| 1 |
/* Copyright (C) 2017-2025 Free Software Foundation, Inc. |
| 2 |
Contributed by Mentor Embedded. |
| 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 AMD GCN. */ |
| 27 |
|
| 28 |
#include "libgomp.h" |
| 29 |
#include <stdlib.h> |
| 30 |
#include <string.h> |
| 31 |
|
| 32 |
#define LITTLEENDIAN_CPU |
| 33 |
#include "hsa.h" |
| 34 |
|
| 35 |
/* Defined in basic-allocator.c via config/amdgcn/allocator.c. */ |
| 36 |
void __gcn_lowlat_init (void *heap, size_t size); |
| 37 |
|
| 38 |
static void gomp_thread_start (struct gomp_thread_pool *); |
| 39 |
extern void build_indirect_map (void); |
| 40 |
|
| 41 |
/* This externally visible function handles target region entry. It |
| 42 |
sets up a per-team thread pool and transfers control by returning to |
| 43 |
the kernel in the master thread or gomp_thread_start in other threads. |
| 44 |
|
| 45 |
The name of this function is part of the interface with the compiler: for |
| 46 |
each OpenMP kernel the compiler configures the stack, then calls here. |
| 47 |
|
| 48 |
Likewise, gomp_gcn_exit_kernel is called during the kernel epilogue. */ |
| 49 |
|
| 50 |
void |
| 51 |
gomp_gcn_enter_kernel (void) |
| 52 |
{ |
| 53 |
int threadid = __builtin_gcn_dim_pos (1); |
| 54 |
|
| 55 |
if (threadid == 0) |
| 56 |
{ |
| 57 |
int numthreads = __builtin_gcn_dim_size (1); |
| 58 |
int teamid = __builtin_gcn_dim_pos(0); |
| 59 |
|
| 60 |
/* Initialize indirect function support. */ |
| 61 |
if (teamid == 0) |
| 62 |
build_indirect_map (); |
| 63 |
|
| 64 |
/* Set up the global state. |
| 65 |
Every team will do this, but that should be harmless. */ |
| 66 |
gomp_global_icv.nthreads_var = 16; |
| 67 |
gomp_global_icv.thread_limit_var = numthreads; |
| 68 |
/* Starting additional threads is not supported. */ |
| 69 |
gomp_global_icv.dyn_var = true; |
| 70 |
|
| 71 |
int __lds *gomp_team_num = (int __lds *) GOMP_TEAM_NUM; |
| 72 |
*gomp_team_num = 0; |
| 73 |
|
| 74 |
/* Initialize the team arena for optimized memory allocation. |
| 75 |
The arena has been allocated on the host side, and the address |
| 76 |
passed in via the kernargs. Each team takes a small slice of it. */ |
| 77 |
struct kernargs_abi *kernargs = |
| 78 |
(struct kernargs_abi*) __builtin_gcn_kernarg_ptr (); |
| 79 |
void *team_arena = ((void*)kernargs->arena_ptr |
| 80 |
+ kernargs->arena_size_per_team * teamid); |
| 81 |
void * __lds *arena_start = (void * __lds *)TEAM_ARENA_START; |
| 82 |
void * __lds *arena_free = (void * __lds *)TEAM_ARENA_FREE; |
| 83 |
void * __lds *arena_end = (void * __lds *)TEAM_ARENA_END; |
| 84 |
*arena_start = team_arena; |
| 85 |
*arena_free = team_arena; |
| 86 |
*arena_end = team_arena + kernargs->arena_size_per_team; |
| 87 |
|
| 88 |
/* Initialize the low-latency heap. The header is the size. */ |
| 89 |
void __lds *lowlat = (void __lds *)GCN_LOWLAT_HEAP; |
| 90 |
hsa_kernel_dispatch_packet_t *queue_ptr = __builtin_gcn_dispatch_ptr (); |
| 91 |
__gcn_lowlat_init ((void*)(uintptr_t)(void __flat*)lowlat, |
| 92 |
queue_ptr->group_segment_size - GCN_LOWLAT_HEAP); |
| 93 |
|
| 94 |
/* Allocate and initialize the team-local-storage data. */ |
| 95 |
struct gomp_thread *thrs = team_malloc_cleared (sizeof (*thrs) |
| 96 |
* numthreads); |
| 97 |
set_gcn_thrs (thrs); |
| 98 |
|
| 99 |
/* Allocate and initialize a pool of threads in the team. |
| 100 |
The threads are already running, of course, we just need to manage |
| 101 |
the communication between them. */ |
| 102 |
struct gomp_thread_pool *pool = team_malloc (sizeof (*pool)); |
| 103 |
pool->threads = team_malloc (sizeof (void *) * numthreads); |
| 104 |
for (int tid = 0; tid < numthreads; tid++) |
| 105 |
pool->threads[tid] = &thrs[tid]; |
| 106 |
pool->threads_size = numthreads; |
| 107 |
pool->threads_used = numthreads; |
| 108 |
pool->threads_busy = 1; |
| 109 |
pool->last_team = NULL; |
| 110 |
gomp_simple_barrier_init (&pool->threads_dock, numthreads); |
| 111 |
thrs->thread_pool = pool; |
| 112 |
|
| 113 |
asm ("s_barrier" ::: "memory"); |
| 114 |
return; /* Return to kernel. */ |
| 115 |
} |
| 116 |
else |
| 117 |
{ |
| 118 |
asm ("s_barrier" ::: "memory"); |
| 119 |
gomp_thread_start (gcn_thrs ()[0].thread_pool); |
| 120 |
/* gomp_thread_start does not return. */ |
| 121 |
} |
| 122 |
} |
| 123 |
|
| 124 |
void |
| 125 |
gomp_gcn_exit_kernel (void) |
| 126 |
{ |
| 127 |
gomp_free_thread (gcn_thrs ()); |
| 128 |
team_free (gcn_thrs ()); |
| 129 |
} |
| 130 |
|
| 131 |
/* This function contains the idle loop in which a thread waits |
| 132 |
to be called up to become part of a team. */ |
| 133 |
|
| 134 |
static void |
| 135 |
gomp_thread_start (struct gomp_thread_pool *pool) |
| 136 |
{ |
| 137 |
struct gomp_thread *thr = gomp_thread (); |
| 138 |
|
| 139 |
gomp_sem_init (&thr->release, 0); |
| 140 |
thr->thread_pool = pool; |
| 141 |
|
| 142 |
/* The loop exits only when "fn" is assigned "gomp_free_pool_helper", |
| 143 |
which contains "s_endpgm", or an infinite no-op loop is |
| 144 |
suspected (this happens when the thread master crashes). */ |
| 145 |
int nul_limit = 99; |
| 146 |
do |
| 147 |
{ |
| 148 |
gomp_simple_barrier_wait (&pool->threads_dock); |
| 149 |
if (!thr->fn) |
| 150 |
{ |
| 151 |
if (nul_limit-- > 0) |
| 152 |
continue; |
| 153 |
else |
| 154 |
{ |
| 155 |
const char msg[] = ("team master not responding;" |
| 156 |
" slave thread aborting"); |
| 157 |
write (2, msg, sizeof (msg)-1); |
| 158 |
abort(); |
| 159 |
} |
| 160 |
} |
| 161 |
thr->fn (thr->data); |
| 162 |
thr->fn = NULL; |
| 163 |
|
| 164 |
struct gomp_task *task = thr->task; |
| 165 |
gomp_team_barrier_wait_final (&thr->ts.team->barrier); |
| 166 |
gomp_finish_task (task); |
| 167 |
} |
| 168 |
while (1); |
| 169 |
} |
| 170 |
|
| 171 |
/* Launch a team. */ |
| 172 |
|
| 173 |
void |
| 174 |
gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads, |
| 175 |
unsigned flags, struct gomp_team *team, |
| 176 |
struct gomp_taskgroup *taskgroup) |
| 177 |
{ |
| 178 |
struct gomp_thread *thr, *nthr; |
| 179 |
struct gomp_task *task; |
| 180 |
struct gomp_task_icv *icv; |
| 181 |
struct gomp_thread_pool *pool; |
| 182 |
unsigned long nthreads_var; |
| 183 |
|
| 184 |
thr = gomp_thread (); |
| 185 |
pool = thr->thread_pool; |
| 186 |
task = thr->task; |
| 187 |
icv = task ? &task->icv : &gomp_global_icv; |
| 188 |
|
| 189 |
/* Always save the previous state, even if this isn't a nested team. |
| 190 |
In particular, we should save any work share state from an outer |
| 191 |
orphaned work share construct. */ |
| 192 |
team->prev_ts = thr->ts; |
| 193 |
|
| 194 |
thr->ts.team = team; |
| 195 |
thr->ts.team_id = 0; |
| 196 |
++thr->ts.level; |
| 197 |
if (nthreads > 1) |
| 198 |
++thr->ts.active_level; |
| 199 |
thr->ts.work_share = &team->work_shares[0]; |
| 200 |
thr->ts.last_work_share = NULL; |
| 201 |
thr->ts.single_count = 0; |
| 202 |
thr->ts.static_trip = 0; |
| 203 |
thr->task = &team->implicit_task[0]; |
| 204 |
nthreads_var = icv->nthreads_var; |
| 205 |
gomp_init_task (thr->task, task, icv); |
| 206 |
team->implicit_task[0].icv.nthreads_var = nthreads_var; |
| 207 |
team->implicit_task[0].taskgroup = taskgroup; |
| 208 |
|
| 209 |
if (nthreads == 1) |
| 210 |
return; |
| 211 |
|
| 212 |
/* Release existing idle threads. */ |
| 213 |
for (unsigned i = 1; i < nthreads; ++i) |
| 214 |
{ |
| 215 |
nthr = pool->threads[i]; |
| 216 |
nthr->ts.team = team; |
| 217 |
nthr->ts.work_share = &team->work_shares[0]; |
| 218 |
nthr->ts.last_work_share = NULL; |
| 219 |
nthr->ts.team_id = i; |
| 220 |
nthr->ts.level = team->prev_ts.level + 1; |
| 221 |
nthr->ts.active_level = thr->ts.active_level; |
| 222 |
nthr->ts.single_count = 0; |
| 223 |
nthr->ts.static_trip = 0; |
| 224 |
nthr->task = &team->implicit_task[i]; |
| 225 |
gomp_init_task (nthr->task, task, icv); |
| 226 |
team->implicit_task[i].icv.nthreads_var = nthreads_var; |
| 227 |
team->implicit_task[i].taskgroup = taskgroup; |
| 228 |
nthr->fn = fn; |
| 229 |
nthr->data = data; |
| 230 |
team->ordered_release[i] = &nthr->release; |
| 231 |
} |
| 232 |
|
| 233 |
gomp_simple_barrier_wait (&pool->threads_dock); |
| 234 |
} |
| 235 |
|
| 236 |
#include "../../team.c" |