/* Copyright (C) 2005, 2008, 2009 Free Software Foundation, Inc.
|
/* Copyright (C) 2005, 2008, 2009 Free Software Foundation, Inc.
|
Contributed by Richard Henderson <rth@redhat.com>.
|
Contributed by Richard Henderson <rth@redhat.com>.
|
|
|
This file is part of the GNU OpenMP Library (libgomp).
|
This file is part of the GNU OpenMP Library (libgomp).
|
|
|
Libgomp is free software; you can redistribute it and/or modify it
|
Libgomp is free software; you can redistribute it and/or modify it
|
under the terms of the GNU General Public License as published by
|
under the terms of the GNU General Public License as published by
|
the Free Software Foundation; either version 3, or (at your option)
|
the Free Software Foundation; either version 3, or (at your option)
|
any later version.
|
any later version.
|
|
|
Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
|
Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
|
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
FOR A PARTICULAR PURPOSE. See the GNU General Public License for
|
FOR A PARTICULAR PURPOSE. See the GNU General Public License for
|
more details.
|
more details.
|
|
|
Under Section 7 of GPL version 3, you are granted additional
|
Under Section 7 of GPL version 3, you are granted additional
|
permissions described in the GCC Runtime Library Exception, version
|
permissions described in the GCC Runtime Library Exception, version
|
3.1, as published by the Free Software Foundation.
|
3.1, as published by the Free Software Foundation.
|
|
|
You should have received a copy of the GNU General Public License and
|
You should have received a copy of the GNU General Public License and
|
a copy of the GCC Runtime Library Exception along with this program;
|
a copy of the GCC Runtime Library Exception along with this program;
|
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
<http://www.gnu.org/licenses/>. */
|
<http://www.gnu.org/licenses/>. */
|
|
|
/* This file contains routines for managing work-share iteration, both
|
/* This file contains routines for managing work-share iteration, both
|
for loops and sections. */
|
for loops and sections. */
|
|
|
#include "libgomp.h"
|
#include "libgomp.h"
|
#include <stdlib.h>
|
#include <stdlib.h>
|
|
|
|
|
/* This function implements the STATIC scheduling method. The caller should
|
/* This function implements the STATIC scheduling method. The caller should
|
iterate *pstart <= x < *pend. Return zero if there are more iterations
|
iterate *pstart <= x < *pend. Return zero if there are more iterations
|
to perform; nonzero if not. Return less than 0 if this thread had
|
to perform; nonzero if not. Return less than 0 if this thread had
|
received the absolutely last iteration. */
|
received the absolutely last iteration. */
|
|
|
int
|
int
|
gomp_iter_static_next (long *pstart, long *pend)
|
gomp_iter_static_next (long *pstart, long *pend)
|
{
|
{
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_team *team = thr->ts.team;
|
struct gomp_team *team = thr->ts.team;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
|
|
if (thr->ts.static_trip == -1)
|
if (thr->ts.static_trip == -1)
|
return -1;
|
return -1;
|
|
|
/* Quick test for degenerate teams and orphaned constructs. */
|
/* Quick test for degenerate teams and orphaned constructs. */
|
if (nthreads == 1)
|
if (nthreads == 1)
|
{
|
{
|
*pstart = ws->next;
|
*pstart = ws->next;
|
*pend = ws->end;
|
*pend = ws->end;
|
thr->ts.static_trip = -1;
|
thr->ts.static_trip = -1;
|
return ws->next == ws->end;
|
return ws->next == ws->end;
|
}
|
}
|
|
|
/* We interpret chunk_size zero as "unspecified", which means that we
|
/* We interpret chunk_size zero as "unspecified", which means that we
|
should break up the iterations such that each thread makes only one
|
should break up the iterations such that each thread makes only one
|
trip through the outer loop. */
|
trip through the outer loop. */
|
if (ws->chunk_size == 0)
|
if (ws->chunk_size == 0)
|
{
|
{
|
unsigned long n, q, i;
|
unsigned long n, q, i;
|
unsigned long s0, e0;
|
unsigned long s0, e0;
|
long s, e;
|
long s, e;
|
|
|
if (thr->ts.static_trip > 0)
|
if (thr->ts.static_trip > 0)
|
return 1;
|
return 1;
|
|
|
/* Compute the total number of iterations. */
|
/* Compute the total number of iterations. */
|
s = ws->incr + (ws->incr > 0 ? -1 : 1);
|
s = ws->incr + (ws->incr > 0 ? -1 : 1);
|
n = (ws->end - ws->next + s) / ws->incr;
|
n = (ws->end - ws->next + s) / ws->incr;
|
i = thr->ts.team_id;
|
i = thr->ts.team_id;
|
|
|
/* Compute the "zero-based" start and end points. That is, as
|
/* Compute the "zero-based" start and end points. That is, as
|
if the loop began at zero and incremented by one. */
|
if the loop began at zero and incremented by one. */
|
q = n / nthreads;
|
q = n / nthreads;
|
q += (q * nthreads != n);
|
q += (q * nthreads != n);
|
s0 = q * i;
|
s0 = q * i;
|
e0 = s0 + q;
|
e0 = s0 + q;
|
if (e0 > n)
|
if (e0 > n)
|
e0 = n;
|
e0 = n;
|
|
|
/* Notice when no iterations allocated for this thread. */
|
/* Notice when no iterations allocated for this thread. */
|
if (s0 >= e0)
|
if (s0 >= e0)
|
{
|
{
|
thr->ts.static_trip = 1;
|
thr->ts.static_trip = 1;
|
return 1;
|
return 1;
|
}
|
}
|
|
|
/* Transform these to the actual start and end numbers. */
|
/* Transform these to the actual start and end numbers. */
|
s = (long)s0 * ws->incr + ws->next;
|
s = (long)s0 * ws->incr + ws->next;
|
e = (long)e0 * ws->incr + ws->next;
|
e = (long)e0 * ws->incr + ws->next;
|
|
|
*pstart = s;
|
*pstart = s;
|
*pend = e;
|
*pend = e;
|
thr->ts.static_trip = (e0 == n ? -1 : 1);
|
thr->ts.static_trip = (e0 == n ? -1 : 1);
|
return 0;
|
return 0;
|
}
|
}
|
else
|
else
|
{
|
{
|
unsigned long n, s0, e0, i, c;
|
unsigned long n, s0, e0, i, c;
|
long s, e;
|
long s, e;
|
|
|
/* Otherwise, each thread gets exactly chunk_size iterations
|
/* Otherwise, each thread gets exactly chunk_size iterations
|
(if available) each time through the loop. */
|
(if available) each time through the loop. */
|
|
|
s = ws->incr + (ws->incr > 0 ? -1 : 1);
|
s = ws->incr + (ws->incr > 0 ? -1 : 1);
|
n = (ws->end - ws->next + s) / ws->incr;
|
n = (ws->end - ws->next + s) / ws->incr;
|
i = thr->ts.team_id;
|
i = thr->ts.team_id;
|
c = ws->chunk_size;
|
c = ws->chunk_size;
|
|
|
/* Initial guess is a C sized chunk positioned nthreads iterations
|
/* Initial guess is a C sized chunk positioned nthreads iterations
|
in, offset by our thread number. */
|
in, offset by our thread number. */
|
s0 = (thr->ts.static_trip * nthreads + i) * c;
|
s0 = (thr->ts.static_trip * nthreads + i) * c;
|
e0 = s0 + c;
|
e0 = s0 + c;
|
|
|
/* Detect overflow. */
|
/* Detect overflow. */
|
if (s0 >= n)
|
if (s0 >= n)
|
return 1;
|
return 1;
|
if (e0 > n)
|
if (e0 > n)
|
e0 = n;
|
e0 = n;
|
|
|
/* Transform these to the actual start and end numbers. */
|
/* Transform these to the actual start and end numbers. */
|
s = (long)s0 * ws->incr + ws->next;
|
s = (long)s0 * ws->incr + ws->next;
|
e = (long)e0 * ws->incr + ws->next;
|
e = (long)e0 * ws->incr + ws->next;
|
|
|
*pstart = s;
|
*pstart = s;
|
*pend = e;
|
*pend = e;
|
|
|
if (e0 == n)
|
if (e0 == n)
|
thr->ts.static_trip = -1;
|
thr->ts.static_trip = -1;
|
else
|
else
|
thr->ts.static_trip++;
|
thr->ts.static_trip++;
|
return 0;
|
return 0;
|
}
|
}
|
}
|
}
|
|
|
|
|
/* This function implements the DYNAMIC scheduling method. Arguments are
|
/* This function implements the DYNAMIC scheduling method. Arguments are
|
as for gomp_iter_static_next. This function must be called with ws->lock
|
as for gomp_iter_static_next. This function must be called with ws->lock
|
held. */
|
held. */
|
|
|
bool
|
bool
|
gomp_iter_dynamic_next_locked (long *pstart, long *pend)
|
gomp_iter_dynamic_next_locked (long *pstart, long *pend)
|
{
|
{
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
long start, end, chunk, left;
|
long start, end, chunk, left;
|
|
|
start = ws->next;
|
start = ws->next;
|
if (start == ws->end)
|
if (start == ws->end)
|
return false;
|
return false;
|
|
|
chunk = ws->chunk_size;
|
chunk = ws->chunk_size;
|
left = ws->end - start;
|
left = ws->end - start;
|
if (ws->incr < 0)
|
if (ws->incr < 0)
|
{
|
{
|
if (chunk < left)
|
if (chunk < left)
|
chunk = left;
|
chunk = left;
|
}
|
}
|
else
|
else
|
{
|
{
|
if (chunk > left)
|
if (chunk > left)
|
chunk = left;
|
chunk = left;
|
}
|
}
|
end = start + chunk;
|
end = start + chunk;
|
|
|
ws->next = end;
|
ws->next = end;
|
*pstart = start;
|
*pstart = start;
|
*pend = end;
|
*pend = end;
|
return true;
|
return true;
|
}
|
}
|
|
|
|
|
#ifdef HAVE_SYNC_BUILTINS
|
#ifdef HAVE_SYNC_BUILTINS
|
/* Similar, but doesn't require the lock held, and uses compare-and-swap
|
/* Similar, but doesn't require the lock held, and uses compare-and-swap
|
instead. Note that the only memory value that changes is ws->next. */
|
instead. Note that the only memory value that changes is ws->next. */
|
|
|
bool
|
bool
|
gomp_iter_dynamic_next (long *pstart, long *pend)
|
gomp_iter_dynamic_next (long *pstart, long *pend)
|
{
|
{
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
long start, end, nend, chunk, incr;
|
long start, end, nend, chunk, incr;
|
|
|
end = ws->end;
|
end = ws->end;
|
incr = ws->incr;
|
incr = ws->incr;
|
chunk = ws->chunk_size;
|
chunk = ws->chunk_size;
|
|
|
if (__builtin_expect (ws->mode, 1))
|
if (__builtin_expect (ws->mode, 1))
|
{
|
{
|
long tmp = __sync_fetch_and_add (&ws->next, chunk);
|
long tmp = __sync_fetch_and_add (&ws->next, chunk);
|
if (incr > 0)
|
if (incr > 0)
|
{
|
{
|
if (tmp >= end)
|
if (tmp >= end)
|
return false;
|
return false;
|
nend = tmp + chunk;
|
nend = tmp + chunk;
|
if (nend > end)
|
if (nend > end)
|
nend = end;
|
nend = end;
|
*pstart = tmp;
|
*pstart = tmp;
|
*pend = nend;
|
*pend = nend;
|
return true;
|
return true;
|
}
|
}
|
else
|
else
|
{
|
{
|
if (tmp <= end)
|
if (tmp <= end)
|
return false;
|
return false;
|
nend = tmp + chunk;
|
nend = tmp + chunk;
|
if (nend < end)
|
if (nend < end)
|
nend = end;
|
nend = end;
|
*pstart = tmp;
|
*pstart = tmp;
|
*pend = nend;
|
*pend = nend;
|
return true;
|
return true;
|
}
|
}
|
}
|
}
|
|
|
start = ws->next;
|
start = ws->next;
|
while (1)
|
while (1)
|
{
|
{
|
long left = end - start;
|
long left = end - start;
|
long tmp;
|
long tmp;
|
|
|
if (start == end)
|
if (start == end)
|
return false;
|
return false;
|
|
|
if (incr < 0)
|
if (incr < 0)
|
{
|
{
|
if (chunk < left)
|
if (chunk < left)
|
chunk = left;
|
chunk = left;
|
}
|
}
|
else
|
else
|
{
|
{
|
if (chunk > left)
|
if (chunk > left)
|
chunk = left;
|
chunk = left;
|
}
|
}
|
nend = start + chunk;
|
nend = start + chunk;
|
|
|
tmp = __sync_val_compare_and_swap (&ws->next, start, nend);
|
tmp = __sync_val_compare_and_swap (&ws->next, start, nend);
|
if (__builtin_expect (tmp == start, 1))
|
if (__builtin_expect (tmp == start, 1))
|
break;
|
break;
|
|
|
start = tmp;
|
start = tmp;
|
}
|
}
|
|
|
*pstart = start;
|
*pstart = start;
|
*pend = nend;
|
*pend = nend;
|
return true;
|
return true;
|
}
|
}
|
#endif /* HAVE_SYNC_BUILTINS */
|
#endif /* HAVE_SYNC_BUILTINS */
|
|
|
|
|
/* This function implements the GUIDED scheduling method. Arguments are
|
/* This function implements the GUIDED scheduling method. Arguments are
|
as for gomp_iter_static_next. This function must be called with the
|
as for gomp_iter_static_next. This function must be called with the
|
work share lock held. */
|
work share lock held. */
|
|
|
bool
|
bool
|
gomp_iter_guided_next_locked (long *pstart, long *pend)
|
gomp_iter_guided_next_locked (long *pstart, long *pend)
|
{
|
{
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_team *team = thr->ts.team;
|
struct gomp_team *team = thr->ts.team;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
unsigned long n, q;
|
unsigned long n, q;
|
long start, end;
|
long start, end;
|
|
|
if (ws->next == ws->end)
|
if (ws->next == ws->end)
|
return false;
|
return false;
|
|
|
start = ws->next;
|
start = ws->next;
|
n = (ws->end - start) / ws->incr;
|
n = (ws->end - start) / ws->incr;
|
q = (n + nthreads - 1) / nthreads;
|
q = (n + nthreads - 1) / nthreads;
|
|
|
if (q < ws->chunk_size)
|
if (q < ws->chunk_size)
|
q = ws->chunk_size;
|
q = ws->chunk_size;
|
if (q <= n)
|
if (q <= n)
|
end = start + q * ws->incr;
|
end = start + q * ws->incr;
|
else
|
else
|
end = ws->end;
|
end = ws->end;
|
|
|
ws->next = end;
|
ws->next = end;
|
*pstart = start;
|
*pstart = start;
|
*pend = end;
|
*pend = end;
|
return true;
|
return true;
|
}
|
}
|
|
|
#ifdef HAVE_SYNC_BUILTINS
|
#ifdef HAVE_SYNC_BUILTINS
|
/* Similar, but doesn't require the lock held, and uses compare-and-swap
|
/* Similar, but doesn't require the lock held, and uses compare-and-swap
|
instead. Note that the only memory value that changes is ws->next. */
|
instead. Note that the only memory value that changes is ws->next. */
|
|
|
bool
|
bool
|
gomp_iter_guided_next (long *pstart, long *pend)
|
gomp_iter_guided_next (long *pstart, long *pend)
|
{
|
{
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_thread *thr = gomp_thread ();
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_work_share *ws = thr->ts.work_share;
|
struct gomp_team *team = thr->ts.team;
|
struct gomp_team *team = thr->ts.team;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
unsigned long nthreads = team ? team->nthreads : 1;
|
long start, end, nend, incr;
|
long start, end, nend, incr;
|
unsigned long chunk_size;
|
unsigned long chunk_size;
|
|
|
start = ws->next;
|
start = ws->next;
|
end = ws->end;
|
end = ws->end;
|
incr = ws->incr;
|
incr = ws->incr;
|
chunk_size = ws->chunk_size;
|
chunk_size = ws->chunk_size;
|
|
|
while (1)
|
while (1)
|
{
|
{
|
unsigned long n, q;
|
unsigned long n, q;
|
long tmp;
|
long tmp;
|
|
|
if (start == end)
|
if (start == end)
|
return false;
|
return false;
|
|
|
n = (end - start) / incr;
|
n = (end - start) / incr;
|
q = (n + nthreads - 1) / nthreads;
|
q = (n + nthreads - 1) / nthreads;
|
|
|
if (q < chunk_size)
|
if (q < chunk_size)
|
q = chunk_size;
|
q = chunk_size;
|
if (__builtin_expect (q <= n, 1))
|
if (__builtin_expect (q <= n, 1))
|
nend = start + q * incr;
|
nend = start + q * incr;
|
else
|
else
|
nend = end;
|
nend = end;
|
|
|
tmp = __sync_val_compare_and_swap (&ws->next, start, nend);
|
tmp = __sync_val_compare_and_swap (&ws->next, start, nend);
|
if (__builtin_expect (tmp == start, 1))
|
if (__builtin_expect (tmp == start, 1))
|
break;
|
break;
|
|
|
start = tmp;
|
start = tmp;
|
}
|
}
|
|
|
*pstart = start;
|
*pstart = start;
|
*pend = nend;
|
*pend = nend;
|
return true;
|
return true;
|
}
|
}
|
#endif /* HAVE_SYNC_BUILTINS */
|
#endif /* HAVE_SYNC_BUILTINS */
|
|
|
