On NVPTX, there's 16 hardware barriers for each thread team, each barrier has a variable waiter count. The instruction 'bar.sync N, M;' allows to wait on barrier number N until M threads have arrived. M should be pre-multiplied by warp width. It's also possible to 'post' the barrier without suspending with 'bar.arrive'.
We should be able to provide gomp barrier via a combination of ptx barriers and atomics. This patch is a first step in that direction. It's mostly a copy of Linux implementation, and it's very likely that functions more complex than gomp_barrier_wait_end are implemented incorrectly. I will have to review all of that (and optimize, hopefully). I'm not sure if naked asm()'s are OK. It's possible to implement a builtin instead for a minor beautification. Thoughts? --- libgomp/config/nvptx/bar.c | 210 +++++++++++++++++++++++++++++++++++++++++++++ libgomp/config/nvptx/bar.h | 129 +++++++++++++++++++++++++++- 2 files changed, 338 insertions(+), 1 deletion(-) diff --git a/libgomp/config/nvptx/bar.c b/libgomp/config/nvptx/bar.c index e69de29..5631cb4 100644 --- a/libgomp/config/nvptx/bar.c +++ b/libgomp/config/nvptx/bar.c @@ -0,0 +1,210 @@ +/* Copyright (C) 2015 Free Software Foundation, Inc. + + This file is part of the GNU Offloading and Multi Processing Library + (libgomp). + + Libgomp is free software; you can redistribute it and/or modify it + under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 3, or (at your option) + any later version. + + Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY + WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS + FOR A PARTICULAR PURPOSE. See the GNU General Public License for + more details. + + Under Section 7 of GPL version 3, you are granted additional + permissions described in the GCC Runtime Library Exception, version + 3.1, as published by the Free Software Foundation. + + 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; + see the files COPYING3 and COPYING.RUNTIME respectively. If not, see + <http://www.gnu.org/licenses/>. */ + +/* This is an NVPTX specific implementation of a barrier synchronization + mechanism for libgomp. This type is private to the library. This + implementation uses atomic instructions and bar.sync instruction. */ + +#include <limits.h> +#include "libgomp.h" + + +void +gomp_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) +{ + if (__builtin_expect (state & BAR_WAS_LAST, 0)) + { + /* Next time we'll be awaiting TOTAL threads again. */ + bar->awaited = bar->total; + __atomic_store_n (&bar->generation, bar->generation + BAR_INCR, + MEMMODEL_RELEASE); + } + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); +} + +void +gomp_barrier_wait (gomp_barrier_t *bar) +{ + gomp_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); +} + +/* Like gomp_barrier_wait, except that if the encountering thread + is not the last one to hit the barrier, it returns immediately. + The intended usage is that a thread which intends to gomp_barrier_destroy + this barrier calls gomp_barrier_wait, while all other threads + call gomp_barrier_wait_last. When gomp_barrier_wait returns, + the barrier can be safely destroyed. */ + +void +gomp_barrier_wait_last (gomp_barrier_t *bar) +{ +#if 0 + gomp_barrier_state_t state = gomp_barrier_wait_start (bar); + if (state & BAR_WAS_LAST) + gomp_barrier_wait_end (bar, state); +#else + gomp_barrier_wait (bar); +#endif +} + +void +gomp_team_barrier_wake (gomp_barrier_t *bar, int count) +{ + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); +} + +void +gomp_team_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) +{ + unsigned int generation, gen; + + gomp_barrier_wait_end (bar, state); + + if (__builtin_expect (state & BAR_WAS_LAST, 0)) + { + /* Next time we'll be awaiting TOTAL threads again. */ + struct gomp_thread *thr = gomp_thread (); + struct gomp_team *team = thr->ts.team; + + bar->awaited = bar->total; + team->work_share_cancelled = 0; + if (__builtin_expect (team->task_count, 0)) + { + gomp_barrier_handle_tasks (state); + state &= ~BAR_WAS_LAST; + } + else + { + state &= ~BAR_CANCELLED; + state += BAR_INCR - BAR_WAS_LAST; + __atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); + return; + } + } + + generation = state; + state &= ~BAR_CANCELLED; + do + { + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); + gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) + { + gomp_barrier_handle_tasks (state); + gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + } + generation |= gen & BAR_WAITING_FOR_TASK; + } + while (gen != state + BAR_INCR); +} + +void +gomp_team_barrier_wait (gomp_barrier_t *bar) +{ + gomp_team_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); +} + +void +gomp_team_barrier_wait_final (gomp_barrier_t *bar) +{ + gomp_barrier_state_t state = gomp_barrier_wait_final_start (bar); + if (__builtin_expect (state & BAR_WAS_LAST, 0)) + bar->awaited_final = bar->total; + gomp_team_barrier_wait_end (bar, state); +} + +bool +gomp_team_barrier_wait_cancel_end (gomp_barrier_t *bar, + gomp_barrier_state_t state) +{ + unsigned int generation, gen; + + if (__builtin_expect (state & BAR_WAS_LAST, 0)) + { + /* Next time we'll be awaiting TOTAL threads again. */ + /* BAR_CANCELLED should never be set in state here, because + cancellation means that at least one of the threads has been + cancelled, thus on a cancellable barrier we should never see + all threads to arrive. */ + struct gomp_thread *thr = gomp_thread (); + struct gomp_team *team = thr->ts.team; + + bar->awaited = bar->total; + team->work_share_cancelled = 0; + if (__builtin_expect (team->task_count, 0)) + { + gomp_barrier_handle_tasks (state); + state &= ~BAR_WAS_LAST; + } + else + { + state += BAR_INCR - BAR_WAS_LAST; + __atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); + return false; + } + } + + if (__builtin_expect (state & BAR_CANCELLED, 0)) + return true; + + generation = state; + do + { + asm ("bar.sync 0, %0;" : : "r"(32*bar->total)); + gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + if (__builtin_expect (gen & BAR_CANCELLED, 0)) + return true; + if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) + { + gomp_barrier_handle_tasks (state); + gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + } + generation |= gen & BAR_WAITING_FOR_TASK; + } + while (gen != state + BAR_INCR); + + return false; +} + +bool +gomp_team_barrier_wait_cancel (gomp_barrier_t *bar) +{ + return gomp_team_barrier_wait_cancel_end (bar, gomp_barrier_wait_start (bar)); +} + +void +gomp_team_barrier_cancel (struct gomp_team *team) +{ + gomp_mutex_lock (&team->task_lock); + if (team->barrier.generation & BAR_CANCELLED) + { + gomp_mutex_unlock (&team->task_lock); + return; + } + team->barrier.generation |= BAR_CANCELLED; + gomp_mutex_unlock (&team->task_lock); + gomp_team_barrier_wake (&team->barrier, INT_MAX); +} diff --git a/libgomp/config/nvptx/bar.h b/libgomp/config/nvptx/bar.h index 009d85f..bbdc466 100644 --- a/libgomp/config/nvptx/bar.h +++ b/libgomp/config/nvptx/bar.h @@ -24,15 +24,142 @@ /* This is an NVPTX specific implementation of a barrier synchronization mechanism for libgomp. This type is private to the library. This - implementation is a stub, for now. */ + implementation uses atomic instructions and bar.sync instruction. */ #ifndef GOMP_BARRIER_H #define GOMP_BARRIER_H 1 +#include "mutex.h" + typedef struct { + unsigned total; + unsigned generation; + unsigned awaited; + unsigned awaited_final; } gomp_barrier_t; typedef unsigned int gomp_barrier_state_t; +/* The generation field contains a counter in the high bits, with a few + low bits dedicated to flags. Note that TASK_PENDING and WAS_LAST can + share space because WAS_LAST is never stored back to generation. */ +#define BAR_TASK_PENDING 1 +#define BAR_WAS_LAST 1 +#define BAR_WAITING_FOR_TASK 2 +#define BAR_CANCELLED 4 +#define BAR_INCR 8 + +static inline void gomp_barrier_init (gomp_barrier_t *bar, unsigned count) +{ + bar->total = count; + bar->awaited = count; + bar->awaited_final = count; + bar->generation = 0; +} + +static inline void gomp_barrier_reinit (gomp_barrier_t *bar, unsigned count) +{ + __atomic_add_fetch (&bar->awaited, count - bar->total, MEMMODEL_ACQ_REL); + bar->total = count; +} + +static inline void gomp_barrier_destroy (gomp_barrier_t *bar) +{ +} + +extern void gomp_barrier_wait (gomp_barrier_t *); +extern void gomp_barrier_wait_last (gomp_barrier_t *); +extern void gomp_barrier_wait_end (gomp_barrier_t *, gomp_barrier_state_t); +extern void gomp_team_barrier_wait (gomp_barrier_t *); +extern void gomp_team_barrier_wait_final (gomp_barrier_t *); +extern void gomp_team_barrier_wait_end (gomp_barrier_t *, + gomp_barrier_state_t); +extern bool gomp_team_barrier_wait_cancel (gomp_barrier_t *); +extern bool gomp_team_barrier_wait_cancel_end (gomp_barrier_t *, + gomp_barrier_state_t); +extern void gomp_team_barrier_wake (gomp_barrier_t *, int); +struct gomp_team; +extern void gomp_team_barrier_cancel (struct gomp_team *); + +static inline gomp_barrier_state_t +gomp_barrier_wait_start (gomp_barrier_t *bar) +{ + unsigned int ret = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + ret &= -BAR_INCR | BAR_CANCELLED; + /* A memory barrier is needed before exiting from the various forms + of gomp_barrier_wait, to satisfy OpenMP API version 3.1 section + 2.8.6 flush Construct, which says there is an implicit flush during + a barrier region. This is a convenient place to add the barrier, + so we use MEMMODEL_ACQ_REL here rather than MEMMODEL_ACQUIRE. */ + if (__atomic_add_fetch (&bar->awaited, -1, MEMMODEL_ACQ_REL) == 0) + ret |= BAR_WAS_LAST; + return ret; +} + +static inline gomp_barrier_state_t +gomp_barrier_wait_cancel_start (gomp_barrier_t *bar) +{ + return gomp_barrier_wait_start (bar); +} + +/* This is like gomp_barrier_wait_start, except it decrements + bar->awaited_final rather than bar->awaited and should be used + for the gomp_team_end barrier only. */ +static inline gomp_barrier_state_t +gomp_barrier_wait_final_start (gomp_barrier_t *bar) +{ + unsigned int ret = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); + ret &= -BAR_INCR | BAR_CANCELLED; + /* See above gomp_barrier_wait_start comment. */ + if (__atomic_add_fetch (&bar->awaited_final, -1, MEMMODEL_ACQ_REL) == 0) + ret |= BAR_WAS_LAST; + return ret; +} + +static inline bool +gomp_barrier_last_thread (gomp_barrier_state_t state) +{ + return state & BAR_WAS_LAST; +} + +/* All the inlines below must be called with team->task_lock + held. */ + +static inline void +gomp_team_barrier_set_task_pending (gomp_barrier_t *bar) +{ + bar->generation |= BAR_TASK_PENDING; +} + +static inline void +gomp_team_barrier_clear_task_pending (gomp_barrier_t *bar) +{ + bar->generation &= ~BAR_TASK_PENDING; +} + +static inline void +gomp_team_barrier_set_waiting_for_tasks (gomp_barrier_t *bar) +{ + bar->generation |= BAR_WAITING_FOR_TASK; +} + +static inline bool +gomp_team_barrier_waiting_for_tasks (gomp_barrier_t *bar) +{ + return (bar->generation & BAR_WAITING_FOR_TASK) != 0; +} + +static inline bool +gomp_team_barrier_cancelled (gomp_barrier_t *bar) +{ + return __builtin_expect ((bar->generation & BAR_CANCELLED) != 0, 0); +} + +static inline void +gomp_team_barrier_done (gomp_barrier_t *bar, gomp_barrier_state_t state) +{ + bar->generation = (state & -BAR_INCR) + BAR_INCR; +} + #endif /* GOMP_BARRIER_H */
