LLVM OpenMP* Runtime Library
ompt-specific.h
1 /*
2  * ompt-specific.h - header of OMPT internal functions implementation
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 OMPT_SPECIFIC_H
14 #define OMPT_SPECIFIC_H
15 
16 #include "kmp.h"
17 
18 #if OMPT_SUPPORT
19 /*****************************************************************************
20  * forward declarations
21  ****************************************************************************/
22 
25 void __ompt_force_initialization();
26 
27 void __ompt_team_assign_id(kmp_team_t *team, ompt_data_t ompt_pid);
28 void __ompt_thread_assign_wait_id(void *variable);
29 
30 void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, kmp_info_t *thr, int gtid,
31  ompt_data_t *ompt_pid, void *codeptr);
32 
33 void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, kmp_info_t *thr,
34  int on_heap, bool always = false);
35 
36 void __ompt_lw_taskteam_unlink(kmp_info_t *thr);
37 
38 ompt_team_info_t *__ompt_get_teaminfo(int depth, int *size);
39 
40 ompt_data_t *__ompt_get_task_data();
41 
42 ompt_data_t *__ompt_get_target_task_data();
43 
44 ompt_task_info_t *__ompt_get_task_info_object(int depth);
45 
46 int __ompt_get_parallel_info_internal(int ancestor_level,
47  ompt_data_t **parallel_data,
48  int *team_size);
49 
50 int __ompt_get_task_info_internal(int ancestor_level, int *type,
51  ompt_data_t **task_data,
52  ompt_frame_t **task_frame,
53  ompt_data_t **parallel_data, int *thread_num);
54 
55 ompt_data_t *__ompt_get_thread_data_internal();
56 
57 // __ompt_task_init:
58 // Initialize OMPT fields maintained by a task. This will only be called after
59 // ompt_start_tool, so we already know whether ompt is enabled or not.
60 
61 static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) {
62  // The calls to __ompt_task_init already have the ompt_enabled condition.
63  task->ompt_task_info.task_data.value = 0;
64  task->ompt_task_info.frame.exit_frame = ompt_data_none;
65  task->ompt_task_info.frame.enter_frame = ompt_data_none;
66  task->ompt_task_info.frame.exit_frame_flags =
67  task->ompt_task_info.frame.enter_frame_flags = OMPT_FRAME_FLAGS_RUNTIME;
68  task->ompt_task_info.dispatch_chunk.start = 0;
69  task->ompt_task_info.dispatch_chunk.iterations = 0;
70 }
71 
72 /*
73  * Unused currently
74 static uint64_t __ompt_get_get_unique_id_internal();
75 */
76 
77 ompt_sync_region_t __ompt_get_barrier_kind(enum barrier_type, kmp_info_t *);
78 
79 /*****************************************************************************
80  * macros
81  ****************************************************************************/
82 
83 #define OMPT_CUR_TASK_INFO(thr) (&((thr)->th.th_current_task->ompt_task_info))
84 #define OMPT_CUR_TASK_DATA(thr) \
85  (&((thr)->th.th_current_task->ompt_task_info.task_data))
86 #define OMPT_CUR_TEAM_INFO(thr) (&((thr)->th.th_team->t.ompt_team_info))
87 #define OMPT_CUR_TEAM_DATA(thr) \
88  (&((thr)->th.th_team->t.ompt_team_info.parallel_data))
89 
90 #define OMPT_HAVE_WEAK_ATTRIBUTE KMP_HAVE_WEAK_ATTRIBUTE
91 #define OMPT_HAVE_PSAPI KMP_HAVE_PSAPI
92 #define OMPT_STR_MATCH(haystack, needle) __kmp_str_match(haystack, 0, needle)
93 
94 inline void *__ompt_load_return_address(int gtid) {
95  kmp_info_t *thr = __kmp_threads[gtid];
96  void *return_address = thr->th.ompt_thread_info.return_address;
97  thr->th.ompt_thread_info.return_address = NULL;
98  return return_address;
99 }
100 
101 /*#define OMPT_STORE_RETURN_ADDRESS(gtid) \
102  if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads[gtid] && \
103  !__kmp_threads[gtid]->th.ompt_thread_info.return_address) \
104  __kmp_threads[gtid]->th.ompt_thread_info.return_address = \
105  __builtin_return_address(0)*/
106 #define OMPT_STORE_RETURN_ADDRESS(gtid) \
107  OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address(0)};
108 #define OMPT_LOAD_RETURN_ADDRESS(gtid) __ompt_load_return_address(gtid)
109 #define OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid) \
110  ((ompt_enabled.enabled && gtid >= 0 && __kmp_threads[gtid] && \
111  __kmp_threads[gtid]->th.ompt_thread_info.return_address) \
112  ? __ompt_load_return_address(gtid) \
113  : __builtin_return_address(0))
114 
115 #define OMPT_GET_DISPATCH_CHUNK(chunk, lb, ub, incr) \
116  do { \
117  if (incr > 0) { \
118  chunk.start = static_cast<uint64_t>(lb); \
119  chunk.iterations = static_cast<uint64_t>(((ub) - (lb)) / (incr) + 1); \
120  } else { \
121  chunk.start = static_cast<uint64_t>(ub); \
122  chunk.iterations = static_cast<uint64_t>(((lb) - (ub)) / -(incr) + 1); \
123  } \
124  } while (0)
125 
126 //******************************************************************************
127 // inline functions
128 //******************************************************************************
129 
130 inline kmp_info_t *ompt_get_thread_gtid(int gtid) {
131  return (gtid >= 0) ? __kmp_thread_from_gtid(gtid) : NULL;
132 }
133 
134 inline kmp_info_t *ompt_get_thread() {
135  int gtid = __kmp_get_gtid();
136  return ompt_get_thread_gtid(gtid);
137 }
138 
139 inline void ompt_set_thread_state(kmp_info_t *thread, ompt_state_t state) {
140  if (thread)
141  thread->th.ompt_thread_info.state = state;
142 }
143 
144 inline const char *ompt_get_runtime_version() {
145  return &__kmp_version_lib_ver[KMP_VERSION_MAGIC_LEN];
146 }
147 
148 inline ompt_work_t ompt_get_work_schedule(enum sched_type schedule) {
149  switch (SCHEDULE_WITHOUT_MODIFIERS(schedule)) {
150  case kmp_sch_static_chunked:
151  case kmp_sch_static_balanced:
152  case kmp_sch_static_greedy:
153  return ompt_work_loop_static;
154  case kmp_sch_dynamic_chunked:
155  case kmp_sch_static_steal:
156  return ompt_work_loop_dynamic;
157  case kmp_sch_guided_iterative_chunked:
158  case kmp_sch_guided_analytical_chunked:
160  case kmp_sch_guided_simd:
161  return ompt_work_loop_guided;
162  default:
163  return ompt_work_loop_other;
164  }
165 }
166 
167 class OmptReturnAddressGuard {
168 private:
169  bool SetAddress{false};
170  int Gtid;
171 
172 public:
173  OmptReturnAddressGuard(int Gtid, void *ReturnAddress) : Gtid(Gtid) {
174  if (ompt_enabled.enabled && Gtid >= 0 && __kmp_threads[Gtid] &&
175  !__kmp_threads[Gtid]->th.ompt_thread_info.return_address) {
176  SetAddress = true;
177  __kmp_threads[Gtid]->th.ompt_thread_info.return_address = ReturnAddress;
178  }
179  }
180  ~OmptReturnAddressGuard() {
181  if (SetAddress)
182  __kmp_threads[Gtid]->th.ompt_thread_info.return_address = NULL;
183  }
184 };
185 
186 #endif // OMPT_SUPPORT
187 
188 // macros providing the OMPT callbacks for reduction clause
189 #if OMPT_SUPPORT && OMPT_OPTIONAL
190 #define OMPT_REDUCTION_DECL(this_thr, gtid) \
191  ompt_data_t *my_task_data = OMPT_CUR_TASK_DATA(this_thr); \
192  ompt_data_t *my_parallel_data = OMPT_CUR_TEAM_DATA(this_thr); \
193  void *return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
194 #define OMPT_REDUCTION_BEGIN \
195  if (ompt_enabled.enabled && ompt_enabled.ompt_callback_reduction) { \
196  ompt_callbacks.ompt_callback(ompt_callback_reduction)( \
197  ompt_sync_region_reduction, ompt_scope_begin, my_parallel_data, \
198  my_task_data, return_address); \
199  }
200 #define OMPT_REDUCTION_END \
201  if (ompt_enabled.enabled && ompt_enabled.ompt_callback_reduction) { \
202  ompt_callbacks.ompt_callback(ompt_callback_reduction)( \
203  ompt_sync_region_reduction, ompt_scope_end, my_parallel_data, \
204  my_task_data, return_address); \
205  }
206 #else // OMPT_SUPPORT && OMPT_OPTIONAL
207 #define OMPT_REDUCTION_DECL(this_thr, gtid)
208 #define OMPT_REDUCTION_BEGIN
209 #define OMPT_REDUCTION_END
210 #endif // ! OMPT_SUPPORT && OMPT_OPTIONAL
211 
212 #endif
sched_type
Definition: kmp.h:370
@ kmp_sch_guided_simd
Definition: kmp.h:391
@ kmp_sch_guided_chunked
Definition: kmp.h:375