1 /* Copyright (C) 2015-2021 Free Software Foundation, Inc.
2    Contributed by Alexander Monakov <amonakov@ispras.ru>
3 
4    This file is part of the GNU Offloading and Multi Processing Library
5    (libgomp).
6 
7    Libgomp is free software; you can redistribute it and/or modify it
8    under the terms of the GNU General Public License as published by
9    the Free Software Foundation; either version 3, or (at your option)
10    any later version.
11 
12    Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
13    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
14    FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
15    more details.
16 
17    Under Section 7 of GPL version 3, you are granted additional
18    permissions described in the GCC Runtime Library Exception, version
19    3.1, as published by the Free Software Foundation.
20 
21    You should have received a copy of the GNU General Public License and
22    a copy of the GCC Runtime Library Exception along with this program;
23    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
24    <http://www.gnu.org/licenses/>.  */
25 
26 /* This file handles maintenance of threads on NVPTX.  */
27 
28 #if defined __nvptx_softstack__ && defined __nvptx_unisimt__
29 
30 #include "libgomp.h"
31 #include <stdlib.h>
32 #include <string.h>
33 
34 struct gomp_thread *nvptx_thrs __attribute__((shared,nocommon));
35 
36 static void gomp_thread_start (struct gomp_thread_pool *);
37 
38 
39 /* This externally visible function handles target region entry.  It
40    sets up a per-team thread pool and transfers control by calling FN (FN_DATA)
41    in the master thread or gomp_thread_start in other threads.
42 
43    The name of this function is part of the interface with the compiler: for
44    each target region, GCC emits a PTX .kernel function that sets up soft-stack
45    and uniform-simt state and calls this function, passing in FN the original
46    function outlined for the target region.  */
47 
48 void
gomp_nvptx_main(void (* fn)(void *),void * fn_data)49 gomp_nvptx_main (void (*fn) (void *), void *fn_data)
50 {
51   int tid, ntids;
52   asm ("mov.u32 %0, %%tid.y;" : "=r" (tid));
53   asm ("mov.u32 %0, %%ntid.y;" : "=r" (ntids));
54   if (tid == 0)
55     {
56       gomp_global_icv.nthreads_var = ntids;
57       /* Starting additional threads is not supported.  */
58       gomp_global_icv.dyn_var = true;
59 
60       nvptx_thrs = alloca (ntids * sizeof (*nvptx_thrs));
61       memset (nvptx_thrs, 0, ntids * sizeof (*nvptx_thrs));
62 
63       struct gomp_thread_pool *pool = alloca (sizeof (*pool));
64       pool->threads = alloca (ntids * sizeof (*pool->threads));
65       for (tid = 0; tid < ntids; tid++)
66 	pool->threads[tid] = nvptx_thrs + tid;
67       pool->threads_size = ntids;
68       pool->threads_used = ntids;
69       pool->threads_busy = 1;
70       pool->last_team = NULL;
71       gomp_simple_barrier_init (&pool->threads_dock, ntids);
72 
73       nvptx_thrs[0].thread_pool = pool;
74       asm ("bar.sync 0;");
75       fn (fn_data);
76 
77       gomp_free_thread (nvptx_thrs);
78     }
79   else
80     {
81       asm ("bar.sync 0;");
82       gomp_thread_start (nvptx_thrs[0].thread_pool);
83     }
84 }
85 
86 /* This function contains the idle loop in which a thread waits
87    to be called up to become part of a team.  */
88 
89 static void
gomp_thread_start(struct gomp_thread_pool * pool)90 gomp_thread_start (struct gomp_thread_pool *pool)
91 {
92   struct gomp_thread *thr = gomp_thread ();
93 
94   gomp_sem_init (&thr->release, 0);
95   thr->thread_pool = pool;
96 
97   do
98     {
99       gomp_simple_barrier_wait (&pool->threads_dock);
100       if (!thr->fn)
101 	continue;
102       thr->fn (thr->data);
103       thr->fn = NULL;
104 
105       struct gomp_task *task = thr->task;
106       gomp_team_barrier_wait_final (&thr->ts.team->barrier);
107       gomp_finish_task (task);
108     }
109   /* Work around an NVIDIA driver bug: when generating sm_50 machine code,
110      it can trash stack pointer R1 in loops lacking exit edges.  Add a cheap
111      artificial exit that the driver would not be able to optimize out.  */
112   while (nvptx_thrs);
113 }
114 
115 /* Launch a team.  */
116 
117 void
gomp_team_start(void (* fn)(void *),void * data,unsigned nthreads,unsigned flags,struct gomp_team * team,struct gomp_taskgroup * taskgroup)118 gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
119 		 unsigned flags, struct gomp_team *team,
120 		 struct gomp_taskgroup *taskgroup)
121 {
122   struct gomp_thread *thr, *nthr;
123   struct gomp_task *task;
124   struct gomp_task_icv *icv;
125   struct gomp_thread_pool *pool;
126   unsigned long nthreads_var;
127 
128   thr = gomp_thread ();
129   pool = thr->thread_pool;
130   task = thr->task;
131   icv = task ? &task->icv : &gomp_global_icv;
132 
133   /* Always save the previous state, even if this isn't a nested team.
134      In particular, we should save any work share state from an outer
135      orphaned work share construct.  */
136   team->prev_ts = thr->ts;
137 
138   thr->ts.team = team;
139   thr->ts.team_id = 0;
140   ++thr->ts.level;
141   if (nthreads > 1)
142     ++thr->ts.active_level;
143   thr->ts.work_share = &team->work_shares[0];
144   thr->ts.last_work_share = NULL;
145   thr->ts.single_count = 0;
146   thr->ts.static_trip = 0;
147   thr->task = &team->implicit_task[0];
148   nthreads_var = icv->nthreads_var;
149   gomp_init_task (thr->task, task, icv);
150   team->implicit_task[0].icv.nthreads_var = nthreads_var;
151   team->implicit_task[0].taskgroup = taskgroup;
152 
153   if (nthreads == 1)
154     return;
155 
156   /* Release existing idle threads.  */
157   for (unsigned i = 1; i < nthreads; ++i)
158     {
159       nthr = pool->threads[i];
160       nthr->ts.team = team;
161       nthr->ts.work_share = &team->work_shares[0];
162       nthr->ts.last_work_share = NULL;
163       nthr->ts.team_id = i;
164       nthr->ts.level = team->prev_ts.level + 1;
165       nthr->ts.active_level = thr->ts.active_level;
166       nthr->ts.single_count = 0;
167       nthr->ts.static_trip = 0;
168       nthr->task = &team->implicit_task[i];
169       gomp_init_task (nthr->task, task, icv);
170       team->implicit_task[i].icv.nthreads_var = nthreads_var;
171       team->implicit_task[i].taskgroup = taskgroup;
172       nthr->fn = fn;
173       nthr->data = data;
174       team->ordered_release[i] = &nthr->release;
175     }
176 
177   gomp_simple_barrier_wait (&pool->threads_dock);
178 }
179 
180 int
gomp_pause_host(void)181 gomp_pause_host (void)
182 {
183   return -1;
184 }
185 
186 #include "../../team.c"
187 #endif
188