events below.
void *itt_sync_obj = NULL;
#if USE_ITT_NOTIFY
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
itt_sync_obj = __kmp_itt_barrier_object(gtid, bt, 1);
#endif
#endif /* USE_ITT_BUILD */
if (__kmp_tasking_mode == tskm_extra_barrier) {
__kmp_tasking_barrier(team, this_thr, gtid);
KA_TRACE(15,
("__kmp_barrier: T#%d(%d:%d) past tasking barrier\n", gtid,
__kmp_team_from_gtid(gtid)->t.t_id, __kmp_tid_from_gtid(gtid)));
}
/* Copy the blocktime info to the thread, where __kmp_wait_template() can
access it when the team struct is not guaranteed to exist. */
// See note about the corresponding code in __kmp_join_barrier() being
// performance-critical.
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
#if KMP_USE_MONITOR
this_thr->th.th_team_bt_intervals =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
this_thr->th.th_team_bt_set =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
#else
this_thr->th.th_team_bt_intervals = KMP_BLOCKTIME_INTERVAL(team, tid);
#endif
}
#if USE_ITT_BUILD
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_starting(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
#if USE_DEBUGGER
// Let the debugger know: the thread arrived to the barrier and waiting.
if (KMP_MASTER_TID(tid)) { // Primary thread counter stored in team struct
team->t.t_bar[bt].b_master_arrived += 1;
} else {
this_thr->th.th_bar[bt].bb.b_worker_arrived += 1;
} // if
#endif /* USE_DEBUGGER */
if (reduce != NULL) {
// KMP_DEBUG_ASSERT( is_split == TRUE ); // #C69956
this_thr->th.th_local.reduce_data = reduce_data;
}
if (KMP_MASTER_TID(tid) && __kmp_tasking_mode != tskm_immediate_exec)
// use 0 to only setup the current team if nthreads > 1
__kmp_task_team_setup(this_thr, team, 0);
if (cancellable) {
cancelled = __kmp_linear_barrier_gather_cancellable(
bt, this_thr, gtid, tid, reduce USE_ITT_BUILD_ARG(itt_sync_obj));
} else {
switch (__kmp_barrier_gather_pattern[bt]) {
case bp_dist_bar: {
__kmp_dist_barrier_gather(bt, this_thr, gtid, tid,
reduce USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hyper_bar: {
// don't set branch bits to 0; use linear
KMP_ASSERT(__kmp_barrier_gather_branch_bits[bt]);
__kmp_hyper_barrier_gather(bt, this_thr, gtid, tid,
reduce USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hierarchical_bar: {
__kmp_hierarchical_barrier_gather(
bt, this_thr, gtid, tid, reduce USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_tree_bar: {
// don't set branch bits to 0; use linear
KMP_ASSERT(__kmp_barrier_gather_branch_bits[bt]);
__kmp_tree_barrier_gather(bt, this_thr, gtid, tid,
reduce USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
default: {
__kmp_linear_barrier_gather(bt, this_thr, gtid, tid,
reduce USE_ITT_BUILD_ARG(itt_sync_obj));
}
}
}
KMP_MB();
if (KMP_MASTER_TID(tid)) {
status = 0;
if (__kmp_tasking_mode != tskm_immediate_exec && !cancelled) {
__kmp_task_team_wait(this_thr, team USE_ITT_BUILD_ARG(itt_sync_obj));
}
#if USE_DEBUGGER
// Let the debugger know: All threads are arrived and starting leaving the
// barrier.
team->t.t_bar[bt].b_team_arrived += 1;
#endif
if (__kmp_omp_cancellation) {
kmp_int32 cancel_request = KMP_ATOMIC_LD_RLX(&team->t.t_cancel_request);
// Reset cancellation flag for worksharing constructs
if (cancel_request == cancel_loop ||
cancel_request == cancel_sections) {
KMP_ATOMIC_ST_RLX(&team->t.t_cancel_request, cancel_noreq);
}
}
#if USE_ITT_BUILD
/* TODO: In case of split reduction barrier, primary thread may send
acquired event early, before the final summation into the shared
variable is done (final summation can be a long operation for array
reductions). */
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_middle(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Barrier - report frame end (only if active_level == 1)
if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
__kmp_forkjoin_frames_mode &&
(this_thr->th.th_teams_microtask == NULL || // either not in teams
this_thr->th.th_teams_size.nteams == 1) && // or inside single team
team->t.t_active_level == 1) {
ident_t *loc = __kmp_threads[gtid]->th.th_ident;
kmp_uint64 cur_time = __itt_get_timestamp();
kmp_info_t **other_threads = team->t.t_threads;
int nproc = this_thr->th.th_team_nproc;
int i;
switch (__kmp_forkjoin_frames_mode) {
case 1:
__kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0,
loc, nproc);
this_thr->th.th_frame_time = cur_time;
break;
case 2: // AC 2015-01-19: currently does not work for hierarchical (to
// be fixed)
__kmp_itt_frame_submit(gtid, this_thr->th.th_bar_min_time, cur_time,
1, loc, nproc);
break;
case 3:
if (__itt_metadata_add_ptr) {
// Initialize with primary thread's wait time
kmp_uint64 delta = cur_time - this_thr->th.th_bar_arrive_time;
// Set arrive time to zero to be able to check it in
// __kmp_invoke_task(); the same is done inside the loop below
this_thr->th.th_bar_arrive_time = 0;
for (i = 1; i < nproc; ++i) {
delta += (cur_time - other_threads[i]->th.th_bar_arrive_time);
other_threads[i]->th.th_bar_arrive_time = 0;
}
__kmp_itt_metadata_imbalance(gtid, this_thr->th.th_frame_time,
cur_time, delta,
(kmp_uint64)(reduce != NULL));
}
__kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0,
loc, nproc);
this_thr->th.th_frame_time = cur_time;
break;
}
}
#endif /* USE_ITT_BUILD */
} else {
status = 1;
#if USE_ITT_BUILD
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_middle(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
}
if ((status == 1 || !is_split) && !cancelled) {
if (cancellable) {
cancelled = __kmp_linear_barrier_release_cancellable(
bt, this_thr, gtid, tid, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
} else {
switch (__kmp_barrier_release_pattern[bt]) {
case bp_dist_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
__kmp_dist_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hyper_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
__kmp_hyper_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hierarchical_bar: {
__kmp_hierarchical_barrier_release(
bt, this_thr, gtid, tid, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_tree_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
__kmp_tree_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
default: {
__kmp_linear_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
}
}
}
if (__kmp_tasking_mode != tskm_immediate_exec && !cancelled) {
__kmp_task_team_sync(this_thr, team);
}
}
#if USE_ITT_BUILD
/* GEH: TODO: Move this under if-condition above and also include in
__kmp_end_split_barrier(). This will more accurately represent the actual
release time of the threads for split barriers. */
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_finished(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
} else { // Team is serialized.
status = 0;
if (__kmp_tasking_mode != tskm_immediate_exec) {
if (this_thr->th.th_task_team != NULL) {
#if USE_ITT_NOTIFY
void *itt_sync_obj = NULL;
if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
itt_sync_obj = __kmp_itt_barrier_object(gtid, bt, 1);
__kmp_itt_barrier_starting(gtid, itt_sync_obj);
}
#endif
KMP_DEBUG_ASSERT(
this_thr->th.th_task_team->tt.tt_found_proxy_tasks == TRUE ||
this_thr->th.th_task_team->tt.tt_hidden_helper_task_encountered ==
TRUE);
__kmp_task_team_wait(this_thr, team USE_ITT_BUILD_ARG(itt_sync_obj));
__kmp_task_team_setup(this_thr, team, 0);
#if USE_ITT_BUILD
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_finished(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
}
}
}
KA_TRACE(15, ("__kmp_barrier: T#%d(%d:%d) is leaving with return value %d\n",
gtid, __kmp_team_from_gtid(gtid)->t.t_id,
__kmp_tid_from_gtid(gtid), status));
#if OMPT_SUPPORT
if (ompt_enabled.enabled) {
#if OMPT_OPTIONAL
if (ompt_enabled.ompt_callback_sync_region_wait) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
barrier_kind, ompt_scope_end, my_parallel_data, my_task_data,
return_address);
}
if (ompt_enabled.ompt_callback_sync_region) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
barrier_kind, ompt_scope_end, my_parallel_data, my_task_data,
return_address);
}
#endif
this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
}
#endif
if (cancellable)
return (int)cancelled;
return status;
}
// Returns 0 if primary thread, 1 if worker thread.
int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
size_t reduce_size, void *reduce_data,
void (*reduce)(void *, void *)) {
return __kmp_barrier_template<>(bt, gtid, is_split, reduce_size, reduce_data,
reduce);
}
#if defined(KMP_GOMP_COMPAT)
// Returns 1 if cancelled, 0 otherwise
int __kmp_barrier_gomp_cancel(int gtid) {
if (__kmp_omp_cancellation) {
int cancelled = __kmp_barrier_template<true>(bs_plain_barrier, gtid, FALSE,
0, NULL, NULL);
if (cancelled) {
int tid = __kmp_tid_from_gtid(gtid);
kmp_info_t *this_thr = __kmp_threads[gtid];
if (KMP_MASTER_TID(tid)) {
// Primary thread does not need to revert anything
} else {
// Workers need to revert their private b_arrived flag
this_thr->th.th_bar[bs_plain_barrier].bb.b_arrived -=
KMP_BARRIER_STATE_BUMP;
}
}
return cancelled;
}
__kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
return FALSE;
}
#endif
void __kmp_end_split_barrier(enum barrier_type bt, int gtid) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_end_split_barrier);
KMP_SET_THREAD_STATE_BLOCK(PLAIN_BARRIER);
KMP_DEBUG_ASSERT(bt < bs_last_barrier);
int tid = __kmp_tid_from_gtid(gtid);
kmp_info_t *this_thr = __kmp_threads[gtid];
kmp_team_t *team = this_thr->th.th_team;
if (!team->t.t_serialized) {
if (KMP_MASTER_GTID(gtid)) {
switch (__kmp_barrier_release_pattern[bt]) {
case bp_dist_bar: {
__kmp_dist_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(NULL));
break;
}
case bp_hyper_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
__kmp_hyper_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(NULL));
break;
}
case bp_hierarchical_bar: {
__kmp_hierarchical_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(NULL));
break;
}
case bp_tree_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
__kmp_tree_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(NULL));
break;
}
default: {
__kmp_linear_barrier_release(bt, this_thr, gtid, tid,
FALSE USE_ITT_BUILD_ARG(NULL));
}
}
if (__kmp_tasking_mode != tskm_immediate_exec) {
__kmp_task_team_sync(this_thr, team);
} // if
}
}
}
void __kmp_join_barrier(int gtid) {
KMP_TIME_PARTITIONED_BLOCK(OMP_join_barrier);
KMP_SET_THREAD_STATE_BLOCK(FORK_JOIN_BARRIER);
KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
kmp_info_t *this_thr = __kmp_threads[gtid];
kmp_team_t *team;
int tid;
#ifdef KMP_DEBUG
int team_id;
#endif /* KMP_DEBUG */
#if USE_ITT_BUILD
void *itt_sync_obj = NULL;
#if USE_ITT_NOTIFY
if (__itt_sync_create_ptr || KMP_ITT_DEBUG) // Don't call routine without need
// Get object created at fork_barrier
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
#endif
#endif /* USE_ITT_BUILD */
#if ((USE_ITT_BUILD && USE_ITT_NOTIFY) || defined KMP_DEBUG)
int nproc = this_thr->th.th_team_nproc;
#endif
KMP_MB();
// Get current info
team = this_thr->th.th_team;
KMP_DEBUG_ASSERT(nproc == team->t.t_nproc);
tid = __kmp_tid_from_gtid(gtid);
#ifdef KMP_DEBUG
team_id = team->t.t_id;
kmp_info_t *master_thread = this_thr->th.th_team_master;
if (master_thread != team->t.t_threads[0]) {
__kmp_print_structure();
}
#endif /* KMP_DEBUG */
KMP_DEBUG_ASSERT(master_thread == team->t.t_threads[0]);
KMP_MB();
// Verify state
KMP_DEBUG_ASSERT(TCR_PTR(this_thr->th.th_team));
KMP_DEBUG_ASSERT(TCR_PTR(this_thr->th.th_root));
KMP_DEBUG_ASSERT(this_thr == team->t.t_threads[tid]);
KA_TRACE(10, ("__kmp_join_barrier: T#%d(%d:%d) arrived at join barrier\n",
gtid, team_id, tid));
#if OMPT_SUPPORT
if (ompt_enabled.enabled) {
#if OMPT_OPTIONAL
ompt_data_t *my_task_data;
ompt_data_t *my_parallel_data;
void *codeptr = NULL;
int ds_tid = this_thr->th.th_info.ds.ds_tid;
if (KMP_MASTER_TID(ds_tid) &&
(ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
codeptr = team->t.ompt_team_info.master_return_address;
my_task_data = OMPT_CUR_TASK_DATA(this_thr);
my_parallel_data = OMPT_CUR_TEAM_DATA(this_thr);
if (ompt_enabled.ompt_callback_sync_region) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
ompt_sync_region_barrier_implicit, ompt_scope_begin, my_parallel_data,
my_task_data, codeptr);
}
if (ompt_enabled.ompt_callback_sync_region_wait) {
ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
ompt_sync_region_barrier_implicit, ompt_scope_begin, my_parallel_data,
my_task_data, codeptr);
}
if (!KMP_MASTER_TID(ds_tid))
this_thr->th.ompt_thread_info.task_data = *OMPT_CUR_TASK_DATA(this_thr);
#endif
this_thr->th.ompt_thread_info.state = ompt_state_wait_barrier_implicit;
}
#endif
if (__kmp_tasking_mode == tskm_extra_barrier) {
__kmp_tasking_barrier(team, this_thr, gtid);
KA_TRACE(10, ("__kmp_join_barrier: T#%d(%d:%d) past tasking barrier\n",
gtid, team_id, tid));
}
#ifdef KMP_DEBUG
if (__kmp_tasking_mode != tskm_immediate_exec) {
KA_TRACE(20, ("__kmp_join_barrier: T#%d, old team = %d, old task_team = "
"%p, th_task_team = %p\n",
__kmp_gtid_from_thread(this_thr), team_id,
team->t.t_task_team[this_thr->th.th_task_state],
this_thr->th.th_task_team));
if (this_thr->th.th_task_team)
KMP_DEBUG_ASSERT(this_thr->th.th_task_team ==
team->t.t_task_team[this_thr->th.th_task_state]);
}
#endif /* KMP_DEBUG */
/* Copy the blocktime info to the thread, where __kmp_wait_template() can
access it when the team struct is not guaranteed to exist. Doing these
loads causes a cache miss slows down EPCC parallel by 2x. As a workaround,
we do not perform the copy if blocktime=infinite, since the values are not
used by __kmp_wait_template() in that case. */
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
#if KMP_USE_MONITOR
this_thr->th.th_team_bt_intervals =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
this_thr->th.th_team_bt_set =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
#else
this_thr->th.th_team_bt_intervals = KMP_BLOCKTIME_INTERVAL(team, tid);
#endif
}
#if USE_ITT_BUILD
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_starting(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
switch (__kmp_barrier_gather_pattern[bs_forkjoin_barrier]) {
case bp_dist_bar: {
__kmp_dist_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid,
NULL USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hyper_bar: {
KMP_ASSERT(__kmp_barrier_gather_branch_bits[bs_forkjoin_barrier]);
__kmp_hyper_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid,
NULL USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hierarchical_bar: {
__kmp_hierarchical_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid,
NULL USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_tree_bar: {
KMP_ASSERT(__kmp_barrier_gather_branch_bits[bs_forkjoin_barrier]);
__kmp_tree_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid,
NULL USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
default: {
__kmp_linear_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid,
NULL USE_ITT_BUILD_ARG(itt_sync_obj));
}
}
/* From this point on, the team data structure may be deallocated at any time
by the primary thread - it is unsafe to reference it in any of the worker
threads. Any per-team data items that need to be referenced before the
end of the barrier should be moved to the kmp_task_team_t structs. */
if (KMP_MASTER_TID(tid)) {
if (__kmp_tasking_mode != tskm_immediate_exec) {
__kmp_task_team_wait(this_thr, team USE_ITT_BUILD_ARG(itt_sync_obj));
}
if (__kmp_display_affinity) {
KMP_CHECK_UPDATE(team->t.t_display_affinity, 0);
}
#if KMP_STATS_ENABLED
// Have primary thread flag the workers to indicate they are now waiting for
// next parallel region, Also wake them up so they switch their timers to
// idle.
for (int i = 0; i < team->t.t_nproc; ++i) {
kmp_info_t *team_thread = team->t.t_threads[i];
if (team_thread == this_thr)
continue;
team_thread->th.th_stats->setIdleFlag();
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME &&
team_thread->th.th_sleep_loc != NULL)
__kmp_null_resume_wrapper(team_thread);
}
#endif
#if USE_ITT_BUILD
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_middle(gtid, itt_sync_obj);
#endif /* USE_ITT_BUILD */
#if USE_ITT_BUILD && USE_ITT_NOTIFY
// Join barrier - report frame end
if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
__kmp_forkjoin_frames_mode &&
(this_thr->th.th_teams_microtask == NULL || // either not in teams
this_thr->th.th_teams_size.nteams == 1) && // or inside single team
team->t.t_active_level == 1) {
kmp_uint64 cur_time = __itt_get_timestamp();
ident_t *loc = team->t.t_ident;
kmp_info_t **other_threads = team->t.t_threads;
switch (__kmp_forkjoin_frames_mode) {
case 1:
__kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0,
loc, nproc);
break;
case 2:
__kmp_itt_frame_submit(gtid, this_thr->th.th_bar_min_time, cur_time, 1,
loc, nproc);
break;
case 3:
if (__itt_metadata_add_ptr) {
// Initialize with primary thread's wait time
kmp_uint64 delta = cur_time - this_thr->th.th_bar_arrive_time;
// Set arrive time to zero to be able to check it in
// __kmp_invoke_task(); the same is done inside the loop below
this_thr->th.th_bar_arrive_time = 0;
for (int i = 1; i < nproc; ++i) {
delta += (cur_time - other_threads[i]->th.th_bar_arrive_time);
other_threads[i]->th.th_bar_arrive_time = 0;
}
__kmp_itt_metadata_imbalance(gtid, this_thr->th.th_frame_time,
cur_time, delta, 0);
}
__kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0,
loc, nproc);
this_thr->th.th_frame_time = cur_time;
break;
}
}
#endif /* USE_ITT_BUILD */
}
#if USE_ITT_BUILD
else {
if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
__kmp_itt_barrier_middle(gtid, itt_sync_obj);
}
#endif /* USE_ITT_BUILD */
#if KMP_DEBUG
if (KMP_MASTER_TID(tid)) {
KA_TRACE(
15,
("__kmp_join_barrier: T#%d(%d:%d) says all %d team threads arrived\n",
gtid, team_id, tid, nproc));
}
#endif /* KMP_DEBUG */
// TODO now, mark worker threads as done so they may be disbanded
KMP_MB(); // Flush all pending memory write invalidates.
KA_TRACE(10,
("__kmp_join_barrier: T#%d(%d:%d) leaving\n", gtid, team_id, tid));
}
// TODO release worker threads' fork barriers as we are ready instead of all at
// once
void __kmp_fork_barrier(int gtid, int tid) {
KMP_TIME_PARTITIONED_BLOCK(OMP_fork_barrier);
KMP_SET_THREAD_STATE_BLOCK(FORK_JOIN_BARRIER);
kmp_info_t *this_thr = __kmp_threads[gtid];
kmp_team_t *team = (tid == 0) ? this_thr->th.th_team : NULL;
#if USE_ITT_BUILD
void *itt_sync_obj = NULL;
#endif /* USE_ITT_BUILD */
if (team)
KA_TRACE(10, ("__kmp_fork_barrier: T#%d(%d:%d) has arrived\n", gtid,
(team != NULL) ? team->t.t_id : -1, tid));
// th_team pointer only valid for primary thread here
if (KMP_MASTER_TID(tid)) {
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
// Create itt barrier object
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 1);
__kmp_itt_barrier_middle(gtid, itt_sync_obj); // Call acquired/releasing
}
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
#ifdef KMP_DEBUG
KMP_DEBUG_ASSERT(team);
kmp_info_t **other_threads = team->t.t_threads;
int i;
// Verify state
KMP_MB();
for (i = 1; i < team->t.t_nproc; ++i) {
KA_TRACE(500,
("__kmp_fork_barrier: T#%d(%d:0) checking T#%d(%d:%d) fork go "
"== %u.\n",
gtid, team->t.t_id, other_threads[i]->th.th_info.ds.ds_gtid,
team->t.t_id, other_threads[i]->th.th_info.ds.ds_tid,
other_threads[i]->th.th_bar[bs_forkjoin_barrier].bb.b_go));
KMP_DEBUG_ASSERT(
(TCR_4(other_threads[i]->th.th_bar[bs_forkjoin_barrier].bb.b_go) &
~(KMP_BARRIER_SLEEP_STATE)) == KMP_INIT_BARRIER_STATE);
KMP_DEBUG_ASSERT(other_threads[i]->th.th_team == team);
}
#endif
if (__kmp_tasking_mode != tskm_immediate_exec) {
// 0 indicates setup current task team if nthreads > 1
__kmp_task_team_setup(this_thr, team, 0);
}
/* The primary thread may have changed its blocktime between join barrier
and fork barrier. Copy the blocktime info to the thread, where
__kmp_wait_template() can access it when the team struct is not
guaranteed to exist. */
// See note about the corresponding code in __kmp_join_barrier() being
// performance-critical
if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
#if KMP_USE_MONITOR
this_thr->th.th_team_bt_intervals =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
this_thr->th.th_team_bt_set =
team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
#else
this_thr->th.th_team_bt_intervals = KMP_BLOCKTIME_INTERVAL(team, tid);
#endif
}
} // primary thread
switch (__kmp_barrier_release_pattern[bs_forkjoin_barrier]) {
case bp_dist_bar: {
__kmp_dist_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid,
TRUE USE_ITT_BUILD_ARG(NULL));
break;
}
case bp_hyper_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bs_forkjoin_barrier]);
__kmp_hyper_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid,
TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_hierarchical_bar: {
__kmp_hierarchical_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid,
TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
case bp_tree_bar: {
KMP_ASSERT(__kmp_barrier_release_branch_bits[bs_forkjoin_barrier]);
__kmp_tree_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid,
TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
break;
}
default: {
__kmp_linear_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid,
TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
}
}
#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 = (team)
? OMPT_CUR_TASK_DATA(this_thr)
: &(this_thr->th.ompt_thread_info.task_data);
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 = team ? team->t.ompt_team_info.master_return_address : NULL;
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
// Early exit for reaping threads releasing forkjoin barrier
if (TCR_4(__kmp_global.g.g_done)) {
this_thr->th.th_task_team = NULL;
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
if (!KMP_MASTER_TID(tid)) {
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
if (itt_sync_obj)
__kmp_itt_barrier_finished(gtid, itt_sync_obj);
}
}
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
KA_TRACE(10, ("__kmp_fork_barrier: T#%d is leaving early\n", gtid));
return;
}
/* We can now assume that a valid team structure has been allocated by the
primary thread and propagated to all worker threads. The current thread,
however, may not be part of the team, so we can't blindly assume that the
team pointer is non-null. */
team = (kmp_team_t *)TCR_PTR(this_thr->th.th_team);
KMP_DEBUG_ASSERT(team != NULL);
tid = __kmp_tid_from_gtid(gtid);
#if KMP_BARRIER_ICV_PULL
/* Primary thread's copy of the ICVs was set up on the implicit taskdata in
__kmp_reinitialize_team. __kmp_fork_call() assumes the primary thread's
implicit task has this data before this function is called. We cannot
modify __kmp_fork_call() to look at the fixed ICVs in the primary thread's
thread struct, because it is not always the case that the threads arrays
have been allocated when __kmp_fork_call() is executed. */
{
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_icv_copy);
if (!KMP_MASTER_TID(tid)) { // primary thread already has ICVs
// Copy the initial ICVs from the primary thread's thread struct to the
// implicit task for this tid.
KA_TRACE(10,
("__kmp_fork_barrier: T#%d(%d) is PULLing ICVs\n", gtid, tid));
__kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team,
tid, FALSE);
copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
&team->t.t_threads[0]
->th.th_bar[bs_forkjoin_barrier]
.bb.th_fixed_icvs);
}
}
#endif // KMP_BARRIER_ICV_PULL
if (__kmp_tasking_mode != tskm_immediate_exec) {
__kmp_task_team_sync(this_thr, team);
}
#if KMP_AFFINITY_SUPPORTED
kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
if (proc_bind == proc_bind_intel) {
// Call dynamic affinity settings
if (__kmp_affinity.type == affinity_balanced && team->t.t_size_changed) {
__kmp_balanced_affinity(this_thr, team->t.t_nproc);
}
} else if (proc_bind != proc_bind_false) {
if (this_thr->th.th_new_place == this_thr->th.th_current_place) {
KA_TRACE(100, ("__kmp_fork_barrier: T#%d already in correct place %d\n",
__kmp_gtid_from_thread(this_thr),
this_thr->th.th_current_place));
} else {
__kmp_affinity_bind_place(gtid);
}
}
#endif // KMP_AFFINITY_SUPPORTED
// Perform the display affinity functionality
if (__kmp_display_affinity) {
if (team->t.t_display_affinity
#if KMP_AFFINITY_SUPPORTED
|| (__kmp_affinity.type == affinity_balanced && team->t.t_size_changed)
#endif
) {
// NULL means use the affinity-format-var ICV
__kmp_aux_display_affinity(gtid, NULL);
this_thr->th.th_prev_num_threads = team->t.t_nproc;
this_thr->th.th_prev_level = team->t.t_level;
}
}
if (!KMP_MASTER_TID(tid))
KMP_CHECK_UPDATE(this_thr->th.th_def_allocator, team->t.t_def_allocator);
#if USE_ITT_BUILD && USE_ITT_NOTIFY
if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
if (!KMP_MASTER_TID(tid)) {
// Get correct barrier object
itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
__kmp_itt_barrier_finished(gtid, itt_sync_obj); // Workers call acquired
} // (prepare called inside barrier_release)
}
#endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
KA_TRACE(10, ("__kmp_fork_barrier: T#%d(%d:%d) is leaving\n", gtid,
team->t.t_id, tid));
}
void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
kmp_internal_control_t *new_icvs, ident_t *loc) {
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_setup_icv_copy);
KMP_DEBUG_ASSERT(team && new_nproc && new_icvs);
KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
/* Primary thread's copy of the ICVs was set up on the implicit taskdata in
__kmp_reinitialize_team. __kmp_fork_call() assumes the primary thread's
implicit task has this data before this function is called. */
#if KMP_BARRIER_ICV_PULL
/* Copy ICVs to primary thread's thread structure into th_fixed_icvs (which
remains untouched), where all of the worker threads can access them and
make their own copies after the barrier. */
KMP_DEBUG_ASSERT(team->t.t_threads[0]); // The threads arrays should be
// allocated at this point
copy_icvs(
&team->t.t_threads[0]->th.th_bar[bs_forkjoin_barrier].bb.th_fixed_icvs,
new_icvs);
KF_TRACE(10, ("__kmp_setup_icv_copy: PULL: T#%d this_thread=%p team=%p\n", 0,
team->t.t_threads[0], team));
#elif KMP_BARRIER_ICV_PUSH
// The ICVs will be propagated in the fork barrier, so nothing needs to be
// done here.
KF_TRACE(10, ("__kmp_setup_icv_copy: PUSH: T#%d this_thread=%p team=%p\n", 0,
team->t.t_threads[0], team));
#else
// Copy the ICVs to each of the non-primary threads. This takes O(nthreads)
// time.
ngo_load(new_icvs);
KMP_DEBUG_ASSERT(team->t.t_threads[0]); // The threads arrays should be
// allocated at this point
for (int f = 1; f < new_nproc; ++f) { // Skip the primary thread
// TODO: GEH - pass in better source location info since usually NULL here
KF_TRACE(10, ("__kmp_setup_icv_copy: LINEAR: T#%d this_thread=%p team=%p\n",
f, team->t.t_threads[f], team));
__kmp_init_implicit_task(loc, team->t.t_threads[f], team, f, FALSE);
ngo_store_icvs(&team->t.t_implicit_task_taskdata[f].td_icvs, new_icvs);
KF_TRACE(10, ("__kmp_setup_icv_copy: LINEAR: T#%d this_thread=%p team=%p\n",
f, team->t.t_threads[f], team));
}
ngo_sync();
#endif // KMP_BARRIER_ICV_PULL
}
Reference in New Issue View Git Blame Copy Permalink