Home | History | Annotate | Line # | Download | only in nvptx
      1 /* Copyright (C) 2015-2024 Free Software Foundation, Inc.
      2    Contributed by Alexander Monakov <amonakov (at) 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 is an NVPTX specific implementation of a barrier synchronization
     27    mechanism for libgomp.  This type is private to the library.  This
     28    implementation uses atomic instructions and bar.sync instruction.  */
     29 
     30 #ifndef GOMP_BARRIER_H
     31 #define GOMP_BARRIER_H 1
     32 
     33 #include "mutex.h"
     34 
     35 typedef struct
     36 {
     37   unsigned total;
     38   unsigned generation;
     39   unsigned awaited;
     40   unsigned awaited_final;
     41 } gomp_barrier_t;
     42 
     43 typedef unsigned int gomp_barrier_state_t;
     44 
     45 /* The generation field contains a counter in the high bits, with a few
     46    low bits dedicated to flags.  Note that TASK_PENDING and WAS_LAST can
     47    share space because WAS_LAST is never stored back to generation.  */
     48 #define BAR_TASK_PENDING	1
     49 #define BAR_WAS_LAST		1
     50 #define BAR_WAITING_FOR_TASK	2
     51 #define BAR_CANCELLED		4
     52 #define BAR_INCR		8
     53 
     54 static inline void gomp_barrier_init (gomp_barrier_t *bar, unsigned count)
     55 {
     56   bar->total = count;
     57   bar->awaited = count;
     58   bar->awaited_final = count;
     59   bar->generation = 0;
     60 }
     61 
     62 static inline void gomp_barrier_reinit (gomp_barrier_t *bar, unsigned count)
     63 {
     64   __atomic_add_fetch (&bar->awaited, count - bar->total, MEMMODEL_ACQ_REL);
     65   bar->total = count;
     66 }
     67 
     68 static inline void gomp_barrier_destroy (gomp_barrier_t *bar)
     69 {
     70 }
     71 
     72 extern void gomp_barrier_wait (gomp_barrier_t *);
     73 extern void gomp_barrier_wait_last (gomp_barrier_t *);
     74 extern void gomp_barrier_wait_end (gomp_barrier_t *, gomp_barrier_state_t);
     75 extern void gomp_team_barrier_wait (gomp_barrier_t *);
     76 extern void gomp_team_barrier_wait_final (gomp_barrier_t *);
     77 extern void gomp_team_barrier_wait_end (gomp_barrier_t *,
     78 					gomp_barrier_state_t);
     79 extern bool gomp_team_barrier_wait_cancel (gomp_barrier_t *);
     80 extern bool gomp_team_barrier_wait_cancel_end (gomp_barrier_t *,
     81 					       gomp_barrier_state_t);
     82 struct gomp_team;
     83 extern void gomp_team_barrier_cancel (struct gomp_team *);
     84 
     85 static inline void
     86 gomp_team_barrier_wake (gomp_barrier_t *bar, int count)
     87 {
     88   /* We never "wake up" threads on nvptx.  Threads wait at barrier
     89      instructions till barrier fullfilled.  Do nothing here.  */
     90 }
     91 
     92 static inline gomp_barrier_state_t
     93 gomp_barrier_wait_start (gomp_barrier_t *bar)
     94 {
     95   unsigned int ret = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE);
     96   ret &= -BAR_INCR | BAR_CANCELLED;
     97   /* A memory barrier is needed before exiting from the various forms
     98      of gomp_barrier_wait, to satisfy OpenMP API version 3.1 section
     99      2.8.6 flush Construct, which says there is an implicit flush during
    100      a barrier region.  This is a convenient place to add the barrier,
    101      so we use MEMMODEL_ACQ_REL here rather than MEMMODEL_ACQUIRE.  */
    102   if (__atomic_add_fetch (&bar->awaited, -1, MEMMODEL_ACQ_REL) == 0)
    103     ret |= BAR_WAS_LAST;
    104   return ret;
    105 }
    106 
    107 static inline gomp_barrier_state_t
    108 gomp_barrier_wait_cancel_start (gomp_barrier_t *bar)
    109 {
    110   return gomp_barrier_wait_start (bar);
    111 }
    112 
    113 /* This is like gomp_barrier_wait_start, except it decrements
    114    bar->awaited_final rather than bar->awaited and should be used
    115    for the gomp_team_end barrier only.  */
    116 static inline gomp_barrier_state_t
    117 gomp_barrier_wait_final_start (gomp_barrier_t *bar)
    118 {
    119   unsigned int ret = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE);
    120   ret &= -BAR_INCR | BAR_CANCELLED;
    121   /* See above gomp_barrier_wait_start comment.  */
    122   if (__atomic_add_fetch (&bar->awaited_final, -1, MEMMODEL_ACQ_REL) == 0)
    123     ret |= BAR_WAS_LAST;
    124   return ret;
    125 }
    126 
    127 static inline bool
    128 gomp_barrier_last_thread (gomp_barrier_state_t state)
    129 {
    130   return state & BAR_WAS_LAST;
    131 }
    132 
    133 /* All the inlines below must be called with team->task_lock
    134    held.  */
    135 
    136 static inline void
    137 gomp_team_barrier_set_task_pending (gomp_barrier_t *bar)
    138 {
    139   bar->generation |= BAR_TASK_PENDING;
    140 }
    141 
    142 static inline void
    143 gomp_team_barrier_clear_task_pending (gomp_barrier_t *bar)
    144 {
    145   bar->generation &= ~BAR_TASK_PENDING;
    146 }
    147 
    148 static inline void
    149 gomp_team_barrier_set_waiting_for_tasks (gomp_barrier_t *bar)
    150 {
    151   bar->generation |= BAR_WAITING_FOR_TASK;
    152 }
    153 
    154 static inline bool
    155 gomp_team_barrier_waiting_for_tasks (gomp_barrier_t *bar)
    156 {
    157   return (bar->generation & BAR_WAITING_FOR_TASK) != 0;
    158 }
    159 
    160 static inline bool
    161 gomp_team_barrier_cancelled (gomp_barrier_t *bar)
    162 {
    163   return __builtin_expect ((bar->generation & BAR_CANCELLED) != 0, 0);
    164 }
    165 
    166 static inline void
    167 gomp_team_barrier_done (gomp_barrier_t *bar, gomp_barrier_state_t state)
    168 {
    169   bar->generation = (state & -BAR_INCR) + BAR_INCR;
    170 }
    171 
    172 #endif /* GOMP_BARRIER_H */
    173