blob: d6e4ff536e382c16e4cdc59961e72f43cd9bdbb9 [file] [log] [blame]
/*
* kmp_csupport.cpp -- kfront linkage support for OpenMP.
*/
//===----------------------------------------------------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is dual licensed under the MIT and the University of Illinois Open
// Source Licenses. See LICENSE.txt for details.
//
//===----------------------------------------------------------------------===//
#include "omp.h" /* extern "C" declarations of user-visible routines */
#include "kmp.h"
#include "kmp_error.h"
#include "kmp_i18n.h"
#include "kmp_itt.h"
#include "kmp_lock.h"
#include "kmp_stats.h"
#if OMPT_SUPPORT
#include "ompt-internal.h"
#include "ompt-specific.h"
#endif
#define MAX_MESSAGE 512
// flags will be used in future, e.g. to implement openmp_strict library
// restrictions
/*!
* @ingroup STARTUP_SHUTDOWN
* @param loc in source location information
* @param flags in for future use (currently ignored)
*
* Initialize the runtime library. This call is optional; if it is not made then
* it will be implicitly called by attempts to use other library functions.
*/
void __kmpc_begin(ident_t *loc, kmp_int32 flags) {
// By default __kmpc_begin() is no-op.
char *env;
if ((env = getenv("KMP_INITIAL_THREAD_BIND")) != NULL &&
__kmp_str_match_true(env)) {
__kmp_middle_initialize();
KC_TRACE(10, ("__kmpc_begin: middle initialization called\n"));
} else if (__kmp_ignore_mppbeg() == FALSE) {
// By default __kmp_ignore_mppbeg() returns TRUE.
__kmp_internal_begin();
KC_TRACE(10, ("__kmpc_begin: called\n"));
}
}
/*!
* @ingroup STARTUP_SHUTDOWN
* @param loc source location information
*
* Shutdown the runtime library. This is also optional, and even if called will
* not do anything unless the `KMP_IGNORE_MPPEND` environment variable is set to
* zero.
*/
void __kmpc_end(ident_t *loc) {
// By default, __kmp_ignore_mppend() returns TRUE which makes __kmpc_end()
// call no-op. However, this can be overridden with KMP_IGNORE_MPPEND
// environment variable. If KMP_IGNORE_MPPEND is 0, __kmp_ignore_mppend()
// returns FALSE and __kmpc_end() will unregister this root (it can cause
// library shut down).
if (__kmp_ignore_mppend() == FALSE) {
KC_TRACE(10, ("__kmpc_end: called\n"));
KA_TRACE(30, ("__kmpc_end\n"));
__kmp_internal_end_thread(-1);
}
}
/*!
@ingroup THREAD_STATES
@param loc Source location information.
@return The global thread index of the active thread.
This function can be called in any context.
If the runtime has ony been entered at the outermost level from a
single (necessarily non-OpenMP<sup>*</sup>) thread, then the thread number is
that which would be returned by omp_get_thread_num() in the outermost
active parallel construct. (Or zero if there is no active parallel
construct, since the master thread is necessarily thread zero).
If multiple non-OpenMP threads all enter an OpenMP construct then this
will be a unique thread identifier among all the threads created by
the OpenMP runtime (but the value cannote be defined in terms of
OpenMP thread ids returned by omp_get_thread_num()).
*/
kmp_int32 __kmpc_global_thread_num(ident_t *loc) {
kmp_int32 gtid = __kmp_entry_gtid();
KC_TRACE(10, ("__kmpc_global_thread_num: T#%d\n", gtid));
return gtid;
}
/*!
@ingroup THREAD_STATES
@param loc Source location information.
@return The number of threads under control of the OpenMP<sup>*</sup> runtime
This function can be called in any context.
It returns the total number of threads under the control of the OpenMP runtime.
That is not a number that can be determined by any OpenMP standard calls, since
the library may be called from more than one non-OpenMP thread, and this
reflects the total over all such calls. Similarly the runtime maintains
underlying threads even when they are not active (since the cost of creating
and destroying OS threads is high), this call counts all such threads even if
they are not waiting for work.
*/
kmp_int32 __kmpc_global_num_threads(ident_t *loc) {
KC_TRACE(10,
("__kmpc_global_num_threads: num_threads = %d\n", __kmp_all_nth));
return TCR_4(__kmp_all_nth);
}
/*!
@ingroup THREAD_STATES
@param loc Source location information.
@return The thread number of the calling thread in the innermost active parallel
construct.
*/
kmp_int32 __kmpc_bound_thread_num(ident_t *loc) {
KC_TRACE(10, ("__kmpc_bound_thread_num: called\n"));
return __kmp_tid_from_gtid(__kmp_entry_gtid());
}
/*!
@ingroup THREAD_STATES
@param loc Source location information.
@return The number of threads in the innermost active parallel construct.
*/
kmp_int32 __kmpc_bound_num_threads(ident_t *loc) {
KC_TRACE(10, ("__kmpc_bound_num_threads: called\n"));
return __kmp_entry_thread()->th.th_team->t.t_nproc;
}
/*!
* @ingroup DEPRECATED
* @param loc location description
*
* This function need not be called. It always returns TRUE.
*/
kmp_int32 __kmpc_ok_to_fork(ident_t *loc) {
#ifndef KMP_DEBUG
return TRUE;
#else
const char *semi2;
const char *semi3;
int line_no;
if (__kmp_par_range == 0) {
return TRUE;
}
semi2 = loc->psource;
if (semi2 == NULL) {
return TRUE;
}
semi2 = strchr(semi2, ';');
if (semi2 == NULL) {
return TRUE;
}
semi2 = strchr(semi2 + 1, ';');
if (semi2 == NULL) {
return TRUE;
}
if (__kmp_par_range_filename[0]) {
const char *name = semi2 - 1;
while ((name > loc->psource) && (*name != '/') && (*name != ';')) {
name--;
}
if ((*name == '/') || (*name == ';')) {
name++;
}
if (strncmp(__kmp_par_range_filename, name, semi2 - name)) {
return __kmp_par_range < 0;
}
}
semi3 = strchr(semi2 + 1, ';');
if (__kmp_par_range_routine[0]) {
if ((semi3 != NULL) && (semi3 > semi2) &&
(strncmp(__kmp_par_range_routine, semi2 + 1, semi3 - semi2 - 1))) {
return __kmp_par_range < 0;
}
}
if (KMP_SSCANF(semi3 + 1, "%d", &line_no) == 1) {
if ((line_no >= __kmp_par_range_lb) && (line_no <= __kmp_par_range_ub)) {
return __kmp_par_range > 0;
}
return __kmp_par_range < 0;
}
return TRUE;
#endif /* KMP_DEBUG */
}
/*!
@ingroup THREAD_STATES
@param loc Source location information.
@return 1 if this thread is executing inside an active parallel region, zero if
not.
*/
kmp_int32 __kmpc_in_parallel(ident_t *loc) {
return __kmp_entry_thread()->th.th_root->r.r_active;
}
/*!
@ingroup PARALLEL
@param loc source location information
@param global_tid global thread number
@param num_threads number of threads requested for this parallel construct
Set the number of threads to be used by the next fork spawned by this thread.
This call is only required if the parallel construct has a `num_threads` clause.
*/
void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
kmp_int32 num_threads) {
KA_TRACE(20, ("__kmpc_push_num_threads: enter T#%d num_threads=%d\n",
global_tid, num_threads));
__kmp_push_num_threads(loc, global_tid, num_threads);
}
void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid) {
KA_TRACE(20, ("__kmpc_pop_num_threads: enter\n"));
/* the num_threads are automatically popped */
}
#if OMP_40_ENABLED
void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
kmp_int32 proc_bind) {
KA_TRACE(20, ("__kmpc_push_proc_bind: enter T#%d proc_bind=%d\n", global_tid,
proc_bind));
__kmp_push_proc_bind(loc, global_tid, (kmp_proc_bind_t)proc_bind);
}
#endif /* OMP_40_ENABLED */
/*!
@ingroup PARALLEL
@param loc source location information
@param argc total number of arguments in the ellipsis
@param microtask pointer to callback routine consisting of outlined parallel
construct
@param ... pointers to shared variables that aren't global
Do the actual fork and call the microtask in the relevant number of threads.
*/
void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, kmpc_micro microtask, ...) {
int gtid = __kmp_entry_gtid();
#if (KMP_STATS_ENABLED)
int inParallel = __kmpc_in_parallel(loc);
if (inParallel) {
KMP_COUNT_BLOCK(OMP_NESTED_PARALLEL);
} else {
KMP_COUNT_BLOCK(OMP_PARALLEL);
}
#endif
// maybe to save thr_state is enough here
{
va_list ap;
va_start(ap, microtask);
#if OMPT_SUPPORT
ompt_frame_t *ompt_frame;
if (ompt_enabled) {
kmp_info_t *master_th = __kmp_threads[gtid];
kmp_team_t *parent_team = master_th->th.th_team;
ompt_lw_taskteam_t *lwt = parent_team->t.ompt_serialized_team_info;
if (lwt)
ompt_frame = &(lwt->ompt_task_info.frame);
else {
int tid = __kmp_tid_from_gtid(gtid);
ompt_frame = &(
parent_team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame);
}
ompt_frame->reenter_runtime_frame = __builtin_frame_address(1);
}
#endif
#if INCLUDE_SSC_MARKS
SSC_MARK_FORKING();
#endif
__kmp_fork_call(loc, gtid, fork_context_intel, argc,
#if OMPT_SUPPORT
VOLATILE_CAST(void *) microtask, // "unwrapped" task
#endif
VOLATILE_CAST(microtask_t) microtask, // "wrapped" task
VOLATILE_CAST(launch_t) __kmp_invoke_task_func,
/* TODO: revert workaround for Intel(R) 64 tracker #96 */
#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
&ap
#else
ap
#endif
);
#if INCLUDE_SSC_MARKS
SSC_MARK_JOINING();
#endif
__kmp_join_call(loc, gtid
#if OMPT_SUPPORT
,
fork_context_intel
#endif
);
va_end(ap);
}
}
#if OMP_40_ENABLED
/*!
@ingroup PARALLEL
@param loc source location information
@param global_tid global thread number
@param num_teams number of teams requested for the teams construct
@param num_threads number of threads per team requested for the teams construct
Set the number of teams to be used by the teams construct.
This call is only required if the teams construct has a `num_teams` clause
or a `thread_limit` clause (or both).
*/
void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
kmp_int32 num_teams, kmp_int32 num_threads) {
KA_TRACE(20,
("__kmpc_push_num_teams: enter T#%d num_teams=%d num_threads=%d\n",
global_tid, num_teams, num_threads));
__kmp_push_num_teams(loc, global_tid, num_teams, num_threads);
}
/*!
@ingroup PARALLEL
@param loc source location information
@param argc total number of arguments in the ellipsis
@param microtask pointer to callback routine consisting of outlined teams
construct
@param ... pointers to shared variables that aren't global
Do the actual fork and call the microtask in the relevant number of threads.
*/
void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,
...) {
int gtid = __kmp_entry_gtid();
kmp_info_t *this_thr = __kmp_threads[gtid];
va_list ap;
va_start(ap, microtask);
KMP_COUNT_BLOCK(OMP_TEAMS);
// remember teams entry point and nesting level
this_thr->th.th_teams_microtask = microtask;
this_thr->th.th_teams_level =
this_thr->th.th_team->t.t_level; // AC: can be >0 on host
#if OMPT_SUPPORT
kmp_team_t *parent_team = this_thr->th.th_team;
int tid = __kmp_tid_from_gtid(gtid);
if (ompt_enabled) {
parent_team->t.t_implicit_task_taskdata[tid]
.ompt_task_info.frame.reenter_runtime_frame =
__builtin_frame_address(1);
}
#endif
// check if __kmpc_push_num_teams called, set default number of teams
// otherwise
if (this_thr->th.th_teams_size.nteams == 0) {
__kmp_push_num_teams(loc, gtid, 0, 0);
}
KMP_DEBUG_ASSERT(this_thr->th.th_set_nproc >= 1);
KMP_DEBUG_ASSERT(this_thr->th.th_teams_size.nteams >= 1);
KMP_DEBUG_ASSERT(this_thr->th.th_teams_size.nth >= 1);
__kmp_fork_call(loc, gtid, fork_context_intel, argc,
#if OMPT_SUPPORT
VOLATILE_CAST(void *) microtask, // "unwrapped" task
#endif
VOLATILE_CAST(microtask_t)
__kmp_teams_master, // "wrapped" task
VOLATILE_CAST(launch_t) __kmp_invoke_teams_master,
#if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
&ap
#else
ap
#endif
);
__kmp_join_call(loc, gtid
#if OMPT_SUPPORT
,
fork_context_intel
#endif
);
this_thr->th.th_teams_microtask = NULL;
this_thr->th.th_teams_level = 0;
*(kmp_int64 *)(&this_thr->th.th_teams_size) = 0L;
va_end(ap);
}
#endif /* OMP_40_ENABLED */
// I don't think this function should ever have been exported.
// The __kmpc_ prefix was misapplied. I'm fairly certain that no generated
// openmp code ever called it, but it's been exported from the RTL for so
// long that I'm afraid to remove the definition.
int __kmpc_invoke_task_func(int gtid) { return __kmp_invoke_task_func(gtid); }
/*!
@ingroup PARALLEL
@param loc source location information
@param global_tid global thread number
Enter a serialized parallel construct. This interface is used to handle a
conditional parallel region, like this,
@code
#pragma omp parallel if (condition)
@endcode
when the condition is false.
*/
void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
// The implementation is now in kmp_runtime.cpp so that it can share static
// functions with kmp_fork_call since the tasks to be done are similar in
// each case.
__kmp_serialized_parallel(loc, global_tid);
}
/*!
@ingroup PARALLEL
@param loc source location information
@param global_tid global thread number
Leave a serialized parallel construct.
*/
void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
kmp_internal_control_t *top;
kmp_info_t *this_thr;
kmp_team_t *serial_team;
KC_TRACE(10,
("__kmpc_end_serialized_parallel: called by T#%d\n", global_tid));
/* skip all this code for autopar serialized loops since it results in
unacceptable overhead */
if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
return;
// Not autopar code
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
this_thr = __kmp_threads[global_tid];
serial_team = this_thr->th.th_serial_team;
#if OMP_45_ENABLED
kmp_task_team_t *task_team = this_thr->th.th_task_team;
// we need to wait for the proxy tasks before finishing the thread
if (task_team != NULL && task_team->tt.tt_found_proxy_tasks)
__kmp_task_team_wait(this_thr, serial_team USE_ITT_BUILD_ARG(NULL));
#endif
KMP_MB();
KMP_DEBUG_ASSERT(serial_team);
KMP_ASSERT(serial_team->t.t_serialized);
KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
KMP_DEBUG_ASSERT(serial_team != this_thr->th.th_root->r.r_root_team);
KMP_DEBUG_ASSERT(serial_team->t.t_threads);
KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
/* If necessary, pop the internal control stack values and replace the team
* values */
top = serial_team->t.t_control_stack_top;
if (top && top->serial_nesting_level == serial_team->t.t_serialized) {
copy_icvs(&serial_team->t.t_threads[0]->th.th_current_task->td_icvs, top);
serial_team->t.t_control_stack_top = top->next;
__kmp_free(top);
}
// if( serial_team -> t.t_serialized > 1 )
serial_team->t.t_level--;
/* pop dispatch buffers stack */
KMP_DEBUG_ASSERT(serial_team->t.t_dispatch->th_disp_buffer);
{
dispatch_private_info_t *disp_buffer =
serial_team->t.t_dispatch->th_disp_buffer;
serial_team->t.t_dispatch->th_disp_buffer =
serial_team->t.t_dispatch->th_disp_buffer->next;
__kmp_free(disp_buffer);
}
--serial_team->t.t_serialized;
if (serial_team->t.t_serialized == 0) {
/* return to the parallel section */
#if KMP_ARCH_X86 || KMP_ARCH_X86_64
if (__kmp_inherit_fp_control && serial_team->t.t_fp_control_saved) {
__kmp_clear_x87_fpu_status_word();
__kmp_load_x87_fpu_control_word(&serial_team->t.t_x87_fpu_control_word);
__kmp_load_mxcsr(&serial_team->t.t_mxcsr);
}
#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
this_thr->th.th_team = serial_team->t.t_parent;
this_thr->th.th_info.ds.ds_tid = serial_team->t.t_master_tid;
/* restore values cached in the thread */
this_thr->th.th_team_nproc = serial_team->t.t_parent->t.t_nproc; /* JPH */
this_thr->th.th_team_master =
serial_team->t.t_parent->t.t_threads[0]; /* JPH */
this_thr->th.th_team_serialized = this_thr->th.th_team->t.t_serialized;
/* TODO the below shouldn't need to be adjusted for serialized teams */
this_thr->th.th_dispatch =
&this_thr->th.th_team->t.t_dispatch[serial_team->t.t_master_tid];
__kmp_pop_current_task_from_thread(this_thr);
KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 0);
this_thr->th.th_current_task->td_flags.executing = 1;
if (__kmp_tasking_mode != tskm_immediate_exec) {
// Copy the task team from the new child / old parent team to the thread.
this_thr->th.th_task_team =
this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state];
KA_TRACE(20,
("__kmpc_end_serialized_parallel: T#%d restoring task_team %p / "
"team %p\n",
global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
}
} else {
if (__kmp_tasking_mode != tskm_immediate_exec) {
KA_TRACE(20, ("__kmpc_end_serialized_parallel: T#%d decreasing nesting "
"depth of serial team %p to %d\n",
global_tid, serial_team, serial_team->t.t_serialized));
}
}
if (__kmp_env_consistency_check)
__kmp_pop_parallel(global_tid, NULL);
}
/*!
@ingroup SYNCHRONIZATION
@param loc source location information.
Execute <tt>flush</tt>. This is implemented as a full memory fence. (Though
depending on the memory ordering convention obeyed by the compiler
even that may not be necessary).
*/
void __kmpc_flush(ident_t *loc) {
KC_TRACE(10, ("__kmpc_flush: called\n"));
/* need explicit __mf() here since use volatile instead in library */
KMP_MB(); /* Flush all pending memory write invalidates. */
#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
#if KMP_MIC
// fence-style instructions do not exist, but lock; xaddl $0,(%rsp) can be used.
// We shouldn't need it, though, since the ABI rules require that
// * If the compiler generates NGO stores it also generates the fence
// * If users hand-code NGO stores they should insert the fence
// therefore no incomplete unordered stores should be visible.
#else
// C74404
// This is to address non-temporal store instructions (sfence needed).
// The clflush instruction is addressed either (mfence needed).
// Probably the non-temporal load monvtdqa instruction should also be
// addressed.
// mfence is a SSE2 instruction. Do not execute it if CPU is not SSE2.
if (!__kmp_cpuinfo.initialized) {
__kmp_query_cpuid(&__kmp_cpuinfo);
}; // if
if (!__kmp_cpuinfo.sse2) {
// CPU cannot execute SSE2 instructions.
} else {
#if KMP_COMPILER_ICC
_mm_mfence();
#elif KMP_COMPILER_MSVC
MemoryBarrier();
#else
__sync_synchronize();
#endif // KMP_COMPILER_ICC
}; // if
#endif // KMP_MIC
#elif (KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || KMP_ARCH_MIPS64)
// Nothing to see here move along
#elif KMP_ARCH_PPC64
// Nothing needed here (we have a real MB above).
#if KMP_OS_CNK
// The flushing thread needs to yield here; this prevents a
// busy-waiting thread from saturating the pipeline. flush is
// often used in loops like this:
// while (!flag) {
// #pragma omp flush(flag)
// }
// and adding the yield here is good for at least a 10x speedup
// when running >2 threads per core (on the NAS LU benchmark).
__kmp_yield(TRUE);
#endif
#else
#error Unknown or unsupported architecture
#endif
}
/* -------------------------------------------------------------------------- */
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid thread id.
Execute a barrier.
*/
void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid) {
KMP_COUNT_BLOCK(OMP_BARRIER);
KC_TRACE(10, ("__kmpc_barrier: called T#%d\n", global_tid));
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
if (__kmp_env_consistency_check) {
if (loc == 0) {
KMP_WARNING(ConstructIdentInvalid); // ??? What does it mean for the user?
}; // if
__kmp_check_barrier(global_tid, ct_barrier, loc);
}
#if OMPT_SUPPORT && OMPT_TRACE
ompt_frame_t *ompt_frame;
if (ompt_enabled) {
ompt_frame = __ompt_get_task_frame_internal(0);
if (ompt_frame->reenter_runtime_frame == NULL)
ompt_frame->reenter_runtime_frame = __builtin_frame_address(1);
}
#endif
__kmp_threads[global_tid]->th.th_ident = loc;
// TODO: explicit barrier_wait_id:
// this function is called when 'barrier' directive is present or
// implicit barrier at the end of a worksharing construct.
// 1) better to add a per-thread barrier counter to a thread data structure
// 2) set to 0 when a new team is created
// 4) no sync is required
__kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
ompt_frame->reenter_runtime_frame = NULL;
}
#endif
}
/* The BARRIER for a MASTER section is always explicit */
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param global_tid global thread number .
@return 1 if this thread should execute the <tt>master</tt> block, 0 otherwise.
*/
kmp_int32 __kmpc_master(ident_t *loc, kmp_int32 global_tid) {
int status = 0;
KC_TRACE(10, ("__kmpc_master: called T#%d\n", global_tid));
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
if (KMP_MASTER_GTID(global_tid)) {
KMP_COUNT_BLOCK(OMP_MASTER);
KMP_PUSH_PARTITIONED_TIMER(OMP_master);
status = 1;
}
#if OMPT_SUPPORT && OMPT_TRACE
if (status) {
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_master_begin)) {
kmp_info_t *this_thr = __kmp_threads[global_tid];
kmp_team_t *team = this_thr->th.th_team;
int tid = __kmp_tid_from_gtid(global_tid);
ompt_callbacks.ompt_callback(ompt_event_master_begin)(
team->t.ompt_team_info.parallel_id,
team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
}
}
#endif
if (__kmp_env_consistency_check) {
#if KMP_USE_DYNAMIC_LOCK
if (status)
__kmp_push_sync(global_tid, ct_master, loc, NULL, 0);
else
__kmp_check_sync(global_tid, ct_master, loc, NULL, 0);
#else
if (status)
__kmp_push_sync(global_tid, ct_master, loc, NULL);
else
__kmp_check_sync(global_tid, ct_master, loc, NULL);
#endif
}
return status;
}
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param global_tid global thread number .
Mark the end of a <tt>master</tt> region. This should only be called by the
thread that executes the <tt>master</tt> region.
*/
void __kmpc_end_master(ident_t *loc, kmp_int32 global_tid) {
KC_TRACE(10, ("__kmpc_end_master: called T#%d\n", global_tid));
KMP_DEBUG_ASSERT(KMP_MASTER_GTID(global_tid));
KMP_POP_PARTITIONED_TIMER();
#if OMPT_SUPPORT && OMPT_TRACE
kmp_info_t *this_thr = __kmp_threads[global_tid];
kmp_team_t *team = this_thr->th.th_team;
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_master_end)) {
int tid = __kmp_tid_from_gtid(global_tid);
ompt_callbacks.ompt_callback(ompt_event_master_end)(
team->t.ompt_team_info.parallel_id,
team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
}
#endif
if (__kmp_env_consistency_check) {
if (global_tid < 0)
KMP_WARNING(ThreadIdentInvalid);
if (KMP_MASTER_GTID(global_tid))
__kmp_pop_sync(global_tid, ct_master, loc);
}
}
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param gtid global thread number.
Start execution of an <tt>ordered</tt> construct.
*/
void __kmpc_ordered(ident_t *loc, kmp_int32 gtid) {
int cid = 0;
kmp_info_t *th;
KMP_DEBUG_ASSERT(__kmp_init_serial);
KC_TRACE(10, ("__kmpc_ordered: called T#%d\n", gtid));
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
#if USE_ITT_BUILD
__kmp_itt_ordered_prep(gtid);
// TODO: ordered_wait_id
#endif /* USE_ITT_BUILD */
th = __kmp_threads[gtid];
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
/* OMPT state update */
th->th.ompt_thread_info.wait_id = (uint64_t)loc;
th->th.ompt_thread_info.state = ompt_state_wait_ordered;
/* OMPT event callback */
if (ompt_callbacks.ompt_callback(ompt_event_wait_ordered)) {
ompt_callbacks.ompt_callback(ompt_event_wait_ordered)(
th->th.ompt_thread_info.wait_id);
}
}
#endif
if (th->th.th_dispatch->th_deo_fcn != 0)
(*th->th.th_dispatch->th_deo_fcn)(&gtid, &cid, loc);
else
__kmp_parallel_deo(&gtid, &cid, loc);
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
/* OMPT state update */
th->th.ompt_thread_info.state = ompt_state_work_parallel;
th->th.ompt_thread_info.wait_id = 0;
/* OMPT event callback */
if (ompt_callbacks.ompt_callback(ompt_event_acquired_ordered)) {
ompt_callbacks.ompt_callback(ompt_event_acquired_ordered)(
th->th.ompt_thread_info.wait_id);
}
}
#endif
#if USE_ITT_BUILD
__kmp_itt_ordered_start(gtid);
#endif /* USE_ITT_BUILD */
}
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param gtid global thread number.
End execution of an <tt>ordered</tt> construct.
*/
void __kmpc_end_ordered(ident_t *loc, kmp_int32 gtid) {
int cid = 0;
kmp_info_t *th;
KC_TRACE(10, ("__kmpc_end_ordered: called T#%d\n", gtid));
#if USE_ITT_BUILD
__kmp_itt_ordered_end(gtid);
// TODO: ordered_wait_id
#endif /* USE_ITT_BUILD */
th = __kmp_threads[gtid];
if (th->th.th_dispatch->th_dxo_fcn != 0)
(*th->th.th_dispatch->th_dxo_fcn)(&gtid, &cid, loc);
else
__kmp_parallel_dxo(&gtid, &cid, loc);
#if OMPT_SUPPORT && OMPT_BLAME
if (ompt_enabled &&
ompt_callbacks.ompt_callback(ompt_event_release_ordered)) {
ompt_callbacks.ompt_callback(ompt_event_release_ordered)(
th->th.ompt_thread_info.wait_id);
}
#endif
}
#if KMP_USE_DYNAMIC_LOCK
static __forceinline void
__kmp_init_indirect_csptr(kmp_critical_name *crit, ident_t const *loc,
kmp_int32 gtid, kmp_indirect_locktag_t tag) {
// Pointer to the allocated indirect lock is written to crit, while indexing
// is ignored.
void *idx;
kmp_indirect_lock_t **lck;
lck = (kmp_indirect_lock_t **)crit;
kmp_indirect_lock_t *ilk = __kmp_allocate_indirect_lock(&idx, gtid, tag);
KMP_I_LOCK_FUNC(ilk, init)(ilk->lock);
KMP_SET_I_LOCK_LOCATION(ilk, loc);
KMP_SET_I_LOCK_FLAGS(ilk, kmp_lf_critical_section);
KA_TRACE(20,
("__kmp_init_indirect_csptr: initialized indirect lock #%d\n", tag));
#if USE_ITT_BUILD
__kmp_itt_critical_creating(ilk->lock, loc);
#endif
int status = KMP_COMPARE_AND_STORE_PTR(lck, nullptr, ilk);
if (status == 0) {
#if USE_ITT_BUILD
__kmp_itt_critical_destroyed(ilk->lock);
#endif
// We don't really need to destroy the unclaimed lock here since it will be
// cleaned up at program exit.
// KMP_D_LOCK_FUNC(&idx, destroy)((kmp_dyna_lock_t *)&idx);
}
KMP_DEBUG_ASSERT(*lck != NULL);
}
// Fast-path acquire tas lock
#define KMP_ACQUIRE_TAS_LOCK(lock, gtid) \
{ \
kmp_tas_lock_t *l = (kmp_tas_lock_t *)lock; \
if (l->lk.poll != KMP_LOCK_FREE(tas) || \
!KMP_COMPARE_AND_STORE_ACQ32(&(l->lk.poll), KMP_LOCK_FREE(tas), \
KMP_LOCK_BUSY(gtid + 1, tas))) { \
kmp_uint32 spins; \
KMP_FSYNC_PREPARE(l); \
KMP_INIT_YIELD(spins); \
if (TCR_4(__kmp_nth) > \
(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) { \
KMP_YIELD(TRUE); \
} else { \
KMP_YIELD_SPIN(spins); \
} \
kmp_backoff_t backoff = __kmp_spin_backoff_params; \
while (l->lk.poll != KMP_LOCK_FREE(tas) || \
!KMP_COMPARE_AND_STORE_ACQ32(&(l->lk.poll), KMP_LOCK_FREE(tas), \
KMP_LOCK_BUSY(gtid + 1, tas))) { \
__kmp_spin_backoff(&backoff); \
if (TCR_4(__kmp_nth) > \
(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) { \
KMP_YIELD(TRUE); \
} else { \
KMP_YIELD_SPIN(spins); \
} \
} \
} \
KMP_FSYNC_ACQUIRED(l); \
}
// Fast-path test tas lock
#define KMP_TEST_TAS_LOCK(lock, gtid, rc) \
{ \
kmp_tas_lock_t *l = (kmp_tas_lock_t *)lock; \
rc = l->lk.poll == KMP_LOCK_FREE(tas) && \
KMP_COMPARE_AND_STORE_ACQ32(&(l->lk.poll), KMP_LOCK_FREE(tas), \
KMP_LOCK_BUSY(gtid + 1, tas)); \
}
// Fast-path release tas lock
#define KMP_RELEASE_TAS_LOCK(lock, gtid) \
{ \
TCW_4(((kmp_tas_lock_t *)lock)->lk.poll, KMP_LOCK_FREE(tas)); \
KMP_MB(); \
}
#if KMP_USE_FUTEX
#include <sys/syscall.h>
#include <unistd.h>
#ifndef FUTEX_WAIT
#define FUTEX_WAIT 0
#endif
#ifndef FUTEX_WAKE
#define FUTEX_WAKE 1
#endif
// Fast-path acquire futex lock
#define KMP_ACQUIRE_FUTEX_LOCK(lock, gtid) \
{ \
kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock; \
kmp_int32 gtid_code = (gtid + 1) << 1; \
KMP_MB(); \
KMP_FSYNC_PREPARE(ftx); \
kmp_int32 poll_val; \
while ((poll_val = KMP_COMPARE_AND_STORE_RET32( \
&(ftx->lk.poll), KMP_LOCK_FREE(futex), \
KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) { \
kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1; \
if (!cond) { \
if (!KMP_COMPARE_AND_STORE_RET32(&(ftx->lk.poll), poll_val, \
poll_val | \
KMP_LOCK_BUSY(1, futex))) { \
continue; \
} \
poll_val |= KMP_LOCK_BUSY(1, futex); \
} \
kmp_int32 rc; \
if ((rc = syscall(__NR_futex, &(ftx->lk.poll), FUTEX_WAIT, poll_val, \
NULL, NULL, 0)) != 0) { \
continue; \
} \
gtid_code |= 1; \
} \
KMP_FSYNC_ACQUIRED(ftx); \
}
// Fast-path test futex lock
#define KMP_TEST_FUTEX_LOCK(lock, gtid, rc) \
{ \
kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock; \
if (KMP_COMPARE_AND_STORE_ACQ32(&(ftx->lk.poll), KMP_LOCK_FREE(futex), \
KMP_LOCK_BUSY(gtid + 1 << 1, futex))) { \
KMP_FSYNC_ACQUIRED(ftx); \
rc = TRUE; \
} else { \
rc = FALSE; \
} \
}
// Fast-path release futex lock
#define KMP_RELEASE_FUTEX_LOCK(lock, gtid) \
{ \
kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock; \
KMP_MB(); \
KMP_FSYNC_RELEASING(ftx); \
kmp_int32 poll_val = \
KMP_XCHG_FIXED32(&(ftx->lk.poll), KMP_LOCK_FREE(futex)); \
if (KMP_LOCK_STRIP(poll_val) & 1) { \
syscall(__NR_futex, &(ftx->lk.poll), FUTEX_WAKE, \
KMP_LOCK_BUSY(1, futex), NULL, NULL, 0); \
} \
KMP_MB(); \
KMP_YIELD(TCR_4(__kmp_nth) > \
(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)); \
}
#endif // KMP_USE_FUTEX
#else // KMP_USE_DYNAMIC_LOCK
static kmp_user_lock_p __kmp_get_critical_section_ptr(kmp_critical_name *crit,
ident_t const *loc,
kmp_int32 gtid) {
kmp_user_lock_p *lck_pp = (kmp_user_lock_p *)crit;
// Because of the double-check, the following load doesn't need to be volatile
kmp_user_lock_p lck = (kmp_user_lock_p)TCR_PTR(*lck_pp);
if (lck == NULL) {
void *idx;
// Allocate & initialize the lock.
// Remember alloc'ed locks in table in order to free them in __kmp_cleanup()
lck = __kmp_user_lock_allocate(&idx, gtid, kmp_lf_critical_section);
__kmp_init_user_lock_with_checks(lck);
__kmp_set_user_lock_location(lck, loc);
#if USE_ITT_BUILD
__kmp_itt_critical_creating(lck);
// __kmp_itt_critical_creating() should be called *before* the first usage
// of underlying lock. It is the only place where we can guarantee it. There
// are chances the lock will destroyed with no usage, but it is not a
// problem, because this is not real event seen by user but rather setting
// name for object (lock). See more details in kmp_itt.h.
#endif /* USE_ITT_BUILD */
// Use a cmpxchg instruction to slam the start of the critical section with
// the lock pointer. If another thread beat us to it, deallocate the lock,
// and use the lock that the other thread allocated.
int status = KMP_COMPARE_AND_STORE_PTR(lck_pp, 0, lck);
if (status == 0) {
// Deallocate the lock and reload the value.
#if USE_ITT_BUILD
__kmp_itt_critical_destroyed(lck);
// Let ITT know the lock is destroyed and the same memory location may be reused
// for another purpose.
#endif /* USE_ITT_BUILD */
__kmp_destroy_user_lock_with_checks(lck);
__kmp_user_lock_free(&idx, gtid, lck);
lck = (kmp_user_lock_p)TCR_PTR(*lck_pp);
KMP_DEBUG_ASSERT(lck != NULL);
}
}
return lck;
}
#endif // KMP_USE_DYNAMIC_LOCK
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param global_tid global thread number .
@param crit identity of the critical section. This could be a pointer to a lock
associated with the critical section, or some other suitably unique value.
Enter code protected by a `critical` construct.
This function blocks until the executing thread can enter the critical section.
*/
void __kmpc_critical(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *crit) {
#if KMP_USE_DYNAMIC_LOCK
__kmpc_critical_with_hint(loc, global_tid, crit, omp_lock_hint_none);
#else
KMP_COUNT_BLOCK(OMP_CRITICAL);
KMP_TIME_PARTITIONED_BLOCK(
OMP_critical_wait); /* Time spent waiting to enter the critical section */
kmp_user_lock_p lck;
KC_TRACE(10, ("__kmpc_critical: called T#%d\n", global_tid));
// TODO: add THR_OVHD_STATE
KMP_CHECK_USER_LOCK_INIT();
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_CRITICAL_SIZE)) {
lck = (kmp_user_lock_p)crit;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_CRITICAL_SIZE)) {
lck = (kmp_user_lock_p)crit;
}
#endif
else { // ticket, queuing or drdpa
lck = __kmp_get_critical_section_ptr(crit, loc, global_tid);
}
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_critical, loc, lck);
// since the critical directive binds to all threads, not just the current
// team we have to check this even if we are in a serialized team.
// also, even if we are the uber thread, we still have to conduct the lock,
// as we have to contend with sibling threads.
#if USE_ITT_BUILD
__kmp_itt_critical_acquiring(lck);
#endif /* USE_ITT_BUILD */
// Value of 'crit' should be good for using as a critical_id of the critical
// section directive.
__kmp_acquire_user_lock_with_checks(lck, global_tid);
#if USE_ITT_BUILD
__kmp_itt_critical_acquired(lck);
#endif /* USE_ITT_BUILD */
KMP_START_EXPLICIT_TIMER(OMP_critical);
KA_TRACE(15, ("__kmpc_critical: done T#%d\n", global_tid));
#endif // KMP_USE_DYNAMIC_LOCK
}
#if KMP_USE_DYNAMIC_LOCK
// Converts the given hint to an internal lock implementation
static __forceinline kmp_dyna_lockseq_t __kmp_map_hint_to_lock(uintptr_t hint) {
#if KMP_USE_TSX
#define KMP_TSX_LOCK(seq) lockseq_##seq
#else
#define KMP_TSX_LOCK(seq) __kmp_user_lock_seq
#endif
#if KMP_ARCH_X86 || KMP_ARCH_X86_64
#define KMP_CPUINFO_RTM (__kmp_cpuinfo.rtm)
#else
#define KMP_CPUINFO_RTM 0
#endif
// Hints that do not require further logic
if (hint & kmp_lock_hint_hle)
return KMP_TSX_LOCK(hle);
if (hint & kmp_lock_hint_rtm)
return KMP_CPUINFO_RTM ? KMP_TSX_LOCK(rtm) : __kmp_user_lock_seq;
if (hint & kmp_lock_hint_adaptive)
return KMP_CPUINFO_RTM ? KMP_TSX_LOCK(adaptive) : __kmp_user_lock_seq;
// Rule out conflicting hints first by returning the default lock
if ((hint & omp_lock_hint_contended) && (hint & omp_lock_hint_uncontended))
return __kmp_user_lock_seq;
if ((hint & omp_lock_hint_speculative) &&
(hint & omp_lock_hint_nonspeculative))
return __kmp_user_lock_seq;
// Do not even consider speculation when it appears to be contended
if (hint & omp_lock_hint_contended)
return lockseq_queuing;
// Uncontended lock without speculation
if ((hint & omp_lock_hint_uncontended) && !(hint & omp_lock_hint_speculative))
return lockseq_tas;
// HLE lock for speculation
if (hint & omp_lock_hint_speculative)
return KMP_TSX_LOCK(hle);
return __kmp_user_lock_seq;
}
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param global_tid global thread number.
@param crit identity of the critical section. This could be a pointer to a lock
associated with the critical section, or some other suitably unique value.
@param hint the lock hint.
Enter code protected by a `critical` construct with a hint. The hint value is
used to suggest a lock implementation. This function blocks until the executing
thread can enter the critical section unless the hint suggests use of
speculative execution and the hardware supports it.
*/
void __kmpc_critical_with_hint(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *crit, uintptr_t hint) {
KMP_COUNT_BLOCK(OMP_CRITICAL);
kmp_user_lock_p lck;
KC_TRACE(10, ("__kmpc_critical: called T#%d\n", global_tid));
kmp_dyna_lock_t *lk = (kmp_dyna_lock_t *)crit;
// Check if it is initialized.
if (*lk == 0) {
kmp_dyna_lockseq_t lckseq = __kmp_map_hint_to_lock(hint);
if (KMP_IS_D_LOCK(lckseq)) {
KMP_COMPARE_AND_STORE_ACQ32((volatile kmp_int32 *)crit, 0,
KMP_GET_D_TAG(lckseq));
} else {
__kmp_init_indirect_csptr(crit, loc, global_tid, KMP_GET_I_TAG(lckseq));
}
}
// Branch for accessing the actual lock object and set operation. This
// branching is inevitable since this lock initialization does not follow the
// normal dispatch path (lock table is not used).
if (KMP_EXTRACT_D_TAG(lk) != 0) {
lck = (kmp_user_lock_p)lk;
if (__kmp_env_consistency_check) {
__kmp_push_sync(global_tid, ct_critical, loc, lck,
__kmp_map_hint_to_lock(hint));
}
#if USE_ITT_BUILD
__kmp_itt_critical_acquiring(lck);
#endif
#if KMP_USE_INLINED_TAS
if (__kmp_user_lock_seq == lockseq_tas && !__kmp_env_consistency_check) {
KMP_ACQUIRE_TAS_LOCK(lck, global_tid);
} else
#elif KMP_USE_INLINED_FUTEX
if (__kmp_user_lock_seq == lockseq_futex && !__kmp_env_consistency_check) {
KMP_ACQUIRE_FUTEX_LOCK(lck, global_tid);
} else
#endif
{
KMP_D_LOCK_FUNC(lk, set)(lk, global_tid);
}
} else {
kmp_indirect_lock_t *ilk = *((kmp_indirect_lock_t **)lk);
lck = ilk->lock;
if (__kmp_env_consistency_check) {
__kmp_push_sync(global_tid, ct_critical, loc, lck,
__kmp_map_hint_to_lock(hint));
}
#if USE_ITT_BUILD
__kmp_itt_critical_acquiring(lck);
#endif
KMP_I_LOCK_FUNC(ilk, set)(lck, global_tid);
}
#if USE_ITT_BUILD
__kmp_itt_critical_acquired(lck);
#endif /* USE_ITT_BUILD */
KMP_PUSH_PARTITIONED_TIMER(OMP_critical);
KA_TRACE(15, ("__kmpc_critical: done T#%d\n", global_tid));
} // __kmpc_critical_with_hint
#endif // KMP_USE_DYNAMIC_LOCK
/*!
@ingroup WORK_SHARING
@param loc source location information.
@param global_tid global thread number .
@param crit identity of the critical section. This could be a pointer to a lock
associated with the critical section, or some other suitably unique value.
Leave a critical section, releasing any lock that was held during its execution.
*/
void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *crit) {
kmp_user_lock_p lck;
KC_TRACE(10, ("__kmpc_end_critical: called T#%d\n", global_tid));
#if KMP_USE_DYNAMIC_LOCK
if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
lck = (kmp_user_lock_p)crit;
KMP_ASSERT(lck != NULL);
if (__kmp_env_consistency_check) {
__kmp_pop_sync(global_tid, ct_critical, loc);
}
#if USE_ITT_BUILD
__kmp_itt_critical_releasing(lck);
#endif
#if KMP_USE_INLINED_TAS
if (__kmp_user_lock_seq == lockseq_tas && !__kmp_env_consistency_check) {
KMP_RELEASE_TAS_LOCK(lck, global_tid);
} else
#elif KMP_USE_INLINED_FUTEX
if (__kmp_user_lock_seq == lockseq_futex && !__kmp_env_consistency_check) {
KMP_RELEASE_FUTEX_LOCK(lck, global_tid);
} else
#endif
{
KMP_D_LOCK_FUNC(lck, unset)((kmp_dyna_lock_t *)lck, global_tid);
}
} else {
kmp_indirect_lock_t *ilk =
(kmp_indirect_lock_t *)TCR_PTR(*((kmp_indirect_lock_t **)crit));
KMP_ASSERT(ilk != NULL);
lck = ilk->lock;
if (__kmp_env_consistency_check) {
__kmp_pop_sync(global_tid, ct_critical, loc);
}
#if USE_ITT_BUILD
__kmp_itt_critical_releasing(lck);
#endif
KMP_I_LOCK_FUNC(ilk, unset)(lck, global_tid);
}
#else // KMP_USE_DYNAMIC_LOCK
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_CRITICAL_SIZE)) {
lck = (kmp_user_lock_p)crit;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_CRITICAL_SIZE)) {
lck = (kmp_user_lock_p)crit;
}
#endif
else { // ticket, queuing or drdpa
lck = (kmp_user_lock_p)TCR_PTR(*((kmp_user_lock_p *)crit));
}
KMP_ASSERT(lck != NULL);
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_critical, loc);
#if USE_ITT_BUILD
__kmp_itt_critical_releasing(lck);
#endif /* USE_ITT_BUILD */
// Value of 'crit' should be good for using as a critical_id of the critical
// section directive.
__kmp_release_user_lock_with_checks(lck, global_tid);
#if OMPT_SUPPORT && OMPT_BLAME
if (ompt_enabled &&
ompt_callbacks.ompt_callback(ompt_event_release_critical)) {
ompt_callbacks.ompt_callback(ompt_event_release_critical)((uint64_t)lck);
}
#endif
#endif // KMP_USE_DYNAMIC_LOCK
KMP_POP_PARTITIONED_TIMER();
KA_TRACE(15, ("__kmpc_end_critical: done T#%d\n", global_tid));
}
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid thread id.
@return one if the thread should execute the master block, zero otherwise
Start execution of a combined barrier and master. The barrier is executed inside
this function.
*/
kmp_int32 __kmpc_barrier_master(ident_t *loc, kmp_int32 global_tid) {
int status;
KC_TRACE(10, ("__kmpc_barrier_master: called T#%d\n", global_tid));
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
if (__kmp_env_consistency_check)
__kmp_check_barrier(global_tid, ct_barrier, loc);
#if USE_ITT_NOTIFY
__kmp_threads[global_tid]->th.th_ident = loc;
#endif
status = __kmp_barrier(bs_plain_barrier, global_tid, TRUE, 0, NULL, NULL);
return (status != 0) ? 0 : 1;
}
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid thread id.
Complete the execution of a combined barrier and master. This function should
only be called at the completion of the <tt>master</tt> code. Other threads will
still be waiting at the barrier and this call releases them.
*/
void __kmpc_end_barrier_master(ident_t *loc, kmp_int32 global_tid) {
KC_TRACE(10, ("__kmpc_end_barrier_master: called T#%d\n", global_tid));
__kmp_end_split_barrier(bs_plain_barrier, global_tid);
}
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid thread id.
@return one if the thread should execute the master block, zero otherwise
Start execution of a combined barrier and master(nowait) construct.
The barrier is executed inside this function.
There is no equivalent "end" function, since the
*/
kmp_int32 __kmpc_barrier_master_nowait(ident_t *loc, kmp_int32 global_tid) {
kmp_int32 ret;
KC_TRACE(10, ("__kmpc_barrier_master_nowait: called T#%d\n", global_tid));
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
if (__kmp_env_consistency_check) {
if (loc == 0) {
KMP_WARNING(ConstructIdentInvalid); // ??? What does it mean for the user?
}
__kmp_check_barrier(global_tid, ct_barrier, loc);
}
#if USE_ITT_NOTIFY
__kmp_threads[global_tid]->th.th_ident = loc;
#endif
__kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
ret = __kmpc_master(loc, global_tid);
if (__kmp_env_consistency_check) {
/* there's no __kmpc_end_master called; so the (stats) */
/* actions of __kmpc_end_master are done here */
if (global_tid < 0) {
KMP_WARNING(ThreadIdentInvalid);
}
if (ret) {
/* only one thread should do the pop since only */
/* one did the push (see __kmpc_master()) */
__kmp_pop_sync(global_tid, ct_master, loc);
}
}
return (ret);
}
/* The BARRIER for a SINGLE process section is always explicit */
/*!
@ingroup WORK_SHARING
@param loc source location information
@param global_tid global thread number
@return One if this thread should execute the single construct, zero otherwise.
Test whether to execute a <tt>single</tt> construct.
There are no implicit barriers in the two "single" calls, rather the compiler
should introduce an explicit barrier if it is required.
*/
kmp_int32 __kmpc_single(ident_t *loc, kmp_int32 global_tid) {
kmp_int32 rc = __kmp_enter_single(global_tid, loc, TRUE);
if (rc) {
// We are going to execute the single statement, so we should count it.
KMP_COUNT_BLOCK(OMP_SINGLE);
KMP_PUSH_PARTITIONED_TIMER(OMP_single);
}
#if OMPT_SUPPORT && OMPT_TRACE
kmp_info_t *this_thr = __kmp_threads[global_tid];
kmp_team_t *team = this_thr->th.th_team;
int tid = __kmp_tid_from_gtid(global_tid);
if (ompt_enabled) {
if (rc) {
if (ompt_callbacks.ompt_callback(ompt_event_single_in_block_begin)) {
ompt_callbacks.ompt_callback(ompt_event_single_in_block_begin)(
team->t.ompt_team_info.parallel_id,
team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id,
team->t.ompt_team_info.microtask);
}
} else {
if (ompt_callbacks.ompt_callback(ompt_event_single_others_begin)) {
ompt_callbacks.ompt_callback(ompt_event_single_others_begin)(
team->t.ompt_team_info.parallel_id,
team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
}
this_thr->th.ompt_thread_info.state = ompt_state_wait_single;
}
}
#endif
return rc;
}
/*!
@ingroup WORK_SHARING
@param loc source location information
@param global_tid global thread number
Mark the end of a <tt>single</tt> construct. This function should
only be called by the thread that executed the block of code protected
by the `single` construct.
*/
void __kmpc_end_single(ident_t *loc, kmp_int32 global_tid) {
__kmp_exit_single(global_tid);
KMP_POP_PARTITIONED_TIMER();
#if OMPT_SUPPORT && OMPT_TRACE
kmp_info_t *this_thr = __kmp_threads[global_tid];
kmp_team_t *team = this_thr->th.th_team;
int tid = __kmp_tid_from_gtid(global_tid);
if (ompt_enabled &&
ompt_callbacks.ompt_callback(ompt_event_single_in_block_end)) {
ompt_callbacks.ompt_callback(ompt_event_single_in_block_end)(
team->t.ompt_team_info.parallel_id,
team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
}
#endif
}
/*!
@ingroup WORK_SHARING
@param loc Source location
@param global_tid Global thread id
Mark the end of a statically scheduled loop.
*/
void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid) {
KE_TRACE(10, ("__kmpc_for_static_fini called T#%d\n", global_tid));
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_loop_end)) {
ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
ompt_callbacks.ompt_callback(ompt_event_loop_end)(team_info->parallel_id,
task_info->task_id);
}
#endif
if (__kmp_env_consistency_check)
__kmp_pop_workshare(global_tid, ct_pdo, loc);
}
// User routines which take C-style arguments (call by value)
// different from the Fortran equivalent routines
void ompc_set_num_threads(int arg) {
// !!!!! TODO: check the per-task binding
__kmp_set_num_threads(arg, __kmp_entry_gtid());
}
void ompc_set_dynamic(int flag) {
kmp_info_t *thread;
/* For the thread-private implementation of the internal controls */
thread = __kmp_entry_thread();
__kmp_save_internal_controls(thread);
set__dynamic(thread, flag ? TRUE : FALSE);
}
void ompc_set_nested(int flag) {
kmp_info_t *thread;
/* For the thread-private internal controls implementation */
thread = __kmp_entry_thread();
__kmp_save_internal_controls(thread);
set__nested(thread, flag ? TRUE : FALSE);
}
void ompc_set_max_active_levels(int max_active_levels) {
/* TO DO */
/* we want per-task implementation of this internal control */
/* For the per-thread internal controls implementation */
__kmp_set_max_active_levels(__kmp_entry_gtid(), max_active_levels);
}
void ompc_set_schedule(omp_sched_t kind, int modifier) {
// !!!!! TODO: check the per-task binding
__kmp_set_schedule(__kmp_entry_gtid(), (kmp_sched_t)kind, modifier);
}
int ompc_get_ancestor_thread_num(int level) {
return __kmp_get_ancestor_thread_num(__kmp_entry_gtid(), level);
}
int ompc_get_team_size(int level) {
return __kmp_get_team_size(__kmp_entry_gtid(), level);
}
void kmpc_set_stacksize(int arg) {
// __kmp_aux_set_stacksize initializes the library if needed
__kmp_aux_set_stacksize(arg);
}
void kmpc_set_stacksize_s(size_t arg) {
// __kmp_aux_set_stacksize initializes the library if needed
__kmp_aux_set_stacksize(arg);
}
void kmpc_set_blocktime(int arg) {
int gtid, tid;
kmp_info_t *thread;
gtid = __kmp_entry_gtid();
tid = __kmp_tid_from_gtid(gtid);
thread = __kmp_thread_from_gtid(gtid);
__kmp_aux_set_blocktime(arg, thread, tid);
}
void kmpc_set_library(int arg) {
// __kmp_user_set_library initializes the library if needed
__kmp_user_set_library((enum library_type)arg);
}
void kmpc_set_defaults(char const *str) {
// __kmp_aux_set_defaults initializes the library if needed
__kmp_aux_set_defaults(str, KMP_STRLEN(str));
}
void kmpc_set_disp_num_buffers(int arg) {
// ignore after initialization because some teams have already
// allocated dispatch buffers
if (__kmp_init_serial == 0 && arg > 0)
__kmp_dispatch_num_buffers = arg;
}
int kmpc_set_affinity_mask_proc(int proc, void **mask) {
#if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
return -1;
#else
if (!TCR_4(__kmp_init_middle)) {
__kmp_middle_initialize();
}
return __kmp_aux_set_affinity_mask_proc(proc, mask);
#endif
}
int kmpc_unset_affinity_mask_proc(int proc, void **mask) {
#if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
return -1;
#else
if (!TCR_4(__kmp_init_middle)) {
__kmp_middle_initialize();
}
return __kmp_aux_unset_affinity_mask_proc(proc, mask);
#endif
}
int kmpc_get_affinity_mask_proc(int proc, void **mask) {
#if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
return -1;
#else
if (!TCR_4(__kmp_init_middle)) {
__kmp_middle_initialize();
}
return __kmp_aux_get_affinity_mask_proc(proc, mask);
#endif
}
/* -------------------------------------------------------------------------- */
/*!
@ingroup THREADPRIVATE
@param loc source location information
@param gtid global thread number
@param cpy_size size of the cpy_data buffer
@param cpy_data pointer to data to be copied
@param cpy_func helper function to call for copying data
@param didit flag variable: 1=single thread; 0=not single thread
__kmpc_copyprivate implements the interface for the private data broadcast
needed for the copyprivate clause associated with a single region in an
OpenMP<sup>*</sup> program (both C and Fortran).
All threads participating in the parallel region call this routine.
One of the threads (called the single thread) should have the <tt>didit</tt>
variable set to 1 and all other threads should have that variable set to 0.
All threads pass a pointer to a data buffer (cpy_data) that they have built.
The OpenMP specification forbids the use of nowait on the single region when a
copyprivate clause is present. However, @ref __kmpc_copyprivate implements a
barrier internally to avoid race conditions, so the code generation for the
single region should avoid generating a barrier after the call to @ref
__kmpc_copyprivate.
The <tt>gtid</tt> parameter is the global thread id for the current thread.
The <tt>loc</tt> parameter is a pointer to source location information.
Internal implementation: The single thread will first copy its descriptor
address (cpy_data) to a team-private location, then the other threads will each
call the function pointed to by the parameter cpy_func, which carries out the
copy by copying the data using the cpy_data buffer.
The cpy_func routine used for the copy and the contents of the data area defined
by cpy_data and cpy_size may be built in any fashion that will allow the copy
to be done. For instance, the cpy_data buffer can hold the actual data to be
copied or it may hold a list of pointers to the data. The cpy_func routine must
interpret the cpy_data buffer appropriately.
The interface to cpy_func is as follows:
@code
void cpy_func( void *destination, void *source )
@endcode
where void *destination is the cpy_data pointer for the thread being copied to
and void *source is the cpy_data pointer for the thread being copied from.
*/
void __kmpc_copyprivate(ident_t *loc, kmp_int32 gtid, size_t cpy_size,
void *cpy_data, void (*cpy_func)(void *, void *),
kmp_int32 didit) {
void **data_ptr;
KC_TRACE(10, ("__kmpc_copyprivate: called T#%d\n", gtid));
KMP_MB();
data_ptr = &__kmp_team_from_gtid(gtid)->t.t_copypriv_data;
if (__kmp_env_consistency_check) {
if (loc == 0) {
KMP_WARNING(ConstructIdentInvalid);
}
}
// ToDo: Optimize the following two barriers into some kind of split barrier
if (didit)
*data_ptr = cpy_data;
/* This barrier is not a barrier region boundary */
#if USE_ITT_NOTIFY
__kmp_threads[gtid]->th.th_ident = loc;
#endif
__kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
if (!didit)
(*cpy_func)(cpy_data, *data_ptr);
// Consider next barrier a user-visible barrier for barrier region boundaries
// Nesting checks are already handled by the single construct checks
#if USE_ITT_NOTIFY
__kmp_threads[gtid]->th.th_ident = loc; // TODO: check if it is needed (e.g.
// tasks can overwrite the location)
#endif
__kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
}
/* -------------------------------------------------------------------------- */
#define INIT_LOCK __kmp_init_user_lock_with_checks
#define INIT_NESTED_LOCK __kmp_init_nested_user_lock_with_checks
#define ACQUIRE_LOCK __kmp_acquire_user_lock_with_checks
#define ACQUIRE_LOCK_TIMED __kmp_acquire_user_lock_with_checks_timed
#define ACQUIRE_NESTED_LOCK __kmp_acquire_nested_user_lock_with_checks
#define ACQUIRE_NESTED_LOCK_TIMED \
__kmp_acquire_nested_user_lock_with_checks_timed
#define RELEASE_LOCK __kmp_release_user_lock_with_checks
#define RELEASE_NESTED_LOCK __kmp_release_nested_user_lock_with_checks
#define TEST_LOCK __kmp_test_user_lock_with_checks
#define TEST_NESTED_LOCK __kmp_test_nested_user_lock_with_checks
#define DESTROY_LOCK __kmp_destroy_user_lock_with_checks
#define DESTROY_NESTED_LOCK __kmp_destroy_nested_user_lock_with_checks
// TODO: Make check abort messages use location info & pass it into
// with_checks routines
#if KMP_USE_DYNAMIC_LOCK
// internal lock initializer
static __forceinline void __kmp_init_lock_with_hint(ident_t *loc, void **lock,
kmp_dyna_lockseq_t seq) {
if (KMP_IS_D_LOCK(seq)) {
KMP_INIT_D_LOCK(lock, seq);
#if USE_ITT_BUILD
__kmp_itt_lock_creating((kmp_user_lock_p)lock, NULL);
#endif
} else {
KMP_INIT_I_LOCK(lock, seq);
#if USE_ITT_BUILD
kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(lock);
__kmp_itt_lock_creating(ilk->lock, loc);
#endif
}
}
// internal nest lock initializer
static __forceinline void
__kmp_init_nest_lock_with_hint(ident_t *loc, void **lock,
kmp_dyna_lockseq_t seq) {
#if KMP_USE_TSX
// Don't have nested lock implementation for speculative locks
if (seq == lockseq_hle || seq == lockseq_rtm || seq == lockseq_adaptive)
seq = __kmp_user_lock_seq;
#endif
switch (seq) {
case lockseq_tas:
seq = lockseq_nested_tas;
break;
#if KMP_USE_FUTEX
case lockseq_futex:
seq = lockseq_nested_futex;
break;
#endif
case lockseq_ticket:
seq = lockseq_nested_ticket;
break;
case lockseq_queuing:
seq = lockseq_nested_queuing;
break;
case lockseq_drdpa:
seq = lockseq_nested_drdpa;
break;
default:
seq = lockseq_nested_queuing;
}
KMP_INIT_I_LOCK(lock, seq);
#if USE_ITT_BUILD
kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(lock);
__kmp_itt_lock_creating(ilk->lock, loc);
#endif
}
/* initialize the lock with a hint */
void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, void **user_lock,
uintptr_t hint) {
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check && user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, "omp_init_lock_with_hint");
}
__kmp_init_lock_with_hint(loc, user_lock, __kmp_map_hint_to_lock(hint));
}
/* initialize the lock with a hint */
void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
void **user_lock, uintptr_t hint) {
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check && user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, "omp_init_nest_lock_with_hint");
}
__kmp_init_nest_lock_with_hint(loc, user_lock, __kmp_map_hint_to_lock(hint));
}
#endif // KMP_USE_DYNAMIC_LOCK
/* initialize the lock */
void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check && user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, "omp_init_lock");
}
__kmp_init_lock_with_hint(loc, user_lock, __kmp_user_lock_seq);
#else // KMP_USE_DYNAMIC_LOCK
static char const *const func = "omp_init_lock";
kmp_user_lock_p lck;
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check) {
if (user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, func);
}
}
KMP_CHECK_USER_LOCK_INIT();
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_user_lock_allocate(user_lock, gtid, 0);
}
INIT_LOCK(lck);
__kmp_set_user_lock_location(lck, loc);
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_init_lock)) {
ompt_callbacks.ompt_callback(ompt_event_init_lock)((uint64_t)lck);
}
#endif
#if USE_ITT_BUILD
__kmp_itt_lock_creating(lck);
#endif /* USE_ITT_BUILD */
#endif // KMP_USE_DYNAMIC_LOCK
} // __kmpc_init_lock
/* initialize the lock */
void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check && user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, "omp_init_nest_lock");
}
__kmp_init_nest_lock_with_hint(loc, user_lock, __kmp_user_lock_seq);
#else // KMP_USE_DYNAMIC_LOCK
static char const *const func = "omp_init_nest_lock";
kmp_user_lock_p lck;
KMP_DEBUG_ASSERT(__kmp_init_serial);
if (__kmp_env_consistency_check) {
if (user_lock == NULL) {
KMP_FATAL(LockIsUninitialized, func);
}
}
KMP_CHECK_USER_LOCK_INIT();
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_user_lock_allocate(user_lock, gtid, 0);
}
INIT_NESTED_LOCK(lck);
__kmp_set_user_lock_location(lck, loc);
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_init_nest_lock)) {
ompt_callbacks.ompt_callback(ompt_event_init_nest_lock)((uint64_t)lck);
}
#endif
#if USE_ITT_BUILD
__kmp_itt_lock_creating(lck);
#endif /* USE_ITT_BUILD */
#endif // KMP_USE_DYNAMIC_LOCK
} // __kmpc_init_nest_lock
void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
#if USE_ITT_BUILD
kmp_user_lock_p lck;
if (KMP_EXTRACT_D_TAG(user_lock) == 0) {
lck = ((kmp_indirect_lock_t *)KMP_LOOKUP_I_LOCK(user_lock))->lock;
} else {
lck = (kmp_user_lock_p)user_lock;
}
__kmp_itt_lock_destroyed(lck);
#endif
KMP_D_LOCK_FUNC(user_lock, destroy)((kmp_dyna_lock_t *)user_lock);
#else
kmp_user_lock_p lck;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_destroy_lock");
}
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_destroy_lock)) {
ompt_callbacks.ompt_callback(ompt_event_destroy_lock)((uint64_t)lck);
}
#endif
#if USE_ITT_BUILD
__kmp_itt_lock_destroyed(lck);
#endif /* USE_ITT_BUILD */
DESTROY_LOCK(lck);
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
;
}
#endif
else {
__kmp_user_lock_free(user_lock, gtid, lck);
}
#endif // KMP_USE_DYNAMIC_LOCK
} // __kmpc_destroy_lock
/* destroy the lock */
void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
#if USE_ITT_BUILD
kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(user_lock);
__kmp_itt_lock_destroyed(ilk->lock);
#endif
KMP_D_LOCK_FUNC(user_lock, destroy)((kmp_dyna_lock_t *)user_lock);
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_destroy_nest_lock");
}
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled &&
ompt_callbacks.ompt_callback(ompt_event_destroy_nest_lock)) {
ompt_callbacks.ompt_callback(ompt_event_destroy_nest_lock)((uint64_t)lck);
}
#endif
#if USE_ITT_BUILD
__kmp_itt_lock_destroyed(lck);
#endif /* USE_ITT_BUILD */
DESTROY_NESTED_LOCK(lck);
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
;
}
#endif
else {
__kmp_user_lock_free(user_lock, gtid, lck);
}
#endif // KMP_USE_DYNAMIC_LOCK
} // __kmpc_destroy_nest_lock
void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
KMP_COUNT_BLOCK(OMP_set_lock);
#if KMP_USE_DYNAMIC_LOCK
int tag = KMP_EXTRACT_D_TAG(user_lock);
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring(
(kmp_user_lock_p)
user_lock); // itt function will get to the right lock object.
#endif
#if KMP_USE_INLINED_TAS
if (tag == locktag_tas && !__kmp_env_consistency_check) {
KMP_ACQUIRE_TAS_LOCK(user_lock, gtid);
} else
#elif KMP_USE_INLINED_FUTEX
if (tag == locktag_futex && !__kmp_env_consistency_check) {
KMP_ACQUIRE_FUTEX_LOCK(user_lock, gtid);
} else
#endif
{
__kmp_direct_set[tag]((kmp_dyna_lock_t *)user_lock, gtid);
}
#if USE_ITT_BUILD
__kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
#endif
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_set_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring(lck);
#endif /* USE_ITT_BUILD */
ACQUIRE_LOCK(lck, gtid);
#if USE_ITT_BUILD
__kmp_itt_lock_acquired(lck);
#endif /* USE_ITT_BUILD */
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_acquired_lock)) {
ompt_callbacks.ompt_callback(ompt_event_acquired_lock)((uint64_t)lck);
}
#endif
#endif // KMP_USE_DYNAMIC_LOCK
}
void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
#endif
KMP_D_LOCK_FUNC(user_lock, set)((kmp_dyna_lock_t *)user_lock, gtid);
#if USE_ITT_BUILD
__kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
#endif
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
// missing support here: need to know whether acquired first or not
}
#endif
#else // KMP_USE_DYNAMIC_LOCK
int acquire_status;
kmp_user_lock_p lck;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_set_nest_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring(lck);
#endif /* USE_ITT_BUILD */
ACQUIRE_NESTED_LOCK(lck, gtid, &acquire_status);
#if USE_ITT_BUILD
__kmp_itt_lock_acquired(lck);
#endif /* USE_ITT_BUILD */
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
if (acquire_status == KMP_LOCK_ACQUIRED_FIRST) {
if (ompt_callbacks.ompt_callback(ompt_event_acquired_nest_lock_first))
ompt_callbacks.ompt_callback(ompt_event_acquired_nest_lock_first)(
(uint64_t)lck);
} else {
if (ompt_callbacks.ompt_callback(ompt_event_acquired_nest_lock_next))
ompt_callbacks.ompt_callback(ompt_event_acquired_nest_lock_next)(
(uint64_t)lck);
}
}
#endif
#endif // KMP_USE_DYNAMIC_LOCK
}
void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
int tag = KMP_EXTRACT_D_TAG(user_lock);
#if USE_ITT_BUILD
__kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
#endif
#if KMP_USE_INLINED_TAS
if (tag == locktag_tas && !__kmp_env_consistency_check) {
KMP_RELEASE_TAS_LOCK(user_lock, gtid);
} else
#elif KMP_USE_INLINED_FUTEX
if (tag == locktag_futex && !__kmp_env_consistency_check) {
KMP_RELEASE_FUTEX_LOCK(user_lock, gtid);
} else
#endif
{
__kmp_direct_unset[tag]((kmp_dyna_lock_t *)user_lock, gtid);
}
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
/* Can't use serial interval since not block structured */
/* release the lock */
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
#if KMP_OS_LINUX && \
(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
// "fast" path implemented to fix customer performance issue
#if USE_ITT_BUILD
__kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
#endif /* USE_ITT_BUILD */
TCW_4(((kmp_user_lock_p)user_lock)->tas.lk.poll, 0);
KMP_MB();
return;
#else
lck = (kmp_user_lock_p)user_lock;
#endif
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_unset_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_releasing(lck);
#endif /* USE_ITT_BUILD */
RELEASE_LOCK(lck, gtid);
#if OMPT_SUPPORT && OMPT_BLAME
if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_release_lock)) {
ompt_callbacks.ompt_callback(ompt_event_release_lock)((uint64_t)lck);
}
#endif
#endif // KMP_USE_DYNAMIC_LOCK
}
/* release the lock */
void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
#if USE_ITT_BUILD
__kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
#endif
KMP_D_LOCK_FUNC(user_lock, unset)((kmp_dyna_lock_t *)user_lock, gtid);
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
/* Can't use serial interval since not block structured */
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
#if KMP_OS_LINUX && \
(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
// "fast" path implemented to fix customer performance issue
kmp_tas_lock_t *tl = (kmp_tas_lock_t *)user_lock;
#if USE_ITT_BUILD
__kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
#endif /* USE_ITT_BUILD */
if (--(tl->lk.depth_locked) == 0) {
TCW_4(tl->lk.poll, 0);
}
KMP_MB();
return;
#else
lck = (kmp_user_lock_p)user_lock;
#endif
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_unset_nest_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_releasing(lck);
#endif /* USE_ITT_BUILD */
int release_status;
release_status = RELEASE_NESTED_LOCK(lck, gtid);
#if OMPT_SUPPORT && OMPT_BLAME
if (ompt_enabled) {
if (release_status == KMP_LOCK_RELEASED) {
if (ompt_callbacks.ompt_callback(ompt_event_release_nest_lock_last)) {
ompt_callbacks.ompt_callback(ompt_event_release_nest_lock_last)(
(uint64_t)lck);
}
} else if (ompt_callbacks.ompt_callback(
ompt_event_release_nest_lock_prev)) {
ompt_callbacks.ompt_callback(ompt_event_release_nest_lock_prev)(
(uint64_t)lck);
}
}
#endif
#endif // KMP_USE_DYNAMIC_LOCK
}
/* try to acquire the lock */
int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
KMP_COUNT_BLOCK(OMP_test_lock);
#if KMP_USE_DYNAMIC_LOCK
int rc;
int tag = KMP_EXTRACT_D_TAG(user_lock);
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
#endif
#if KMP_USE_INLINED_TAS
if (tag == locktag_tas && !__kmp_env_consistency_check) {
KMP_TEST_TAS_LOCK(user_lock, gtid, rc);
} else
#elif KMP_USE_INLINED_FUTEX
if (tag == locktag_futex && !__kmp_env_consistency_check) {
KMP_TEST_FUTEX_LOCK(user_lock, gtid, rc);
} else
#endif
{
rc = __kmp_direct_test[tag]((kmp_dyna_lock_t *)user_lock, gtid);
}
if (rc) {
#if USE_ITT_BUILD
__kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
#endif
return FTN_TRUE;
} else {
#if USE_ITT_BUILD
__kmp_itt_lock_cancelled((kmp_user_lock_p)user_lock);
#endif
return FTN_FALSE;
}
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
int rc;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_test_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring(lck);
#endif /* USE_ITT_BUILD */
rc = TEST_LOCK(lck, gtid);
#if USE_ITT_BUILD
if (rc) {
__kmp_itt_lock_acquired(lck);
} else {
__kmp_itt_lock_cancelled(lck);
}
#endif /* USE_ITT_BUILD */
return (rc ? FTN_TRUE : FTN_FALSE);
/* Can't use serial interval since not block structured */
#endif // KMP_USE_DYNAMIC_LOCK
}
/* try to acquire the lock */
int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
#if KMP_USE_DYNAMIC_LOCK
int rc;
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
#endif
rc = KMP_D_LOCK_FUNC(user_lock, test)((kmp_dyna_lock_t *)user_lock, gtid);
#if USE_ITT_BUILD
if (rc) {
__kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
} else {
__kmp_itt_lock_cancelled((kmp_user_lock_p)user_lock);
}
#endif
return rc;
#else // KMP_USE_DYNAMIC_LOCK
kmp_user_lock_p lck;
int rc;
if ((__kmp_user_lock_kind == lk_tas) &&
(sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#if KMP_USE_FUTEX
else if ((__kmp_user_lock_kind == lk_futex) &&
(sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
OMP_NEST_LOCK_T_SIZE)) {
lck = (kmp_user_lock_p)user_lock;
}
#endif
else {
lck = __kmp_lookup_user_lock(user_lock, "omp_test_nest_lock");
}
#if USE_ITT_BUILD
__kmp_itt_lock_acquiring(lck);
#endif /* USE_ITT_BUILD */
rc = TEST_NESTED_LOCK(lck, gtid);
#if USE_ITT_BUILD
if (rc) {
__kmp_itt_lock_acquired(lck);
} else {
__kmp_itt_lock_cancelled(lck);
}
#endif /* USE_ITT_BUILD */
return rc;
/* Can't use serial interval since not block structured */
#endif // KMP_USE_DYNAMIC_LOCK
}
// Interface to fast scalable reduce methods routines
// keep the selected method in a thread local structure for cross-function
// usage: will be used in __kmpc_end_reduce* functions;
// another solution: to re-determine the method one more time in
// __kmpc_end_reduce* functions (new prototype required then)
// AT: which solution is better?
#define __KMP_SET_REDUCTION_METHOD(gtid, rmethod) \
((__kmp_threads[(gtid)]->th.th_local.packed_reduction_method) = (rmethod))
#define __KMP_GET_REDUCTION_METHOD(gtid) \
(__kmp_threads[(gtid)]->th.th_local.packed_reduction_method)
// description of the packed_reduction_method variable: look at the macros in
// kmp.h
// used in a critical section reduce block
static __forceinline void
__kmp_enter_critical_section_reduce_block(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *crit) {
// this lock was visible to a customer and to the threading profile tool as a
// serial overhead span (although it's used for an internal purpose only)
// why was it visible in previous implementation?
// should we keep it visible in new reduce block?
kmp_user_lock_p lck;
#if KMP_USE_DYNAMIC_LOCK
kmp_dyna_lock_t *lk = (kmp_dyna_lock_t *)crit;
// Check if it is initialized.
if (*lk == 0) {
if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
KMP_COMPARE_AND_STORE_ACQ32((volatile kmp_int32 *)crit, 0,
KMP_GET_D_TAG(__kmp_user_lock_seq));
} else {
__kmp_init_indirect_csptr(crit, loc, global_tid,
KMP_GET_I_TAG(__kmp_user_lock_seq));
}
}
// Branch for accessing the actual lock object and set operation. This
// branching is inevitable since this lock initialization does not follow the
// normal dispatch path (lock table is not used).
if (KMP_EXTRACT_D_TAG(lk) != 0) {
lck = (kmp_user_lock_p)lk;
KMP_DEBUG_ASSERT(lck != NULL);
if (__kmp_env_consistency_check) {
__kmp_push_sync(global_tid, ct_critical, loc, lck, __kmp_user_lock_seq);
}
KMP_D_LOCK_FUNC(lk, set)(lk, global_tid);
} else {
kmp_indirect_lock_t *ilk = *((kmp_indirect_lock_t **)lk);
lck = ilk->lock;
KMP_DEBUG_ASSERT(lck != NULL);
if (__kmp_env_consistency_check) {
__kmp_push_sync(global_tid, ct_critical, loc, lck, __kmp_user_lock_seq);
}
KMP_I_LOCK_FUNC(ilk, set)(lck, global_tid);
}
#else // KMP_USE_DYNAMIC_LOCK
// We know that the fast reduction code is only emitted by Intel compilers
// with 32 byte critical sections. If there isn't enough space, then we
// have to use a pointer.
if (__kmp_base_user_lock_size <= INTEL_CRITICAL_SIZE) {
lck = (kmp_user_lock_p)crit;
} else {
lck = __kmp_get_critical_section_ptr(crit, loc, global_tid);
}
KMP_DEBUG_ASSERT(lck != NULL);
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_critical, loc, lck);
__kmp_acquire_user_lock_with_checks(lck, global_tid);
#endif // KMP_USE_DYNAMIC_LOCK
}
// used in a critical section reduce block
static __forceinline void
__kmp_end_critical_section_reduce_block(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *crit) {
kmp_user_lock_p lck;
#if KMP_USE_DYNAMIC_LOCK
if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
lck = (kmp_user_lock_p)crit;
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_critical, loc);
KMP_D_LOCK_FUNC(lck, unset)((kmp_dyna_lock_t *)lck, global_tid);
} else {
kmp_indirect_lock_t *ilk =
(kmp_indirect_lock_t *)TCR_PTR(*((kmp_indirect_lock_t **)crit));
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_critical, loc);
KMP_I_LOCK_FUNC(ilk, unset)(ilk->lock, global_tid);
}
#else // KMP_USE_DYNAMIC_LOCK
// We know that the fast reduction code is only emitted by Intel compilers
// with 32 byte critical sections. If there isn't enough space, then we have
// to use a pointer.
if (__kmp_base_user_lock_size > 32) {
lck = *((kmp_user_lock_p *)crit);
KMP_ASSERT(lck != NULL);
} else {
lck = (kmp_user_lock_p)crit;
}
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_critical, loc);
__kmp_release_user_lock_with_checks(lck, global_tid);
#endif // KMP_USE_DYNAMIC_LOCK
} // __kmp_end_critical_section_reduce_block
/* 2.a.i. Reduce Block without a terminating barrier */
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid global thread number
@param num_vars number of items (variables) to be reduced
@param reduce_size size of data in bytes to be reduced
@param reduce_data pointer to data to be reduced
@param reduce_func callback function providing reduction operation on two
operands and returning result of reduction in lhs_data
@param lck pointer to the unique lock data structure
@result 1 for the master thread, 0 for all other team threads, 2 for all team
threads if atomic reduction needed
The nowait version is used for a reduce clause with the nowait argument.
*/
kmp_int32
__kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size, void *reduce_data,
void (*reduce_func)(void *lhs_data, void *rhs_data),
kmp_critical_name *lck) {
KMP_COUNT_BLOCK(REDUCE_nowait);
int retval = 0;
PACKED_REDUCTION_METHOD_T packed_reduction_method;
#if OMP_40_ENABLED
kmp_team_t *team;
kmp_info_t *th;
int teams_swapped = 0, task_state;
#endif
KA_TRACE(10, ("__kmpc_reduce_nowait() enter: called T#%d\n", global_tid));
// why do we need this initialization here at all?
// Reduction clause can not be used as a stand-alone directive.
// do not call __kmp_serial_initialize(), it will be called by
// __kmp_parallel_initialize() if needed
// possible detection of false-positive race by the threadchecker ???
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
// check correctness of reduce block nesting
#if KMP_USE_DYNAMIC_LOCK
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_reduce, loc, NULL, 0);
#else
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_reduce, loc, NULL);
#endif
#if OMP_40_ENABLED
th = __kmp_thread_from_gtid(global_tid);
if (th->th.th_teams_microtask) { // AC: check if we are inside the teams
// construct?
team = th->th.th_team;
if (team->t.t_level == th->th.th_teams_level) {
// this is reduction at teams construct
KMP_DEBUG_ASSERT(!th->th.th_info.ds.ds_tid); // AC: check that tid == 0
// Let's swap teams temporarily for the reduction barrier
teams_swapped = 1;
th->th.th_info.ds.ds_tid = team->t.t_master_tid;
th->th.th_team = team->t.t_parent;
th->th.th_team_nproc = th->th.th_team->t.t_nproc;
th->th.th_task_team = th->th.th_team->t.t_task_team[0];
task_state = th->th.th_task_state;
th->th.th_task_state = 0;
}
}
#endif // OMP_40_ENABLED
// packed_reduction_method value will be reused by __kmp_end_reduce* function,
// the value should be kept in a variable
// the variable should be either a construct-specific or thread-specific
// property, not a team specific property
// (a thread can reach the next reduce block on the next construct, reduce
// method may differ on the next construct)
// an ident_t "loc" parameter could be used as a construct-specific property
// (what if loc == 0?)
// (if both construct-specific and team-specific variables were shared,
// then unness extra syncs should be needed)
// a thread-specific variable is better regarding two issues above (next
// construct and extra syncs)
// a thread-specific "th_local.reduction_method" variable is used currently
// each thread executes 'determine' and 'set' lines (no need to execute by one
// thread, to avoid unness extra syncs)
packed_reduction_method = __kmp_determine_reduction_method(
loc, global_tid, num_vars, reduce_size, reduce_data, reduce_func, lck);
__KMP_SET_REDUCTION_METHOD(global_tid, packed_reduction_method);
if (packed_reduction_method == critical_reduce_block) {
__kmp_enter_critical_section_reduce_block(loc, global_tid, lck);
retval = 1;
} else if (packed_reduction_method == empty_reduce_block) {
// usage: if team size == 1, no synchronization is required ( Intel
// platforms only )
retval = 1;
} else if (packed_reduction_method == atomic_reduce_block) {
retval = 2;
// all threads should do this pop here (because __kmpc_end_reduce_nowait()
// won't be called by the code gen)
// (it's not quite good, because the checking block has been closed by
// this 'pop',
// but atomic operation has not been executed yet, will be executed
// slightly later, literally on next instruction)
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_reduce, loc);
} else if (TEST_REDUCTION_METHOD(packed_reduction_method,
tree_reduce_block)) {
// AT: performance issue: a real barrier here
// AT: (if master goes slow, other threads are blocked here waiting for the
// master to come and release them)
// AT: (it's not what a customer might expect specifying NOWAIT clause)
// AT: (specifying NOWAIT won't result in improvement of performance, it'll
// be confusing to a customer)
// AT: another implementation of *barrier_gather*nowait() (or some other design)
// might go faster and be more in line with sense of NOWAIT
// AT: TO DO: do epcc test and compare times
// this barrier should be invisible to a customer and to the threading profile
// tool (it's neither a terminating barrier nor customer's code, it's
// used for an internal purpose)
#if USE_ITT_NOTIFY
__kmp_threads[global_tid]->th.th_ident = loc;
#endif
retval =
__kmp_barrier(UNPACK_REDUCTION_BARRIER(packed_reduction_method),
global_tid, FALSE, reduce_size, reduce_data, reduce_func);
retval = (retval != 0) ? (0) : (1);
// all other workers except master should do this pop here
// ( none of other workers will get to __kmpc_end_reduce_nowait() )
if (__kmp_env_consistency_check) {
if (retval == 0) {
__kmp_pop_sync(global_tid, ct_reduce, loc);
}
}
} else {
// should never reach this block
KMP_ASSERT(0); // "unexpected method"
}
#if OMP_40_ENABLED
if (teams_swapped) {
// Restore thread structure
th->th.th_info.ds.ds_tid = 0;
th->th.th_team = team;
th->th.th_team_nproc = team->t.t_nproc;
th->th.th_task_team = team->t.t_task_team[task_state];
th->th.th_task_state = task_state;
}
#endif
KA_TRACE(
10,
("__kmpc_reduce_nowait() exit: called T#%d: method %08x, returns %08x\n",
global_tid, packed_reduction_method, retval));
return retval;
}
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid global thread id.
@param lck pointer to the unique lock data structure
Finish the execution of a reduce nowait.
*/
void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *lck) {
PACKED_REDUCTION_METHOD_T packed_reduction_method;
KA_TRACE(10, ("__kmpc_end_reduce_nowait() enter: called T#%d\n", global_tid));
packed_reduction_method = __KMP_GET_REDUCTION_METHOD(global_tid);
if (packed_reduction_method == critical_reduce_block) {
__kmp_end_critical_section_reduce_block(loc, global_tid, lck);
} else if (packed_reduction_method == empty_reduce_block) {
// usage: if team size == 1, no synchronization is required ( on Intel
// platforms only )
} else if (packed_reduction_method == atomic_reduce_block) {
// neither master nor other workers should get here
// (code gen does not generate this call in case 2: atomic reduce block)
// actually it's better to remove this elseif at all;
// after removal this value will checked by the 'else' and will assert
} else if (TEST_REDUCTION_METHOD(packed_reduction_method,
tree_reduce_block)) {
// only master gets here
} else {
// should never reach this block
KMP_ASSERT(0); // "unexpected method"
}
if (__kmp_env_consistency_check)
__kmp_pop_sync(global_tid, ct_reduce, loc);
KA_TRACE(10, ("__kmpc_end_reduce_nowait() exit: called T#%d: method %08x\n",
global_tid, packed_reduction_method));
return;
}
/* 2.a.ii. Reduce Block with a terminating barrier */
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid global thread number
@param num_vars number of items (variables) to be reduced
@param reduce_size size of data in bytes to be reduced
@param reduce_data pointer to data to be reduced
@param reduce_func callback function providing reduction operation on two
operands and returning result of reduction in lhs_data
@param lck pointer to the unique lock data structure
@result 1 for the master thread, 0 for all other team threads, 2 for all team
threads if atomic reduction needed
A blocking reduce that includes an implicit barrier.
*/
kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size, void *reduce_data,
void (*reduce_func)(void *lhs_data, void *rhs_data),
kmp_critical_name *lck) {
KMP_COUNT_BLOCK(REDUCE_wait);
int retval = 0;
PACKED_REDUCTION_METHOD_T packed_reduction_method;
KA_TRACE(10, ("__kmpc_reduce() enter: called T#%d\n", global_tid));
// why do we need this initialization here at all?
// Reduction clause can not be a stand-alone directive.
// do not call __kmp_serial_initialize(), it will be called by
// __kmp_parallel_initialize() if needed
// possible detection of false-positive race by the threadchecker ???
if (!TCR_4(__kmp_init_parallel))
__kmp_parallel_initialize();
// check correctness of reduce block nesting
#if KMP_USE_DYNAMIC_LOCK
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_reduce, loc, NULL, 0);
#else
if (__kmp_env_consistency_check)
__kmp_push_sync(global_tid, ct_reduce, loc, NULL);
#endif
packed_reduction_method = __kmp_determine_reduction_method(
loc, global_tid, num_vars, reduce_size, reduce_data, reduce_func, lck);
__KMP_SET_REDUCTION_METHOD(global_tid, packed_reduction_method);
if (packed_reduction_method == critical_reduce_block) {
__kmp_enter_critical_section_reduce_block(loc, global_tid, lck);
retval = 1;
} else if (packed_reduction_method == empty_reduce_block) {
// usage: if team size == 1, no synchronization is required ( Intel
// platforms only )
retval = 1;
} else if (packed_reduction_method == atomic_reduce_block) {
retval = 2;
} else if (TEST_REDUCTION_METHOD(packed_reduction_method,
tree_reduce_block)) {
// case tree_reduce_block:
// this barrier should be visible to a customer and to the threading profile
// tool (it's a terminating barrier on constructs if NOWAIT not specified)
#if USE_ITT_NOTIFY
__kmp_threads[global_tid]->th.th_ident =
loc; // needed for correct notification of frames
#endif
retval =
__kmp_barrier(UNPACK_REDUCTION_BARRIER(packed_reduction_method),
global_tid, TRUE, reduce_size, reduce_data, reduce_func);
retval = (retval != 0) ? (0)