1 /*
2 * kmp_barrier.h
3 */
4
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12
13 #ifndef KMP_BARRIER_H
14 #define KMP_BARRIER_H
15
16 #include "kmp.h"
17 #include "kmp_i18n.h"
18
19 #if KMP_HAVE_XMMINTRIN_H && KMP_HAVE__MM_MALLOC
20 #include <xmmintrin.h>
21 #define KMP_ALIGNED_ALLOCATE(size, alignment) _mm_malloc(size, alignment)
22 #define KMP_ALIGNED_FREE(ptr) _mm_free(ptr)
23 #elif KMP_HAVE_ALIGNED_ALLOC
24 #define KMP_ALGIN_UP(val, alignment) \
25 (((val) + (alignment)-1) / (alignment) * (alignment))
26 #define KMP_ALIGNED_ALLOCATE(size, alignment) \
27 aligned_alloc(alignment, KMP_ALGIN_UP(size, alignment))
28 #define KMP_ALIGNED_FREE(ptr) free(ptr)
29 #elif KMP_HAVE_POSIX_MEMALIGN
KMP_ALIGNED_ALLOCATE(size_t size,size_t alignment)30 static inline void *KMP_ALIGNED_ALLOCATE(size_t size, size_t alignment) {
31 void *ptr;
32 int n = posix_memalign(&ptr, alignment, size);
33 if (n != 0) {
34 if (ptr)
35 free(ptr);
36 return nullptr;
37 }
38 return ptr;
39 }
40 #define KMP_ALIGNED_FREE(ptr) free(ptr)
41 #elif KMP_HAVE__ALIGNED_MALLOC
42 #include <malloc.h>
43 #define KMP_ALIGNED_ALLOCATE(size, alignment) _aligned_malloc(size, alignment)
44 #define KMP_ALIGNED_FREE(ptr) _aligned_free(ptr)
45 #else
46 #define KMP_ALIGNED_ALLOCATE(size, alignment) KMP_INTERNAL_MALLOC(size)
47 #define KMP_ALIGNED_FREE(ptr) KMP_INTERNAL_FREE(ptr)
48 #endif
49
50 // Use four cache lines: MLC tends to prefetch the next or previous cache line
51 // creating a possible fake conflict between cores, so this is the only way to
52 // guarantee that no such prefetch can happen.
53 #ifndef KMP_FOURLINE_ALIGN_CACHE
54 #define KMP_FOURLINE_ALIGN_CACHE KMP_ALIGN(4 * CACHE_LINE)
55 #endif
56
57 #define KMP_OPTIMIZE_FOR_REDUCTIONS 0
58
59 class distributedBarrier {
60 struct flags_s {
61 kmp_uint32 volatile KMP_FOURLINE_ALIGN_CACHE stillNeed;
62 };
63
64 struct go_s {
65 std::atomic<kmp_uint64> KMP_FOURLINE_ALIGN_CACHE go;
66 };
67
68 struct iter_s {
69 kmp_uint64 volatile KMP_FOURLINE_ALIGN_CACHE iter;
70 };
71
72 struct sleep_s {
73 std::atomic<bool> KMP_FOURLINE_ALIGN_CACHE sleep;
74 };
75
76 void init(size_t nthr);
77 void resize(size_t nthr);
78 void computeGo(size_t n);
79 void computeVarsForN(size_t n);
80
81 public:
82 enum {
83 MAX_ITERS = 3,
84 MAX_GOS = 8,
85 IDEAL_GOS = 4,
86 IDEAL_CONTENTION = 16,
87 };
88
89 flags_s *flags[MAX_ITERS];
90 go_s *go;
91 iter_s *iter;
92 sleep_s *sleep;
93
94 size_t KMP_ALIGN_CACHE num_threads; // number of threads in barrier
95 size_t KMP_ALIGN_CACHE max_threads; // size of arrays in data structure
96 // number of go signals each requiring one write per iteration
97 size_t KMP_ALIGN_CACHE num_gos;
98 // number of groups of gos
99 size_t KMP_ALIGN_CACHE num_groups;
100 // threads per go signal
101 size_t KMP_ALIGN_CACHE threads_per_go;
102 bool KMP_ALIGN_CACHE fix_threads_per_go;
103 // threads per group
104 size_t KMP_ALIGN_CACHE threads_per_group;
105 // number of go signals in a group
106 size_t KMP_ALIGN_CACHE gos_per_group;
107 void *team_icvs;
108
109 distributedBarrier() = delete;
110 ~distributedBarrier() = delete;
111
112 // Used instead of constructor to create aligned data
allocate(int nThreads)113 static distributedBarrier *allocate(int nThreads) {
114 distributedBarrier *d = (distributedBarrier *)KMP_ALIGNED_ALLOCATE(
115 sizeof(distributedBarrier), 4 * CACHE_LINE);
116 if (!d) {
117 KMP_FATAL(MemoryAllocFailed);
118 }
119 d->num_threads = 0;
120 d->max_threads = 0;
121 for (int i = 0; i < MAX_ITERS; ++i)
122 d->flags[i] = NULL;
123 d->go = NULL;
124 d->iter = NULL;
125 d->sleep = NULL;
126 d->team_icvs = NULL;
127 d->fix_threads_per_go = false;
128 // calculate gos and groups ONCE on base size
129 d->computeGo(nThreads);
130 d->init(nThreads);
131 return d;
132 }
133
deallocate(distributedBarrier * db)134 static void deallocate(distributedBarrier *db) { KMP_ALIGNED_FREE(db); }
135
update_num_threads(size_t nthr)136 void update_num_threads(size_t nthr) { init(nthr); }
137
need_resize(size_t new_nthr)138 bool need_resize(size_t new_nthr) { return (new_nthr > max_threads); }
get_num_threads()139 size_t get_num_threads() { return num_threads; }
140 kmp_uint64 go_release();
141 void go_reset();
142 };
143
144 #endif // KMP_BARRIER_H
145