201 lines · c
1#include <dlfcn.h>2#include <omp-tools.h>3#include <omp.h>4#include <stdio.h>5#include <stdlib.h>6#include <sys/types.h>7#include <unistd.h>8 9typedef struct omp_t_data {10 // Thread data11 ompt_state_t ompt_state;12 ompt_wait_id_t ompt_wait_id;13 int omp_thread_num;14 ompt_data_t *ompt_thread_data;15 // Parallel data16 int omp_num_threads;17 int omp_level;18 int omp_active_level;19 ompt_data_t *ompt_parallel_data;20 // Task data21 int omp_max_threads;22 int omp_parallel;23 int omp_final;24 int omp_dynamic;25 int omp_nested;26 int omp_max_active_levels;27 omp_sched_t omp_kind;28 int omp_modifier;29 omp_proc_bind_t omp_proc_bind;30 ompt_frame_t *ompt_frame_list;31 ompt_data_t *ompt_task_data;32} omp_t_data_t;33 34static __thread omp_t_data_t thread_data;35 36static ompt_function_lookup_t ompt_lookup;37// NOLINTNEXTLINE "Used in Macro:register_callback_t below."38static ompt_set_callback_t ompt_set_callback;39static ompt_get_callback_t ompt_get_callback;40static ompt_get_state_t ompt_get_state;41static ompt_get_task_info_t ompt_get_task_info;42static ompt_get_thread_data_t ompt_get_thread_data;43static ompt_get_parallel_info_t ompt_get_parallel_info;44static ompt_get_unique_id_t ompt_get_unique_id;45static ompt_get_num_procs_t ompt_get_num_procs;46static ompt_get_num_places_t ompt_get_num_places;47static ompt_get_place_proc_ids_t ompt_get_place_proc_ids;48static ompt_get_place_num_t ompt_get_place_num;49static ompt_get_partition_place_nums_t ompt_get_partition_place_nums;50static ompt_get_proc_id_t ompt_get_proc_id;51static ompt_enumerate_states_t ompt_enumerate_states;52static ompt_enumerate_mutex_impls_t ompt_enumerate_mutex_impls;53static int checks = 0;54 55static void on_ompt_callback_implicit_task(ompt_scope_endpoint_t endpoint,56 ompt_data_t *parallel_data,57 ompt_data_t *task_data,58 unsigned int team_size,59 unsigned int thread_num, int flags) {60 if (endpoint == ompt_scope_begin)61 task_data->value = ompt_get_unique_id();62}63 64static void on_ompt_callback_thread_begin(ompt_thread_t thread_type,65 ompt_data_t *t_data) {66 t_data->value = ompt_get_unique_id();67}68 69static void on_ompt_callback_parallel_begin(70 ompt_data_t *encountering_task_data,71 const ompt_frame_t *encountering_task_frame, ompt_data_t *parallel_data,72 uint32_t requested_team_size, int flag, const void *codeptr_ra) {73 parallel_data->value = ompt_get_unique_id();74}75 76#define register_callback_t(name, type) \77 do { \78 type f_##name = &on_##name; \79 if (ompt_set_callback(name, (ompt_callback_t)f_##name) == ompt_set_never) \80 printf("0: Could not register callback '" #name "'\n"); \81 } while (0)82 83#define register_callback(name) register_callback_t(name, name##_t)84 85static int ompt_initialize(ompt_function_lookup_t lookup,86 int initial_device_num, ompt_data_t *tool_data) {87 ompt_lookup = lookup;88 // TODO: remove: printf("runtime_version: %s, omp_version: %i\n",89 // runtime_version, omp_version);90 91 // TODO: remove macro92 // #define declare_inquery_fn(F) F = (F##_t)lookup(#F);93 // FOREACH_OMPT_INQUIRY_FN(declare_inquery_fn)94 // #undef declare_inquery_fn95 96 ompt_set_callback_t ompt_set_callback =97 (ompt_set_callback_t)lookup("ompt_set_callback");98 ompt_get_callback = (ompt_get_callback_t)lookup("ompt_get_callback");99 ompt_get_state = (ompt_get_state_t)lookup("ompt_get_state");100 ompt_get_task_info = (ompt_get_task_info_t)lookup("ompt_get_task_info");101 ompt_get_thread_data = (ompt_get_thread_data_t)lookup("ompt_get_thread_data");102 ompt_get_parallel_info =103 (ompt_get_parallel_info_t)lookup("ompt_get_parallel_info");104 ompt_get_unique_id = (ompt_get_unique_id_t)lookup("ompt_get_unique_id");105 106 ompt_get_num_procs = (ompt_get_num_procs_t)lookup("ompt_get_num_procs");107 ompt_get_num_places = (ompt_get_num_places_t)lookup("ompt_get_num_places");108 ompt_get_place_proc_ids =109 (ompt_get_place_proc_ids_t)lookup("ompt_get_place_proc_ids");110 ompt_get_place_num = (ompt_get_place_num_t)lookup("ompt_get_place_num");111 ompt_get_partition_place_nums =112 (ompt_get_partition_place_nums_t)lookup("ompt_get_partition_place_nums");113 ompt_get_proc_id = (ompt_get_proc_id_t)lookup("ompt_get_proc_id");114 ompt_enumerate_states =115 (ompt_enumerate_states_t)lookup("ompt_enumerate_states");116 ompt_enumerate_mutex_impls =117 (ompt_enumerate_mutex_impls_t)lookup("ompt_enumerate_mutex_impls");118 119 register_callback(ompt_callback_implicit_task);120 register_callback(ompt_callback_thread_begin);121 register_callback(ompt_callback_parallel_begin);122 123 return 1; // activate tool124}125 126static void ompt_finalize(ompt_data_t *tool_data) {}127 128// "This func will be invoked by OpenMP implementation, refer spec: 4.2.1"129// NOLINTNEXTLINE130ompt_start_tool_result_t *ompt_start_tool(unsigned int omp_version,131 const char *runtime_version) {132 static ompt_start_tool_result_t ompt_start_tool_result = {133 &ompt_initialize, &ompt_finalize, {0}};134 return &ompt_start_tool_result;135}136 137static void collectParallelData(omp_t_data_t *data) {138 data->omp_num_threads = omp_get_num_threads();139 data->omp_level = omp_get_level();140 data->omp_active_level = omp_get_active_level();141 ompt_get_parallel_info(0, &(data->ompt_parallel_data), NULL);142}143 144static void collectTaskData(omp_t_data_t *data) {145 data->omp_max_threads = omp_get_max_threads();146 data->omp_parallel = omp_in_parallel();147 data->omp_final = omp_in_final();148 data->omp_dynamic = omp_get_dynamic();149 data->omp_nested = omp_get_max_active_levels() > 1;150 data->omp_max_active_levels = omp_get_max_active_levels();151 omp_get_schedule(&(data->omp_kind), &(data->omp_modifier));152 data->omp_proc_bind = omp_get_proc_bind();153 ompt_get_task_info(0, NULL, &(data->ompt_task_data), &(data->ompt_frame_list),154 NULL, NULL);155}156 157static void collectThreadData(omp_t_data_t *data) {158 data->omp_thread_num = omp_get_thread_num();159 data->ompt_state = (ompt_state_t)ompt_get_state(&(data->ompt_wait_id));160 data->ompt_thread_data = ompt_get_thread_data();161}162 163#ifdef __cplusplus164extern "C" {165#endif166__attribute__((noinline)) static void *breakToolTest(omp_t_data_t *data) {167 return data;168}169#ifdef __cplusplus170}171#endif172 173#ifdef __cplusplus174extern "C" {175#endif176static void *ompd_tool_break(void *n) {177 (void)n;178 asm("");179 return NULL;180}181#ifdef __cplusplus182}183#endif184 185// NOLINTNEXTLINE "This func will be invoked in testcases."186static void *ompd_tool_test(void *n) {187 collectThreadData(&thread_data);188 collectParallelData(&thread_data);189 collectTaskData(&thread_data);190 breakToolTest(&thread_data);191 checks++;192 ompd_tool_break(NULL);193 return NULL;194}195 196__attribute__((__constructor__)) static void init(void) {}197 198__attribute__((__destructor__)) static void fini(void) {199 printf("Finished %i testsuites.\n", checks);200}201