1 // RUN: %libomp-cxx-compile-and-run
2 // RUN: %libomp-cxx-compile && env OMP_NUM_THREADS=1 %libomp-run
3 
4 /*
5  * This test aims to check whether hidden helper thread has right gtid. We also
6  * test if there is mixed dependences between regular tasks and hidden helper
7  * tasks, the tasks are executed by right set of threads. It is equivalent to
8  * the following code:
9  *
10  * #pragma omp parallel for
11  * for (int i = 0; i < N; ++i) {
12  *   int data1 = -1, data2 = -1, data3 = -1;
13  *   int depvar;
14  * #pragma omp task shared(data1) depend(inout: depvar)
15  *   {
16  *     data1 = omp_get_global_thread_id();
17  *   }
18  * #pragma omp task hidden helper shared(data2) depend(inout: depvar)
19  *   {
20  *     data2 = omp_get_global_thread_id();
21  *   }
22  * #pragma omp task shared(data3) depend(inout: depvar)
23  *   {
24  *     data3 = omp_get_global_thread_id();
25  *   }
26  * #pragma omp taskwait
27  *   assert(data1 == 0 || data1 > __kmp_num_hidden_helper_threads);
28  *   assert(data2 > 0 && data2 <= __kmp_num_hidden_helper_threads);
29  *   assert(data3 == 0 || data3 > __kmp_num_hidden_helper_threads);
30  * }
31  */
32 
33 #include "common.h"
34 
35 extern "C" {
36 struct kmp_task_t_with_privates {
37   kmp_task_t task;
38 };
39 
40 struct anon {
41   int32_t *data;
42 };
43 }
44 
45 kmp_int32 __kmp_hidden_helper_threads_num;
46 
47 kmp_int32 omp_task_entry(kmp_int32 gtid, kmp_task_t_with_privates *task) {
48   auto shareds = reinterpret_cast<anon *>(task->task.shareds);
49   auto p = shareds->data;
50   *p = __kmpc_global_thread_num(nullptr);
51   return 0;
52 }
53 
54 template <bool hidden_helper_task> void assert_gtid(int v) {
55   if (__kmp_hidden_helper_threads_num) {
56     if (hidden_helper_task) {
57       assert(v > 0 && v <= __kmp_hidden_helper_threads_num);
58     } else {
59       assert(v == 0 || v > __kmp_hidden_helper_threads_num);
60     }
61   } else {
62     assert(v >= 0);
63   }
64 }
65 
66 int main(int argc, char *argv[]) {
67   __kmp_hidden_helper_threads_num = get_num_hidden_helper_threads();
68 
69   constexpr const int N = 1024;
70 #pragma omp parallel for
71   for (int i = 0; i < N; ++i) {
72     int32_t data1 = -1, data2 = -1, data3 = -1;
73     int depvar;
74     int32_t gtid = __kmpc_global_thread_num(nullptr);
75 
76     // Task 1, regular task
77     auto task1 = __kmpc_omp_task_alloc(
78         nullptr, gtid, 1, sizeof(kmp_task_t_with_privates), sizeof(anon),
79         reinterpret_cast<kmp_routine_entry_t>(omp_task_entry));
80     auto shareds = reinterpret_cast<anon *>(task1->shareds);
81     shareds->data = &data1;
82 
83     kmp_depend_info_t depinfo1;
84     depinfo1.base_addr = reinterpret_cast<intptr_t>(&depvar);
85     depinfo1.flag = 3; // INOUT
86     depinfo1.len = 4;
87 
88     __kmpc_omp_task_with_deps(nullptr, gtid, task1, 1, &depinfo1, 0, nullptr);
89 
90     // Task 2, hidden helper task
91     auto task2 = __kmpc_omp_target_task_alloc(
92         nullptr, gtid, 1, sizeof(kmp_task_t_with_privates), sizeof(anon),
93         reinterpret_cast<kmp_routine_entry_t>(omp_task_entry), -1);
94     shareds = reinterpret_cast<anon *>(task2->shareds);
95     shareds->data = &data2;
96 
97     kmp_depend_info_t depinfo2;
98     depinfo2.base_addr = reinterpret_cast<intptr_t>(&depvar);
99     depinfo2.flag = 3; // INOUT
100     depinfo2.len = 4;
101 
102     __kmpc_omp_task_with_deps(nullptr, gtid, task2, 1, &depinfo2, 0, nullptr);
103 
104     // Task 3, regular task
105     auto task3 = __kmpc_omp_task_alloc(
106         nullptr, gtid, 1, sizeof(kmp_task_t_with_privates), sizeof(anon),
107         reinterpret_cast<kmp_routine_entry_t>(omp_task_entry));
108     shareds = reinterpret_cast<anon *>(task3->shareds);
109     shareds->data = &data3;
110 
111     kmp_depend_info_t depinfo3;
112     depinfo3.base_addr = reinterpret_cast<intptr_t>(&depvar);
113     depinfo3.flag = 3; // INOUT
114     depinfo3.len = 4;
115 
116     __kmpc_omp_task_with_deps(nullptr, gtid, task3, 1, &depinfo3, 0, nullptr);
117 
118     __kmpc_omp_taskwait(nullptr, gtid);
119 
120     // FIXME: 8 here is not accurate
121     assert_gtid<false>(data1);
122     assert_gtid<true>(data2);
123     assert_gtid<false>(data3);
124   }
125 
126   std::cout << "PASS\n";
127   return 0;
128 }
129 
130 // CHECK: PASS
131