| 1 | // RUN: %libomp-compile-and-run |
| 2 | // The runtime currently does not get dependency information from GCC. |
| 3 | // UNSUPPORTED: gcc |
| 4 | |
| 5 | // Tests OMP 5.x task dependence "omp_all_memory", |
| 6 | // emulates compiler codegen versions for new dep kind |
| 7 | // |
| 8 | // Task tree created: |
| 9 | // task0 - task1 (in: i1, i2) |
| 10 | // \ |
| 11 | // task2 (inoutset: i2), (in: i1) |
| 12 | // / |
| 13 | // task3 (omp_all_memory) via flag=0x80 |
| 14 | // / |
| 15 | // task4 - task5 (in: i1, i2) |
| 16 | // / |
| 17 | // task6 (omp_all_memory) via addr=-1 |
| 18 | // / |
| 19 | // task7 (omp_all_memory) via flag=0x80 |
| 20 | // / |
| 21 | // task8 (in: i3) |
| 22 | // / |
| 23 | // task9 - no dependences |
| 24 | // / |
| 25 | // taskwait (omp_all_memory) (should not wait for task9, see prints) |
| 26 | // |
| 27 | #include <stdio.h> |
| 28 | #include <omp.h> |
| 29 | |
| 30 | #ifdef _WIN32 |
| 31 | #include <windows.h> |
| 32 | #define mysleep(n) Sleep(n) |
| 33 | #else |
| 34 | #include <unistd.h> |
| 35 | #define mysleep(n) usleep((n)*1000) |
| 36 | #endif |
| 37 | |
| 38 | // to check the # of concurrent tasks (must be 1 for MTX, <3 for other kinds) |
| 39 | static int checker = 0; |
| 40 | static int err = 0; |
| 41 | static int taskwait_flag = 0; |
| 42 | #ifndef DELAY |
| 43 | // set delay interval in ms for dependent tasks |
| 44 | #define DELAY 100 |
| 45 | #endif |
| 46 | |
| 47 | // --------------------------------------------------------------------------- |
| 48 | // internal data to emulate compiler codegen |
| 49 | typedef struct DEP { |
| 50 | size_t addr; |
| 51 | size_t len; |
| 52 | unsigned char flags; |
| 53 | } dep; |
| 54 | #define DEP_ALL_MEM 0x80 |
| 55 | typedef struct task { |
| 56 | void** shareds; |
| 57 | void* entry; |
| 58 | int part_id; |
| 59 | void* destr_thunk; |
| 60 | int priority; |
| 61 | long long device_id; |
| 62 | int f_priv; |
| 63 | } task_t; |
| 64 | #define TIED 1 |
| 65 | typedef int(*entry_t)(int, task_t*); |
| 66 | typedef struct ID { |
| 67 | int reserved_1; |
| 68 | int flags; |
| 69 | int reserved_2; |
| 70 | int reserved_3; |
| 71 | char *psource; |
| 72 | } id; |
| 73 | // thunk routine for tasks with ALL dependency |
| 74 | int thunk_m(int gtid, task_t* ptask) { |
| 75 | int lcheck, th; |
| 76 | #pragma omp atomic capture |
| 77 | lcheck = ++checker; |
| 78 | th = omp_get_thread_num(); |
| 79 | printf(format: "task m_%d, th %d, checker %d\n" , ptask->f_priv, th, lcheck); |
| 80 | if (lcheck != 1) { // no more than 1 task at a time |
| 81 | err++; |
| 82 | printf(format: "Error m1, checker %d != 1\n" , lcheck); |
| 83 | } |
| 84 | mysleep(DELAY); |
| 85 | #pragma omp atomic read |
| 86 | lcheck = checker; // must still be equal to 1 |
| 87 | if (lcheck != 1) { |
| 88 | err++; |
| 89 | printf(format: "Error m2, checker %d != 1\n" , lcheck); |
| 90 | } |
| 91 | #pragma omp atomic |
| 92 | --checker; |
| 93 | return 0; |
| 94 | } |
| 95 | // thunk routine for tasks with inoutset dependency |
| 96 | int thunk_s(int gtid, task_t* ptask) { |
| 97 | int lcheck, th; |
| 98 | #pragma omp atomic capture |
| 99 | lcheck = ++checker; // 1 |
| 100 | th = omp_get_thread_num(); |
| 101 | printf(format: "task 2_%d, th %d, checker %d\n" , ptask->f_priv, th, lcheck); |
| 102 | if (lcheck != 1) { // no more than 1 task at a time |
| 103 | err++; |
| 104 | printf(format: "Error s1, checker %d != 1\n" , lcheck); |
| 105 | } |
| 106 | mysleep(DELAY); |
| 107 | #pragma omp atomic read |
| 108 | lcheck = checker; // must still be equal to 1 |
| 109 | if (lcheck != 1) { |
| 110 | err++; |
| 111 | printf(format: "Error s2, checker %d != 1\n" , lcheck); |
| 112 | } |
| 113 | #pragma omp atomic |
| 114 | --checker; |
| 115 | return 0; |
| 116 | } |
| 117 | |
| 118 | #ifdef __cplusplus |
| 119 | extern "C" { |
| 120 | #endif |
| 121 | int __kmpc_global_thread_num(id*); |
| 122 | task_t *__kmpc_omp_task_alloc(id *loc, int gtid, int flags, |
| 123 | size_t sz, size_t shar, entry_t rtn); |
| 124 | int __kmpc_omp_task_with_deps(id *loc, int gtid, task_t *task, int ndeps, |
| 125 | dep *dep_lst, int nd_noalias, dep *noalias_lst); |
| 126 | void __kmpc_omp_wait_deps(id *loc, int gtid, int ndeps, dep *dep_lst, |
| 127 | int ndeps_noalias, dep *noalias_dep_lst); |
| 128 | static id loc = {0, 2, 0, 0, ";file;func;0;0;;" }; |
| 129 | #ifdef __cplusplus |
| 130 | } // extern "C" |
| 131 | #endif |
| 132 | // End of internal data |
| 133 | // --------------------------------------------------------------------------- |
| 134 | |
| 135 | int main() |
| 136 | { |
| 137 | int i1,i2,i3; |
| 138 | omp_set_num_threads(8); |
| 139 | omp_set_dynamic(0); |
| 140 | #pragma omp parallel |
| 141 | { |
| 142 | #pragma omp single nowait |
| 143 | { |
| 144 | dep sdep[2]; |
| 145 | task_t *ptr; |
| 146 | int gtid = __kmpc_global_thread_num(&loc); |
| 147 | int t = omp_get_thread_num(); |
| 148 | // Create longest task first to ensure it is stolen. |
| 149 | // The test may hang if the task created last and |
| 150 | // executed by a thread which executes taskwait. |
| 151 | #pragma omp task |
| 152 | { // task 9 - long running task |
| 153 | int flag; |
| 154 | int th = omp_get_thread_num(); |
| 155 | printf(format: "signalled independent task 9_%d, th %d started....\n" , t, th); |
| 156 | // Wait for taskwait depend() to finish |
| 157 | // If the taskwait depend() improperly depends on this task |
| 158 | // to finish, then the test will hang and a timeout should trigger |
| 159 | while (1) { |
| 160 | #pragma omp atomic read |
| 161 | flag = taskwait_flag; |
| 162 | if (flag == 1) |
| 163 | break; |
| 164 | } |
| 165 | printf(format: "signalled independent task 9_%d, th %d finished....\n" , t, th); |
| 166 | } |
| 167 | #pragma omp task depend(in: i1, i2) |
| 168 | { // task 0 |
| 169 | int lcheck, th; |
| 170 | #pragma omp atomic capture |
| 171 | lcheck = ++checker; // 1 or 2 |
| 172 | th = omp_get_thread_num(); |
| 173 | printf(format: "task 0_%d, th %d, checker %d\n" , t, th, lcheck); |
| 174 | if (lcheck > 2 || lcheck < 1) { |
| 175 | err++; // no more than 2 tasks concurrently |
| 176 | printf(format: "Error1, checker %d, not 1 or 2\n" , lcheck); |
| 177 | } |
| 178 | mysleep(DELAY); |
| 179 | #pragma omp atomic read |
| 180 | lcheck = checker; // 1 or 2 |
| 181 | if (lcheck > 2 || lcheck < 1) { |
| 182 | #pragma omp atomic |
| 183 | err++; |
| 184 | printf(format: "Error2, checker %d, not 1 or 2\n" , lcheck); |
| 185 | } |
| 186 | #pragma omp atomic |
| 187 | --checker; |
| 188 | } |
| 189 | #pragma omp task depend(in: i1, i2) |
| 190 | { // task 1 |
| 191 | int lcheck, th; |
| 192 | #pragma omp atomic capture |
| 193 | lcheck = ++checker; // 1 or 2 |
| 194 | th = omp_get_thread_num(); |
| 195 | printf(format: "task 1_%d, th %d, checker %d\n" , t, th, lcheck); |
| 196 | if (lcheck > 2 || lcheck < 1) { |
| 197 | err++; // no more than 2 tasks concurrently |
| 198 | printf(format: "Error3, checker %d, not 1 or 2\n" , lcheck); |
| 199 | } |
| 200 | mysleep(DELAY); |
| 201 | #pragma omp atomic read |
| 202 | lcheck = checker; // 1 or 2 |
| 203 | if (lcheck > 2 || lcheck < 1) { |
| 204 | err++; |
| 205 | printf(format: "Error4, checker %d, not 1 or 2\n" , lcheck); |
| 206 | } |
| 207 | #pragma omp atomic |
| 208 | --checker; |
| 209 | } |
| 210 | // compiler codegen start |
| 211 | // task2 |
| 212 | ptr = __kmpc_omp_task_alloc(loc: &loc, gtid, TIED, sz: sizeof(task_t), shar: 0, rtn: thunk_s); |
| 213 | sdep[0].addr = (size_t)&i1; |
| 214 | sdep[0].len = 0; // not used |
| 215 | sdep[0].flags = 1; // IN |
| 216 | sdep[1].addr = (size_t)&i2; |
| 217 | sdep[1].len = 0; // not used |
| 218 | sdep[1].flags = 8; // INOUTSET |
| 219 | ptr->f_priv = t + 10; // init single first-private variable |
| 220 | __kmpc_omp_task_with_deps(loc: &loc, gtid, task: ptr, ndeps: 2, dep_lst: sdep, nd_noalias: 0, noalias_lst: 0); |
| 221 | |
| 222 | // task3 |
| 223 | ptr = __kmpc_omp_task_alloc(loc: &loc, gtid, TIED, sz: sizeof(task_t), shar: 0, rtn: thunk_m); |
| 224 | sdep[0].addr = (size_t)&i1; // to be ignored |
| 225 | sdep[0].len = 0; // not used |
| 226 | sdep[0].flags = 1; // IN |
| 227 | sdep[1].addr = 0; |
| 228 | sdep[1].len = 0; // not used |
| 229 | sdep[1].flags = DEP_ALL_MEM; // omp_all_memory |
| 230 | ptr->f_priv = t + 20; // init single first-private variable |
| 231 | __kmpc_omp_task_with_deps(loc: &loc, gtid, task: ptr, ndeps: 2, dep_lst: sdep, nd_noalias: 0, noalias_lst: 0); |
| 232 | // compiler codegen end |
| 233 | #pragma omp task depend(in: i1, i2) |
| 234 | { // task 4 |
| 235 | int lcheck, th; |
| 236 | #pragma omp atomic capture |
| 237 | lcheck = ++checker; // 1 or 2 |
| 238 | th = omp_get_thread_num(); |
| 239 | printf(format: "task 4_%d, th %d, checker %d\n" , t, th, lcheck); |
| 240 | if (lcheck > 2 || lcheck < 1) { |
| 241 | err++; // no more than 2 tasks concurrently |
| 242 | printf(format: "Error5, checker %d, not 1 or 2\n" , lcheck); |
| 243 | } |
| 244 | mysleep(DELAY); |
| 245 | #pragma omp atomic read |
| 246 | lcheck = checker; // 1 or 2 |
| 247 | if (lcheck > 2 || lcheck < 1) { |
| 248 | err++; |
| 249 | printf(format: "Error6, checker %d, not 1 or 2\n" , lcheck); |
| 250 | } |
| 251 | #pragma omp atomic |
| 252 | --checker; |
| 253 | } |
| 254 | #pragma omp task depend(in: i1, i2) |
| 255 | { // task 5 |
| 256 | int lcheck, th; |
| 257 | #pragma omp atomic capture |
| 258 | lcheck = ++checker; // 1 or 2 |
| 259 | th = omp_get_thread_num(); |
| 260 | printf(format: "task 5_%d, th %d, checker %d\n" , t, th, lcheck); |
| 261 | if (lcheck > 2 || lcheck < 1) { |
| 262 | err++; // no more than 2 tasks concurrently |
| 263 | printf(format: "Error7, checker %d, not 1 or 2\n" , lcheck); |
| 264 | } |
| 265 | mysleep(DELAY); |
| 266 | #pragma omp atomic read |
| 267 | lcheck = checker; // 1 or 2 |
| 268 | if (lcheck > 2 || lcheck < 1) { |
| 269 | err++; |
| 270 | printf(format: "Error8, checker %d, not 1 or 2\n" , lcheck); |
| 271 | } |
| 272 | #pragma omp atomic |
| 273 | --checker; |
| 274 | } |
| 275 | // compiler codegen start |
| 276 | // task6 |
| 277 | ptr = __kmpc_omp_task_alloc(loc: &loc, gtid, TIED, sz: sizeof(task_t), shar: 0, rtn: thunk_m); |
| 278 | sdep[0].addr = (size_t)(-1); // omp_all_memory |
| 279 | sdep[0].len = 0; // not used |
| 280 | sdep[0].flags = 2; // OUT |
| 281 | ptr->f_priv = t + 30; // init single first-private variable |
| 282 | __kmpc_omp_task_with_deps(loc: &loc, gtid, task: ptr, ndeps: 1, dep_lst: sdep, nd_noalias: 0, noalias_lst: 0); |
| 283 | |
| 284 | // task7 |
| 285 | ptr = __kmpc_omp_task_alloc(loc: &loc, gtid, TIED, sz: sizeof(task_t), shar: 0, rtn: thunk_m); |
| 286 | sdep[0].addr = 0; |
| 287 | sdep[0].len = 0; // not used |
| 288 | sdep[0].flags = DEP_ALL_MEM; // omp_all_memory |
| 289 | sdep[1].addr = (size_t)&i3; // to be ignored |
| 290 | sdep[1].len = 0; // not used |
| 291 | sdep[1].flags = 4; // MUTEXINOUTSET |
| 292 | ptr->f_priv = t + 40; // init single first-private variable |
| 293 | __kmpc_omp_task_with_deps(loc: &loc, gtid, task: ptr, ndeps: 2, dep_lst: sdep, nd_noalias: 0, noalias_lst: 0); |
| 294 | // compiler codegen end |
| 295 | #pragma omp task depend(in: i3) |
| 296 | { // task 8 |
| 297 | int lcheck, th; |
| 298 | #pragma omp atomic capture |
| 299 | lcheck = ++checker; // 1 |
| 300 | th = omp_get_thread_num(); |
| 301 | printf(format: "task 8_%d, th %d, checker %d\n" , t, th, lcheck); |
| 302 | if (lcheck != 1) { |
| 303 | err++; |
| 304 | printf(format: "Error9, checker %d, != 1\n" , lcheck); |
| 305 | } |
| 306 | mysleep(DELAY); |
| 307 | #pragma omp atomic read |
| 308 | lcheck = checker; |
| 309 | if (lcheck != 1) { |
| 310 | err++; |
| 311 | printf(format: "Error10, checker %d, != 1\n" , lcheck); |
| 312 | } |
| 313 | #pragma omp atomic |
| 314 | --checker; |
| 315 | } |
| 316 | mysleep(1); // wait a bit to ensure at least first task is stolen |
| 317 | // #pragma omp taskwait depend(omp_all_memory: out) |
| 318 | printf(format: "all 10 tasks generated;\n" |
| 319 | "taskwait depend(omp_all_memory: out) started, th %d\n" , t); |
| 320 | __kmpc_omp_wait_deps(loc: &loc, gtid, ndeps: 1, dep_lst: sdep, ndeps_noalias: 0, noalias_dep_lst: 0); |
| 321 | #pragma omp atomic write |
| 322 | taskwait_flag = 1; |
| 323 | printf(format: "taskwait depend(omp_all_memory: out) passed, th %d\n" , t); |
| 324 | fflush(stream: 0); |
| 325 | } // single |
| 326 | } // parallel |
| 327 | if (err == 0 && checker == 0) { |
| 328 | printf(format: "passed\n" ); |
| 329 | return 0; |
| 330 | } else { |
| 331 | printf(format: "failed, err = %d, checker = %d\n" , err, checker); |
| 332 | return 1; |
| 333 | } |
| 334 | } |
| 335 | |