Page MenuHomePhabricator

Please use GitHub pull requests for new patches. Phabricator shutdown timeline

Differential D47709 Diff 150122 runtime/src/kmp_tasking.cpp

Changeset View

Standalone View

View Options

runtime/src/kmp_tasking.cpp

Show First 20 Lines • Show All 1,593 Lines • ▼ Show 20 Linesstatic kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid,
kmp_info_t *thread; kmp_info_t *thread;
int thread_finished = FALSE; int thread_finished = FALSE;
KMP_SET_THREAD_STATE_BLOCK(TASKWAIT); KMP_SET_THREAD_STATE_BLOCK(TASKWAIT);
KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref)); KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref));
if (__kmp_tasking_mode != tskm_immediate_exec) { if (__kmp_tasking_mode != tskm_immediate_exec) {
thread = __kmp_threads[gtid]; thread = __kmp_threads[gtid];
taskdata = thread->th.th_current_task; taskdata = thread->th.th_current_task;
hbaeUnsubmitted
Not Done
ReplyInline Actions

Can we change the function name to __kmp_omp_taskloop_task?
__kmpc_* is usually for ABI functions.

hbae: Can we change the function name to `__kmp_omp_taskloop_task`? __kmpc_* is usually for ABI…
#if OMPT_SUPPORT && OMPT_OPTIONAL #if OMPT_SUPPORT && OMPT_OPTIONAL
ompt_data_t *my_task_data; ompt_data_t *my_task_data;
ompt_data_t *my_parallel_data; ompt_data_t *my_parallel_data;
if (ompt) { if (ompt) {
my_task_data = &(taskdata->ompt_task_info.task_data); my_task_data = &(taskdata->ompt_task_info.task_data);
my_parallel_data = OMPT_CUR_TEAM_DATA(thread); my_parallel_data = OMPT_CUR_TEAM_DATA(thread);
▲ Show 20 Lines • Show All 2,135 Lines • ▼ Show 20 Lines for (i = 0; i < num_tasks; ++i) {
if (ptask_dup != NULL) // set lastprivate flag, construct fistprivates, etc. if (ptask_dup != NULL) // set lastprivate flag, construct fistprivates, etc.
ptask_dup(next_task, task, lastpriv); ptask_dup(next_task, task, lastpriv);
KA_TRACE(40, KA_TRACE(40,
("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, "
"upper %lld stride %lld, (offsets %p %p)\n", "upper %lld stride %lld, (offsets %p %p)\n",
gtid, i, next_task, lower, upper, st, gtid, i, next_task, lower, upper, st,
next_task_bounds.get_lower_offset(), next_task_bounds.get_lower_offset(),
next_task_bounds.get_upper_offset())); next_task_bounds.get_upper_offset()));
__kmp_omp_task(gtid, next_task, true); // schedule new task __kmpc_omp_task(NULL, gtid, next_task); // schedule new task
protze.joachimUnsubmitted
Done
ReplyInline Actions

same as below

protze.joachim: same as below
lower = upper + st; // adjust lower bound for the next iteration lower = upper + st; // adjust lower bound for the next iteration
} }
// free the pattern task and exit // free the pattern task and exit
__kmp_task_start(gtid, task, current_task); // make internal bookkeeping __kmp_task_start(gtid, task, current_task); // make internal bookkeeping
// do not execute the pattern task, just do internal bookkeeping // do not execute the pattern task, just do internal bookkeeping
__kmp_task_finish(gtid, task, current_task); __kmp_task_finish(gtid, task, current_task);
} }
▲ Show 20 Lines • Show All 139 Lines • ▼ Show 20 Lines#endif
p->task_dup = task_dup; p->task_dup = task_dup;
p->st = st; p->st = st;
p->ub_glob = ub_glob; p->ub_glob = ub_glob;
p->num_tasks = n_tsk1; p->num_tasks = n_tsk1;
p->grainsize = grainsize; p->grainsize = grainsize;
p->extras = ext1; p->extras = ext1;
p->tc = tc1; p->tc = tc1;
p->num_t_min = num_t_min; p->num_t_min = num_t_min;
__kmp_omp_task(gtid, new_task, true); // schedule new task __kmpc_omp_task(NULL, gtid, new_task); // schedule new task
protze.joachimUnsubmitted
Done
ReplyInline Actions

We somehow need to pass a return-address in here. So probably a new variant of __kmpc_omp_task, which takes the address and adds it to the ompt-callback invocation?

protze.joachim: We somehow need to pass a return-address in here. So probably a new variant of __kmpc_omp_task…
// execute the 1st half of current subrange // execute the 1st half of current subrange
if (n_tsk0 > num_t_min) if (n_tsk0 > num_t_min)
__kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0, __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0,
ext0, tc0, num_t_min, task_dup); ext0, tc0, num_t_min, task_dup);
else else
__kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0,
gr_size0, ext0, tc0, task_dup); gr_size0, ext0, tc0, task_dup);
hbaeUnsubmitted
Not Done
ReplyInline Actions

I think it is better to keep the original code (__kmp_omp_task() call) at this line.

hbae: I think it is better to keep the original code (`__kmp_omp_task()` call) at this line.
KA_TRACE(40, ("__kmpc_taskloop_recur(exit): T#%d\n", gtid)); KA_TRACE(40, ("__kmpc_taskloop_recur(exit): T#%d\n", gtid));
} }
/*! /*!
@ingroup TASKING @ingroup TASKING
@param loc Source location information @param loc Source location information
@param gtid Global thread ID @param gtid Global thread ID
Show All 10 Lines
Execute the taskloop construct. Execute the taskloop construct.
*/ */
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup,
int sched, kmp_uint64 grainsize, void *task_dup) { int sched, kmp_uint64 grainsize, void *task_dup) {
kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task); kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
KMP_DEBUG_ASSERT(task != NULL); KMP_DEBUG_ASSERT(task != NULL);
#if OMPT_SUPPORT && OMPT_OPTIONAL
ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
if (ompt_enabled.ompt_callback_work) {
ompt_callbacks.ompt_callback(ompt_callback_work)(
ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data),
&(task_info->task_data), 0, OMPT_GET_RETURN_ADDRESS(0));
}
#endif
if (nogroup == 0) { if (nogroup == 0) {
protze.joachimUnsubmitted
Done
ReplyInline Actions

Shouldn't there be st(ride) in the formula?

protze.joachim: Shouldn't there be st(ride) in the formula?
AndreyChurbanovUnsubmitted
Done
ReplyInline Actions

Why not use calculated below iterations count - tc? It is computed at lines 3980 - 3995, including the check against zero.

AndreyChurbanov: Why not use calculated below iterations count - tc? It is computed at lines 3980 - 3995…
sconventAuthorUnsubmitted
Done
ReplyInline Actions

There is a difference between the iteration count and the task count.
From the spec (TR6 page 423):

For a taskloop construct, count represents the number of iterations in the iteration space, which may be the result of collapsing several associated loops.

tc seems to be the task count, not the iteration count. And st(ride) is not needed for the calculation of the iteration count, right?

sconvent: There is a difference between the iteration count and the task count. From the spec (TR6 page…
AndreyChurbanovUnsubmitted
Done
ReplyInline Actions

No, the tc states for "trace count" which is exactly iterations count. And the stride is important here.
E.g. for(i=0; i<10; i+=3) causes the iteration count to be 4 not 10, because of stride=3.
For task count another parameter used - it is grainsize.

AndreyChurbanov: No, the tc states for "trace count" which is exactly iterations count. And the stride is…
sconventAuthorUnsubmitted
Done
ReplyInline Actions

You're right, I updated the diff to use tc.

sconvent: You're right, I updated the diff to use `tc`.
#if OMPT_SUPPORT && OMPT_OPTIONAL #if OMPT_SUPPORT && OMPT_OPTIONAL
OMPT_STORE_RETURN_ADDRESS(gtid); OMPT_STORE_RETURN_ADDRESS(gtid);
#endif #endif
__kmpc_taskgroup(loc, gtid); __kmpc_taskgroup(loc, gtid);
} }
// ========================================================================= // =========================================================================
// calculate loop parameters // calculate loop parameters
Show All 23 Lines#endif
if (tc == 0) { if (tc == 0) {
KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d zero-trip loop\n", gtid)); KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d zero-trip loop\n", gtid));
// free the pattern task and exit // free the pattern task and exit
__kmp_task_start(gtid, task, current_task); __kmp_task_start(gtid, task, current_task);
// do not execute anything for zero-trip loop // do not execute anything for zero-trip loop
__kmp_task_finish(gtid, task, current_task); __kmp_task_finish(gtid, task, current_task);
return; return;
} }
#if OMPT_SUPPORT && OMPT_OPTIONAL
ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
if (ompt_enabled.ompt_callback_work) {
ompt_callbacks.ompt_callback(ompt_callback_work)(
ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data),
&(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
}
#endif
if (num_tasks_min == 0) if (num_tasks_min == 0)
// TODO: can we choose better default heuristic? // TODO: can we choose better default heuristic?
num_tasks_min = num_tasks_min =
KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE); KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE);
// compute num_tasks/grainsize based on the input provided // compute num_tasks/grainsize based on the input provided
switch (sched) { switch (sched) {
case 0: // no schedule clause specified, we can choose the default case 0: // no schedule clause specified, we can choose the default
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines KA_TRACE(20, ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu"
gtid, tc, num_tasks, num_tasks_min, grainsize, extras)); gtid, tc, num_tasks, num_tasks_min, grainsize, extras));
#if OMPT_SUPPORT && OMPT_OPTIONAL #if OMPT_SUPPORT && OMPT_OPTIONAL
OMPT_STORE_RETURN_ADDRESS(gtid); OMPT_STORE_RETURN_ADDRESS(gtid);
#endif #endif
__kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
grainsize, extras, tc, task_dup); grainsize, extras, tc, task_dup);
} }
if (nogroup == 0) {
#if OMPT_SUPPORT && OMPT_OPTIONAL
OMPT_STORE_RETURN_ADDRESS(gtid);
#endif
__kmpc_end_taskgroup(loc, gtid);
}
#if OMPT_SUPPORT && OMPT_OPTIONAL #if OMPT_SUPPORT && OMPT_OPTIONAL
if (ompt_enabled.ompt_callback_work) { if (ompt_enabled.ompt_callback_work) {
ompt_callbacks.ompt_callback(ompt_callback_work)( ompt_callbacks.ompt_callback(ompt_callback_work)(
ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data), ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data),
&(task_info->task_data), 0, OMPT_GET_RETURN_ADDRESS(0)); &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
} }
#endif #endif
if (nogroup == 0) {
#if OMPT_SUPPORT && OMPT_OPTIONAL
OMPT_STORE_RETURN_ADDRESS(gtid);
#endif
__kmpc_end_taskgroup(loc, gtid);
}
KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid)); KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid));
} }
#endif #endif