1 | // RUN: %libomp-compile && env LIBOMP_NUM_HIDDEN_HELPER_THREADS=0 OMP_PROC_BIND=close OMP_PLACES=cores KMP_AFFINITY=verbose %libomp-run 8 1 4 |
2 | // REQUIRES: linux |
3 | // |
4 | // This test pthread_creates 8 root threads before any OpenMP |
5 | // runtime entry is ever called. We have all the root threads |
6 | // register with the runtime by calling omp_set_num_threads(), |
7 | // but this does not initialize their affinity. The fourth root thread |
8 | // then calls a parallel region and we make sure its affinity |
9 | // is correct. We also make sure all the other root threads are |
10 | // free-floating since they have not called into a parallel region. |
11 | |
12 | #define _GNU_SOURCE |
13 | #include <stdio.h> |
14 | #include <stdlib.h> |
15 | #include <omp.h> |
16 | #include <pthread.h> |
17 | #include <unistd.h> |
18 | #include <assert.h> |
19 | #include <sys/types.h> |
20 | #include <sys/syscall.h> |
21 | #include "libomp_test_affinity.h" |
22 | |
23 | volatile int entry_flag = 0; |
24 | volatile int flag = 0; |
25 | volatile int num_roots_arrived = 0; |
26 | int num_roots; |
27 | int spawner = 0; |
28 | pthread_mutex_t lock; |
29 | int register_workers = 0; // boolean |
30 | affinity_mask_t *full_mask; |
31 | |
32 | int __kmpc_global_thread_num(void*); |
33 | |
34 | int get_os_thread_id() { |
35 | return (int)syscall(SYS_gettid); |
36 | } |
37 | |
38 | int place_and_affinity_match() { |
39 | int i, max_cpu; |
40 | char buf[512]; |
41 | affinity_mask_t *mask = affinity_mask_alloc(); |
42 | int place = omp_get_place_num(); |
43 | int num_procs = omp_get_place_num_procs(place); |
44 | int *ids = (int*)malloc(size: sizeof(int) * num_procs); |
45 | omp_get_place_proc_ids(place, ids); |
46 | get_thread_affinity(mask); |
47 | affinity_mask_snprintf(buf, bufsize: sizeof(buf), mask); |
48 | printf(format: "Primary Thread Place: %d\n" , place); |
49 | printf(format: "Primary Thread mask: %s\n" , buf); |
50 | |
51 | for (i = 0; i < num_procs; ++i) { |
52 | int cpu = ids[i]; |
53 | if (!affinity_mask_isset(mask, cpu)) |
54 | return 0; |
55 | } |
56 | |
57 | max_cpu = AFFINITY_MAX_CPUS; |
58 | for (i = 0; i < max_cpu; ++i) { |
59 | int cpu = i; |
60 | if (affinity_mask_isset(mask, cpu)) { |
61 | int j, found = 0; |
62 | for (j = 0; j < num_procs; ++j) { |
63 | if (ids[j] == cpu) { |
64 | found = 1; |
65 | break; |
66 | } |
67 | } |
68 | if (!found) |
69 | return 0; |
70 | } |
71 | } |
72 | |
73 | affinity_mask_free(mask); |
74 | free(ptr: ids); |
75 | return 1; |
76 | } |
77 | |
78 | void* thread_func(void *arg) { |
79 | int place, nplaces; |
80 | int root_id = *((int*)arg); |
81 | int pid = getpid(); |
82 | int tid = get_os_thread_id(); |
83 | |
84 | // Order how the root threads are assigned a gtid in the runtime |
85 | // i.e., root_id = gtid |
86 | while (1) { |
87 | int v = entry_flag; |
88 | if (v == root_id) |
89 | break; |
90 | } |
91 | |
92 | // If main root thread |
93 | if (root_id == spawner) { |
94 | printf(format: "Initial application thread (pid=%d, tid=%d, spawner=%d) reached thread_func (will call OpenMP)\n" , pid, tid, spawner); |
95 | omp_set_num_threads(4); |
96 | #pragma omp atomic |
97 | entry_flag++; |
98 | // Wait for the workers to signal their arrival before #pragma omp parallel |
99 | while (num_roots_arrived < num_roots - 1) {} |
100 | // This will trigger the output for KMP_AFFINITY in this case |
101 | #pragma omp parallel |
102 | { |
103 | int gtid = __kmpc_global_thread_num(NULL); |
104 | #pragma omp single |
105 | { |
106 | printf(format: "Exactly %d threads in the #pragma omp parallel\n" , |
107 | omp_get_num_threads()); |
108 | } |
109 | #pragma omp critical |
110 | { |
111 | printf(format: "OpenMP thread %d: gtid=%d\n" , omp_get_thread_num(), gtid); |
112 | } |
113 | } |
114 | flag = 1; |
115 | if (!place_and_affinity_match()) { |
116 | fprintf(stderr, format: "error: place and affinity mask do not match for primary thread\n" ); |
117 | exit (EXIT_FAILURE); |
118 | } |
119 | |
120 | } else { // If worker root thread |
121 | // Worker root threads, register with OpenMP through omp_set_num_threads() |
122 | // if designated to, signal their arrival and then wait for the main root |
123 | // thread to signal them to exit. |
124 | printf(format: "New root pthread (pid=%d, tid=%d) reached thread_func\n" , pid, tid); |
125 | if (register_workers) |
126 | omp_set_num_threads(4); |
127 | #pragma omp atomic |
128 | entry_flag++; |
129 | |
130 | pthread_mutex_lock(mutex: &lock); |
131 | num_roots_arrived++; |
132 | pthread_mutex_unlock(mutex: &lock); |
133 | while (flag == 0) {} |
134 | |
135 | // Main check whether root threads' mask is equal to the |
136 | // initial affinity mask |
137 | affinity_mask_t *mask = affinity_mask_alloc(); |
138 | get_thread_affinity(mask); |
139 | if (!affinity_mask_equal(mask1: mask, mask2: full_mask)) { |
140 | char buf[1024]; |
141 | printf(format: "root thread %d mask: " , root_id); |
142 | affinity_mask_snprintf(buf, bufsize: sizeof(buf), mask); |
143 | printf(format: "initial affinity mask: %s\n" , buf); |
144 | fprintf(stderr, format: "error: root thread %d affinity mask not equal" |
145 | " to initial full mask\n" , root_id); |
146 | affinity_mask_free(mask); |
147 | exit(EXIT_FAILURE); |
148 | } |
149 | affinity_mask_free(mask); |
150 | } |
151 | return NULL; |
152 | } |
153 | |
154 | int main(int argc, char** argv) { |
155 | int i; |
156 | if (argc != 3 && argc != 4) { |
157 | fprintf(stderr, format: "usage: %s <num_roots> <register_workers_bool> [<spawn_root_number>]\n" , argv[0]); |
158 | exit(EXIT_FAILURE); |
159 | } |
160 | |
161 | // Initialize pthread mutex |
162 | pthread_mutex_init(mutex: &lock, NULL); |
163 | |
164 | // Get initial full mask |
165 | full_mask = affinity_mask_alloc(); |
166 | get_thread_affinity(mask: full_mask); |
167 | |
168 | // Get the number of root pthreads to create and allocate resources for them |
169 | num_roots = atoi(nptr: argv[1]); |
170 | pthread_t *roots = (pthread_t*)malloc(size: sizeof(pthread_t) * num_roots); |
171 | int *root_ids = (int*)malloc(size: sizeof(int) * num_roots); |
172 | |
173 | // Get the flag indicating whether to have root pthreads call omp_set_num_threads() or not |
174 | register_workers = atoi(nptr: argv[2]); |
175 | |
176 | if (argc == 4) |
177 | spawner = atoi(nptr: argv[3]); |
178 | |
179 | // Spawn worker root threads |
180 | for (i = 1; i < num_roots; ++i) { |
181 | *(root_ids + i) = i; |
182 | pthread_create(newthread: roots + i, NULL, start_routine: thread_func, arg: root_ids + i); |
183 | } |
184 | // Have main root thread (root 0) go into thread_func |
185 | *root_ids = 0; |
186 | thread_func(arg: root_ids); |
187 | |
188 | // Cleanup all resources |
189 | for (i = 1; i < num_roots; ++i) { |
190 | void *status; |
191 | pthread_join(th: roots[i], thread_return: &status); |
192 | } |
193 | free(ptr: roots); |
194 | free(ptr: root_ids); |
195 | pthread_mutex_destroy(mutex: &lock); |
196 | return EXIT_SUCCESS; |
197 | } |
198 | |