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 23 /// Entrypoint used by libomptarget to register callbacks in libomp, if not 24 /// done already 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 /* 58 * Unused currently 59 static uint64_t __ompt_get_get_unique_id_internal(); 60 */ 61 62 ompt_sync_region_t __ompt_get_barrier_kind(enum barrier_type, kmp_info_t *); 63 64 /***************************************************************************** 65 * macros 66 ****************************************************************************/ 67 68 #define OMPT_CUR_TASK_INFO(thr) (&((thr)->th.th_current_task->ompt_task_info)) 69 #define OMPT_CUR_TASK_DATA(thr) \ 70 (&((thr)->th.th_current_task->ompt_task_info.task_data)) 71 #define OMPT_CUR_TEAM_INFO(thr) (&((thr)->th.th_team->t.ompt_team_info)) 72 #define OMPT_CUR_TEAM_DATA(thr) \ 73 (&((thr)->th.th_team->t.ompt_team_info.parallel_data)) 74 75 #define OMPT_HAVE_WEAK_ATTRIBUTE KMP_HAVE_WEAK_ATTRIBUTE 76 #define OMPT_HAVE_PSAPI KMP_HAVE_PSAPI 77 #define OMPT_STR_MATCH(haystack, needle) __kmp_str_match(haystack, 0, needle) 78 79 inline void *__ompt_load_return_address(int gtid) { 80 kmp_info_t *thr = __kmp_threads[gtid]; 81 void *return_address = thr->th.ompt_thread_info.return_address; 82 thr->th.ompt_thread_info.return_address = NULL; 83 return return_address; 84 } 85 86 /*#define OMPT_STORE_RETURN_ADDRESS(gtid) \ 87 if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads[gtid] && \ 88 !__kmp_threads[gtid]->th.ompt_thread_info.return_address) \ 89 __kmp_threads[gtid]->th.ompt_thread_info.return_address = \ 90 __builtin_return_address(0)*/ 91 #define OMPT_STORE_RETURN_ADDRESS(gtid) \ 92 OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address(0)}; 93 #define OMPT_LOAD_RETURN_ADDRESS(gtid) __ompt_load_return_address(gtid) 94 #define OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid) \ 95 ((ompt_enabled.enabled && gtid >= 0 && __kmp_threads[gtid] && \ 96 __kmp_threads[gtid]->th.ompt_thread_info.return_address) \ 97 ? __ompt_load_return_address(gtid) \ 98 : __builtin_return_address(0)) 99 100 #define OMPT_GET_DISPATCH_CHUNK(chunk, lb, ub, incr) \ 101 do { \ 102 if (incr > 0) { \ 103 chunk.start = static_cast<uint64_t>(lb); \ 104 chunk.iterations = static_cast<uint64_t>(((ub) - (lb)) / (incr) + 1); \ 105 } else { \ 106 chunk.start = static_cast<uint64_t>(ub); \ 107 chunk.iterations = static_cast<uint64_t>(((lb) - (ub)) / -(incr) + 1); \ 108 } \ 109 } while (0) 110 111 //****************************************************************************** 112 // inline functions 113 //****************************************************************************** 114 115 inline kmp_info_t *ompt_get_thread_gtid(int gtid) { 116 return (gtid >= 0) ? __kmp_thread_from_gtid(gtid) : NULL; 117 } 118 119 inline kmp_info_t *ompt_get_thread() { 120 int gtid = __kmp_get_gtid(); 121 return ompt_get_thread_gtid(gtid); 122 } 123 124 inline void ompt_set_thread_state(kmp_info_t *thread, ompt_state_t state) { 125 if (thread) 126 thread->th.ompt_thread_info.state = state; 127 } 128 129 inline const char *ompt_get_runtime_version() { 130 return &__kmp_version_lib_ver[KMP_VERSION_MAGIC_LEN]; 131 } 132 133 inline ompt_work_t ompt_get_work_schedule(enum sched_type schedule) { 134 switch (SCHEDULE_WITHOUT_MODIFIERS(schedule)) { 135 case kmp_sch_static_chunked: 136 case kmp_sch_static_balanced: 137 case kmp_sch_static_greedy: 138 return ompt_work_loop_static; 139 case kmp_sch_dynamic_chunked: 140 case kmp_sch_static_steal: 141 return ompt_work_loop_dynamic; 142 case kmp_sch_guided_iterative_chunked: 143 case kmp_sch_guided_analytical_chunked: 144 case kmp_sch_guided_chunked: 145 case kmp_sch_guided_simd: 146 return ompt_work_loop_guided; 147 default: 148 return ompt_work_loop_other; 149 } 150 } 151 152 class OmptReturnAddressGuard { 153 private: 154 bool SetAddress{false}; 155 int Gtid; 156 157 public: 158 OmptReturnAddressGuard(int Gtid, void *ReturnAddress) : Gtid(Gtid) { 159 if (ompt_enabled.enabled && Gtid >= 0 && __kmp_threads[Gtid] && 160 !__kmp_threads[Gtid]->th.ompt_thread_info.return_address) { 161 SetAddress = true; 162 __kmp_threads[Gtid]->th.ompt_thread_info.return_address = ReturnAddress; 163 } 164 } 165 ~OmptReturnAddressGuard() { 166 if (SetAddress) 167 __kmp_threads[Gtid]->th.ompt_thread_info.return_address = NULL; 168 } 169 }; 170 171 #endif // OMPT_SUPPORT 172 173 // macros providing the OMPT callbacks for reduction clause 174 #if OMPT_SUPPORT && OMPT_OPTIONAL 175 #define OMPT_REDUCTION_DECL(this_thr, gtid) \ 176 ompt_data_t *my_task_data = OMPT_CUR_TASK_DATA(this_thr); \ 177 ompt_data_t *my_parallel_data = OMPT_CUR_TEAM_DATA(this_thr); \ 178 void *return_address = OMPT_LOAD_RETURN_ADDRESS(gtid); 179 #define OMPT_REDUCTION_BEGIN \ 180 if (ompt_enabled.enabled && ompt_enabled.ompt_callback_reduction) { \ 181 ompt_callbacks.ompt_callback(ompt_callback_reduction)( \ 182 ompt_sync_region_reduction, ompt_scope_begin, my_parallel_data, \ 183 my_task_data, return_address); \ 184 } 185 #define OMPT_REDUCTION_END \ 186 if (ompt_enabled.enabled && ompt_enabled.ompt_callback_reduction) { \ 187 ompt_callbacks.ompt_callback(ompt_callback_reduction)( \ 188 ompt_sync_region_reduction, ompt_scope_end, my_parallel_data, \ 189 my_task_data, return_address); \ 190 } 191 #else // OMPT_SUPPORT && OMPT_OPTIONAL 192 #define OMPT_REDUCTION_DECL(this_thr, gtid) 193 #define OMPT_REDUCTION_BEGIN 194 #define OMPT_REDUCTION_END 195 #endif // ! OMPT_SUPPORT && OMPT_OPTIONAL 196 197 #endif 198