File: | projects/openmp/runtime/src/kmp_runtime.cpp |
Warning: | line 1390, column 5 Value stored to 'implicit_task_data' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* |
2 | * kmp_runtime.cpp -- KPTS runtime support library |
3 | */ |
4 | |
5 | //===----------------------------------------------------------------------===// |
6 | // |
7 | // The LLVM Compiler Infrastructure |
8 | // |
9 | // This file is dual licensed under the MIT and the University of Illinois Open |
10 | // Source Licenses. See LICENSE.txt for details. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "kmp.h" |
15 | #include "kmp_affinity.h" |
16 | #include "kmp_atomic.h" |
17 | #include "kmp_environment.h" |
18 | #include "kmp_error.h" |
19 | #include "kmp_i18n.h" |
20 | #include "kmp_io.h" |
21 | #include "kmp_itt.h" |
22 | #include "kmp_settings.h" |
23 | #include "kmp_stats.h" |
24 | #include "kmp_str.h" |
25 | #include "kmp_wait_release.h" |
26 | #include "kmp_wrapper_getpid.h" |
27 | #include "kmp_dispatch.h" |
28 | #if KMP_USE_HIER_SCHED0 |
29 | #include "kmp_dispatch_hier.h" |
30 | #endif |
31 | |
32 | #if OMPT_SUPPORT1 |
33 | #include "ompt-specific.h" |
34 | #endif |
35 | |
36 | /* these are temporary issues to be dealt with */ |
37 | #define KMP_USE_PRCTL0 0 |
38 | |
39 | #if KMP_OS_WINDOWS0 |
40 | #include <process.h> |
41 | #endif |
42 | |
43 | #include "tsan_annotations.h" |
44 | |
45 | #if defined(KMP_GOMP_COMPAT) |
46 | char const __kmp_version_alt_comp[] = |
47 | KMP_VERSION_PREFIX"\x00@(#) " "Intel(R) OMP " "alternative compiler support: yes"; |
48 | #endif /* defined(KMP_GOMP_COMPAT) */ |
49 | |
50 | char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX"\x00@(#) " "Intel(R) OMP " "API version: " |
51 | #if OMP_50_ENABLED(50 >= 50) |
52 | "5.0 (201611)"; |
53 | #elif OMP_45_ENABLED(50 >= 45) |
54 | "4.5 (201511)"; |
55 | #elif OMP_40_ENABLED(50 >= 40) |
56 | "4.0 (201307)"; |
57 | #else |
58 | "3.1 (201107)"; |
59 | #endif |
60 | |
61 | #ifdef KMP_DEBUG1 |
62 | char const __kmp_version_lock[] = |
63 | KMP_VERSION_PREFIX"\x00@(#) " "Intel(R) OMP " "lock type: run time selectable"; |
64 | #endif /* KMP_DEBUG */ |
65 | |
66 | #define KMP_MIN(x, y)((x) < (y) ? (x) : (y)) ((x) < (y) ? (x) : (y)) |
67 | |
68 | /* ------------------------------------------------------------------------ */ |
69 | |
70 | #if KMP_USE_MONITOR |
71 | kmp_info_t __kmp_monitor; |
72 | #endif |
73 | |
74 | /* Forward declarations */ |
75 | |
76 | void __kmp_cleanup(void); |
77 | |
78 | static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid, |
79 | int gtid); |
80 | static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, |
81 | kmp_internal_control_t *new_icvs, |
82 | ident_t *loc); |
83 | #if OMP_40_ENABLED(50 >= 40) && KMP_AFFINITY_SUPPORTED1 |
84 | static void __kmp_partition_places(kmp_team_t *team, |
85 | int update_master_only = 0); |
86 | #endif |
87 | static void __kmp_do_serial_initialize(void); |
88 | void __kmp_fork_barrier(int gtid, int tid); |
89 | void __kmp_join_barrier(int gtid); |
90 | void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc, |
91 | kmp_internal_control_t *new_icvs, ident_t *loc); |
92 | |
93 | #ifdef USE_LOAD_BALANCE1 |
94 | static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc); |
95 | #endif |
96 | |
97 | static int __kmp_expand_threads(int nNeed); |
98 | #if KMP_OS_WINDOWS0 |
99 | static int __kmp_unregister_root_other_thread(int gtid); |
100 | #endif |
101 | static void __kmp_unregister_library(void); // called by __kmp_internal_end() |
102 | static void __kmp_reap_thread(kmp_info_t *thread, int is_root); |
103 | kmp_info_t *__kmp_thread_pool_insert_pt = NULL__null; |
104 | |
105 | /* Calculate the identifier of the current thread */ |
106 | /* fast (and somewhat portable) way to get unique identifier of executing |
107 | thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */ |
108 | int __kmp_get_global_thread_id() { |
109 | int i; |
110 | kmp_info_t **other_threads; |
111 | size_t stack_data; |
112 | char *stack_addr; |
113 | size_t stack_size; |
114 | char *stack_base; |
115 | |
116 | KA_TRACE(if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } |
117 | 1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } |
118 | ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n",if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); } |
119 | __kmp_nth, __kmp_all_nth))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n" , __kmp_nth, __kmp_all_nth); }; |
120 | |
121 | /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to |
122 | a parallel region, made it return KMP_GTID_DNE to force serial_initialize |
123 | by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee |
124 | __kmp_init_gtid for this to work. */ |
125 | |
126 | if (!TCR_4(__kmp_init_gtid)(__kmp_init_gtid)) |
127 | return KMP_GTID_DNE(-2); |
128 | |
129 | #ifdef KMP_TDATA_GTID1 |
130 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 3) { |
131 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using TDATA\n" ); }; |
132 | return __kmp_gtid; |
133 | } |
134 | #endif |
135 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 2) { |
136 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using keyed TLS\n" ); }; |
137 | return __kmp_gtid_get_specific(); |
138 | } |
139 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: using internal alg.\n" ); }; |
140 | |
141 | stack_addr = (char *)&stack_data; |
142 | other_threads = __kmp_threads; |
143 | |
144 | /* ATT: The code below is a source of potential bugs due to unsynchronized |
145 | access to __kmp_threads array. For example: |
146 | 1. Current thread loads other_threads[i] to thr and checks it, it is |
147 | non-NULL. |
148 | 2. Current thread is suspended by OS. |
149 | 3. Another thread unregisters and finishes (debug versions of free() |
150 | may fill memory with something like 0xEF). |
151 | 4. Current thread is resumed. |
152 | 5. Current thread reads junk from *thr. |
153 | TODO: Fix it. --ln */ |
154 | |
155 | for (i = 0; i < __kmp_threads_capacity; i++) { |
156 | |
157 | kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i])((void *)(other_threads[i])); |
158 | if (!thr) |
159 | continue; |
160 | |
161 | stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize)((void *)(thr->th.th_info.ds.ds_stacksize)); |
162 | stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase)((void *)(thr->th.th_info.ds.ds_stackbase)); |
163 | |
164 | /* stack grows down -- search through all of the active threads */ |
165 | |
166 | if (stack_addr <= stack_base) { |
167 | size_t stack_diff = stack_base - stack_addr; |
168 | |
169 | if (stack_diff <= stack_size) { |
170 | /* The only way we can be closer than the allocated */ |
171 | /* stack size is if we are running on this thread. */ |
172 | KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i)((__kmp_gtid_get_specific() == i) ? 0 : __kmp_debug_assert("__kmp_gtid_get_specific() == i" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 172)); |
173 | return i; |
174 | } |
175 | } |
176 | } |
177 | |
178 | /* get specific to try and determine our gtid */ |
179 | KA_TRACE(1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); } |
180 | ("*** __kmp_get_global_thread_id: internal alg. failed to find "if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); } |
181 | "thread, using TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id: internal alg. failed to find " "thread, using TLS\n"); }; |
182 | i = __kmp_gtid_get_specific(); |
183 | |
184 | /*fprintf( stderr, "=== %d\n", i ); */ /* GROO */ |
185 | |
186 | /* if we havn't been assigned a gtid, then return code */ |
187 | if (i < 0) |
188 | return i; |
189 | |
190 | /* dynamically updated stack window for uber threads to avoid get_specific |
191 | call */ |
192 | if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)(other_threads[i]->th.th_info.ds.ds_stackgrow)) { |
193 | KMP_FATAL(StackOverflow, i)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_StackOverflow, i), __kmp_msg_null ); |
194 | } |
195 | |
196 | stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; |
197 | if (stack_addr > stack_base) { |
198 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr)((other_threads[i]->th.th_info.ds.ds_stackbase)) = ((stack_addr )); |
199 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ) |
200 | other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ) |
201 | stack_base)((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((other_threads [i]->th.th_info.ds.ds_stacksize + stack_addr - stack_base) ); |
202 | } else { |
203 | TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((stack_base - stack_addr)) |
204 | stack_base - stack_addr)((other_threads[i]->th.th_info.ds.ds_stacksize)) = ((stack_base - stack_addr)); |
205 | } |
206 | |
207 | /* Reprint stack bounds for ubermaster since they have been refined */ |
208 | if (__kmp_storage_map) { |
209 | char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; |
210 | char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize; |
211 | __kmp_print_storage_map_gtid(i, stack_beg, stack_end, |
212 | other_threads[i]->th.th_info.ds.ds_stacksize, |
213 | "th_%d stack (refinement)", i); |
214 | } |
215 | return i; |
216 | } |
217 | |
218 | int __kmp_get_global_thread_id_reg() { |
219 | int gtid; |
220 | |
221 | if (!__kmp_init_serial) { |
222 | gtid = KMP_GTID_DNE(-2); |
223 | } else |
224 | #ifdef KMP_TDATA_GTID1 |
225 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 3) { |
226 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using TDATA\n" ); }; |
227 | gtid = __kmp_gtid; |
228 | } else |
229 | #endif |
230 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) >= 2) { |
231 | KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n" ); }; |
232 | gtid = __kmp_gtid_get_specific(); |
233 | } else { |
234 | KA_TRACE(1000,if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using internal alg.\n" ); } |
235 | ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"))if (kmp_a_debug >= 1000) { __kmp_debug_printf ("*** __kmp_get_global_thread_id_reg: using internal alg.\n" ); }; |
236 | gtid = __kmp_get_global_thread_id(); |
237 | } |
238 | |
239 | /* we must be a new uber master sibling thread */ |
240 | if (gtid == KMP_GTID_DNE(-2)) { |
241 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); } |
242 | ("__kmp_get_global_thread_id_reg: Encountered new root thread. "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); } |
243 | "Registering a new gtid.\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_global_thread_id_reg: Encountered new root thread. " "Registering a new gtid.\n"); }; |
244 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
245 | if (!__kmp_init_serial) { |
246 | __kmp_do_serial_initialize(); |
247 | gtid = __kmp_gtid_get_specific(); |
248 | } else { |
249 | gtid = __kmp_register_root(FALSE0); |
250 | } |
251 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
252 | /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */ |
253 | } |
254 | |
255 | KMP_DEBUG_ASSERT(gtid >= 0)((gtid >= 0) ? 0 : __kmp_debug_assert("gtid >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 255)); |
256 | |
257 | return gtid; |
258 | } |
259 | |
260 | /* caller must hold forkjoin_lock */ |
261 | void __kmp_check_stack_overlap(kmp_info_t *th) { |
262 | int f; |
263 | char *stack_beg = NULL__null; |
264 | char *stack_end = NULL__null; |
265 | int gtid; |
266 | |
267 | KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: called\n" ); }; |
268 | if (__kmp_storage_map) { |
269 | stack_end = (char *)th->th.th_info.ds.ds_stackbase; |
270 | stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; |
271 | |
272 | gtid = __kmp_gtid_from_thread(th)((((th) != __null) ? 0 : __kmp_debug_assert("(th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 272)), (th)->th.th_info.ds.ds_gtid); |
273 | |
274 | if (gtid == KMP_GTID_MONITOR(-4)) { |
275 | __kmp_print_storage_map_gtid( |
276 | gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, |
277 | "th_%s stack (%s)", "mon", |
278 | (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); |
279 | } else { |
280 | __kmp_print_storage_map_gtid( |
281 | gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, |
282 | "th_%d stack (%s)", gtid, |
283 | (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); |
284 | } |
285 | } |
286 | |
287 | /* No point in checking ubermaster threads since they use refinement and |
288 | * cannot overlap */ |
289 | gtid = __kmp_gtid_from_thread(th)((((th) != __null) ? 0 : __kmp_debug_assert("(th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 289)), (th)->th.th_info.ds.ds_gtid); |
290 | if (__kmp_env_checks == TRUE(!0) && !KMP_UBER_GTID(gtid)((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 290)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 290)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) { |
291 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: performing extensive checking\n" ); } |
292 | ("__kmp_check_stack_overlap: performing extensive checking\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: performing extensive checking\n" ); }; |
293 | if (stack_beg == NULL__null) { |
294 | stack_end = (char *)th->th.th_info.ds.ds_stackbase; |
295 | stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; |
296 | } |
297 | |
298 | for (f = 0; f < __kmp_threads_capacity; f++) { |
299 | kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f])((void *)(__kmp_threads[f])); |
300 | |
301 | if (f_th && f_th != th) { |
302 | char *other_stack_end = |
303 | (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase)((void *)(f_th->th.th_info.ds.ds_stackbase)); |
304 | char *other_stack_beg = |
305 | other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize)((void *)(f_th->th.th_info.ds.ds_stacksize)); |
306 | if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) || |
307 | (stack_end > other_stack_beg && stack_end < other_stack_end)) { |
308 | |
309 | /* Print the other stack values before the abort */ |
310 | if (__kmp_storage_map) |
311 | __kmp_print_storage_map_gtid( |
312 | -1, other_stack_beg, other_stack_end, |
313 | (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize)((void *)(f_th->th.th_info.ds.ds_stacksize)), |
314 | "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th)((((f_th) != __null) ? 0 : __kmp_debug_assert("(f_th) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 314)), (f_th)->th.th_info.ds.ds_gtid)); |
315 | |
316 | __kmp_fatal(KMP_MSG(StackOverlap)__kmp_msg_format(kmp_i18n_msg_StackOverlap), KMP_HNT(ChangeStackLimit)__kmp_msg_format(kmp_i18n_hnt_ChangeStackLimit), |
317 | __kmp_msg_null); |
318 | } |
319 | } |
320 | } |
321 | } |
322 | KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_check_stack_overlap: returning\n" ); }; |
323 | } |
324 | |
325 | /* ------------------------------------------------------------------------ */ |
326 | |
327 | void __kmp_infinite_loop(void) { |
328 | static int done = FALSE0; |
329 | |
330 | while (!done) { |
331 | KMP_YIELD(1){ __kmp_x86_pause(); __kmp_yield((1)); }; |
332 | } |
333 | } |
334 | |
335 | #define MAX_MESSAGE512 512 |
336 | |
337 | void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size, |
338 | char const *format, ...) { |
339 | char buffer[MAX_MESSAGE512]; |
340 | va_list ap; |
341 | |
342 | va_start(ap, format)__builtin_va_start(ap, format); |
343 | KMP_SNPRINTFsnprintf(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1, |
344 | p2, (unsigned long)size, format); |
345 | __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); |
346 | __kmp_vprintf(kmp_err, buffer, ap); |
347 | #if KMP_PRINT_DATA_PLACEMENT |
348 | int node; |
349 | if (gtid >= 0) { |
350 | if (p1 <= p2 && (char *)p2 - (char *)p1 == size) { |
351 | if (__kmp_storage_map_verbose) { |
352 | node = __kmp_get_host_node(p1); |
353 | if (node < 0) /* doesn't work, so don't try this next time */ |
354 | __kmp_storage_map_verbose = FALSE0; |
355 | else { |
356 | char *last; |
357 | int lastNode; |
358 | int localProc = __kmp_get_cpu_from_gtid(gtid); |
359 | |
360 | const int page_size = KMP_GET_PAGE_SIZE()getpagesize(); |
361 | |
362 | p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1)); |
363 | p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1)); |
364 | if (localProc >= 0) |
365 | __kmp_printf_no_lock(" GTID %d localNode %d\n", gtid, |
366 | localProc >> 1); |
367 | else |
368 | __kmp_printf_no_lock(" GTID %d\n", gtid); |
369 | #if KMP_USE_PRCTL0 |
370 | /* The more elaborate format is disabled for now because of the prctl |
371 | * hanging bug. */ |
372 | do { |
373 | last = p1; |
374 | lastNode = node; |
375 | /* This loop collates adjacent pages with the same host node. */ |
376 | do { |
377 | (char *)p1 += page_size; |
378 | } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode); |
379 | __kmp_printf_no_lock(" %p-%p memNode %d\n", last, (char *)p1 - 1, |
380 | lastNode); |
381 | } while (p1 <= p2); |
382 | #else |
383 | __kmp_printf_no_lock(" %p-%p memNode %d\n", p1, |
384 | (char *)p1 + (page_size - 1), |
385 | __kmp_get_host_node(p1)); |
386 | if (p1 < p2) { |
387 | __kmp_printf_no_lock(" %p-%p memNode %d\n", p2, |
388 | (char *)p2 + (page_size - 1), |
389 | __kmp_get_host_node(p2)); |
390 | } |
391 | #endif |
392 | } |
393 | } |
394 | } else |
395 | __kmp_printf_no_lock(" %s\n", KMP_I18N_STR(StorageMapWarning)__kmp_i18n_catgets(kmp_i18n_str_StorageMapWarning)); |
396 | } |
397 | #endif /* KMP_PRINT_DATA_PLACEMENT */ |
398 | __kmp_release_bootstrap_lock(&__kmp_stdio_lock); |
399 | } |
400 | |
401 | void __kmp_warn(char const *format, ...) { |
402 | char buffer[MAX_MESSAGE512]; |
403 | va_list ap; |
404 | |
405 | if (__kmp_generate_warnings == kmp_warnings_off) { |
406 | return; |
407 | } |
408 | |
409 | va_start(ap, format)__builtin_va_start(ap, format); |
410 | |
411 | KMP_SNPRINTFsnprintf(buffer, sizeof(buffer), "OMP warning: %s\n", format); |
412 | __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); |
413 | __kmp_vprintf(kmp_err, buffer, ap); |
414 | __kmp_release_bootstrap_lock(&__kmp_stdio_lock); |
415 | |
416 | va_end(ap)__builtin_va_end(ap); |
417 | } |
418 | |
419 | void __kmp_abort_process() { |
420 | // Later threads may stall here, but that's ok because abort() will kill them. |
421 | __kmp_acquire_bootstrap_lock(&__kmp_exit_lock); |
422 | |
423 | if (__kmp_debug_buf) { |
424 | __kmp_dump_debug_buffer(); |
425 | } |
426 | |
427 | if (KMP_OS_WINDOWS0) { |
428 | // Let other threads know of abnormal termination and prevent deadlock |
429 | // if abort happened during library initialization or shutdown |
430 | __kmp_global.g.g_abort = SIGABRT6; |
431 | |
432 | /* On Windows* OS by default abort() causes pop-up error box, which stalls |
433 | nightly testing. Unfortunately, we cannot reliably suppress pop-up error |
434 | boxes. _set_abort_behavior() works well, but this function is not |
435 | available in VS7 (this is not problem for DLL, but it is a problem for |
436 | static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not |
437 | help, at least in some versions of MS C RTL. |
438 | |
439 | It seems following sequence is the only way to simulate abort() and |
440 | avoid pop-up error box. */ |
441 | raise(SIGABRT6); |
442 | _exit(3); // Just in case, if signal ignored, exit anyway. |
443 | } else { |
444 | abort(); |
445 | } |
446 | |
447 | __kmp_infinite_loop(); |
448 | __kmp_release_bootstrap_lock(&__kmp_exit_lock); |
449 | |
450 | } // __kmp_abort_process |
451 | |
452 | void __kmp_abort_thread(void) { |
453 | // TODO: Eliminate g_abort global variable and this function. |
454 | // In case of abort just call abort(), it will kill all the threads. |
455 | __kmp_infinite_loop(); |
456 | } // __kmp_abort_thread |
457 | |
458 | /* Print out the storage map for the major kmp_info_t thread data structures |
459 | that are allocated together. */ |
460 | |
461 | static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) { |
462 | __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d", |
463 | gtid); |
464 | |
465 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team, |
466 | sizeof(kmp_desc_t), "th_%d.th_info", gtid); |
467 | |
468 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head, |
469 | sizeof(kmp_local_t), "th_%d.th_local", gtid); |
470 | |
471 | __kmp_print_storage_map_gtid( |
472 | gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier], |
473 | sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid); |
474 | |
475 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier], |
476 | &thr->th.th_bar[bs_plain_barrier + 1], |
477 | sizeof(kmp_balign_t), "th_%d.th_bar[plain]", |
478 | gtid); |
479 | |
480 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier], |
481 | &thr->th.th_bar[bs_forkjoin_barrier + 1], |
482 | sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]", |
483 | gtid); |
484 | |
485 | #if KMP_FAST_REDUCTION_BARRIER1 |
486 | __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier], |
487 | &thr->th.th_bar[bs_reduction_barrier + 1], |
488 | sizeof(kmp_balign_t), "th_%d.th_bar[reduction]", |
489 | gtid); |
490 | #endif // KMP_FAST_REDUCTION_BARRIER |
491 | } |
492 | |
493 | /* Print out the storage map for the major kmp_team_t team data structures |
494 | that are allocated together. */ |
495 | |
496 | static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team, |
497 | int team_id, int num_thr) { |
498 | int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2; |
499 | __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d", |
500 | header, team_id); |
501 | |
502 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0], |
503 | &team->t.t_bar[bs_last_barrier], |
504 | sizeof(kmp_balign_team_t) * bs_last_barrier, |
505 | "%s_%d.t_bar", header, team_id); |
506 | |
507 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier], |
508 | &team->t.t_bar[bs_plain_barrier + 1], |
509 | sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]", |
510 | header, team_id); |
511 | |
512 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier], |
513 | &team->t.t_bar[bs_forkjoin_barrier + 1], |
514 | sizeof(kmp_balign_team_t), |
515 | "%s_%d.t_bar[forkjoin]", header, team_id); |
516 | |
517 | #if KMP_FAST_REDUCTION_BARRIER1 |
518 | __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier], |
519 | &team->t.t_bar[bs_reduction_barrier + 1], |
520 | sizeof(kmp_balign_team_t), |
521 | "%s_%d.t_bar[reduction]", header, team_id); |
522 | #endif // KMP_FAST_REDUCTION_BARRIER |
523 | |
524 | __kmp_print_storage_map_gtid( |
525 | -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr], |
526 | sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id); |
527 | |
528 | __kmp_print_storage_map_gtid( |
529 | -1, &team->t.t_threads[0], &team->t.t_threads[num_thr], |
530 | sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id); |
531 | |
532 | __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0], |
533 | &team->t.t_disp_buffer[num_disp_buff], |
534 | sizeof(dispatch_shared_info_t) * num_disp_buff, |
535 | "%s_%d.t_disp_buffer", header, team_id); |
536 | |
537 | __kmp_print_storage_map_gtid(-1, &team->t.t_taskq, &team->t.t_copypriv_data, |
538 | sizeof(kmp_taskq_t), "%s_%d.t_taskq", header, |
539 | team_id); |
540 | } |
541 | |
542 | static void __kmp_init_allocator() {} |
543 | static void __kmp_fini_allocator() {} |
544 | |
545 | /* ------------------------------------------------------------------------ */ |
546 | |
547 | #ifdef KMP_DYNAMIC_LIB1 |
548 | #if KMP_OS_WINDOWS0 |
549 | |
550 | static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) { |
551 | // TODO: Change to __kmp_break_bootstrap_lock(). |
552 | __kmp_init_bootstrap_lock(lck); // make the lock released |
553 | } |
554 | |
555 | static void __kmp_reset_locks_on_process_detach(int gtid_req) { |
556 | int i; |
557 | int thread_count; |
558 | |
559 | // PROCESS_DETACH is expected to be called by a thread that executes |
560 | // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one |
561 | // calling ProcessExit or FreeLibrary). So, it might be safe to access the |
562 | // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some |
563 | // threads can be still alive here, although being about to be terminated. The |
564 | // threads in the array with ds_thread==0 are most suspicious. Actually, it |
565 | // can be not safe to access the __kmp_threads[]. |
566 | |
567 | // TODO: does it make sense to check __kmp_roots[] ? |
568 | |
569 | // Let's check that there are no other alive threads registered with the OMP |
570 | // lib. |
571 | while (1) { |
572 | thread_count = 0; |
573 | for (i = 0; i < __kmp_threads_capacity; ++i) { |
574 | if (!__kmp_threads) |
575 | continue; |
576 | kmp_info_t *th = __kmp_threads[i]; |
577 | if (th == NULL__null) |
578 | continue; |
579 | int gtid = th->th.th_info.ds.ds_gtid; |
580 | if (gtid == gtid_req) |
581 | continue; |
582 | if (gtid < 0) |
583 | continue; |
584 | DWORD exit_val; |
585 | int alive = __kmp_is_thread_alive(th, &exit_val); |
586 | if (alive) { |
587 | ++thread_count; |
588 | } |
589 | } |
590 | if (thread_count == 0) |
591 | break; // success |
592 | } |
593 | |
594 | // Assume that I'm alone. Now it might be safe to check and reset locks. |
595 | // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset. |
596 | __kmp_reset_lock(&__kmp_forkjoin_lock); |
597 | #ifdef KMP_DEBUG1 |
598 | __kmp_reset_lock(&__kmp_stdio_lock); |
599 | #endif // KMP_DEBUG |
600 | } |
601 | |
602 | BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) { |
603 | //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock ); |
604 | |
605 | switch (fdwReason) { |
606 | |
607 | case DLL_PROCESS_ATTACH: |
608 | KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: PROCESS_ATTACH\n" ); }; |
609 | |
610 | return TRUE(!0); |
611 | |
612 | case DLL_PROCESS_DETACH: |
613 | KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: PROCESS_DETACH T#%d\n" , __kmp_gtid_get_specific()); }; |
614 | |
615 | if (lpReserved != NULL__null) { |
616 | // lpReserved is used for telling the difference: |
617 | // lpReserved == NULL when FreeLibrary() was called, |
618 | // lpReserved != NULL when the process terminates. |
619 | // When FreeLibrary() is called, worker threads remain alive. So they will |
620 | // release the forkjoin lock by themselves. When the process terminates, |
621 | // worker threads disappear triggering the problem of unreleased forkjoin |
622 | // lock as described below. |
623 | |
624 | // A worker thread can take the forkjoin lock. The problem comes up if |
625 | // that worker thread becomes dead before it releases the forkjoin lock. |
626 | // The forkjoin lock remains taken, while the thread executing |
627 | // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try |
628 | // to take the forkjoin lock and will always fail, so that the application |
629 | // will never finish [normally]. This scenario is possible if |
630 | // __kmpc_end() has not been executed. It looks like it's not a corner |
631 | // case, but common cases: |
632 | // - the main function was compiled by an alternative compiler; |
633 | // - the main function was compiled by icl but without /Qopenmp |
634 | // (application with plugins); |
635 | // - application terminates by calling C exit(), Fortran CALL EXIT() or |
636 | // Fortran STOP. |
637 | // - alive foreign thread prevented __kmpc_end from doing cleanup. |
638 | // |
639 | // This is a hack to work around the problem. |
640 | // TODO: !!! figure out something better. |
641 | __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific()); |
642 | } |
643 | |
644 | __kmp_internal_end_library(__kmp_gtid_get_specific()); |
645 | |
646 | return TRUE(!0); |
647 | |
648 | case DLL_THREAD_ATTACH: |
649 | KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: THREAD_ATTACH\n" ); }; |
650 | |
651 | /* if we want to register new siblings all the time here call |
652 | * __kmp_get_gtid(); */ |
653 | return TRUE(!0); |
654 | |
655 | case DLL_THREAD_DETACH: |
656 | KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()))if (kmp_a_debug >= 10) { __kmp_debug_printf ("DllMain: THREAD_DETACH T#%d\n" , __kmp_gtid_get_specific()); }; |
657 | |
658 | __kmp_internal_end_thread(__kmp_gtid_get_specific()); |
659 | return TRUE(!0); |
660 | } |
661 | |
662 | return TRUE(!0); |
663 | } |
664 | |
665 | #endif /* KMP_OS_WINDOWS */ |
666 | #endif /* KMP_DYNAMIC_LIB */ |
667 | |
668 | /* Change the library type to "status" and return the old type */ |
669 | /* called from within initialization routines where __kmp_initz_lock is held */ |
670 | int __kmp_change_library(int status) { |
671 | int old_status; |
672 | |
673 | old_status = __kmp_yield_init & |
674 | 1; // check whether KMP_LIBRARY=throughput (even init count) |
675 | |
676 | if (status) { |
677 | __kmp_yield_init |= 1; // throughput => turnaround (odd init count) |
678 | } else { |
679 | __kmp_yield_init &= ~1; // turnaround => throughput (even init count) |
680 | } |
681 | |
682 | return old_status; // return previous setting of whether |
683 | // KMP_LIBRARY=throughput |
684 | } |
685 | |
686 | /* __kmp_parallel_deo -- Wait until it's our turn. */ |
687 | void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { |
688 | int gtid = *gtid_ref; |
689 | #ifdef BUILD_PARALLEL_ORDERED1 |
690 | kmp_team_t *team = __kmp_team_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 690)), __kmp_threads[(gtid)]->th.th_team); |
691 | #endif /* BUILD_PARALLEL_ORDERED */ |
692 | |
693 | if (__kmp_env_consistency_check) { |
694 | if (__kmp_threads[gtid]->th.th_root->r.r_active) |
695 | #if KMP_USE_DYNAMIC_LOCK1 |
696 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null, 0); |
697 | #else |
698 | __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL__null); |
699 | #endif |
700 | } |
701 | #ifdef BUILD_PARALLEL_ORDERED1 |
702 | if (!team->t.t_serialized) { |
703 | KMP_MB(); |
704 | KMP_WAIT_YIELD__kmp_wait_yield_4(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 704)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), |
705 | KMP_EQ__kmp_eq_4, NULL__null); |
706 | KMP_MB(); |
707 | } |
708 | #endif /* BUILD_PARALLEL_ORDERED */ |
709 | } |
710 | |
711 | /* __kmp_parallel_dxo -- Signal the next task. */ |
712 | void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { |
713 | int gtid = *gtid_ref; |
714 | #ifdef BUILD_PARALLEL_ORDERED1 |
715 | int tid = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 715)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
716 | kmp_team_t *team = __kmp_team_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 716)), __kmp_threads[(gtid)]->th.th_team); |
717 | #endif /* BUILD_PARALLEL_ORDERED */ |
718 | |
719 | if (__kmp_env_consistency_check) { |
720 | if (__kmp_threads[gtid]->th.th_root->r.r_active) |
721 | __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref); |
722 | } |
723 | #ifdef BUILD_PARALLEL_ORDERED1 |
724 | if (!team->t.t_serialized) { |
725 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
726 | |
727 | /* use the tid of the next thread in this team */ |
728 | /* TODO replace with general release procedure */ |
729 | team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc); |
730 | |
731 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
732 | } |
733 | #endif /* BUILD_PARALLEL_ORDERED */ |
734 | } |
735 | |
736 | /* ------------------------------------------------------------------------ */ |
737 | /* The BARRIER for a SINGLE process section is always explicit */ |
738 | |
739 | int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) { |
740 | int status; |
741 | kmp_info_t *th; |
742 | kmp_team_t *team; |
743 | |
744 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) |
745 | __kmp_parallel_initialize(); |
746 | |
747 | th = __kmp_threads[gtid]; |
748 | team = th->th.th_team; |
749 | status = 0; |
750 | |
751 | th->th.th_ident = id_ref; |
752 | |
753 | if (team->t.t_serialized) { |
754 | status = 1; |
755 | } else { |
756 | kmp_int32 old_this = th->th.th_local.this_construct; |
757 | |
758 | ++th->th.th_local.this_construct; |
759 | /* try to set team count to thread count--success means thread got the |
760 | single block */ |
761 | /* TODO: Should this be acquire or release? */ |
762 | if (team->t.t_construct == old_this) { |
763 | status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this, |
764 | th->th.th_local.this_construct); |
765 | } |
766 | #if USE_ITT_BUILD1 |
767 | if (__itt_metadata_add_ptr__kmp_itt_metadata_add_ptr__3_0 && __kmp_forkjoin_frames_mode == 3 && |
768 | KMP_MASTER_GTID(gtid)((((((gtid)) >= 0) ? 0 : __kmp_debug_assert("((gtid)) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 768)), __kmp_threads[((gtid))]->th.th_info.ds.ds_tid) == 0) && |
769 | #if OMP_40_ENABLED(50 >= 40) |
770 | th->th.th_teams_microtask == NULL__null && |
771 | #endif |
772 | team->t.t_active_level == |
773 | 1) { // Only report metadata by master of active team at level 1 |
774 | __kmp_itt_metadata_single(id_ref); |
775 | } |
776 | #endif /* USE_ITT_BUILD */ |
777 | } |
778 | |
779 | if (__kmp_env_consistency_check) { |
780 | if (status && push_ws) { |
781 | __kmp_push_workshare(gtid, ct_psingle, id_ref); |
782 | } else { |
783 | __kmp_check_workshare(gtid, ct_psingle, id_ref); |
784 | } |
785 | } |
786 | #if USE_ITT_BUILD1 |
787 | if (status) { |
788 | __kmp_itt_single_start(gtid); |
789 | } |
790 | #endif /* USE_ITT_BUILD */ |
791 | return status; |
792 | } |
793 | |
794 | void __kmp_exit_single(int gtid) { |
795 | #if USE_ITT_BUILD1 |
796 | __kmp_itt_single_end(gtid); |
797 | #endif /* USE_ITT_BUILD */ |
798 | if (__kmp_env_consistency_check) |
799 | __kmp_pop_workshare(gtid, ct_psingle, NULL__null); |
800 | } |
801 | |
802 | /* determine if we can go parallel or must use a serialized parallel region and |
803 | * how many threads we can use |
804 | * set_nproc is the number of threads requested for the team |
805 | * returns 0 if we should serialize or only use one thread, |
806 | * otherwise the number of threads to use |
807 | * The forkjoin lock is held by the caller. */ |
808 | static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team, |
809 | int master_tid, int set_nthreads |
810 | #if OMP_40_ENABLED(50 >= 40) |
811 | , |
812 | int enter_teams |
813 | #endif /* OMP_40_ENABLED */ |
814 | ) { |
815 | int capacity; |
816 | int new_nthreads; |
817 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 817)); |
818 | KMP_DEBUG_ASSERT(root && parent_team)((root && parent_team) ? 0 : __kmp_debug_assert("root && parent_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 818)); |
819 | |
820 | // If dyn-var is set, dynamically adjust the number of desired threads, |
821 | // according to the method specified by dynamic_mode. |
822 | new_nthreads = set_nthreads; |
823 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic)) { |
824 | ; |
825 | } |
826 | #ifdef USE_LOAD_BALANCE1 |
827 | else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) { |
828 | new_nthreads = __kmp_load_balance_nproc(root, set_nthreads); |
829 | if (new_nthreads == 1) { |
830 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); } |
831 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); } |
832 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to 1 thread\n", master_tid); }; |
833 | return 1; |
834 | } |
835 | if (new_nthreads < set_nthreads) { |
836 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
837 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
838 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d load balance reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; |
839 | } |
840 | } |
841 | #endif /* USE_LOAD_BALANCE */ |
842 | else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) { |
843 | new_nthreads = __kmp_avail_proc - __kmp_nth + |
844 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); |
845 | if (new_nthreads <= 1) { |
846 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); } |
847 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); } |
848 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to 1 thread\n", master_tid); }; |
849 | return 1; |
850 | } |
851 | if (new_nthreads < set_nthreads) { |
852 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
853 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
854 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d thread limit reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; |
855 | } else { |
856 | new_nthreads = set_nthreads; |
857 | } |
858 | } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) { |
859 | if (set_nthreads > 2) { |
860 | new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]); |
861 | new_nthreads = (new_nthreads % set_nthreads) + 1; |
862 | if (new_nthreads == 1) { |
863 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); } |
864 | "reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); } |
865 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to 1 thread\n", master_tid); }; |
866 | return 1; |
867 | } |
868 | if (new_nthreads < set_nthreads) { |
869 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
870 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); } |
871 | master_tid, new_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d dynamic random reduced " "reservation to %d threads\n", master_tid, new_nthreads); }; |
872 | } |
873 | } |
874 | } else { |
875 | KMP_ASSERT(0)((0) ? 0 : __kmp_debug_assert("0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 875)); |
876 | } |
877 | |
878 | // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT. |
879 | if (__kmp_nth + new_nthreads - |
880 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > |
881 | __kmp_max_nth) { |
882 | int tl_nthreads = __kmp_max_nth - __kmp_nth + |
883 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); |
884 | if (tl_nthreads <= 0) { |
885 | tl_nthreads = 1; |
886 | } |
887 | |
888 | // If dyn-var is false, emit a 1-time warning. |
889 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { |
890 | __kmp_reserve_warn = 1; |
891 | __kmp_msg(kmp_ms_warning, |
892 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), |
893 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); |
894 | } |
895 | if (tl_nthreads == 1) { |
896 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } |
897 | "reduced reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } |
898 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); }; |
899 | return 1; |
900 | } |
901 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } |
902 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } |
903 | master_tid, tl_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); }; |
904 | new_nthreads = tl_nthreads; |
905 | } |
906 | |
907 | // Respect OMP_THREAD_LIMIT |
908 | if (root->r.r_cg_nthreads + new_nthreads - |
909 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > |
910 | __kmp_cg_max_nth) { |
911 | int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads + |
912 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); |
913 | if (tl_nthreads <= 0) { |
914 | tl_nthreads = 1; |
915 | } |
916 | |
917 | // If dyn-var is false, emit a 1-time warning. |
918 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { |
919 | __kmp_reserve_warn = 1; |
920 | __kmp_msg(kmp_ms_warning, |
921 | KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, tl_nthreads ), |
922 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); |
923 | } |
924 | if (tl_nthreads == 1) { |
925 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } |
926 | "reduced reservation to 1 thread\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); } |
927 | master_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " "reduced reservation to 1 thread\n", master_tid); }; |
928 | return 1; |
929 | } |
930 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } |
931 | "reservation to %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); } |
932 | master_tid, tl_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " "reservation to %d threads\n", master_tid, tl_nthreads); }; |
933 | new_nthreads = tl_nthreads; |
934 | } |
935 | |
936 | // Check if the threads array is large enough, or needs expanding. |
937 | // See comment in __kmp_register_root() about the adjustment if |
938 | // __kmp_threads[0] == NULL. |
939 | capacity = __kmp_threads_capacity; |
940 | if (TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { |
941 | --capacity; |
942 | } |
943 | if (__kmp_nth + new_nthreads - |
944 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > |
945 | capacity) { |
946 | // Expand the threads array. |
947 | int slotsRequired = __kmp_nth + new_nthreads - |
948 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) - |
949 | capacity; |
950 | int slotsAdded = __kmp_expand_threads(slotsRequired); |
951 | if (slotsAdded < slotsRequired) { |
952 | // The threads array was not expanded enough. |
953 | new_nthreads -= (slotsRequired - slotsAdded); |
954 | KMP_ASSERT(new_nthreads >= 1)((new_nthreads >= 1) ? 0 : __kmp_debug_assert("new_nthreads >= 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 954)); |
955 | |
956 | // If dyn-var is false, emit a 1-time warning. |
957 | if (!get__dynamic_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.dynamic) && (!__kmp_reserve_warn)) { |
958 | __kmp_reserve_warn = 1; |
959 | if (__kmp_tp_cached) { |
960 | __kmp_msg(kmp_ms_warning, |
961 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), |
962 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), |
963 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); |
964 | } else { |
965 | __kmp_msg(kmp_ms_warning, |
966 | KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, set_nthreads, new_nthreads ), |
967 | KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), __kmp_msg_null); |
968 | } |
969 | } |
970 | } |
971 | } |
972 | |
973 | #ifdef KMP_DEBUG1 |
974 | if (new_nthreads == 1) { |
975 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } |
976 | ("__kmp_reserve_threads: T#%d serializing team after reclaiming "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } |
977 | "dead roots and rechecking; requested %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); } |
978 | __kmp_get_gtid(), set_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d serializing team after reclaiming " "dead roots and rechecking; requested %d threads\n", __kmp_get_global_thread_id (), set_nthreads); }; |
979 | } else { |
980 | KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); } |
981 | " %d threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); } |
982 | __kmp_get_gtid(), new_nthreads, set_nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_reserve_threads: T#%d allocating %d threads; requested" " %d threads\n", __kmp_get_global_thread_id(), new_nthreads, set_nthreads); }; |
983 | } |
984 | #endif // KMP_DEBUG |
985 | return new_nthreads; |
986 | } |
987 | |
988 | /* Allocate threads from the thread pool and assign them to the new team. We are |
989 | assured that there are enough threads available, because we checked on that |
990 | earlier within critical section forkjoin */ |
991 | static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team, |
992 | kmp_info_t *master_th, int master_gtid) { |
993 | int i; |
994 | int use_hot_team; |
995 | |
996 | KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_fork_team_threads: new_nprocs = %d\n" , team->t.t_nproc); }; |
997 | KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid())((master_gtid == __kmp_get_global_thread_id()) ? 0 : __kmp_debug_assert ("master_gtid == __kmp_get_global_thread_id()", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 997)); |
998 | KMP_MB(); |
999 | |
1000 | /* first, let's setup the master thread */ |
1001 | master_th->th.th_info.ds.ds_tid = 0; |
1002 | master_th->th.th_team = team; |
1003 | master_th->th.th_team_nproc = team->t.t_nproc; |
1004 | master_th->th.th_team_master = master_th; |
1005 | master_th->th.th_team_serialized = FALSE0; |
1006 | master_th->th.th_dispatch = &team->t.t_dispatch[0]; |
1007 | |
1008 | /* make sure we are not the optimized hot team */ |
1009 | #if KMP_NESTED_HOT_TEAMS1 |
1010 | use_hot_team = 0; |
1011 | kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams; |
1012 | if (hot_teams) { // hot teams array is not allocated if |
1013 | // KMP_HOT_TEAMS_MAX_LEVEL=0 |
1014 | int level = team->t.t_active_level - 1; // index in array of hot teams |
1015 | if (master_th->th.th_teams_microtask) { // are we inside the teams? |
1016 | if (master_th->th.th_teams_size.nteams > 1) { |
1017 | ++level; // level was not increased in teams construct for |
1018 | // team_of_masters |
1019 | } |
1020 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && |
1021 | master_th->th.th_teams_level == team->t.t_level) { |
1022 | ++level; // level was not increased in teams construct for |
1023 | // team_of_workers before the parallel |
1024 | } // team->t.t_level will be increased inside parallel |
1025 | } |
1026 | if (level < __kmp_hot_teams_max_level) { |
1027 | if (hot_teams[level].hot_team) { |
1028 | // hot team has already been allocated for given level |
1029 | KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team)((hot_teams[level].hot_team == team) ? 0 : __kmp_debug_assert ("hot_teams[level].hot_team == team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1029)); |
1030 | use_hot_team = 1; // the team is ready to use |
1031 | } else { |
1032 | use_hot_team = 0; // AC: threads are not allocated yet |
1033 | hot_teams[level].hot_team = team; // remember new hot team |
1034 | hot_teams[level].hot_team_nth = team->t.t_nproc; |
1035 | } |
1036 | } else { |
1037 | use_hot_team = 0; |
1038 | } |
1039 | } |
1040 | #else |
1041 | use_hot_team = team == root->r.r_hot_team; |
1042 | #endif |
1043 | if (!use_hot_team) { |
1044 | |
1045 | /* install the master thread */ |
1046 | team->t.t_threads[0] = master_th; |
1047 | __kmp_initialize_info(master_th, team, 0, master_gtid); |
1048 | |
1049 | /* now, install the worker threads */ |
1050 | for (i = 1; i < team->t.t_nproc; i++) { |
1051 | |
1052 | /* fork or reallocate a new thread and install it in team */ |
1053 | kmp_info_t *thr = __kmp_allocate_thread(root, team, i); |
1054 | team->t.t_threads[i] = thr; |
1055 | KMP_DEBUG_ASSERT(thr)((thr) ? 0 : __kmp_debug_assert("thr", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1055)); |
1056 | KMP_DEBUG_ASSERT(thr->th.th_team == team)((thr->th.th_team == team) ? 0 : __kmp_debug_assert("thr->th.th_team == team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1056)); |
1057 | /* align team and thread arrived states */ |
1058 | KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
1059 | "T#%d(%d:%d) join =%llu, plain=%llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
1060 | __kmp_gtid_from_tid(0, team), team->t.t_id, 0,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
1061 | __kmp_gtid_from_tid(i, team), team->t.t_id, i,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
1062 | team->t.t_bar[bs_forkjoin_barrier].b_arrived,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
1063 | team->t.t_bar[bs_plain_barrier].b_arrived))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " "T#%d(%d:%d) join =%llu, plain=%llu\n", ((((0) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1060)), team->t.t_threads[(0)]->th.th_info.ds.ds_gtid ), team->t.t_id, 0, ((((i) >= 0 && (team) != __null ) ? 0 : __kmp_debug_assert("(i) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1061)), team->t.t_threads[(i)]->th.th_info.ds.ds_gtid ), team->t.t_id, i, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); }; |
1064 | #if OMP_40_ENABLED(50 >= 40) |
1065 | thr->th.th_teams_microtask = master_th->th.th_teams_microtask; |
1066 | thr->th.th_teams_level = master_th->th.th_teams_level; |
1067 | thr->th.th_teams_size = master_th->th.th_teams_size; |
1068 | #endif |
1069 | { // Initialize threads' barrier data. |
1070 | int b; |
1071 | kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar; |
1072 | for (b = 0; b < bs_last_barrier; ++b) { |
1073 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; |
1074 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1074)); |
1075 | #if USE_DEBUGGER0 |
1076 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; |
1077 | #endif |
1078 | } |
1079 | } |
1080 | } |
1081 | |
1082 | #if OMP_40_ENABLED(50 >= 40) && KMP_AFFINITY_SUPPORTED1 |
1083 | __kmp_partition_places(team); |
1084 | #endif |
1085 | } |
1086 | |
1087 | KMP_MB(); |
1088 | } |
1089 | |
1090 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
1091 | // Propagate any changes to the floating point control registers out to the team |
1092 | // We try to avoid unnecessary writes to the relevant cache line in the team |
1093 | // structure, so we don't make changes unless they are needed. |
1094 | inline static void propagateFPControl(kmp_team_t *team) { |
1095 | if (__kmp_inherit_fp_control) { |
1096 | kmp_int16 x87_fpu_control_word; |
1097 | kmp_uint32 mxcsr; |
1098 | |
1099 | // Get master values of FPU control flags (both X87 and vector) |
1100 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); |
1101 | __kmp_store_mxcsr(&mxcsr); |
1102 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; |
1103 | |
1104 | // There is no point looking at t_fp_control_saved here. |
1105 | // If it is TRUE, we still have to update the values if they are different |
1106 | // from those we now have. If it is FALSE we didn't save anything yet, but |
1107 | // our objective is the same. We have to ensure that the values in the team |
1108 | // are the same as those we have. |
1109 | // So, this code achieves what we need whether or not t_fp_control_saved is |
1110 | // true. By checking whether the value needs updating we avoid unnecessary |
1111 | // writes that would put the cache-line into a written state, causing all |
1112 | // threads in the team to have to read it again. |
1113 | KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word)if ((team->t.t_x87_fpu_control_word) != (x87_fpu_control_word )) (team->t.t_x87_fpu_control_word) = (x87_fpu_control_word ); |
1114 | KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr)if ((team->t.t_mxcsr) != (mxcsr)) (team->t.t_mxcsr) = ( mxcsr); |
1115 | // Although we don't use this value, other code in the runtime wants to know |
1116 | // whether it should restore them. So we must ensure it is correct. |
1117 | KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE)if ((team->t.t_fp_control_saved) != ((!0))) (team->t.t_fp_control_saved ) = ((!0)); |
1118 | } else { |
1119 | // Similarly here. Don't write to this cache-line in the team structure |
1120 | // unless we have to. |
1121 | KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE)if ((team->t.t_fp_control_saved) != (0)) (team->t.t_fp_control_saved ) = (0); |
1122 | } |
1123 | } |
1124 | |
1125 | // Do the opposite, setting the hardware registers to the updated values from |
1126 | // the team. |
1127 | inline static void updateHWFPControl(kmp_team_t *team) { |
1128 | if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) { |
1129 | // Only reset the fp control regs if they have been changed in the team. |
1130 | // the parallel region that we are exiting. |
1131 | kmp_int16 x87_fpu_control_word; |
1132 | kmp_uint32 mxcsr; |
1133 | __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); |
1134 | __kmp_store_mxcsr(&mxcsr); |
1135 | mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; |
1136 | |
1137 | if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) { |
1138 | __kmp_clear_x87_fpu_status_word(); |
1139 | __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word); |
1140 | } |
1141 | |
1142 | if (team->t.t_mxcsr != mxcsr) { |
1143 | __kmp_load_mxcsr(&team->t.t_mxcsr)_mm_setcsr(*(&team->t.t_mxcsr)); |
1144 | } |
1145 | } |
1146 | } |
1147 | #else |
1148 | #define propagateFPControl(x) ((void)0) |
1149 | #define updateHWFPControl(x) ((void)0) |
1150 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
1151 | |
1152 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, |
1153 | int realloc); // forward declaration |
1154 | |
1155 | /* Run a parallel region that has been serialized, so runs only in a team of the |
1156 | single master thread. */ |
1157 | void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) { |
1158 | kmp_info_t *this_thr; |
1159 | kmp_team_t *serial_team; |
1160 | |
1161 | KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: called by T#%d\n" , global_tid); }; |
1162 | |
1163 | /* Skip all this code for autopar serialized loops since it results in |
1164 | unacceptable overhead */ |
1165 | if (loc != NULL__null && (loc->flags & KMP_IDENT_AUTOPAR0x08)) |
1166 | return; |
1167 | |
1168 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) |
1169 | __kmp_parallel_initialize(); |
1170 | |
1171 | this_thr = __kmp_threads[global_tid]; |
1172 | serial_team = this_thr->th.th_serial_team; |
1173 | |
1174 | /* utilize the serialized team held by this thread */ |
1175 | KMP_DEBUG_ASSERT(serial_team)((serial_team) ? 0 : __kmp_debug_assert("serial_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1175)); |
1176 | KMP_MB(); |
1177 | |
1178 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
1179 | KMP_DEBUG_ASSERT(((this_thr->th.th_task_team == this_thr->th.th_team-> t.t_task_team[this_thr->th.th_task_state]) ? 0 : __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1181)) |
1180 | this_thr->th.th_task_team ==((this_thr->th.th_task_team == this_thr->th.th_team-> t.t_task_team[this_thr->th.th_task_state]) ? 0 : __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1181)) |
1181 | this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state])((this_thr->th.th_task_team == this_thr->th.th_team-> t.t_task_team[this_thr->th.th_task_state]) ? 0 : __kmp_debug_assert ("this_thr->th.th_task_team == this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1181)); |
1182 | KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==((serial_team->t.t_task_team[this_thr->th.th_task_state ] == __null) ? 0 : __kmp_debug_assert("serial_team->t.t_task_team[this_thr->th.th_task_state] == __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1183)) |
1183 | NULL)((serial_team->t.t_task_team[this_thr->th.th_task_state ] == __null) ? 0 : __kmp_debug_assert("serial_team->t.t_task_team[this_thr->th.th_task_state] == __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1183)); |
1184 | KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); } |
1185 | "team %p, new task_team = NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); } |
1186 | global_tid, this_thr->th.th_task_team, this_thr->th.th_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " "team %p, new task_team = NULL\n", global_tid, this_thr-> th.th_task_team, this_thr->th.th_team); }; |
1187 | this_thr->th.th_task_team = NULL__null; |
1188 | } |
1189 | |
1190 | #if OMP_40_ENABLED(50 >= 40) |
1191 | kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind; |
1192 | if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { |
1193 | proc_bind = proc_bind_false; |
1194 | } else if (proc_bind == proc_bind_default) { |
1195 | // No proc_bind clause was specified, so use the current value |
1196 | // of proc-bind-var for this parallel region. |
1197 | proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind; |
1198 | } |
1199 | // Reset for next parallel region |
1200 | this_thr->th.th_set_proc_bind = proc_bind_default; |
1201 | #endif /* OMP_40_ENABLED */ |
1202 | |
1203 | #if OMPT_SUPPORT1 |
1204 | ompt_data_t ompt_parallel_data; |
1205 | ompt_parallel_data.ptr = NULL__null; |
1206 | ompt_data_t *implicit_task_data; |
1207 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid)__ompt_load_return_address(global_tid); |
1208 | if (ompt_enabled.enabled && |
1209 | this_thr->th.ompt_thread_info.state != omp_state_overhead) { |
1210 | |
1211 | ompt_task_info_t *parent_task_info; |
1212 | parent_task_info = OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info)); |
1213 | |
1214 | parent_task_info->frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); |
1215 | if (ompt_enabled.ompt_callback_parallel_begin) { |
1216 | int team_size = 1; |
1217 | |
1218 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( |
1219 | &(parent_task_info->task_data), &(parent_task_info->frame), |
1220 | &ompt_parallel_data, team_size, ompt_invoker_program, codeptr); |
1221 | } |
1222 | } |
1223 | #endif // OMPT_SUPPORT |
1224 | |
1225 | if (this_thr->th.th_team != serial_team) { |
1226 | // Nested level will be an index in the nested nthreads array |
1227 | int level = this_thr->th.th_team->t.t_level; |
1228 | |
1229 | if (serial_team->t.t_serialized) { |
1230 | /* this serial team was already used |
1231 | TODO increase performance by making this locks more specific */ |
1232 | kmp_team_t *new_team; |
1233 | |
1234 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
1235 | |
1236 | new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1, |
1237 | #if OMPT_SUPPORT1 |
1238 | ompt_parallel_data, |
1239 | #endif |
1240 | #if OMP_40_ENABLED(50 >= 40) |
1241 | proc_bind, |
1242 | #endif |
1243 | &this_thr->th.th_current_task->td_icvs, |
1244 | 0 USE_NESTED_HOT_ARG(NULL), __null); |
1245 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
1246 | KMP_ASSERT(new_team)((new_team) ? 0 : __kmp_debug_assert("new_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1246)); |
1247 | |
1248 | /* setup new serialized team and install it */ |
1249 | new_team->t.t_threads[0] = this_thr; |
1250 | new_team->t.t_parent = this_thr->th.th_team; |
1251 | serial_team = new_team; |
1252 | this_thr->th.th_serial_team = serial_team; |
1253 | |
1254 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } |
1255 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } |
1256 | ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); } |
1257 | global_tid, serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n" , global_tid, serial_team); }; |
1258 | |
1259 | /* TODO the above breaks the requirement that if we run out of resources, |
1260 | then we can still guarantee that serialized teams are ok, since we may |
1261 | need to allocate a new one */ |
1262 | } else { |
1263 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } |
1264 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } |
1265 | ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); } |
1266 | global_tid, serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n" , global_tid, serial_team); }; |
1267 | } |
1268 | |
1269 | /* we have to initialize this serial team */ |
1270 | KMP_DEBUG_ASSERT(serial_team->t.t_threads)((serial_team->t.t_threads) ? 0 : __kmp_debug_assert("serial_team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1270)); |
1271 | KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr)((serial_team->t.t_threads[0] == this_thr) ? 0 : __kmp_debug_assert ("serial_team->t.t_threads[0] == this_thr", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1271)); |
1272 | KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team)((this_thr->th.th_team != serial_team) ? 0 : __kmp_debug_assert ("this_thr->th.th_team != serial_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1272)); |
1273 | serial_team->t.t_ident = loc; |
1274 | serial_team->t.t_serialized = 1; |
1275 | serial_team->t.t_nproc = 1; |
1276 | serial_team->t.t_parent = this_thr->th.th_team; |
1277 | serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched; |
1278 | this_thr->th.th_team = serial_team; |
1279 | serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid; |
1280 | |
1281 | KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#d curtask=%p\n" , global_tid, this_thr->th.th_current_task); } |
1282 | this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#d curtask=%p\n" , global_tid, this_thr->th.th_current_task); }; |
1283 | KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1)((this_thr->th.th_current_task->td_flags.executing == 1 ) ? 0 : __kmp_debug_assert("this_thr->th.th_current_task->td_flags.executing == 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1283)); |
1284 | this_thr->th.th_current_task->td_flags.executing = 0; |
1285 | |
1286 | __kmp_push_current_task_to_thread(this_thr, serial_team, 0); |
1287 | |
1288 | /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an |
1289 | implicit task for each serialized task represented by |
1290 | team->t.t_serialized? */ |
1291 | copy_icvs(&this_thr->th.th_current_task->td_icvs, |
1292 | &this_thr->th.th_current_task->td_parent->td_icvs); |
1293 | |
1294 | // Thread value exists in the nested nthreads array for the next nested |
1295 | // level |
1296 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { |
1297 | this_thr->th.th_current_task->td_icvs.nproc = |
1298 | __kmp_nested_nth.nth[level + 1]; |
1299 | } |
1300 | |
1301 | #if OMP_40_ENABLED(50 >= 40) |
1302 | if (__kmp_nested_proc_bind.used && |
1303 | (level + 1 < __kmp_nested_proc_bind.used)) { |
1304 | this_thr->th.th_current_task->td_icvs.proc_bind = |
1305 | __kmp_nested_proc_bind.bind_types[level + 1]; |
1306 | } |
1307 | #endif /* OMP_40_ENABLED */ |
1308 | |
1309 | #if USE_DEBUGGER0 |
1310 | serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger. |
1311 | #endif |
1312 | this_thr->th.th_info.ds.ds_tid = 0; |
1313 | |
1314 | /* set thread cache values */ |
1315 | this_thr->th.th_team_nproc = 1; |
1316 | this_thr->th.th_team_master = this_thr; |
1317 | this_thr->th.th_team_serialized = 1; |
1318 | |
1319 | serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1; |
1320 | serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level; |
1321 | |
1322 | propagateFPControl(serial_team); |
1323 | |
1324 | /* check if we need to allocate dispatch buffers stack */ |
1325 | KMP_DEBUG_ASSERT(serial_team->t.t_dispatch)((serial_team->t.t_dispatch) ? 0 : __kmp_debug_assert("serial_team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1325)); |
1326 | if (!serial_team->t.t_dispatch->th_disp_buffer) { |
1327 | serial_team->t.t_dispatch->th_disp_buffer = |
1328 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1329) |
1329 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1329); |
1330 | } |
1331 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; |
1332 | |
1333 | KMP_MB(); |
1334 | |
1335 | } else { |
1336 | /* this serialized team is already being used, |
1337 | * that's fine, just add another nested level */ |
1338 | KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team)((this_thr->th.th_team == serial_team) ? 0 : __kmp_debug_assert ("this_thr->th.th_team == serial_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1338)); |
1339 | KMP_DEBUG_ASSERT(serial_team->t.t_threads)((serial_team->t.t_threads) ? 0 : __kmp_debug_assert("serial_team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1339)); |
1340 | KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr)((serial_team->t.t_threads[0] == this_thr) ? 0 : __kmp_debug_assert ("serial_team->t.t_threads[0] == this_thr", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1340)); |
1341 | ++serial_team->t.t_serialized; |
1342 | this_thr->th.th_team_serialized = serial_team->t.t_serialized; |
1343 | |
1344 | // Nested level will be an index in the nested nthreads array |
1345 | int level = this_thr->th.th_team->t.t_level; |
1346 | // Thread value exists in the nested nthreads array for the next nested |
1347 | // level |
1348 | if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { |
1349 | this_thr->th.th_current_task->td_icvs.nproc = |
1350 | __kmp_nested_nth.nth[level + 1]; |
1351 | } |
1352 | serial_team->t.t_level++; |
1353 | KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); } |
1354 | "of serial team %p to %d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); } |
1355 | global_tid, serial_team, serial_team->t.t_level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmpc_serialized_parallel: T#%d increasing nesting level " "of serial team %p to %d\n", global_tid, serial_team, serial_team ->t.t_level); }; |
1356 | |
1357 | /* allocate/push dispatch buffers stack */ |
1358 | KMP_DEBUG_ASSERT(serial_team->t.t_dispatch)((serial_team->t.t_dispatch) ? 0 : __kmp_debug_assert("serial_team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1358)); |
1359 | { |
1360 | dispatch_private_info_t *disp_buffer = |
1361 | (dispatch_private_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_private_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1362) |
1362 | sizeof(dispatch_private_info_t))___kmp_allocate((sizeof(dispatch_private_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1362); |
1363 | disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer; |
1364 | serial_team->t.t_dispatch->th_disp_buffer = disp_buffer; |
1365 | } |
1366 | this_thr->th.th_dispatch = serial_team->t.t_dispatch; |
1367 | |
1368 | KMP_MB(); |
1369 | } |
1370 | #if OMP_40_ENABLED(50 >= 40) |
1371 | KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq)if ((serial_team->t.t_cancel_request) != (cancel_noreq)) ( serial_team->t.t_cancel_request) = (cancel_noreq); |
1372 | #endif |
1373 | |
1374 | if (__kmp_env_consistency_check) |
1375 | __kmp_push_parallel(global_tid, NULL__null); |
1376 | #if OMPT_SUPPORT1 |
1377 | serial_team->t.ompt_team_info.master_return_address = codeptr; |
1378 | if (ompt_enabled.enabled && |
1379 | this_thr->th.ompt_thread_info.state != omp_state_overhead) { |
1380 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); |
1381 | |
1382 | ompt_lw_taskteam_t lw_taskteam; |
1383 | __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid, |
1384 | &ompt_parallel_data, codeptr); |
1385 | |
1386 | __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1); |
1387 | // don't use lw_taskteam after linking. content was swaped |
1388 | |
1389 | /* OMPT implicit task begin */ |
1390 | implicit_task_data = OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )); |
Value stored to 'implicit_task_data' is never read | |
1391 | if (ompt_enabled.ompt_callback_implicit_task) { |
1392 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1393 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr)(&(this_thr->th.th_team->t.ompt_team_info.parallel_data )), |
1394 | OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )), 1, __kmp_tid_from_gtid(global_tid)((((global_tid) >= 0) ? 0 : __kmp_debug_assert("(global_tid) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1394)), __kmp_threads[(global_tid)]->th.th_info.ds.ds_tid )); |
1395 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info)) |
1396 | ->thread_num = __kmp_tid_from_gtid(global_tid)((((global_tid) >= 0) ? 0 : __kmp_debug_assert("(global_tid) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1396)), __kmp_threads[(global_tid)]->th.th_info.ds.ds_tid ); |
1397 | } |
1398 | |
1399 | /* OMPT state */ |
1400 | this_thr->th.ompt_thread_info.state = omp_state_work_parallel; |
1401 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); |
1402 | } |
1403 | #endif |
1404 | } |
1405 | |
1406 | /* most of the work for a fork */ |
1407 | /* return true if we really went parallel, false if serialized */ |
1408 | int __kmp_fork_call(ident_t *loc, int gtid, |
1409 | enum fork_context_e call_context, // Intel, GNU, ... |
1410 | kmp_int32 argc, microtask_t microtask, launch_t invoker, |
1411 | /* TODO: revert workaround for Intel(R) 64 tracker #96 */ |
1412 | #if (KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) && KMP_OS_LINUX1 |
1413 | va_list *ap |
1414 | #else |
1415 | va_list ap |
1416 | #endif |
1417 | ) { |
1418 | void **argv; |
1419 | int i; |
1420 | int master_tid; |
1421 | int master_this_cons; |
1422 | kmp_team_t *team; |
1423 | kmp_team_t *parent_team; |
1424 | kmp_info_t *master_th; |
1425 | kmp_root_t *root; |
1426 | int nthreads; |
1427 | int master_active; |
1428 | int master_set_numthreads; |
1429 | int level; |
1430 | #if OMP_40_ENABLED(50 >= 40) |
1431 | int active_level; |
1432 | int teams_level; |
1433 | #endif |
1434 | #if KMP_NESTED_HOT_TEAMS1 |
1435 | kmp_hot_team_ptr_t **p_hot_teams; |
1436 | #endif |
1437 | { // KMP_TIME_BLOCK |
1438 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call)((void)0); |
1439 | KMP_COUNT_VALUE(OMP_PARALLEL_args, argc)((void)0); |
1440 | |
1441 | KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: enter T#%d\n" , gtid); }; |
1442 | if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL__null) { |
1443 | /* Some systems prefer the stack for the root thread(s) to start with */ |
1444 | /* some gap from the parent stack to prevent false sharing. */ |
1445 | void *dummy = KMP_ALLOCA(__kmp_stkpadding)__builtin_alloca (__kmp_stkpadding); |
1446 | /* These 2 lines below are so this does not get optimized out */ |
1447 | if (__kmp_stkpadding > KMP_MAX_STKPADDING(2 * 1024 * 1024)) |
1448 | __kmp_stkpadding += (short)((kmp_int64)dummy); |
1449 | } |
1450 | |
1451 | /* initialize if needed */ |
1452 | KMP_DEBUG_ASSERT(((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1453)) |
1453 | __kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1453)); // AC: potentially unsafe, not in sync with shutdown |
1454 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) |
1455 | __kmp_parallel_initialize(); |
1456 | |
1457 | /* setup current data */ |
1458 | master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with |
1459 | // shutdown |
1460 | parent_team = master_th->th.th_team; |
1461 | master_tid = master_th->th.th_info.ds.ds_tid; |
1462 | master_this_cons = master_th->th.th_local.this_construct; |
1463 | root = master_th->th.th_root; |
1464 | master_active = root->r.r_active; |
1465 | master_set_numthreads = master_th->th.th_set_nproc; |
1466 | |
1467 | #if OMPT_SUPPORT1 |
1468 | ompt_data_t ompt_parallel_data; |
1469 | ompt_parallel_data.ptr = NULL__null; |
1470 | ompt_data_t *parent_task_data; |
1471 | omp_frame_t *ompt_frame; |
1472 | ompt_data_t *implicit_task_data; |
1473 | void *return_address = NULL__null; |
1474 | |
1475 | if (ompt_enabled.enabled) { |
1476 | __ompt_get_task_info_internal(0, NULL__null, &parent_task_data, &ompt_frame, |
1477 | NULL__null, NULL__null); |
1478 | return_address = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); |
1479 | } |
1480 | #endif |
1481 | |
1482 | // Nested level will be an index in the nested nthreads array |
1483 | level = parent_team->t.t_level; |
1484 | // used to launch non-serial teams even if nested is not allowed |
1485 | active_level = parent_team->t.t_active_level; |
1486 | #if OMP_40_ENABLED(50 >= 40) |
1487 | // needed to check nesting inside the teams |
1488 | teams_level = master_th->th.th_teams_level; |
1489 | #endif |
1490 | #if KMP_NESTED_HOT_TEAMS1 |
1491 | p_hot_teams = &master_th->th.th_hot_teams; |
1492 | if (*p_hot_teams == NULL__null && __kmp_hot_teams_max_level > 0) { |
1493 | *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1494) |
1494 | sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level)___kmp_allocate((sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1494); |
1495 | (*p_hot_teams)[0].hot_team = root->r.r_hot_team; |
1496 | // it is either actual or not needed (when active_level > 0) |
1497 | (*p_hot_teams)[0].hot_team_nth = 1; |
1498 | } |
1499 | #endif |
1500 | |
1501 | #if OMPT_SUPPORT1 |
1502 | if (ompt_enabled.enabled) { |
1503 | if (ompt_enabled.ompt_callback_parallel_begin) { |
1504 | int team_size = master_set_numthreads |
1505 | ? master_set_numthreads |
1506 | : get__nproc_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); |
1507 | ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)ompt_callback_parallel_begin_callback( |
1508 | parent_task_data, ompt_frame, &ompt_parallel_data, team_size, |
1509 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime ), return_address); |
1510 | } |
1511 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
1512 | } |
1513 | #endif |
1514 | |
1515 | master_th->th.th_ident = loc; |
1516 | |
1517 | #if OMP_40_ENABLED(50 >= 40) |
1518 | if (master_th->th.th_teams_microtask && ap && |
1519 | microtask != (microtask_t)__kmp_teams_master && level == teams_level) { |
1520 | // AC: This is start of parallel that is nested inside teams construct. |
1521 | // The team is actual (hot), all workers are ready at the fork barrier. |
1522 | // No lock needed to initialize the team a bit, then free workers. |
1523 | parent_team->t.t_ident = loc; |
1524 | __kmp_alloc_argv_entries(argc, parent_team, TRUE(!0)); |
1525 | parent_team->t.t_argc = argc; |
1526 | argv = (void **)parent_team->t.t_argv; |
1527 | for (i = argc - 1; i >= 0; --i) |
1528 | /* TODO: revert workaround for Intel(R) 64 tracker #96 */ |
1529 | #if (KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) && KMP_OS_LINUX1 |
1530 | *argv++ = va_arg(*ap, void *)__builtin_va_arg(*ap, void *); |
1531 | #else |
1532 | *argv++ = va_arg(ap, void *)__builtin_va_arg(ap, void *); |
1533 | #endif |
1534 | // Increment our nested depth levels, but not increase the serialization |
1535 | if (parent_team == master_th->th.th_serial_team) { |
1536 | // AC: we are in serialized parallel |
1537 | __kmpc_serialized_parallel(loc, gtid); |
1538 | KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1)((parent_team->t.t_serialized > 1) ? 0 : __kmp_debug_assert ("parent_team->t.t_serialized > 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1538)); |
1539 | // AC: need this in order enquiry functions work |
1540 | // correctly, will restore at join time |
1541 | parent_team->t.t_serialized--; |
1542 | #if OMPT_SUPPORT1 |
1543 | void *dummy; |
1544 | void **exit_runtime_p; |
1545 | |
1546 | ompt_lw_taskteam_t lw_taskteam; |
1547 | |
1548 | if (ompt_enabled.enabled) { |
1549 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, |
1550 | &ompt_parallel_data, return_address); |
1551 | exit_runtime_p = &(lw_taskteam.ompt_task_info.frame.exit_frame); |
1552 | |
1553 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); |
1554 | // don't use lw_taskteam after linking. content was swaped |
1555 | |
1556 | /* OMPT implicit task begin */ |
1557 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); |
1558 | if (ompt_enabled.ompt_callback_implicit_task) { |
1559 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1560 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), |
1561 | implicit_task_data, 1, __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1561)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)); |
1562 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)) |
1563 | ->thread_num = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1563)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
1564 | } |
1565 | |
1566 | /* OMPT state */ |
1567 | master_th->th.ompt_thread_info.state = omp_state_work_parallel; |
1568 | } else { |
1569 | exit_runtime_p = &dummy; |
1570 | } |
1571 | #endif |
1572 | |
1573 | { |
1574 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
1575 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
1576 | __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv |
1577 | #if OMPT_SUPPORT1 |
1578 | , |
1579 | exit_runtime_p |
1580 | #endif |
1581 | ); |
1582 | } |
1583 | |
1584 | #if OMPT_SUPPORT1 |
1585 | *exit_runtime_p = NULL__null; |
1586 | if (ompt_enabled.enabled) { |
1587 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->frame.exit_frame = NULL__null; |
1588 | if (ompt_enabled.ompt_callback_implicit_task) { |
1589 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1590 | ompt_scope_end, NULL__null, implicit_task_data, 1, |
1591 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num); |
1592 | } |
1593 | __ompt_lw_taskteam_unlink(master_th); |
1594 | |
1595 | if (ompt_enabled.ompt_callback_parallel_end) { |
1596 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( |
1597 | OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )), |
1598 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime ), return_address); |
1599 | } |
1600 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
1601 | } |
1602 | #endif |
1603 | return TRUE(!0); |
1604 | } |
1605 | |
1606 | parent_team->t.t_pkfn = microtask; |
1607 | parent_team->t.t_invoke = invoker; |
1608 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); |
1609 | parent_team->t.t_active_level++; |
1610 | parent_team->t.t_level++; |
1611 | |
1612 | /* Change number of threads in the team if requested */ |
1613 | if (master_set_numthreads) { // The parallel has num_threads clause |
1614 | if (master_set_numthreads < master_th->th.th_teams_size.nth) { |
1615 | // AC: only can reduce number of threads dynamically, can't increase |
1616 | kmp_info_t **other_threads = parent_team->t.t_threads; |
1617 | parent_team->t.t_nproc = master_set_numthreads; |
1618 | for (i = 0; i < master_set_numthreads; ++i) { |
1619 | other_threads[i]->th.th_team_nproc = master_set_numthreads; |
1620 | } |
1621 | // Keep extra threads hot in the team for possible next parallels |
1622 | } |
1623 | master_th->th.th_set_nproc = 0; |
1624 | } |
1625 | |
1626 | #if USE_DEBUGGER0 |
1627 | if (__kmp_debugging) { // Let debugger override number of threads. |
1628 | int nth = __kmp_omp_num_threads(loc); |
1629 | if (nth > 0) { // 0 means debugger doesn't want to change num threads |
1630 | master_set_numthreads = nth; |
1631 | } |
1632 | } |
1633 | #endif |
1634 | |
1635 | KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } |
1636 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } |
1637 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; |
1638 | __kmp_internal_fork(loc, gtid, parent_team); |
1639 | KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } |
1640 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); } |
1641 | root, parent_team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after internal fork: root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, parent_team, master_th, gtid ); }; |
1642 | |
1643 | /* Invoke microtask for MASTER thread */ |
1644 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); } |
1645 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; |
1646 | |
1647 | { |
1648 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
1649 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
1650 | if (!parent_team->t.t_invoke(gtid)) { |
1651 | KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread")((0) ? 0 : __kmp_debug_assert(("cannot invoke microtask for MASTER thread" ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1651)); |
1652 | } |
1653 | } |
1654 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); } |
1655 | parent_team->t.t_id, parent_team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, parent_team->t.t_id, parent_team->t.t_pkfn); }; |
1656 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
1657 | |
1658 | KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: parallel exit T#%d\n" , gtid); }; |
1659 | |
1660 | return TRUE(!0); |
1661 | } // Parallel closely nested in teams construct |
1662 | #endif /* OMP_40_ENABLED */ |
1663 | |
1664 | #if KMP_DEBUG1 |
1665 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
1666 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==((master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1667)) |
1667 | parent_team->t.t_task_team[master_th->th.th_task_state])((master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1667)); |
1668 | } |
1669 | #endif |
1670 | |
1671 | if (parent_team->t.t_active_level >= |
1672 | master_th->th.th_current_task->td_icvs.max_active_levels) { |
1673 | nthreads = 1; |
1674 | } else { |
1675 | #if OMP_40_ENABLED(50 >= 40) |
1676 | int enter_teams = ((ap == NULL__null && active_level == 0) || |
1677 | (ap && teams_level > 0 && teams_level == level)); |
1678 | #endif |
1679 | nthreads = |
1680 | master_set_numthreads |
1681 | ? master_set_numthreads |
1682 | : get__nproc_2(((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc) |
1683 | parent_team,((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc) |
1684 | master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.nproc); // TODO: get nproc directly from current task |
1685 | |
1686 | // Check if we need to take forkjoin lock? (no need for serialized |
1687 | // parallel out of teams construct). This code moved here from |
1688 | // __kmp_reserve_threads() to speedup nested serialized parallels. |
1689 | if (nthreads > 1) { |
1690 | if ((!get__nested(master_th)(((master_th)->th.th_current_task->td_icvs.nested) ? (( !0)) : (0)) && (root->r.r_in_parallel |
1691 | #if OMP_40_ENABLED(50 >= 40) |
1692 | && !enter_teams |
1693 | #endif /* OMP_40_ENABLED */ |
1694 | )) || |
1695 | (__kmp_library == library_serial)) { |
1696 | KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); } |
1697 | " threads\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); } |
1698 | gtid, nthreads))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing team; requested %d" " threads\n", gtid, nthreads); }; |
1699 | nthreads = 1; |
1700 | } |
1701 | } |
1702 | if (nthreads > 1) { |
1703 | /* determine how many new threads we can use */ |
1704 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
1705 | nthreads = __kmp_reserve_threads( |
1706 | root, parent_team, master_tid, nthreads |
1707 | #if OMP_40_ENABLED(50 >= 40) |
1708 | /* AC: If we execute teams from parallel region (on host), then |
1709 | teams should be created but each can only have 1 thread if |
1710 | nesting is disabled. If teams called from serial region, then |
1711 | teams and their threads should be created regardless of the |
1712 | nesting setting. */ |
1713 | , |
1714 | enter_teams |
1715 | #endif /* OMP_40_ENABLED */ |
1716 | ); |
1717 | if (nthreads == 1) { |
1718 | // Free lock for single thread execution here; for multi-thread |
1719 | // execution it will be freed later after team of threads created |
1720 | // and initialized |
1721 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
1722 | } |
1723 | } |
1724 | } |
1725 | KMP_DEBUG_ASSERT(nthreads > 0)((nthreads > 0) ? 0 : __kmp_debug_assert("nthreads > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1725)); |
1726 | |
1727 | // If we temporarily changed the set number of threads then restore it now |
1728 | master_th->th.th_set_nproc = 0; |
1729 | |
1730 | /* create a serialized parallel region? */ |
1731 | if (nthreads == 1) { |
1732 | /* josh todo: hypothetical question: what do we do for OS X*? */ |
1733 | #if KMP_OS_LINUX1 && \ |
1734 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) |
1735 | void *args[argc]; |
1736 | #else |
1737 | void **args = (void **)KMP_ALLOCA(argc * sizeof(void *))__builtin_alloca (argc * sizeof(void *)); |
1738 | #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \ |
1739 | KMP_ARCH_AARCH64) */ |
1740 | |
1741 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing parallel region\n" , gtid); } |
1742 | ("__kmp_fork_call: T#%d serializing parallel region\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serializing parallel region\n" , gtid); }; |
1743 | |
1744 | __kmpc_serialized_parallel(loc, gtid); |
1745 | |
1746 | if (call_context == fork_context_intel) { |
1747 | /* TODO this sucks, use the compiler itself to pass args! :) */ |
1748 | master_th->th.th_serial_team->t.t_ident = loc; |
1749 | #if OMP_40_ENABLED(50 >= 40) |
1750 | if (!ap) { |
1751 | // revert change made in __kmpc_serialized_parallel() |
1752 | master_th->th.th_serial_team->t.t_level--; |
1753 | // Get args from parent team for teams construct |
1754 | |
1755 | #if OMPT_SUPPORT1 |
1756 | void *dummy; |
1757 | void **exit_runtime_p; |
1758 | ompt_task_info_t *task_info; |
1759 | |
1760 | ompt_lw_taskteam_t lw_taskteam; |
1761 | |
1762 | if (ompt_enabled.enabled) { |
1763 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, |
1764 | &ompt_parallel_data, return_address); |
1765 | |
1766 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); |
1767 | // don't use lw_taskteam after linking. content was swaped |
1768 | |
1769 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); |
1770 | exit_runtime_p = &(task_info->frame.exit_frame); |
1771 | if (ompt_enabled.ompt_callback_implicit_task) { |
1772 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1773 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), |
1774 | &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1774)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)); |
1775 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)) |
1776 | ->thread_num = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1776)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
1777 | } |
1778 | |
1779 | /* OMPT state */ |
1780 | master_th->th.ompt_thread_info.state = omp_state_work_parallel; |
1781 | } else { |
1782 | exit_runtime_p = &dummy; |
1783 | } |
1784 | #endif |
1785 | |
1786 | { |
1787 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
1788 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
1789 | __kmp_invoke_microtask(microtask, gtid, 0, argc, |
1790 | parent_team->t.t_argv |
1791 | #if OMPT_SUPPORT1 |
1792 | , |
1793 | exit_runtime_p |
1794 | #endif |
1795 | ); |
1796 | } |
1797 | |
1798 | #if OMPT_SUPPORT1 |
1799 | if (ompt_enabled.enabled) { |
1800 | exit_runtime_p = NULL__null; |
1801 | if (ompt_enabled.ompt_callback_implicit_task) { |
1802 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1803 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, |
1804 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num); |
1805 | } |
1806 | |
1807 | __ompt_lw_taskteam_unlink(master_th); |
1808 | if (ompt_enabled.ompt_callback_parallel_end) { |
1809 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( |
1810 | OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), parent_task_data, |
1811 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime ), return_address); |
1812 | } |
1813 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
1814 | } |
1815 | #endif |
1816 | } else if (microtask == (microtask_t)__kmp_teams_master) { |
1817 | KMP_DEBUG_ASSERT(master_th->th.th_team ==((master_th->th.th_team == master_th->th.th_serial_team ) ? 0 : __kmp_debug_assert("master_th->th.th_team == master_th->th.th_serial_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1818)) |
1818 | master_th->th.th_serial_team)((master_th->th.th_team == master_th->th.th_serial_team ) ? 0 : __kmp_debug_assert("master_th->th.th_team == master_th->th.th_serial_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1818)); |
1819 | team = master_th->th.th_team; |
1820 | // team->t.t_pkfn = microtask; |
1821 | team->t.t_invoke = invoker; |
1822 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); |
1823 | team->t.t_argc = argc; |
1824 | argv = (void **)team->t.t_argv; |
1825 | if (ap) { |
1826 | for (i = argc - 1; i >= 0; --i) |
1827 | // TODO: revert workaround for Intel(R) 64 tracker #96 |
1828 | #if (KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) && KMP_OS_LINUX1 |
1829 | *argv++ = va_arg(*ap, void *)__builtin_va_arg(*ap, void *); |
1830 | #else |
1831 | *argv++ = va_arg(ap, void *)__builtin_va_arg(ap, void *); |
1832 | #endif |
1833 | } else { |
1834 | for (i = 0; i < argc; ++i) |
1835 | // Get args from parent team for teams construct |
1836 | argv[i] = parent_team->t.t_argv[i]; |
1837 | } |
1838 | // AC: revert change made in __kmpc_serialized_parallel() |
1839 | // because initial code in teams should have level=0 |
1840 | team->t.t_level--; |
1841 | // AC: call special invoker for outer "parallel" of teams construct |
1842 | { |
1843 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
1844 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
1845 | invoker(gtid); |
1846 | } |
1847 | } else { |
1848 | #endif /* OMP_40_ENABLED */ |
1849 | argv = args; |
1850 | for (i = argc - 1; i >= 0; --i) |
1851 | // TODO: revert workaround for Intel(R) 64 tracker #96 |
1852 | #if (KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) && KMP_OS_LINUX1 |
1853 | *argv++ = va_arg(*ap, void *)__builtin_va_arg(*ap, void *); |
1854 | #else |
1855 | *argv++ = va_arg(ap, void *)__builtin_va_arg(ap, void *); |
1856 | #endif |
1857 | KMP_MB(); |
1858 | |
1859 | #if OMPT_SUPPORT1 |
1860 | void *dummy; |
1861 | void **exit_runtime_p; |
1862 | ompt_task_info_t *task_info; |
1863 | |
1864 | ompt_lw_taskteam_t lw_taskteam; |
1865 | |
1866 | if (ompt_enabled.enabled) { |
1867 | __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, |
1868 | &ompt_parallel_data, return_address); |
1869 | __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0); |
1870 | // don't use lw_taskteam after linking. content was swaped |
1871 | task_info = OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)); |
1872 | exit_runtime_p = &(task_info->frame.exit_frame); |
1873 | |
1874 | /* OMPT implicit task begin */ |
1875 | implicit_task_data = OMPT_CUR_TASK_DATA(master_th)(&(master_th->th.th_current_task->ompt_task_info.task_data )); |
1876 | if (ompt_enabled.ompt_callback_implicit_task) { |
1877 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1878 | ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )), |
1879 | implicit_task_data, 1, __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1879)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)); |
1880 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info)) |
1881 | ->thread_num = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1881)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
1882 | } |
1883 | |
1884 | /* OMPT state */ |
1885 | master_th->th.ompt_thread_info.state = omp_state_work_parallel; |
1886 | } else { |
1887 | exit_runtime_p = &dummy; |
1888 | } |
1889 | #endif |
1890 | |
1891 | { |
1892 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
1893 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
1894 | __kmp_invoke_microtask(microtask, gtid, 0, argc, args |
1895 | #if OMPT_SUPPORT1 |
1896 | , |
1897 | exit_runtime_p |
1898 | #endif |
1899 | ); |
1900 | } |
1901 | |
1902 | #if OMPT_SUPPORT1 |
1903 | if (ompt_enabled.enabled) { |
1904 | *exit_runtime_p = NULL__null; |
1905 | if (ompt_enabled.ompt_callback_implicit_task) { |
1906 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
1907 | ompt_scope_end, NULL__null, &(task_info->task_data), 1, |
1908 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num); |
1909 | } |
1910 | |
1911 | ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th)(&(master_th->th.th_team->t.ompt_team_info.parallel_data )); |
1912 | __ompt_lw_taskteam_unlink(master_th); |
1913 | if (ompt_enabled.ompt_callback_parallel_end) { |
1914 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( |
1915 | &ompt_parallel_data, parent_task_data, |
1916 | OMPT_INVOKER(call_context)((call_context == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime ), return_address); |
1917 | } |
1918 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
1919 | } |
1920 | #endif |
1921 | #if OMP_40_ENABLED(50 >= 40) |
1922 | } |
1923 | #endif /* OMP_40_ENABLED */ |
1924 | } else if (call_context == fork_context_gnu) { |
1925 | #if OMPT_SUPPORT1 |
1926 | ompt_lw_taskteam_t lwt; |
1927 | __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data, |
1928 | return_address); |
1929 | |
1930 | lwt.ompt_task_info.frame.exit_frame = NULL__null; |
1931 | __ompt_lw_taskteam_link(&lwt, master_th, 1); |
1932 | // don't use lw_taskteam after linking. content was swaped |
1933 | #endif |
1934 | |
1935 | // we were called from GNU native code |
1936 | KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serial exit\n" , gtid); }; |
1937 | return FALSE0; |
1938 | } else { |
1939 | KMP_ASSERT2(call_context < fork_context_last,((call_context < fork_context_last) ? 0 : __kmp_debug_assert (("__kmp_fork_call: unknown fork_context parameter"), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1940)) |
1940 | "__kmp_fork_call: unknown fork_context parameter")((call_context < fork_context_last) ? 0 : __kmp_debug_assert (("__kmp_fork_call: unknown fork_context parameter"), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 1940)); |
1941 | } |
1942 | |
1943 | KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d serial exit\n" , gtid); }; |
1944 | KMP_MB(); |
1945 | return FALSE0; |
1946 | } |
1947 | |
1948 | // GEH: only modify the executing flag in the case when not serialized |
1949 | // serialized case is handled in kmpc_serialized_parallel |
1950 | KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } |
1951 | "curtask=%p, curtask_max_aclevel=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } |
1952 | parent_team->t.t_active_level, master_th,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } |
1953 | master_th->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); } |
1954 | master_th->th.th_current_task->td_icvs.max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " "curtask=%p, curtask_max_aclevel=%d\n", parent_team->t.t_active_level , master_th, master_th->th.th_current_task, master_th-> th.th_current_task->td_icvs.max_active_levels); }; |
1955 | // TODO: GEH - cannot do this assertion because root thread not set up as |
1956 | // executing |
1957 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 ); |
1958 | master_th->th.th_current_task->td_flags.executing = 0; |
1959 | |
1960 | #if OMP_40_ENABLED(50 >= 40) |
1961 | if (!master_th->th.th_teams_microtask || level > teams_level) |
1962 | #endif /* OMP_40_ENABLED */ |
1963 | { |
1964 | /* Increment our nested depth level */ |
1965 | KMP_ATOMIC_INC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_add(1, std::memory_order_acq_rel ); |
1966 | } |
1967 | |
1968 | // See if we need to make a copy of the ICVs. |
1969 | int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc; |
1970 | if ((level + 1 < __kmp_nested_nth.used) && |
1971 | (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) { |
1972 | nthreads_icv = __kmp_nested_nth.nth[level + 1]; |
1973 | } else { |
1974 | nthreads_icv = 0; // don't update |
1975 | } |
1976 | |
1977 | #if OMP_40_ENABLED(50 >= 40) |
1978 | // Figure out the proc_bind_policy for the new team. |
1979 | kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; |
1980 | kmp_proc_bind_t proc_bind_icv = |
1981 | proc_bind_default; // proc_bind_default means don't update |
1982 | if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { |
1983 | proc_bind = proc_bind_false; |
1984 | } else { |
1985 | if (proc_bind == proc_bind_default) { |
1986 | // No proc_bind clause specified; use current proc-bind-var for this |
1987 | // parallel region |
1988 | proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; |
1989 | } |
1990 | /* else: The proc_bind policy was specified explicitly on parallel clause. |
1991 | This overrides proc-bind-var for this parallel region, but does not |
1992 | change proc-bind-var. */ |
1993 | // Figure the value of proc-bind-var for the child threads. |
1994 | if ((level + 1 < __kmp_nested_proc_bind.used) && |
1995 | (__kmp_nested_proc_bind.bind_types[level + 1] != |
1996 | master_th->th.th_current_task->td_icvs.proc_bind)) { |
1997 | proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; |
1998 | } |
1999 | } |
2000 | |
2001 | // Reset for next parallel region |
2002 | master_th->th.th_set_proc_bind = proc_bind_default; |
2003 | #endif /* OMP_40_ENABLED */ |
2004 | |
2005 | if ((nthreads_icv > 0) |
2006 | #if OMP_40_ENABLED(50 >= 40) |
2007 | || (proc_bind_icv != proc_bind_default) |
2008 | #endif /* OMP_40_ENABLED */ |
2009 | ) { |
2010 | kmp_internal_control_t new_icvs; |
2011 | copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs); |
2012 | new_icvs.next = NULL__null; |
2013 | if (nthreads_icv > 0) { |
2014 | new_icvs.nproc = nthreads_icv; |
2015 | } |
2016 | |
2017 | #if OMP_40_ENABLED(50 >= 40) |
2018 | if (proc_bind_icv != proc_bind_default) { |
2019 | new_icvs.proc_bind = proc_bind_icv; |
2020 | } |
2021 | #endif /* OMP_40_ENABLED */ |
2022 | |
2023 | /* allocate a new parallel team */ |
2024 | KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before __kmp_allocate_team\n" ); }; |
2025 | team = __kmp_allocate_team(root, nthreads, nthreads, |
2026 | #if OMPT_SUPPORT1 |
2027 | ompt_parallel_data, |
2028 | #endif |
2029 | #if OMP_40_ENABLED(50 >= 40) |
2030 | proc_bind, |
2031 | #endif |
2032 | &new_icvs, argc USE_NESTED_HOT_ARG(master_th), master_th); |
2033 | } else { |
2034 | /* allocate a new parallel team */ |
2035 | KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: before __kmp_allocate_team\n" ); }; |
2036 | team = __kmp_allocate_team(root, nthreads, nthreads, |
2037 | #if OMPT_SUPPORT1 |
2038 | ompt_parallel_data, |
2039 | #endif |
2040 | #if OMP_40_ENABLED(50 >= 40) |
2041 | proc_bind, |
2042 | #endif |
2043 | &master_th->th.th_current_task->td_icvs, |
2044 | argc USE_NESTED_HOT_ARG(master_th), master_th); |
2045 | } |
2046 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n" , team); } |
2047 | 10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n" , team); }; |
2048 | |
2049 | /* setup the new team */ |
2050 | KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid)if ((team->t.t_master_tid) != (master_tid)) (team->t.t_master_tid ) = (master_tid); |
2051 | KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons)if ((team->t.t_master_this_cons) != (master_this_cons)) (team ->t.t_master_this_cons) = (master_this_cons); |
2052 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); |
2053 | KMP_CHECK_UPDATE(team->t.t_parent, parent_team)if ((team->t.t_parent) != (parent_team)) (team->t.t_parent ) = (parent_team); |
2054 | KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask)if ((team->t.t_pkfn) != (microtask)) (((team->t.t_pkfn) )) = (((microtask))); |
2055 | #if OMPT_SUPPORT1 |
2056 | KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,if ((team->t.ompt_team_info.master_return_address) != (return_address )) (((team->t.ompt_team_info.master_return_address))) = (( (return_address))) |
2057 | return_address)if ((team->t.ompt_team_info.master_return_address) != (return_address )) (((team->t.ompt_team_info.master_return_address))) = (( (return_address))); |
2058 | #endif |
2059 | KMP_CHECK_UPDATE(team->t.t_invoke, invoker)if ((team->t.t_invoke) != (invoker)) (team->t.t_invoke) = (invoker); // TODO move to root, maybe |
2060 | // TODO: parent_team->t.t_level == INT_MAX ??? |
2061 | #if OMP_40_ENABLED(50 >= 40) |
2062 | if (!master_th->th.th_teams_microtask || level > teams_level) { |
2063 | #endif /* OMP_40_ENABLED */ |
2064 | int new_level = parent_team->t.t_level + 1; |
2065 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); |
2066 | new_level = parent_team->t.t_active_level + 1; |
2067 | KMP_CHECK_UPDATE(team->t.t_active_level, new_level)if ((team->t.t_active_level) != (new_level)) (team->t.t_active_level ) = (new_level); |
2068 | #if OMP_40_ENABLED(50 >= 40) |
2069 | } else { |
2070 | // AC: Do not increase parallel level at start of the teams construct |
2071 | int new_level = parent_team->t.t_level; |
2072 | KMP_CHECK_UPDATE(team->t.t_level, new_level)if ((team->t.t_level) != (new_level)) (team->t.t_level) = (new_level); |
2073 | new_level = parent_team->t.t_active_level; |
2074 | KMP_CHECK_UPDATE(team->t.t_active_level, new_level)if ((team->t.t_active_level) != (new_level)) (team->t.t_active_level ) = (new_level); |
2075 | } |
2076 | #endif /* OMP_40_ENABLED */ |
2077 | kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid)((parent_team)->t.t_threads[(master_tid)]->th.th_current_task ->td_icvs.sched); |
2078 | // set master's schedule as new run-time schedule |
2079 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched)if ((team->t.t_sched.sched) != (new_sched.sched)) (team-> t.t_sched.sched) = (new_sched.sched); |
2080 | |
2081 | #if OMP_40_ENABLED(50 >= 40) |
2082 | KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq)if ((team->t.t_cancel_request) != (cancel_noreq)) (team-> t.t_cancel_request) = (cancel_noreq); |
2083 | #endif |
2084 | |
2085 | // Update the floating point rounding in the team if required. |
2086 | propagateFPControl(team); |
2087 | |
2088 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
2089 | // Set master's task team to team's task team. Unless this is hot team, it |
2090 | // should be NULL. |
2091 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==((master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2092)) |
2092 | parent_team->t.t_task_team[master_th->th.th_task_state])((master_th->th.th_task_team == parent_team->t.t_task_team [master_th->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == parent_team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2092)); |
2093 | KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Master T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", ((((master_th) != __null ) ? 0 : __kmp_debug_assert("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2095)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team, team->t.t_task_team[master_th ->th.th_task_state], team); } |
2094 | "%p, new task_team %p / team %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Master T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", ((((master_th) != __null ) ? 0 : __kmp_debug_assert("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2095)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team, team->t.t_task_team[master_th ->th.th_task_state], team); } |
2095 | __kmp_gtid_from_thread(master_th),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Master T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", ((((master_th) != __null ) ? 0 : __kmp_debug_assert("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2095)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team, team->t.t_task_team[master_th ->th.th_task_state], team); } |
2096 | master_th->th.th_task_team, parent_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Master T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", ((((master_th) != __null ) ? 0 : __kmp_debug_assert("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2095)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team, team->t.t_task_team[master_th ->th.th_task_state], team); } |
2097 | team->t.t_task_team[master_th->th.th_task_state], team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: Master T#%d pushing task_team %p / team " "%p, new task_team %p / team %p\n", ((((master_th) != __null ) ? 0 : __kmp_debug_assert("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2095)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team, team->t.t_task_team[master_th ->th.th_task_state], team); }; |
2098 | |
2099 | if (active_level || master_th->th.th_task_team) { |
2100 | // Take a memo of master's task_state |
2101 | KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack)((master_th->th.th_task_state_memo_stack) ? 0 : __kmp_debug_assert ("master_th->th.th_task_state_memo_stack", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2101)); |
2102 | if (master_th->th.th_task_state_top >= |
2103 | master_th->th.th_task_state_stack_sz) { // increase size |
2104 | kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz; |
2105 | kmp_uint8 *old_stack, *new_stack; |
2106 | kmp_uint32 i; |
2107 | new_stack = (kmp_uint8 *)__kmp_allocate(new_size)___kmp_allocate((new_size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2107); |
2108 | for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) { |
2109 | new_stack[i] = master_th->th.th_task_state_memo_stack[i]; |
2110 | } |
2111 | for (i = master_th->th.th_task_state_stack_sz; i < new_size; |
2112 | ++i) { // zero-init rest of stack |
2113 | new_stack[i] = 0; |
2114 | } |
2115 | old_stack = master_th->th.th_task_state_memo_stack; |
2116 | master_th->th.th_task_state_memo_stack = new_stack; |
2117 | master_th->th.th_task_state_stack_sz = new_size; |
2118 | __kmp_free(old_stack)___kmp_free((old_stack), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2118); |
2119 | } |
2120 | // Store master's task_state on stack |
2121 | master_th->th |
2122 | .th_task_state_memo_stack[master_th->th.th_task_state_top] = |
2123 | master_th->th.th_task_state; |
2124 | master_th->th.th_task_state_top++; |
2125 | #if KMP_NESTED_HOT_TEAMS1 |
2126 | if (team == master_th->th.th_hot_teams[active_level].hot_team) { |
2127 | // Restore master's nested state if nested hot team |
2128 | master_th->th.th_task_state = |
2129 | master_th->th |
2130 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; |
2131 | } else { |
2132 | #endif |
2133 | master_th->th.th_task_state = 0; |
2134 | #if KMP_NESTED_HOT_TEAMS1 |
2135 | } |
2136 | #endif |
2137 | } |
2138 | #if !KMP_NESTED_HOT_TEAMS1 |
2139 | KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||(((master_th->th.th_task_team == __null) || (team == root-> r.r_hot_team)) ? 0 : __kmp_debug_assert("(master_th->th.th_task_team == __null) || (team == root->r.r_hot_team)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2140)) |
2140 | (team == root->r.r_hot_team))(((master_th->th.th_task_team == __null) || (team == root-> r.r_hot_team)) ? 0 : __kmp_debug_assert("(master_th->th.th_task_team == __null) || (team == root->r.r_hot_team)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2140)); |
2141 | #endif |
2142 | } |
2143 | |
2144 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } |
2145 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } |
2146 | ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } |
2147 | gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); } |
2148 | team->t.t_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n" , gtid, parent_team->t.t_id, team->t.t_master_tid, team ->t.t_id, team->t.t_nproc); }; |
2149 | KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||((team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))) ? 0 : __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2152)) |
2150 | (team->t.t_master_tid == 0 &&((team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))) ? 0 : __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2152)) |
2151 | (team->t.t_parent == root->r.r_root_team ||((team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))) ? 0 : __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2152)) |
2152 | team->t.t_parent->t.t_serialized)))((team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))) ? 0 : __kmp_debug_assert ("team != root->r.r_hot_team || (team->t.t_master_tid == 0 && (team->t.t_parent == root->r.r_root_team || team->t.t_parent->t.t_serialized))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2152)); |
2153 | KMP_MB(); |
2154 | |
2155 | /* now, setup the arguments */ |
2156 | argv = (void **)team->t.t_argv; |
2157 | #if OMP_40_ENABLED(50 >= 40) |
2158 | if (ap) { |
2159 | #endif /* OMP_40_ENABLED */ |
2160 | for (i = argc - 1; i >= 0; --i) { |
2161 | // TODO: revert workaround for Intel(R) 64 tracker #96 |
2162 | #if (KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) && KMP_OS_LINUX1 |
2163 | void *new_argv = va_arg(*ap, void *)__builtin_va_arg(*ap, void *); |
2164 | #else |
2165 | void *new_argv = va_arg(ap, void *)__builtin_va_arg(ap, void *); |
2166 | #endif |
2167 | KMP_CHECK_UPDATE(*argv, new_argv)if ((*argv) != (new_argv)) (*argv) = (new_argv); |
2168 | argv++; |
2169 | } |
2170 | #if OMP_40_ENABLED(50 >= 40) |
2171 | } else { |
2172 | for (i = 0; i < argc; ++i) { |
2173 | // Get args from parent team for teams construct |
2174 | KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i])if ((argv[i]) != (team->t.t_parent->t.t_argv[i])) (argv [i]) = (team->t.t_parent->t.t_argv[i]); |
2175 | } |
2176 | } |
2177 | #endif /* OMP_40_ENABLED */ |
2178 | |
2179 | /* now actually fork the threads */ |
2180 | KMP_CHECK_UPDATE(team->t.t_master_active, master_active)if ((team->t.t_master_active) != (master_active)) (team-> t.t_master_active) = (master_active); |
2181 | if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong |
2182 | root->r.r_active = TRUE(!0); |
2183 | |
2184 | __kmp_fork_team_threads(root, team, master_th, gtid); |
2185 | __kmp_setup_icv_copy(team, nthreads, |
2186 | &master_th->th.th_current_task->td_icvs, loc); |
2187 | |
2188 | #if OMPT_SUPPORT1 |
2189 | master_th->th.ompt_thread_info.state = omp_state_work_parallel; |
2190 | #endif |
2191 | |
2192 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
2193 | |
2194 | #if USE_ITT_BUILD1 |
2195 | if (team->t.t_active_level == 1 // only report frames at level 1 |
2196 | #if OMP_40_ENABLED(50 >= 40) |
2197 | && !master_th->th.th_teams_microtask // not in teams construct |
2198 | #endif /* OMP_40_ENABLED */ |
2199 | ) { |
2200 | #if USE_ITT_NOTIFY1 |
2201 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && |
2202 | (__kmp_forkjoin_frames_mode == 3 || |
2203 | __kmp_forkjoin_frames_mode == 1)) { |
2204 | kmp_uint64 tmp_time = 0; |
2205 | if (__itt_get_timestamp_ptr__kmp_itt_get_timestamp_ptr__3_0) |
2206 | tmp_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); |
2207 | // Internal fork - report frame begin |
2208 | master_th->th.th_frame_time = tmp_time; |
2209 | if (__kmp_forkjoin_frames_mode == 3) |
2210 | team->t.t_region_time = tmp_time; |
2211 | } else |
2212 | // only one notification scheme (either "submit" or "forking/joined", not both) |
2213 | #endif /* USE_ITT_NOTIFY */ |
2214 | if ((__itt_frame_begin_v3_ptr__kmp_itt_frame_begin_v3_ptr__3_0 || KMP_ITT_DEBUG0) && |
2215 | __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) { |
2216 | // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer. |
2217 | __kmp_itt_region_forking(gtid, team->t.t_nproc, 0); |
2218 | } |
2219 | } |
2220 | #endif /* USE_ITT_BUILD */ |
2221 | |
2222 | /* now go on and do the work */ |
2223 | KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team)((team == __kmp_threads[gtid]->th.th_team) ? 0 : __kmp_debug_assert ("team == __kmp_threads[gtid]->th.th_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2223)); |
2224 | KMP_MB(); |
2225 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); } |
2226 | ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); } |
2227 | root, team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n" , root, team, master_th, gtid); }; |
2228 | |
2229 | #if USE_ITT_BUILD1 |
2230 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { |
2231 | team->t.t_stack_id = |
2232 | __kmp_itt_stack_caller_create(); // create new stack stitching id |
2233 | // before entering fork barrier |
2234 | } |
2235 | #endif /* USE_ITT_BUILD */ |
2236 | |
2237 | #if OMP_40_ENABLED(50 >= 40) |
2238 | // AC: skip __kmp_internal_fork at teams construct, let only master |
2239 | // threads execute |
2240 | if (ap) |
2241 | #endif /* OMP_40_ENABLED */ |
2242 | { |
2243 | __kmp_internal_fork(loc, gtid, team); |
2244 | KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); } |
2245 | "master_th=%p, gtid=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); } |
2246 | root, team, master_th, gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_internal_fork : after : root=%p, team=%p, " "master_th=%p, gtid=%d\n", root, team, master_th, gtid); }; |
2247 | } |
2248 | |
2249 | if (call_context == fork_context_gnu) { |
2250 | KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: parallel exit T#%d\n" , gtid); }; |
2251 | return TRUE(!0); |
2252 | } |
2253 | |
2254 | /* Invoke microtask for MASTER thread */ |
2255 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); } |
2256 | team->t.t_id, team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); }; |
2257 | } // END of timer KMP_fork_call block |
2258 | |
2259 | { |
2260 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
2261 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
2262 | if (!team->t.t_invoke(gtid)) { |
2263 | KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread")((0) ? 0 : __kmp_debug_assert(("cannot invoke microtask for MASTER thread" ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2263)); |
2264 | } |
2265 | } |
2266 | KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); } |
2267 | team->t.t_id, team->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n" , gtid, team->t.t_id, team->t.t_pkfn); }; |
2268 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
2269 | |
2270 | KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_fork_call: parallel exit T#%d\n" , gtid); }; |
2271 | |
2272 | #if OMPT_SUPPORT1 |
2273 | if (ompt_enabled.enabled) { |
2274 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
2275 | } |
2276 | #endif |
2277 | |
2278 | return TRUE(!0); |
2279 | } |
2280 | |
2281 | #if OMPT_SUPPORT1 |
2282 | static inline void __kmp_join_restore_state(kmp_info_t *thread, |
2283 | kmp_team_t *team) { |
2284 | // restore state outside the region |
2285 | thread->th.ompt_thread_info.state = |
2286 | ((team->t.t_serialized) ? omp_state_work_serial |
2287 | : omp_state_work_parallel); |
2288 | } |
2289 | |
2290 | static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread, |
2291 | kmp_team_t *team, ompt_data_t *parallel_data, |
2292 | fork_context_e fork_context, void *codeptr) { |
2293 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); |
2294 | if (ompt_enabled.ompt_callback_parallel_end) { |
2295 | ompt_callbacks.ompt_callback(ompt_callback_parallel_end)ompt_callback_parallel_end_callback( |
2296 | parallel_data, &(task_info->task_data), OMPT_INVOKER(fork_context)((fork_context == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime ), |
2297 | codeptr); |
2298 | } |
2299 | |
2300 | task_info->frame.enter_frame = NULL__null; |
2301 | __kmp_join_restore_state(thread, team); |
2302 | } |
2303 | #endif |
2304 | |
2305 | void __kmp_join_call(ident_t *loc, int gtid |
2306 | #if OMPT_SUPPORT1 |
2307 | , |
2308 | enum fork_context_e fork_context |
2309 | #endif |
2310 | #if OMP_40_ENABLED(50 >= 40) |
2311 | , |
2312 | int exit_teams |
2313 | #endif /* OMP_40_ENABLED */ |
2314 | ) { |
2315 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call)((void)0); |
2316 | kmp_team_t *team; |
2317 | kmp_team_t *parent_team; |
2318 | kmp_info_t *master_th; |
2319 | kmp_root_t *root; |
2320 | int master_active; |
2321 | int i; |
2322 | |
2323 | KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: enter T#%d\n" , gtid); }; |
2324 | |
2325 | /* setup current data */ |
2326 | master_th = __kmp_threads[gtid]; |
2327 | root = master_th->th.th_root; |
2328 | team = master_th->th.th_team; |
2329 | parent_team = team->t.t_parent; |
2330 | |
2331 | master_th->th.th_ident = loc; |
2332 | |
2333 | #if OMPT_SUPPORT1 |
2334 | if (ompt_enabled.enabled) { |
2335 | master_th->th.ompt_thread_info.state = omp_state_overhead; |
2336 | } |
2337 | #endif |
2338 | |
2339 | #if KMP_DEBUG1 |
2340 | if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) { |
2341 | KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", ((((master_th) != __null) ? 0 : __kmp_debug_assert ("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2343)), (master_th)->th.th_info.ds.ds_gtid), team, team-> t.t_task_team[master_th->th.th_task_state], master_th-> th.th_task_team); } |
2342 | "th_task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", ((((master_th) != __null) ? 0 : __kmp_debug_assert ("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2343)), (master_th)->th.th_info.ds.ds_gtid), team, team-> t.t_task_team[master_th->th.th_task_state], master_th-> th.th_task_team); } |
2343 | __kmp_gtid_from_thread(master_th), team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", ((((master_th) != __null) ? 0 : __kmp_debug_assert ("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2343)), (master_th)->th.th_info.ds.ds_gtid), team, team-> t.t_task_team[master_th->th.th_task_state], master_th-> th.th_task_team); } |
2344 | team->t.t_task_team[master_th->th.th_task_state],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", ((((master_th) != __null) ? 0 : __kmp_debug_assert ("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2343)), (master_th)->th.th_info.ds.ds_gtid), team, team-> t.t_task_team[master_th->th.th_task_state], master_th-> th.th_task_team); } |
2345 | master_th->th.th_task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " "th_task_team = %p\n", ((((master_th) != __null) ? 0 : __kmp_debug_assert ("(master_th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2343)), (master_th)->th.th_info.ds.ds_gtid), team, team-> t.t_task_team[master_th->th.th_task_state], master_th-> th.th_task_team); }; |
2346 | KMP_DEBUG_ASSERT(master_th->th.th_task_team ==((master_th->th.th_task_team == team->t.t_task_team[master_th ->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2347)) |
2347 | team->t.t_task_team[master_th->th.th_task_state])((master_th->th.th_task_team == team->t.t_task_team[master_th ->th.th_task_state]) ? 0 : __kmp_debug_assert("master_th->th.th_task_team == team->t.t_task_team[master_th->th.th_task_state]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2347)); |
2348 | } |
2349 | #endif |
2350 | |
2351 | if (team->t.t_serialized) { |
2352 | #if OMP_40_ENABLED(50 >= 40) |
2353 | if (master_th->th.th_teams_microtask) { |
2354 | // We are in teams construct |
2355 | int level = team->t.t_level; |
2356 | int tlevel = master_th->th.th_teams_level; |
2357 | if (level == tlevel) { |
2358 | // AC: we haven't incremented it earlier at start of teams construct, |
2359 | // so do it here - at the end of teams construct |
2360 | team->t.t_level++; |
2361 | } else if (level == tlevel + 1) { |
2362 | // AC: we are exiting parallel inside teams, need to increment |
2363 | // serialization in order to restore it in the next call to |
2364 | // __kmpc_end_serialized_parallel |
2365 | team->t.t_serialized++; |
2366 | } |
2367 | } |
2368 | #endif /* OMP_40_ENABLED */ |
2369 | __kmpc_end_serialized_parallel(loc, gtid); |
2370 | |
2371 | #if OMPT_SUPPORT1 |
2372 | if (ompt_enabled.enabled) { |
2373 | __kmp_join_restore_state(master_th, parent_team); |
2374 | } |
2375 | #endif |
2376 | |
2377 | return; |
2378 | } |
2379 | |
2380 | master_active = team->t.t_master_active; |
2381 | |
2382 | #if OMP_40_ENABLED(50 >= 40) |
2383 | if (!exit_teams) |
2384 | #endif /* OMP_40_ENABLED */ |
2385 | { |
2386 | // AC: No barrier for internal teams at exit from teams construct. |
2387 | // But there is barrier for external team (league). |
2388 | __kmp_internal_join(loc, gtid, team); |
2389 | } |
2390 | #if OMP_40_ENABLED(50 >= 40) |
2391 | else { |
2392 | master_th->th.th_task_state = |
2393 | 0; // AC: no tasking in teams (out of any parallel) |
2394 | } |
2395 | #endif /* OMP_40_ENABLED */ |
2396 | |
2397 | KMP_MB(); |
2398 | |
2399 | #if OMPT_SUPPORT1 |
2400 | ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data); |
2401 | void *codeptr = team->t.ompt_team_info.master_return_address; |
2402 | #endif |
2403 | |
2404 | #if USE_ITT_BUILD1 |
2405 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { |
2406 | __kmp_itt_stack_caller_destroy( |
2407 | (__itt_caller)team->t |
2408 | .t_stack_id); // destroy the stack stitching id after join barrier |
2409 | } |
2410 | |
2411 | // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer. |
2412 | if (team->t.t_active_level == 1 |
2413 | #if OMP_40_ENABLED(50 >= 40) |
2414 | && !master_th->th.th_teams_microtask /* not in teams construct */ |
2415 | #endif /* OMP_40_ENABLED */ |
2416 | ) { |
2417 | master_th->th.th_ident = loc; |
2418 | // only one notification scheme (either "submit" or "forking/joined", not |
2419 | // both) |
2420 | if ((__itt_frame_submit_v3_ptr__kmp_itt_frame_submit_v3_ptr__3_0 || KMP_ITT_DEBUG0) && |
2421 | __kmp_forkjoin_frames_mode == 3) |
2422 | __kmp_itt_frame_submit(gtid, team->t.t_region_time, |
2423 | master_th->th.th_frame_time, 0, loc, |
2424 | master_th->th.th_team_nproc, 1); |
2425 | else if ((__itt_frame_end_v3_ptr__kmp_itt_frame_end_v3_ptr__3_0 || KMP_ITT_DEBUG0) && |
2426 | !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames) |
2427 | __kmp_itt_region_joined(gtid); |
2428 | } // active_level == 1 |
2429 | #endif /* USE_ITT_BUILD */ |
2430 | |
2431 | #if OMP_40_ENABLED(50 >= 40) |
2432 | if (master_th->th.th_teams_microtask && !exit_teams && |
2433 | team->t.t_pkfn != (microtask_t)__kmp_teams_master && |
2434 | team->t.t_level == master_th->th.th_teams_level + 1) { |
2435 | // AC: We need to leave the team structure intact at the end of parallel |
2436 | // inside the teams construct, so that at the next parallel same (hot) team |
2437 | // works, only adjust nesting levels |
2438 | |
2439 | /* Decrement our nested depth level */ |
2440 | team->t.t_level--; |
2441 | team->t.t_active_level--; |
2442 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); |
2443 | |
2444 | /* Restore number of threads in the team if needed */ |
2445 | if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) { |
2446 | int old_num = master_th->th.th_team_nproc; |
2447 | int new_num = master_th->th.th_teams_size.nth; |
2448 | kmp_info_t **other_threads = team->t.t_threads; |
2449 | team->t.t_nproc = new_num; |
2450 | for (i = 0; i < old_num; ++i) { |
2451 | other_threads[i]->th.th_team_nproc = new_num; |
2452 | } |
2453 | // Adjust states of non-used threads of the team |
2454 | for (i = old_num; i < new_num; ++i) { |
2455 | // Re-initialize thread's barrier data. |
2456 | int b; |
2457 | kmp_balign_t *balign = other_threads[i]->th.th_bar; |
2458 | for (b = 0; b < bs_last_barrier; ++b) { |
2459 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; |
2460 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2460)); |
2461 | #if USE_DEBUGGER0 |
2462 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; |
2463 | #endif |
2464 | } |
2465 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
2466 | // Synchronize thread's task state |
2467 | other_threads[i]->th.th_task_state = master_th->th.th_task_state; |
2468 | } |
2469 | } |
2470 | } |
2471 | |
2472 | #if OMPT_SUPPORT1 |
2473 | if (ompt_enabled.enabled) { |
2474 | __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context, |
2475 | codeptr); |
2476 | } |
2477 | #endif |
2478 | |
2479 | return; |
2480 | } |
2481 | #endif /* OMP_40_ENABLED */ |
2482 | |
2483 | /* do cleanup and restore the parent team */ |
2484 | master_th->th.th_info.ds.ds_tid = team->t.t_master_tid; |
2485 | master_th->th.th_local.this_construct = team->t.t_master_this_cons; |
2486 | |
2487 | master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid]; |
2488 | |
2489 | /* jc: The following lock has instructions with REL and ACQ semantics, |
2490 | separating the parallel user code called in this parallel region |
2491 | from the serial user code called after this function returns. */ |
2492 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
2493 | |
2494 | #if OMP_40_ENABLED(50 >= 40) |
2495 | if (!master_th->th.th_teams_microtask || |
2496 | team->t.t_level > master_th->th.th_teams_level) |
2497 | #endif /* OMP_40_ENABLED */ |
2498 | { |
2499 | /* Decrement our nested depth level */ |
2500 | KMP_ATOMIC_DEC(&root->r.r_in_parallel)(&root->r.r_in_parallel)->fetch_sub(1, std::memory_order_acq_rel ); |
2501 | } |
2502 | KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0)((root->r.r_in_parallel >= 0) ? 0 : __kmp_debug_assert( "root->r.r_in_parallel >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2502)); |
2503 | |
2504 | #if OMPT_SUPPORT1 |
2505 | if (ompt_enabled.enabled) { |
2506 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); |
2507 | if (ompt_enabled.ompt_callback_implicit_task) { |
2508 | int ompt_team_size = team->t.t_nproc; |
2509 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
2510 | ompt_scope_end, NULL__null, &(task_info->task_data), ompt_team_size, |
2511 | OMPT_CUR_TASK_INFO(master_th)(&(master_th->th.th_current_task->ompt_task_info))->thread_num); |
2512 | } |
2513 | |
2514 | task_info->frame.exit_frame = NULL__null; |
2515 | task_info->task_data = ompt_data_none; |
2516 | } |
2517 | #endif |
2518 | |
2519 | KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n" , 0, master_th, team); } |
2520 | master_th, team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n" , 0, master_th, team); }; |
2521 | __kmp_pop_current_task_from_thread(master_th); |
2522 | |
2523 | #if OMP_40_ENABLED(50 >= 40) && KMP_AFFINITY_SUPPORTED1 |
2524 | // Restore master thread's partition. |
2525 | master_th->th.th_first_place = team->t.t_first_place; |
2526 | master_th->th.th_last_place = team->t.t_last_place; |
2527 | #endif /* OMP_40_ENABLED */ |
2528 | |
2529 | updateHWFPControl(team); |
2530 | |
2531 | if (root->r.r_active != master_active) |
2532 | root->r.r_active = master_active; |
2533 | |
2534 | __kmp_free_team(root, team USE_NESTED_HOT_ARG(, master_th |
2535 | master_th), master_th); // this will free worker threads |
2536 | |
2537 | /* this race was fun to find. make sure the following is in the critical |
2538 | region otherwise assertions may fail occasionally since the old team may be |
2539 | reallocated and the hierarchy appears inconsistent. it is actually safe to |
2540 | run and won't cause any bugs, but will cause those assertion failures. it's |
2541 | only one deref&assign so might as well put this in the critical region */ |
2542 | master_th->th.th_team = parent_team; |
2543 | master_th->th.th_team_nproc = parent_team->t.t_nproc; |
2544 | master_th->th.th_team_master = parent_team->t.t_threads[0]; |
2545 | master_th->th.th_team_serialized = parent_team->t.t_serialized; |
2546 | |
2547 | /* restore serialized team, if need be */ |
2548 | if (parent_team->t.t_serialized && |
2549 | parent_team != master_th->th.th_serial_team && |
2550 | parent_team != root->r.r_root_team) { |
2551 | __kmp_free_team(root, |
2552 | master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL), __null); |
2553 | master_th->th.th_serial_team = parent_team; |
2554 | } |
2555 | |
2556 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
2557 | if (master_th->th.th_task_state_top > |
2558 | 0) { // Restore task state from memo stack |
2559 | KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack)((master_th->th.th_task_state_memo_stack) ? 0 : __kmp_debug_assert ("master_th->th.th_task_state_memo_stack", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2559)); |
2560 | // Remember master's state if we re-use this nested hot team |
2561 | master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] = |
2562 | master_th->th.th_task_state; |
2563 | --master_th->th.th_task_state_top; // pop |
2564 | // Now restore state at this level |
2565 | master_th->th.th_task_state = |
2566 | master_th->th |
2567 | .th_task_state_memo_stack[master_th->th.th_task_state_top]; |
2568 | } |
2569 | // Copy the task team from the parent team to the master thread |
2570 | master_th->th.th_task_team = |
2571 | parent_team->t.t_task_team[master_th->th.th_task_state]; |
2572 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n" , ((((master_th) != __null) ? 0 : __kmp_debug_assert("(master_th) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2574)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team); } |
2573 | ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n" , ((((master_th) != __null) ? 0 : __kmp_debug_assert("(master_th) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2574)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team); } |
2574 | __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n" , ((((master_th) != __null) ? 0 : __kmp_debug_assert("(master_th) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2574)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team); } |
2575 | parent_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n" , ((((master_th) != __null) ? 0 : __kmp_debug_assert("(master_th) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2574)), (master_th)->th.th_info.ds.ds_gtid), master_th-> th.th_task_team, parent_team); }; |
2576 | } |
2577 | |
2578 | // TODO: GEH - cannot do this assertion because root thread not set up as |
2579 | // executing |
2580 | // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 ); |
2581 | master_th->th.th_current_task->td_flags.executing = 1; |
2582 | |
2583 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
2584 | |
2585 | #if OMPT_SUPPORT1 |
2586 | if (ompt_enabled.enabled) { |
2587 | __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context, |
2588 | codeptr); |
2589 | } |
2590 | #endif |
2591 | |
2592 | KMP_MB(); |
2593 | KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_join_call: exit T#%d\n" , gtid); }; |
2594 | } |
2595 | |
2596 | /* Check whether we should push an internal control record onto the |
2597 | serial team stack. If so, do it. */ |
2598 | void __kmp_save_internal_controls(kmp_info_t *thread) { |
2599 | |
2600 | if (thread->th.th_team != thread->th.th_serial_team) { |
2601 | return; |
2602 | } |
2603 | if (thread->th.th_team->t.t_serialized > 1) { |
2604 | int push = 0; |
2605 | |
2606 | if (thread->th.th_team->t.t_control_stack_top == NULL__null) { |
2607 | push = 1; |
2608 | } else { |
2609 | if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level != |
2610 | thread->th.th_team->t.t_serialized) { |
2611 | push = 1; |
2612 | } |
2613 | } |
2614 | if (push) { /* push a record on the serial team's stack */ |
2615 | kmp_internal_control_t *control = |
2616 | (kmp_internal_control_t *)__kmp_allocate(___kmp_allocate((sizeof(kmp_internal_control_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2617) |
2617 | sizeof(kmp_internal_control_t))___kmp_allocate((sizeof(kmp_internal_control_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2617); |
2618 | |
2619 | copy_icvs(control, &thread->th.th_current_task->td_icvs); |
2620 | |
2621 | control->serial_nesting_level = thread->th.th_team->t.t_serialized; |
2622 | |
2623 | control->next = thread->th.th_team->t.t_control_stack_top; |
2624 | thread->th.th_team->t.t_control_stack_top = control; |
2625 | } |
2626 | } |
2627 | } |
2628 | |
2629 | /* Changes set_nproc */ |
2630 | void __kmp_set_num_threads(int new_nth, int gtid) { |
2631 | kmp_info_t *thread; |
2632 | kmp_root_t *root; |
2633 | |
2634 | KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_num_threads: new __kmp_nth = %d\n" , new_nth); }; |
2635 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2635)); |
2636 | |
2637 | if (new_nth < 1) |
2638 | new_nth = 1; |
2639 | else if (new_nth > __kmp_max_nth) |
2640 | new_nth = __kmp_max_nth; |
2641 | |
2642 | KMP_COUNT_VALUE(OMP_set_numthreads, new_nth)((void)0); |
2643 | thread = __kmp_threads[gtid]; |
2644 | |
2645 | __kmp_save_internal_controls(thread); |
2646 | |
2647 | set__nproc(thread, new_nth)(((thread)->th.th_current_task->td_icvs.nproc) = (new_nth )); |
2648 | |
2649 | // If this omp_set_num_threads() call will cause the hot team size to be |
2650 | // reduced (in the absence of a num_threads clause), then reduce it now, |
2651 | // rather than waiting for the next parallel region. |
2652 | root = thread->th.th_root; |
2653 | if (__kmp_init_parallel && (!root->r.r_active) && |
2654 | (root->r.r_hot_team->t.t_nproc > new_nth) |
2655 | #if KMP_NESTED_HOT_TEAMS1 |
2656 | && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode |
2657 | #endif |
2658 | ) { |
2659 | kmp_team_t *hot_team = root->r.r_hot_team; |
2660 | int f; |
2661 | |
2662 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
2663 | |
2664 | // Release the extra threads we don't need any more. |
2665 | for (f = new_nth; f < hot_team->t.t_nproc; f++) { |
2666 | KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL)((hot_team->t.t_threads[f] != __null) ? 0 : __kmp_debug_assert ("hot_team->t.t_threads[f] != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2666)); |
2667 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
2668 | // When decreasing team size, threads no longer in the team should unref |
2669 | // task team. |
2670 | hot_team->t.t_threads[f]->th.th_task_team = NULL__null; |
2671 | } |
2672 | __kmp_free_thread(hot_team->t.t_threads[f]); |
2673 | hot_team->t.t_threads[f] = NULL__null; |
2674 | } |
2675 | hot_team->t.t_nproc = new_nth; |
2676 | #if KMP_NESTED_HOT_TEAMS1 |
2677 | if (thread->th.th_hot_teams) { |
2678 | KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team)((hot_team == thread->th.th_hot_teams[0].hot_team) ? 0 : __kmp_debug_assert ("hot_team == thread->th.th_hot_teams[0].hot_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2678)); |
2679 | thread->th.th_hot_teams[0].hot_team_nth = new_nth; |
2680 | } |
2681 | #endif |
2682 | |
2683 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
2684 | |
2685 | // Update the t_nproc field in the threads that are still active. |
2686 | for (f = 0; f < new_nth; f++) { |
2687 | KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL)((hot_team->t.t_threads[f] != __null) ? 0 : __kmp_debug_assert ("hot_team->t.t_threads[f] != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2687)); |
2688 | hot_team->t.t_threads[f]->th.th_team_nproc = new_nth; |
2689 | } |
2690 | // Special flag in case omp_set_num_threads() call |
2691 | hot_team->t.t_size_changed = -1; |
2692 | } |
2693 | } |
2694 | |
2695 | /* Changes max_active_levels */ |
2696 | void __kmp_set_max_active_levels(int gtid, int max_active_levels) { |
2697 | kmp_info_t *thread; |
2698 | |
2699 | KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); } |
2700 | "%d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); } |
2701 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: new max_active_levels for thread " "%d = (%d)\n", gtid, max_active_levels); }; |
2702 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2702)); |
2703 | |
2704 | // validate max_active_levels |
2705 | if (max_active_levels < 0) { |
2706 | KMP_WARNING(ActiveLevelsNegative, max_active_levels)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsNegative , max_active_levels), __kmp_msg_null); |
2707 | // We ignore this call if the user has specified a negative value. |
2708 | // The current setting won't be changed. The last valid setting will be |
2709 | // used. A warning will be issued (if warnings are allowed as controlled by |
2710 | // the KMP_WARNINGS env var). |
2711 | KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } |
2712 | "max_active_levels for thread %d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } |
2713 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: the call is ignored: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); }; |
2714 | return; |
2715 | } |
2716 | if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647) { |
2717 | // it's OK, the max_active_levels is within the valid range: [ 0; |
2718 | // KMP_MAX_ACTIVE_LEVELS_LIMIT ] |
2719 | // We allow a zero value. (implementation defined behavior) |
2720 | } else { |
2721 | KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsExceedLimit , max_active_levels, 2147483647), __kmp_msg_null) |
2722 | KMP_MAX_ACTIVE_LEVELS_LIMIT)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_ActiveLevelsExceedLimit , max_active_levels, 2147483647), __kmp_msg_null); |
2723 | max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT2147483647; |
2724 | // Current upper limit is MAX_INT. (implementation defined behavior) |
2725 | // If the input exceeds the upper limit, we correct the input to be the |
2726 | // upper limit. (implementation defined behavior) |
2727 | // Actually, the flow should never get here until we use MAX_INT limit. |
2728 | } |
2729 | KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } |
2730 | "max_active_levels for thread %d = (%d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); } |
2731 | gtid, max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_max_active_levels: after validation: new " "max_active_levels for thread %d = (%d)\n", gtid, max_active_levels ); }; |
2732 | |
2733 | thread = __kmp_threads[gtid]; |
2734 | |
2735 | __kmp_save_internal_controls(thread); |
2736 | |
2737 | set__max_active_levels(thread, max_active_levels)(((thread)->th.th_current_task->td_icvs.max_active_levels ) = (max_active_levels)); |
2738 | } |
2739 | |
2740 | /* Gets max_active_levels */ |
2741 | int __kmp_get_max_active_levels(int gtid) { |
2742 | kmp_info_t *thread; |
2743 | |
2744 | KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d\n" , gtid); }; |
2745 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2745)); |
2746 | |
2747 | thread = __kmp_threads[gtid]; |
2748 | KMP_DEBUG_ASSERT(thread->th.th_current_task)((thread->th.th_current_task) ? 0 : __kmp_debug_assert("thread->th.th_current_task" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2748)); |
2749 | KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } |
2750 | "curtask_maxaclevel=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } |
2751 | gtid, thread->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); } |
2752 | thread->th.th_current_task->td_icvs.max_active_levels))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_max_active_levels: thread %d, curtask=%p, " "curtask_maxaclevel=%d\n", gtid, thread->th.th_current_task , thread->th.th_current_task->td_icvs.max_active_levels ); }; |
2753 | return thread->th.th_current_task->td_icvs.max_active_levels; |
2754 | } |
2755 | |
2756 | /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */ |
2757 | void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) { |
2758 | kmp_info_t *thread; |
2759 | // kmp_team_t *team; |
2760 | |
2761 | KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n" , gtid, (int)kind, chunk); } |
2762 | gtid, (int)kind, chunk))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n" , gtid, (int)kind, chunk); }; |
2763 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2763)); |
2764 | |
2765 | // Check if the kind parameter is valid, correct if needed. |
2766 | // Valid parameters should fit in one of two intervals - standard or extended: |
2767 | // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper> |
2768 | // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103 |
2769 | if (kind <= kmp_sched_lower || kind >= kmp_sched_upper || |
2770 | (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) { |
2771 | // TODO: Hint needs attention in case we change the default schedule. |
2772 | __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind)__kmp_msg_format(kmp_i18n_msg_ScheduleKindOutOfRange, kind), |
2773 | KMP_HNT(DefaultScheduleKindUsed, "static, no chunk")__kmp_msg_format(kmp_i18n_hnt_DefaultScheduleKindUsed, "static, no chunk" ), |
2774 | __kmp_msg_null); |
2775 | kind = kmp_sched_default; |
2776 | chunk = 0; // ignore chunk value in case of bad kind |
2777 | } |
2778 | |
2779 | thread = __kmp_threads[gtid]; |
2780 | |
2781 | __kmp_save_internal_controls(thread); |
2782 | |
2783 | if (kind < kmp_sched_upper_std) { |
2784 | if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK1) { |
2785 | // differ static chunked vs. unchunked: chunk should be invalid to |
2786 | // indicate unchunked schedule (which is the default) |
2787 | thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static; |
2788 | } else { |
2789 | thread->th.th_current_task->td_icvs.sched.r_sched_type = |
2790 | __kmp_sch_map[kind - kmp_sched_lower - 1]; |
2791 | } |
2792 | } else { |
2793 | // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std - |
2794 | // kmp_sched_lower - 2 ]; |
2795 | thread->th.th_current_task->td_icvs.sched.r_sched_type = |
2796 | __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std - |
2797 | kmp_sched_lower - 2]; |
2798 | } |
2799 | if (kind == kmp_sched_auto || chunk < 1) { |
2800 | // ignore parameter chunk for schedule auto |
2801 | thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK1; |
2802 | } else { |
2803 | thread->th.th_current_task->td_icvs.sched.chunk = chunk; |
2804 | } |
2805 | } |
2806 | |
2807 | /* Gets def_sched_var ICV values */ |
2808 | void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) { |
2809 | kmp_info_t *thread; |
2810 | enum sched_type th_type; |
2811 | |
2812 | KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_schedule: thread %d\n" , gtid); }; |
2813 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2813)); |
2814 | |
2815 | thread = __kmp_threads[gtid]; |
2816 | |
2817 | th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type; |
2818 | |
2819 | switch (th_type) { |
2820 | case kmp_sch_static: |
2821 | case kmp_sch_static_greedy: |
2822 | case kmp_sch_static_balanced: |
2823 | *kind = kmp_sched_static; |
2824 | *chunk = 0; // chunk was not set, try to show this fact via zero value |
2825 | return; |
2826 | case kmp_sch_static_chunked: |
2827 | *kind = kmp_sched_static; |
2828 | break; |
2829 | case kmp_sch_dynamic_chunked: |
2830 | *kind = kmp_sched_dynamic; |
2831 | break; |
2832 | case kmp_sch_guided_chunked: |
2833 | case kmp_sch_guided_iterative_chunked: |
2834 | case kmp_sch_guided_analytical_chunked: |
2835 | *kind = kmp_sched_guided; |
2836 | break; |
2837 | case kmp_sch_auto: |
2838 | *kind = kmp_sched_auto; |
2839 | break; |
2840 | case kmp_sch_trapezoidal: |
2841 | *kind = kmp_sched_trapezoidal; |
2842 | break; |
2843 | #if KMP_STATIC_STEAL_ENABLED1 |
2844 | case kmp_sch_static_steal: |
2845 | *kind = kmp_sched_static_steal; |
2846 | break; |
2847 | #endif |
2848 | default: |
2849 | KMP_FATAL(UnknownSchedulingType, th_type)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownSchedulingType , th_type), __kmp_msg_null); |
2850 | } |
2851 | |
2852 | *chunk = thread->th.th_current_task->td_icvs.sched.chunk; |
2853 | } |
2854 | |
2855 | int __kmp_get_ancestor_thread_num(int gtid, int level) { |
2856 | |
2857 | int ii, dd; |
2858 | kmp_team_t *team; |
2859 | kmp_info_t *thr; |
2860 | |
2861 | KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_ancestor_thread_num: thread %d %d\n" , gtid, level); }; |
2862 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2862)); |
2863 | |
2864 | // validate level |
2865 | if (level == 0) |
2866 | return 0; |
2867 | if (level < 0) |
2868 | return -1; |
2869 | thr = __kmp_threads[gtid]; |
2870 | team = thr->th.th_team; |
2871 | ii = team->t.t_level; |
2872 | if (level > ii) |
2873 | return -1; |
2874 | |
2875 | #if OMP_40_ENABLED(50 >= 40) |
2876 | if (thr->th.th_teams_microtask) { |
2877 | // AC: we are in teams region where multiple nested teams have same level |
2878 | int tlevel = thr->th.th_teams_level; // the level of the teams construct |
2879 | if (level <= |
2880 | tlevel) { // otherwise usual algorithm works (will not touch the teams) |
2881 | KMP_DEBUG_ASSERT(ii >= tlevel)((ii >= tlevel) ? 0 : __kmp_debug_assert("ii >= tlevel" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2881)); |
2882 | // AC: As we need to pass by the teams league, we need to artificially |
2883 | // increase ii |
2884 | if (ii == tlevel) { |
2885 | ii += 2; // three teams have same level |
2886 | } else { |
2887 | ii++; // two teams have same level |
2888 | } |
2889 | } |
2890 | } |
2891 | #endif |
2892 | |
2893 | if (ii == level) |
2894 | return __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2894)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
2895 | |
2896 | dd = team->t.t_serialized; |
2897 | level++; |
2898 | while (ii > level) { |
2899 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { |
2900 | } |
2901 | if ((team->t.t_serialized) && (!dd)) { |
2902 | team = team->t.t_parent; |
2903 | continue; |
2904 | } |
2905 | if (ii > level) { |
2906 | team = team->t.t_parent; |
2907 | dd = team->t.t_serialized; |
2908 | ii--; |
2909 | } |
2910 | } |
2911 | |
2912 | return (dd > 1) ? (0) : (team->t.t_master_tid); |
2913 | } |
2914 | |
2915 | int __kmp_get_team_size(int gtid, int level) { |
2916 | |
2917 | int ii, dd; |
2918 | kmp_team_t *team; |
2919 | kmp_info_t *thr; |
2920 | |
2921 | KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_get_team_size: thread %d %d\n" , gtid, level); }; |
2922 | KMP_DEBUG_ASSERT(__kmp_init_serial)((__kmp_init_serial) ? 0 : __kmp_debug_assert("__kmp_init_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2922)); |
2923 | |
2924 | // validate level |
2925 | if (level == 0) |
2926 | return 1; |
2927 | if (level < 0) |
2928 | return -1; |
2929 | thr = __kmp_threads[gtid]; |
2930 | team = thr->th.th_team; |
2931 | ii = team->t.t_level; |
2932 | if (level > ii) |
2933 | return -1; |
2934 | |
2935 | #if OMP_40_ENABLED(50 >= 40) |
2936 | if (thr->th.th_teams_microtask) { |
2937 | // AC: we are in teams region where multiple nested teams have same level |
2938 | int tlevel = thr->th.th_teams_level; // the level of the teams construct |
2939 | if (level <= |
2940 | tlevel) { // otherwise usual algorithm works (will not touch the teams) |
2941 | KMP_DEBUG_ASSERT(ii >= tlevel)((ii >= tlevel) ? 0 : __kmp_debug_assert("ii >= tlevel" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 2941)); |
2942 | // AC: As we need to pass by the teams league, we need to artificially |
2943 | // increase ii |
2944 | if (ii == tlevel) { |
2945 | ii += 2; // three teams have same level |
2946 | } else { |
2947 | ii++; // two teams have same level |
2948 | } |
2949 | } |
2950 | } |
2951 | #endif |
2952 | |
2953 | while (ii > level) { |
2954 | for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { |
2955 | } |
2956 | if (team->t.t_serialized && (!dd)) { |
2957 | team = team->t.t_parent; |
2958 | continue; |
2959 | } |
2960 | if (ii > level) { |
2961 | team = team->t.t_parent; |
2962 | ii--; |
2963 | } |
2964 | } |
2965 | |
2966 | return team->t.t_nproc; |
2967 | } |
2968 | |
2969 | kmp_r_sched_t __kmp_get_schedule_global() { |
2970 | // This routine created because pairs (__kmp_sched, __kmp_chunk) and |
2971 | // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults |
2972 | // independently. So one can get the updated schedule here. |
2973 | |
2974 | kmp_r_sched_t r_sched; |
2975 | |
2976 | // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static, |
2977 | // __kmp_guided. __kmp_sched should keep original value, so that user can set |
2978 | // KMP_SCHEDULE multiple times, and thus have different run-time schedules in |
2979 | // different roots (even in OMP 2.5) |
2980 | if (__kmp_sched == kmp_sch_static) { |
2981 | // replace STATIC with more detailed schedule (balanced or greedy) |
2982 | r_sched.r_sched_type = __kmp_static; |
2983 | } else if (__kmp_sched == kmp_sch_guided_chunked) { |
2984 | // replace GUIDED with more detailed schedule (iterative or analytical) |
2985 | r_sched.r_sched_type = __kmp_guided; |
2986 | } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other |
2987 | r_sched.r_sched_type = __kmp_sched; |
2988 | } |
2989 | |
2990 | if (__kmp_chunk < KMP_DEFAULT_CHUNK1) { |
2991 | // __kmp_chunk may be wrong here (if it was not ever set) |
2992 | r_sched.chunk = KMP_DEFAULT_CHUNK1; |
2993 | } else { |
2994 | r_sched.chunk = __kmp_chunk; |
2995 | } |
2996 | |
2997 | return r_sched; |
2998 | } |
2999 | |
3000 | /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE) |
3001 | at least argc number of *t_argv entries for the requested team. */ |
3002 | static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) { |
3003 | |
3004 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3004)); |
3005 | if (!realloc || argc > team->t.t_max_argc) { |
3006 | |
3007 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); } |
3008 | "current entries=%d\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); } |
3009 | team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " "current entries=%d\n", team->t.t_id, argc, (realloc) ? team ->t.t_max_argc : 0); }; |
3010 | /* if previously allocated heap space for args, free them */ |
3011 | if (realloc && team->t.t_argv != &team->t.t_inline_argv[0]) |
3012 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3012); |
3013 | |
3014 | if (argc <= KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))) { |
3015 | /* use unused space in the cache line for arguments */ |
3016 | team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *))); |
3017 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } |
3018 | "argv entries\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } |
3019 | team->t.t_id, team->t.t_max_argc))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: inline allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); }; |
3020 | team->t.t_argv = &team->t.t_inline_argv[0]; |
3021 | if (__kmp_storage_map) { |
3022 | __kmp_print_storage_map_gtid( |
3023 | -1, &team->t.t_inline_argv[0], |
3024 | &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))], |
3025 | (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES(int)((4 * 64 - ((3 * (sizeof(void *)) + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + sizeof(kmp_int16) + sizeof(kmp_uint32)) % 64)) / (sizeof(void *)))), "team_%d.t_inline_argv", |
3026 | team->t.t_id); |
3027 | } |
3028 | } else { |
3029 | /* allocate space for arguments in the heap */ |
3030 | team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES100 >> 1)) |
3031 | ? KMP_MIN_MALLOC_ARGV_ENTRIES100 |
3032 | : 2 * argc; |
3033 | KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } |
3034 | "argv entries\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); } |
3035 | team->t.t_id, team->t.t_max_argc))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " "argv entries\n", team->t.t_id, team->t.t_max_argc); }; |
3036 | team->t.t_argv = |
3037 | (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc)___kmp_page_allocate((sizeof(void *) * team->t.t_max_argc) , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3037); |
3038 | if (__kmp_storage_map) { |
3039 | __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0], |
3040 | &team->t.t_argv[team->t.t_max_argc], |
3041 | sizeof(void *) * team->t.t_max_argc, |
3042 | "team_%d.t_argv", team->t.t_id); |
3043 | } |
3044 | } |
3045 | } |
3046 | } |
3047 | |
3048 | static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) { |
3049 | int i; |
3050 | int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2; |
3051 | team->t.t_threads = |
3052 | (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth)___kmp_allocate((sizeof(kmp_info_t *) * max_nth), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3052); |
3053 | team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(___kmp_allocate((sizeof(dispatch_shared_info_t) * num_disp_buff ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3054) |
3054 | sizeof(dispatch_shared_info_t) * num_disp_buff)___kmp_allocate((sizeof(dispatch_shared_info_t) * num_disp_buff ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3054); |
3055 | team->t.t_dispatch = |
3056 | (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth)___kmp_allocate((sizeof(kmp_disp_t) * max_nth), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3056); |
3057 | team->t.t_implicit_task_taskdata = |
3058 | (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth)___kmp_allocate((sizeof(kmp_taskdata_t) * max_nth), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3058); |
3059 | team->t.t_max_nproc = max_nth; |
3060 | |
3061 | /* setup dispatch buffers */ |
3062 | for (i = 0; i < num_disp_buff; ++i) { |
3063 | team->t.t_disp_buffer[i].buffer_index = i; |
3064 | #if OMP_45_ENABLED(50 >= 45) |
3065 | team->t.t_disp_buffer[i].doacross_buf_idx = i; |
3066 | #endif |
3067 | } |
3068 | } |
3069 | |
3070 | static void __kmp_free_team_arrays(kmp_team_t *team) { |
3071 | /* Note: this does not free the threads in t_threads (__kmp_free_threads) */ |
3072 | int i; |
3073 | for (i = 0; i < team->t.t_max_nproc; ++i) { |
3074 | if (team->t.t_dispatch[i].th_disp_buffer != NULL__null) { |
3075 | __kmp_free(team->t.t_dispatch[i].th_disp_buffer)___kmp_free((team->t.t_dispatch[i].th_disp_buffer), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3075); |
3076 | team->t.t_dispatch[i].th_disp_buffer = NULL__null; |
3077 | } |
3078 | } |
3079 | #if KMP_USE_HIER_SCHED0 |
3080 | __kmp_dispatch_free_hierarchies(team); |
3081 | #endif |
3082 | __kmp_free(team->t.t_threads)___kmp_free((team->t.t_threads), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3082); |
3083 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3083); |
3084 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3084); |
3085 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3085); |
3086 | team->t.t_threads = NULL__null; |
3087 | team->t.t_disp_buffer = NULL__null; |
3088 | team->t.t_dispatch = NULL__null; |
3089 | team->t.t_implicit_task_taskdata = 0; |
3090 | } |
3091 | |
3092 | static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) { |
3093 | kmp_info_t **oldThreads = team->t.t_threads; |
3094 | |
3095 | __kmp_free(team->t.t_disp_buffer)___kmp_free((team->t.t_disp_buffer), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3095); |
3096 | __kmp_free(team->t.t_dispatch)___kmp_free((team->t.t_dispatch), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3096); |
3097 | __kmp_free(team->t.t_implicit_task_taskdata)___kmp_free((team->t.t_implicit_task_taskdata), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3097); |
3098 | __kmp_allocate_team_arrays(team, max_nth); |
3099 | |
3100 | KMP_MEMCPYmemcpy(team->t.t_threads, oldThreads, |
3101 | team->t.t_nproc * sizeof(kmp_info_t *)); |
3102 | |
3103 | __kmp_free(oldThreads)___kmp_free((oldThreads), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3103); |
3104 | } |
3105 | |
3106 | static kmp_internal_control_t __kmp_get_global_icvs(void) { |
3107 | |
3108 | kmp_r_sched_t r_sched = |
3109 | __kmp_get_schedule_global(); // get current state of scheduling globals |
3110 | |
3111 | #if OMP_40_ENABLED(50 >= 40) |
3112 | KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0)((__kmp_nested_proc_bind.used > 0) ? 0 : __kmp_debug_assert ("__kmp_nested_proc_bind.used > 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3112)); |
3113 | #endif /* OMP_40_ENABLED */ |
3114 | |
3115 | kmp_internal_control_t g_icvs = { |
3116 | 0, // int serial_nesting_level; //corresponds to value of th_team_serialized |
3117 | (kmp_int8)__kmp_dflt_nested, // int nested; //internal control |
3118 | // for nested parallelism (per thread) |
3119 | (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic |
3120 | // adjustment of threads (per thread) |
3121 | (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for |
3122 | // whether blocktime is explicitly set |
3123 | __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime |
3124 | #if KMP_USE_MONITOR |
3125 | __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime |
3126 | // intervals |
3127 | #endif |
3128 | __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for |
3129 | // next parallel region (per thread) |
3130 | // (use a max ub on value if __kmp_parallel_initialize not called yet) |
3131 | __kmp_dflt_max_active_levels, // int max_active_levels; //internal control |
3132 | // for max_active_levels |
3133 | r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule |
3134 | // {sched,chunk} pair |
3135 | #if OMP_40_ENABLED(50 >= 40) |
3136 | __kmp_nested_proc_bind.bind_types[0], |
3137 | __kmp_default_device, |
3138 | #endif /* OMP_40_ENABLED */ |
3139 | NULL__null // struct kmp_internal_control *next; |
3140 | }; |
3141 | |
3142 | return g_icvs; |
3143 | } |
3144 | |
3145 | static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) { |
3146 | |
3147 | kmp_internal_control_t gx_icvs; |
3148 | gx_icvs.serial_nesting_level = |
3149 | 0; // probably =team->t.t_serial like in save_inter_controls |
3150 | copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs); |
3151 | gx_icvs.next = NULL__null; |
3152 | |
3153 | return gx_icvs; |
3154 | } |
3155 | |
3156 | static void __kmp_initialize_root(kmp_root_t *root) { |
3157 | int f; |
3158 | kmp_team_t *root_team; |
3159 | kmp_team_t *hot_team; |
3160 | int hot_team_max_nth; |
3161 | kmp_r_sched_t r_sched = |
3162 | __kmp_get_schedule_global(); // get current state of scheduling globals |
3163 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); |
3164 | KMP_DEBUG_ASSERT(root)((root) ? 0 : __kmp_debug_assert("root", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3164)); |
3165 | KMP_ASSERT(!root->r.r_begin)((!root->r.r_begin) ? 0 : __kmp_debug_assert("!root->r.r_begin" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3165)); |
3166 | |
3167 | /* setup the root state structure */ |
3168 | __kmp_init_lock(&root->r.r_begin_lock); |
3169 | root->r.r_begin = FALSE0; |
3170 | root->r.r_active = FALSE0; |
3171 | root->r.r_in_parallel = 0; |
3172 | root->r.r_blocktime = __kmp_dflt_blocktime; |
3173 | root->r.r_nested = __kmp_dflt_nested; |
3174 | root->r.r_cg_nthreads = 1; |
3175 | |
3176 | /* setup the root team for this task */ |
3177 | /* allocate the root team structure */ |
3178 | KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: before root_team\n" ); }; |
3179 | |
3180 | root_team = |
3181 | __kmp_allocate_team(root, |
3182 | 1, // new_nproc |
3183 | 1, // max_nproc |
3184 | #if OMPT_SUPPORT1 |
3185 | ompt_data_none, // root parallel id |
3186 | #endif |
3187 | #if OMP_40_ENABLED(50 >= 40) |
3188 | __kmp_nested_proc_bind.bind_types[0], |
3189 | #endif |
3190 | &r_icvs, |
3191 | 0 // argc |
3192 | USE_NESTED_HOT_ARG(NULL), __null // master thread is unknown |
3193 | ); |
3194 | #if USE_DEBUGGER0 |
3195 | // Non-NULL value should be assigned to make the debugger display the root |
3196 | // team. |
3197 | TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0))((root_team->t.t_pkfn)) = (((microtask_t)(~0))); |
3198 | #endif |
3199 | |
3200 | KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: after root_team = %p\n" , root_team); }; |
3201 | |
3202 | root->r.r_root_team = root_team; |
3203 | root_team->t.t_control_stack_top = NULL__null; |
3204 | |
3205 | /* initialize root team */ |
3206 | root_team->t.t_threads[0] = NULL__null; |
3207 | root_team->t.t_nproc = 1; |
3208 | root_team->t.t_serialized = 1; |
3209 | // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; |
3210 | root_team->t.t_sched.sched = r_sched.sched; |
3211 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } |
3212 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } |
3213 | ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); } |
3214 | root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n" , root_team->t.t_id, 0, 0); }; |
3215 | |
3216 | /* setup the hot team for this task */ |
3217 | /* allocate the hot team structure */ |
3218 | KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: before hot_team\n" ); }; |
3219 | |
3220 | hot_team = |
3221 | __kmp_allocate_team(root, |
3222 | 1, // new_nproc |
3223 | __kmp_dflt_team_nth_ub * 2, // max_nproc |
3224 | #if OMPT_SUPPORT1 |
3225 | ompt_data_none, // root parallel id |
3226 | #endif |
3227 | #if OMP_40_ENABLED(50 >= 40) |
3228 | __kmp_nested_proc_bind.bind_types[0], |
3229 | #endif |
3230 | &r_icvs, |
3231 | 0 // argc |
3232 | USE_NESTED_HOT_ARG(NULL), __null // master thread is unknown |
3233 | ); |
3234 | KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_root: after hot_team = %p\n" , hot_team); }; |
3235 | |
3236 | root->r.r_hot_team = hot_team; |
3237 | root_team->t.t_control_stack_top = NULL__null; |
3238 | |
3239 | /* first-time initialization */ |
3240 | hot_team->t.t_parent = root_team; |
3241 | |
3242 | /* initialize hot team */ |
3243 | hot_team_max_nth = hot_team->t.t_max_nproc; |
3244 | for (f = 0; f < hot_team_max_nth; ++f) { |
3245 | hot_team->t.t_threads[f] = NULL__null; |
3246 | } |
3247 | hot_team->t.t_nproc = 1; |
3248 | // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; |
3249 | hot_team->t.t_sched.sched = r_sched.sched; |
3250 | hot_team->t.t_size_changed = 0; |
3251 | } |
3252 | |
3253 | #ifdef KMP_DEBUG1 |
3254 | |
3255 | typedef struct kmp_team_list_item { |
3256 | kmp_team_p const *entry; |
3257 | struct kmp_team_list_item *next; |
3258 | } kmp_team_list_item_t; |
3259 | typedef kmp_team_list_item_t *kmp_team_list_t; |
3260 | |
3261 | static void __kmp_print_structure_team_accum( // Add team to list of teams. |
3262 | kmp_team_list_t list, // List of teams. |
3263 | kmp_team_p const *team // Team to add. |
3264 | ) { |
3265 | |
3266 | // List must terminate with item where both entry and next are NULL. |
3267 | // Team is added to the list only once. |
3268 | // List is sorted in ascending order by team id. |
3269 | // Team id is *not* a key. |
3270 | |
3271 | kmp_team_list_t l; |
3272 | |
3273 | KMP_DEBUG_ASSERT(list != NULL)((list != __null) ? 0 : __kmp_debug_assert("list != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3273)); |
3274 | if (team == NULL__null) { |
3275 | return; |
3276 | } |
3277 | |
3278 | __kmp_print_structure_team_accum(list, team->t.t_parent); |
3279 | __kmp_print_structure_team_accum(list, team->t.t_next_pool); |
3280 | |
3281 | // Search list for the team. |
3282 | l = list; |
3283 | while (l->next != NULL__null && l->entry != team) { |
3284 | l = l->next; |
3285 | } |
3286 | if (l->next != NULL__null) { |
3287 | return; // Team has been added before, exit. |
3288 | } |
3289 | |
3290 | // Team is not found. Search list again for insertion point. |
3291 | l = list; |
3292 | while (l->next != NULL__null && l->entry->t.t_id <= team->t.t_id) { |
3293 | l = l->next; |
3294 | } |
3295 | |
3296 | // Insert team. |
3297 | { |
3298 | kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(malloc(sizeof(kmp_team_list_item_t)) |
3299 | sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); |
3300 | *item = *l; |
3301 | l->entry = team; |
3302 | l->next = item; |
3303 | } |
3304 | } |
3305 | |
3306 | static void __kmp_print_structure_team(char const *title, kmp_team_p const *team |
3307 | |
3308 | ) { |
3309 | __kmp_printf("%s", title); |
3310 | if (team != NULL__null) { |
3311 | __kmp_printf("%2x %p\n", team->t.t_id, team); |
3312 | } else { |
3313 | __kmp_printf(" - (nil)\n"); |
3314 | } |
3315 | } |
3316 | |
3317 | static void __kmp_print_structure_thread(char const *title, |
3318 | kmp_info_p const *thread) { |
3319 | __kmp_printf("%s", title); |
3320 | if (thread != NULL__null) { |
3321 | __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread); |
3322 | } else { |
3323 | __kmp_printf(" - (nil)\n"); |
3324 | } |
3325 | } |
3326 | |
3327 | void __kmp_print_structure(void) { |
3328 | |
3329 | kmp_team_list_t list; |
3330 | |
3331 | // Initialize list of teams. |
3332 | list = |
3333 | (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t))malloc(sizeof(kmp_team_list_item_t)); |
3334 | list->entry = NULL__null; |
3335 | list->next = NULL__null; |
3336 | |
3337 | __kmp_printf("\n------------------------------\nGlobal Thread " |
3338 | "Table\n------------------------------\n"); |
3339 | { |
3340 | int gtid; |
3341 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { |
3342 | __kmp_printf("%2d", gtid); |
3343 | if (__kmp_threads != NULL__null) { |
3344 | __kmp_printf(" %p", __kmp_threads[gtid]); |
3345 | } |
3346 | if (__kmp_root != NULL__null) { |
3347 | __kmp_printf(" %p", __kmp_root[gtid]); |
3348 | } |
3349 | __kmp_printf("\n"); |
3350 | } |
3351 | } |
3352 | |
3353 | // Print out __kmp_threads array. |
3354 | __kmp_printf("\n------------------------------\nThreads\n--------------------" |
3355 | "----------\n"); |
3356 | if (__kmp_threads != NULL__null) { |
3357 | int gtid; |
3358 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { |
3359 | kmp_info_t const *thread = __kmp_threads[gtid]; |
3360 | if (thread != NULL__null) { |
3361 | __kmp_printf("GTID %2d %p:\n", gtid, thread); |
3362 | __kmp_printf(" Our Root: %p\n", thread->th.th_root); |
3363 | __kmp_print_structure_team(" Our Team: ", thread->th.th_team); |
3364 | __kmp_print_structure_team(" Serial Team: ", |
3365 | thread->th.th_serial_team); |
3366 | __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc); |
3367 | __kmp_print_structure_thread(" Master: ", |
3368 | thread->th.th_team_master); |
3369 | __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized); |
3370 | __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc); |
3371 | #if OMP_40_ENABLED(50 >= 40) |
3372 | __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind); |
3373 | #endif |
3374 | __kmp_print_structure_thread(" Next in pool: ", |
3375 | thread->th.th_next_pool); |
3376 | __kmp_printf("\n"); |
3377 | __kmp_print_structure_team_accum(list, thread->th.th_team); |
3378 | __kmp_print_structure_team_accum(list, thread->th.th_serial_team); |
3379 | } |
3380 | } |
3381 | } else { |
3382 | __kmp_printf("Threads array is not allocated.\n"); |
3383 | } |
3384 | |
3385 | // Print out __kmp_root array. |
3386 | __kmp_printf("\n------------------------------\nUbers\n----------------------" |
3387 | "--------\n"); |
3388 | if (__kmp_root != NULL__null) { |
3389 | int gtid; |
3390 | for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { |
3391 | kmp_root_t const *root = __kmp_root[gtid]; |
3392 | if (root != NULL__null) { |
3393 | __kmp_printf("GTID %2d %p:\n", gtid, root); |
3394 | __kmp_print_structure_team(" Root Team: ", root->r.r_root_team); |
3395 | __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team); |
3396 | __kmp_print_structure_thread(" Uber Thread: ", |
3397 | root->r.r_uber_thread); |
3398 | __kmp_printf(" Active?: %2d\n", root->r.r_active); |
3399 | __kmp_printf(" Nested?: %2d\n", root->r.r_nested); |
3400 | __kmp_printf(" In Parallel: %2d\n", |
3401 | KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel)(&root->r.r_in_parallel)->load(std::memory_order_relaxed )); |
3402 | __kmp_printf("\n"); |
3403 | __kmp_print_structure_team_accum(list, root->r.r_root_team); |
3404 | __kmp_print_structure_team_accum(list, root->r.r_hot_team); |
3405 | } |
3406 | } |
3407 | } else { |
3408 | __kmp_printf("Ubers array is not allocated.\n"); |
3409 | } |
3410 | |
3411 | __kmp_printf("\n------------------------------\nTeams\n----------------------" |
3412 | "--------\n"); |
3413 | while (list->next != NULL__null) { |
3414 | kmp_team_p const *team = list->entry; |
3415 | int i; |
3416 | __kmp_printf("Team %2x %p:\n", team->t.t_id, team); |
3417 | __kmp_print_structure_team(" Parent Team: ", team->t.t_parent); |
3418 | __kmp_printf(" Master TID: %2d\n", team->t.t_master_tid); |
3419 | __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc); |
3420 | __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized); |
3421 | __kmp_printf(" Number threads: %2d\n", team->t.t_nproc); |
3422 | for (i = 0; i < team->t.t_nproc; ++i) { |
3423 | __kmp_printf(" Thread %2d: ", i); |
3424 | __kmp_print_structure_thread("", team->t.t_threads[i]); |
3425 | } |
3426 | __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool); |
3427 | __kmp_printf("\n"); |
3428 | list = list->next; |
3429 | } |
3430 | |
3431 | // Print out __kmp_thread_pool and __kmp_team_pool. |
3432 | __kmp_printf("\n------------------------------\nPools\n----------------------" |
3433 | "--------\n"); |
3434 | __kmp_print_structure_thread("Thread pool: ", |
3435 | CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool)); |
3436 | __kmp_print_structure_team("Team pool: ", |
3437 | CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool)); |
3438 | __kmp_printf("\n"); |
3439 | |
3440 | // Free team list. |
3441 | while (list != NULL__null) { |
3442 | kmp_team_list_item_t *item = list; |
3443 | list = list->next; |
3444 | KMP_INTERNAL_FREE(item)free(item); |
3445 | } |
3446 | } |
3447 | |
3448 | #endif |
3449 | |
3450 | //--------------------------------------------------------------------------- |
3451 | // Stuff for per-thread fast random number generator |
3452 | // Table of primes |
3453 | static const unsigned __kmp_primes[] = { |
3454 | 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877, |
3455 | 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231, |
3456 | 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201, |
3457 | 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3, |
3458 | 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7, |
3459 | 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9, |
3460 | 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45, |
3461 | 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7, |
3462 | 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363, |
3463 | 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3, |
3464 | 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f}; |
3465 | |
3466 | //--------------------------------------------------------------------------- |
3467 | // __kmp_get_random: Get a random number using a linear congruential method. |
3468 | unsigned short __kmp_get_random(kmp_info_t *thread) { |
3469 | unsigned x = thread->th.th_x; |
3470 | unsigned short r = x >> 16; |
3471 | |
3472 | thread->th.th_x = x * thread->th.th_a + 1; |
3473 | |
3474 | KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_get_random: THREAD: %d, RETURN: %u\n" , thread->th.th_info.ds.ds_tid, r); } |
3475 | thread->th.th_info.ds.ds_tid, r))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_get_random: THREAD: %d, RETURN: %u\n" , thread->th.th_info.ds.ds_tid, r); }; |
3476 | |
3477 | return r; |
3478 | } |
3479 | //-------------------------------------------------------- |
3480 | // __kmp_init_random: Initialize a random number generator |
3481 | void __kmp_init_random(kmp_info_t *thread) { |
3482 | unsigned seed = thread->th.th_info.ds.ds_tid; |
3483 | |
3484 | thread->th.th_a = |
3485 | __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))]; |
3486 | thread->th.th_x = (seed + 1) * thread->th.th_a + 1; |
3487 | KA_TRACE(30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_init_random: THREAD: %u; A: %u\n" , seed, thread->th.th_a); } |
3488 | ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_init_random: THREAD: %u; A: %u\n" , seed, thread->th.th_a); }; |
3489 | } |
3490 | |
3491 | #if KMP_OS_WINDOWS0 |
3492 | /* reclaim array entries for root threads that are already dead, returns number |
3493 | * reclaimed */ |
3494 | static int __kmp_reclaim_dead_roots(void) { |
3495 | int i, r = 0; |
3496 | |
3497 | for (i = 0; i < __kmp_threads_capacity; ++i) { |
3498 | if (KMP_UBER_GTID(i)((((i) >= (-6)) ? 0 : __kmp_debug_assert("(i) >= (-6)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3498)), (((i) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(i) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3498)), (i) >= 0 && __kmp_root[(i)] && __kmp_threads [(i)] && (__kmp_threads[(i)] == __kmp_root[(i)]->r .r_uber_thread)) && |
3499 | !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])((void *)(__kmp_threads[i]))) && |
3500 | !__kmp_root[i] |
3501 | ->r.r_active) { // AC: reclaim only roots died in non-active state |
3502 | r += __kmp_unregister_root_other_thread(i); |
3503 | } |
3504 | } |
3505 | return r; |
3506 | } |
3507 | #endif |
3508 | |
3509 | /* This function attempts to create free entries in __kmp_threads and |
3510 | __kmp_root, and returns the number of free entries generated. |
3511 | |
3512 | For Windows* OS static library, the first mechanism used is to reclaim array |
3513 | entries for root threads that are already dead. |
3514 | |
3515 | On all platforms, expansion is attempted on the arrays __kmp_threads_ and |
3516 | __kmp_root, with appropriate update to __kmp_threads_capacity. Array |
3517 | capacity is increased by doubling with clipping to __kmp_tp_capacity, if |
3518 | threadprivate cache array has been created. Synchronization with |
3519 | __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock. |
3520 | |
3521 | After any dead root reclamation, if the clipping value allows array expansion |
3522 | to result in the generation of a total of nNeed free slots, the function does |
3523 | that expansion. If not, nothing is done beyond the possible initial root |
3524 | thread reclamation. |
3525 | |
3526 | If any argument is negative, the behavior is undefined. */ |
3527 | static int __kmp_expand_threads(int nNeed) { |
3528 | int added = 0; |
3529 | int minimumRequiredCapacity; |
3530 | int newCapacity; |
3531 | kmp_info_t **newThreads; |
3532 | kmp_root_t **newRoot; |
3533 | |
3534 | // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so |
3535 | // resizing __kmp_threads does not need additional protection if foreign |
3536 | // threads are present |
3537 | |
3538 | #if KMP_OS_WINDOWS0 && !defined KMP_DYNAMIC_LIB1 |
3539 | /* only for Windows static library */ |
3540 | /* reclaim array entries for root threads that are already dead */ |
3541 | added = __kmp_reclaim_dead_roots(); |
3542 | |
3543 | if (nNeed) { |
3544 | nNeed -= added; |
3545 | if (nNeed < 0) |
3546 | nNeed = 0; |
3547 | } |
3548 | #endif |
3549 | if (nNeed <= 0) |
3550 | return added; |
3551 | |
3552 | // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If |
3553 | // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the |
3554 | // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become |
3555 | // > __kmp_max_nth in one of two ways: |
3556 | // |
3557 | // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0] |
3558 | // may not be resused by another thread, so we may need to increase |
3559 | // __kmp_threads_capacity to __kmp_max_nth + 1. |
3560 | // |
3561 | // 2) New foreign root(s) are encountered. We always register new foreign |
3562 | // roots. This may cause a smaller # of threads to be allocated at |
3563 | // subsequent parallel regions, but the worker threads hang around (and |
3564 | // eventually go to sleep) and need slots in the __kmp_threads[] array. |
3565 | // |
3566 | // Anyway, that is the reason for moving the check to see if |
3567 | // __kmp_max_nth was exceeded into __kmp_reserve_threads() |
3568 | // instead of having it performed here. -BB |
3569 | |
3570 | KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity)((__kmp_sys_max_nth >= __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("__kmp_sys_max_nth >= __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3570)); |
3571 | |
3572 | /* compute expansion headroom to check if we can expand */ |
3573 | if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) { |
3574 | /* possible expansion too small -- give up */ |
3575 | return added; |
3576 | } |
3577 | minimumRequiredCapacity = __kmp_threads_capacity + nNeed; |
3578 | |
3579 | newCapacity = __kmp_threads_capacity; |
3580 | do { |
3581 | newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1) |
3582 | : __kmp_sys_max_nth; |
3583 | } while (newCapacity < minimumRequiredCapacity); |
3584 | newThreads = (kmp_info_t **)__kmp_allocate(___kmp_allocate(((sizeof(kmp_info_t *) + sizeof(kmp_root_t *) ) * newCapacity + 64), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3585) |
3585 | (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE)___kmp_allocate(((sizeof(kmp_info_t *) + sizeof(kmp_root_t *) ) * newCapacity + 64), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3585); |
3586 | newRoot = |
3587 | (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity); |
3588 | KMP_MEMCPYmemcpy(newThreads, __kmp_threads, |
3589 | __kmp_threads_capacity * sizeof(kmp_info_t *)); |
3590 | KMP_MEMCPYmemcpy(newRoot, __kmp_root, |
3591 | __kmp_threads_capacity * sizeof(kmp_root_t *)); |
3592 | |
3593 | kmp_info_t **temp_threads = __kmp_threads; |
3594 | *(kmp_info_t * *volatile *)&__kmp_threads = newThreads; |
3595 | *(kmp_root_t * *volatile *)&__kmp_root = newRoot; |
3596 | __kmp_free(temp_threads)___kmp_free((temp_threads), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3596); |
3597 | added += newCapacity - __kmp_threads_capacity; |
3598 | *(volatile int *)&__kmp_threads_capacity = newCapacity; |
3599 | |
3600 | if (newCapacity > __kmp_tp_capacity) { |
3601 | __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); |
3602 | if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) { |
3603 | __kmp_threadprivate_resize_cache(newCapacity); |
3604 | } else { // increase __kmp_tp_capacity to correspond with kmp_threads size |
3605 | *(volatile int *)&__kmp_tp_capacity = newCapacity; |
3606 | } |
3607 | __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); |
3608 | } |
3609 | |
3610 | return added; |
3611 | } |
3612 | |
3613 | /* Register the current thread as a root thread and obtain our gtid. We must |
3614 | have the __kmp_initz_lock held at this point. Argument TRUE only if are the |
3615 | thread that calls from __kmp_do_serial_initialize() */ |
3616 | int __kmp_register_root(int initial_thread) { |
3617 | kmp_info_t *root_thread; |
3618 | kmp_root_t *root; |
3619 | int gtid; |
3620 | int capacity; |
3621 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
3622 | KA_TRACE(20, ("__kmp_register_root: entered\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: entered\n" ); }; |
3623 | KMP_MB(); |
3624 | |
3625 | /* 2007-03-02: |
3626 | If initial thread did not invoke OpenMP RTL yet, and this thread is not an |
3627 | initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not |
3628 | work as expected -- it may return false (that means there is at least one |
3629 | empty slot in __kmp_threads array), but it is possible the only free slot |
3630 | is #0, which is reserved for initial thread and so cannot be used for this |
3631 | one. Following code workarounds this bug. |
3632 | |
3633 | However, right solution seems to be not reserving slot #0 for initial |
3634 | thread because: |
3635 | (1) there is no magic in slot #0, |
3636 | (2) we cannot detect initial thread reliably (the first thread which does |
3637 | serial initialization may be not a real initial thread). |
3638 | */ |
3639 | capacity = __kmp_threads_capacity; |
3640 | if (!initial_thread && TCR_PTR(__kmp_threads[0])((void *)(__kmp_threads[0])) == NULL__null) { |
3641 | --capacity; |
3642 | } |
3643 | |
3644 | /* see if there are too many threads */ |
3645 | if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) { |
3646 | if (__kmp_tp_cached) { |
3647 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), |
3648 | KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity)__kmp_msg_format(kmp_i18n_hnt_Set_ALL_THREADPRIVATE, __kmp_tp_capacity ), |
3649 | KMP_HNT(PossibleSystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_PossibleSystemLimitOnThreads), __kmp_msg_null); |
3650 | } else { |
3651 | __kmp_fatal(KMP_MSG(CantRegisterNewThread)__kmp_msg_format(kmp_i18n_msg_CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads)__kmp_msg_format(kmp_i18n_hnt_SystemLimitOnThreads), |
3652 | __kmp_msg_null); |
3653 | } |
3654 | } |
3655 | |
3656 | /* find an available thread slot */ |
3657 | /* Don't reassign the zero slot since we need that to only be used by initial |
3658 | thread */ |
3659 | for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid])((void *)(__kmp_threads[gtid])) != NULL__null; |
3660 | gtid++) |
3661 | ; |
3662 | KA_TRACE(1,if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array: T#%d\n" , gtid); } |
3663 | ("__kmp_register_root: found slot in threads array: T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_register_root: found slot in threads array: T#%d\n" , gtid); }; |
3664 | KMP_ASSERT(gtid < __kmp_threads_capacity)((gtid < __kmp_threads_capacity) ? 0 : __kmp_debug_assert( "gtid < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3664)); |
3665 | |
3666 | /* update global accounting */ |
3667 | __kmp_all_nth++; |
3668 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); |
3669 | |
3670 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low |
3671 | // numbers of procs, and method #2 (keyed API call) for higher numbers. |
3672 | if (__kmp_adjust_gtid_mode) { |
3673 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { |
3674 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { |
3675 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); |
3676 | } |
3677 | } else { |
3678 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { |
3679 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); |
3680 | } |
3681 | } |
3682 | } |
3683 | |
3684 | #ifdef KMP_ADJUST_BLOCKTIME1 |
3685 | /* Adjust blocktime to zero if necessary */ |
3686 | /* Middle initialization might not have occurred yet */ |
3687 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
3688 | if (__kmp_nth > __kmp_avail_proc) { |
3689 | __kmp_zero_bt = TRUE(!0); |
3690 | } |
3691 | } |
3692 | #endif /* KMP_ADJUST_BLOCKTIME */ |
3693 | |
3694 | /* setup this new hierarchy */ |
3695 | if (!(root = __kmp_root[gtid])) { |
3696 | root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t))___kmp_allocate((sizeof(kmp_root_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3696); |
3697 | KMP_DEBUG_ASSERT(!root->r.r_root_team)((!root->r.r_root_team) ? 0 : __kmp_debug_assert("!root->r.r_root_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3697)); |
3698 | } |
3699 | |
3700 | #if KMP_STATS_ENABLED0 |
3701 | // Initialize stats as soon as possible (right after gtid assignment). |
3702 | __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid); |
3703 | KMP_START_EXPLICIT_TIMER(OMP_worker_thread_life)((void)0); |
3704 | KMP_SET_THREAD_STATE(SERIAL_REGION)((void)0); |
3705 | KMP_INIT_PARTITIONED_TIMERS(OMP_serial)((void)0); |
3706 | #endif |
3707 | __kmp_initialize_root(root); |
3708 | |
3709 | /* setup new root thread structure */ |
3710 | if (root->r.r_uber_thread) { |
3711 | root_thread = root->r.r_uber_thread; |
3712 | } else { |
3713 | root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3713); |
3714 | if (__kmp_storage_map) { |
3715 | __kmp_print_thread_storage_map(root_thread, gtid); |
3716 | } |
3717 | root_thread->th.th_info.ds.ds_gtid = gtid; |
3718 | #if OMPT_SUPPORT1 |
3719 | root_thread->th.ompt_thread_info.thread_data.ptr = NULL__null; |
3720 | #endif |
3721 | root_thread->th.th_root = root; |
3722 | if (__kmp_env_consistency_check) { |
3723 | root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid); |
3724 | } |
3725 | #if USE_FAST_MEMORY3 |
3726 | __kmp_initialize_fast_memory(root_thread); |
3727 | #endif /* USE_FAST_MEMORY */ |
3728 | |
3729 | #if KMP_USE_BGET1 |
3730 | KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL)((root_thread->th.th_local.bget_data == __null) ? 0 : __kmp_debug_assert ("root_thread->th.th_local.bget_data == __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3730)); |
3731 | __kmp_initialize_bget(root_thread); |
3732 | #endif |
3733 | __kmp_init_random(root_thread); // Initialize random number generator |
3734 | } |
3735 | |
3736 | /* setup the serial team held in reserve by the root thread */ |
3737 | if (!root_thread->th.th_serial_team) { |
3738 | kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); |
3739 | KF_TRACE(10, ("__kmp_register_root: before serial_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: before serial_team\n" ); }; |
3740 | root_thread->th.th_serial_team = |
3741 | __kmp_allocate_team(root, 1, 1, |
3742 | #if OMPT_SUPPORT1 |
3743 | ompt_data_none, // root parallel id |
3744 | #endif |
3745 | #if OMP_40_ENABLED(50 >= 40) |
3746 | proc_bind_default, |
3747 | #endif |
3748 | &r_icvs, 0 USE_NESTED_HOT_ARG(NULL), __null); |
3749 | } |
3750 | KMP_ASSERT(root_thread->th.th_serial_team)((root_thread->th.th_serial_team) ? 0 : __kmp_debug_assert ("root_thread->th.th_serial_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3750)); |
3751 | KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: after serial_team = %p\n" , root_thread->th.th_serial_team); } |
3752 | root_thread->th.th_serial_team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_register_root: after serial_team = %p\n" , root_thread->th.th_serial_team); }; |
3753 | |
3754 | /* drop root_thread into place */ |
3755 | TCW_SYNC_PTR(__kmp_threads[gtid], root_thread)((__kmp_threads[gtid])) = ((root_thread)); |
3756 | |
3757 | root->r.r_root_team->t.t_threads[0] = root_thread; |
3758 | root->r.r_hot_team->t.t_threads[0] = root_thread; |
3759 | root_thread->th.th_serial_team->t.t_threads[0] = root_thread; |
3760 | // AC: the team created in reserve, not for execution (it is unused for now). |
3761 | root_thread->th.th_serial_team->t.t_serialized = 0; |
3762 | root->r.r_uber_thread = root_thread; |
3763 | |
3764 | /* initialize the thread, get it ready to go */ |
3765 | __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid); |
3766 | TCW_4(__kmp_init_gtid, TRUE)(__kmp_init_gtid) = ((!0)); |
3767 | |
3768 | /* prepare the master thread for get_gtid() */ |
3769 | __kmp_gtid_set_specific(gtid); |
3770 | |
3771 | #if USE_ITT_BUILD1 |
3772 | __kmp_itt_thread_name(gtid); |
3773 | #endif /* USE_ITT_BUILD */ |
3774 | |
3775 | #ifdef KMP_TDATA_GTID1 |
3776 | __kmp_gtid = gtid; |
3777 | #endif |
3778 | __kmp_create_worker(gtid, root_thread, __kmp_stksize); |
3779 | KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid)((__kmp_gtid_get_specific() == gtid) ? 0 : __kmp_debug_assert ("__kmp_gtid_get_specific() == gtid", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3779)); |
3780 | |
3781 | KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, ((((0) >= 0 && (root->r.r_hot_team ) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (root->r.r_hot_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3783)), root->r.r_hot_team->t.t_threads[(0)]->th.th_info .ds.ds_gtid), root->r.r_hot_team->t.t_id, 0, 0, 0); } |
3782 | "plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, ((((0) >= 0 && (root->r.r_hot_team ) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (root->r.r_hot_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3783)), root->r.r_hot_team->t.t_threads[(0)]->th.th_info .ds.ds_gtid), root->r.r_hot_team->t.t_id, 0, 0, 0); } |
3783 | gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, ((((0) >= 0 && (root->r.r_hot_team ) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (root->r.r_hot_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3783)), root->r.r_hot_team->t.t_threads[(0)]->th.th_info .ds.ds_gtid), root->r.r_hot_team->t.t_id, 0, 0, 0); } |
3784 | root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, ((((0) >= 0 && (root->r.r_hot_team ) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (root->r.r_hot_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3783)), root->r.r_hot_team->t.t_threads[(0)]->th.th_info .ds.ds_gtid), root->r.r_hot_team->t.t_id, 0, 0, 0); } |
3785 | KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " "plain=%u\n", gtid, ((((0) >= 0 && (root->r.r_hot_team ) != __null) ? 0 : __kmp_debug_assert("(0) >= 0 && (root->r.r_hot_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3783)), root->r.r_hot_team->t.t_threads[(0)]->th.th_info .ds.ds_gtid), root->r.r_hot_team->t.t_id, 0, 0, 0); }; |
3786 | { // Initialize barrier data. |
3787 | int b; |
3788 | for (b = 0; b < bs_last_barrier; ++b) { |
3789 | root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE0; |
3790 | #if USE_DEBUGGER0 |
3791 | root_thread->th.th_bar[b].bb.b_worker_arrived = 0; |
3792 | #endif |
3793 | } |
3794 | } |
3795 | KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==((root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0) ? 0 : __kmp_debug_assert("root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3796)) |
3796 | KMP_INIT_BARRIER_STATE)((root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0) ? 0 : __kmp_debug_assert("root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3796)); |
3797 | |
3798 | #if KMP_AFFINITY_SUPPORTED1 |
3799 | #if OMP_40_ENABLED(50 >= 40) |
3800 | root_thread->th.th_current_place = KMP_PLACE_UNDEFINED(-2); |
3801 | root_thread->th.th_new_place = KMP_PLACE_UNDEFINED(-2); |
3802 | root_thread->th.th_first_place = KMP_PLACE_UNDEFINED(-2); |
3803 | root_thread->th.th_last_place = KMP_PLACE_UNDEFINED(-2); |
3804 | #endif |
3805 | |
3806 | if (TCR_4(__kmp_init_middle)(__kmp_init_middle)) { |
3807 | __kmp_affinity_set_init_mask(gtid, TRUE(!0)); |
3808 | } |
3809 | #endif /* KMP_AFFINITY_SUPPORTED */ |
3810 | |
3811 | __kmp_root_counter++; |
3812 | |
3813 | #if OMPT_SUPPORT1 |
3814 | if (!initial_thread && ompt_enabled.enabled) { |
3815 | |
3816 | ompt_thread_t *root_thread = ompt_get_thread(); |
3817 | |
3818 | ompt_set_thread_state(root_thread, omp_state_overhead); |
3819 | |
3820 | if (ompt_enabled.ompt_callback_thread_begin) { |
3821 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( |
3822 | ompt_thread_initial, __ompt_get_thread_data_internal()); |
3823 | } |
3824 | ompt_data_t *task_data; |
3825 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, NULL__null, NULL__null); |
3826 | if (ompt_enabled.ompt_callback_task_create) { |
3827 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( |
3828 | NULL__null, NULL__null, task_data, ompt_task_initial, 0, NULL__null); |
3829 | // initial task has nothing to return to |
3830 | } |
3831 | |
3832 | ompt_set_thread_state(root_thread, omp_state_work_serial); |
3833 | } |
3834 | #endif |
3835 | |
3836 | KMP_MB(); |
3837 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
3838 | |
3839 | return gtid; |
3840 | } |
3841 | |
3842 | #if KMP_NESTED_HOT_TEAMS1 |
3843 | static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level, |
3844 | const int max_level) { |
3845 | int i, n, nth; |
3846 | kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams; |
3847 | if (!hot_teams || !hot_teams[level].hot_team) { |
3848 | return 0; |
3849 | } |
3850 | KMP_DEBUG_ASSERT(level < max_level)((level < max_level) ? 0 : __kmp_debug_assert("level < max_level" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3850)); |
3851 | kmp_team_t *team = hot_teams[level].hot_team; |
3852 | nth = hot_teams[level].hot_team_nth; |
3853 | n = nth - 1; // master is not freed |
3854 | if (level < max_level - 1) { |
3855 | for (i = 0; i < nth; ++i) { |
3856 | kmp_info_t *th = team->t.t_threads[i]; |
3857 | n += __kmp_free_hot_teams(root, th, level + 1, max_level); |
3858 | if (i > 0 && th->th.th_hot_teams) { |
3859 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3859); |
3860 | th->th.th_hot_teams = NULL__null; |
3861 | } |
3862 | } |
3863 | } |
3864 | __kmp_free_team(root, team, NULL__null); |
3865 | return n; |
3866 | } |
3867 | #endif |
3868 | |
3869 | // Resets a root thread and clear its root and hot teams. |
3870 | // Returns the number of __kmp_threads entries directly and indirectly freed. |
3871 | static int __kmp_reset_root(int gtid, kmp_root_t *root) { |
3872 | kmp_team_t *root_team = root->r.r_root_team; |
3873 | kmp_team_t *hot_team = root->r.r_hot_team; |
3874 | int n = hot_team->t.t_nproc; |
3875 | int i; |
3876 | |
3877 | KMP_DEBUG_ASSERT(!root->r.r_active)((!root->r.r_active) ? 0 : __kmp_debug_assert("!root->r.r_active" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3877)); |
3878 | |
3879 | root->r.r_root_team = NULL__null; |
3880 | root->r.r_hot_team = NULL__null; |
3881 | // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team |
3882 | // before call to __kmp_free_team(). |
3883 | __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL), __null); |
3884 | #if KMP_NESTED_HOT_TEAMS1 |
3885 | if (__kmp_hot_teams_max_level > |
3886 | 0) { // need to free nested hot teams and their threads if any |
3887 | for (i = 0; i < hot_team->t.t_nproc; ++i) { |
3888 | kmp_info_t *th = hot_team->t.t_threads[i]; |
3889 | if (__kmp_hot_teams_max_level > 1) { |
3890 | n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level); |
3891 | } |
3892 | if (th->th.th_hot_teams) { |
3893 | __kmp_free(th->th.th_hot_teams)___kmp_free((th->th.th_hot_teams), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3893); |
3894 | th->th.th_hot_teams = NULL__null; |
3895 | } |
3896 | } |
3897 | } |
3898 | #endif |
3899 | __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL), __null); |
3900 | |
3901 | // Before we can reap the thread, we need to make certain that all other |
3902 | // threads in the teams that had this root as ancestor have stopped trying to |
3903 | // steal tasks. |
3904 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
3905 | __kmp_wait_to_unref_task_teams(); |
3906 | } |
3907 | |
3908 | #if KMP_OS_WINDOWS0 |
3909 | /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */ |
3910 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } |
3911 | 10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPECif (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } |
3912 | "\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } |
3913 | (LPVOID) & (root->r.r_uber_thread->th),if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); } |
3914 | root->r.r_uber_thread->th.th_info.ds.ds_thread))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_reset_root: free handle, th = %p, handle = %" "lu" "\n", (LPVOID) & (root->r.r_uber_thread->th), root->r.r_uber_thread->th.th_info.ds.ds_thread); }; |
3915 | __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread); |
3916 | #endif /* KMP_OS_WINDOWS */ |
3917 | |
3918 | #if OMPT_SUPPORT1 |
3919 | if (ompt_enabled.ompt_callback_thread_end) { |
3920 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback( |
3921 | &(root->r.r_uber_thread->th.ompt_thread_info.thread_data)); |
3922 | } |
3923 | #endif |
3924 | |
3925 | TCW_4(__kmp_nth,(__kmp_nth) = (__kmp_nth - 1) |
3926 | __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth. |
3927 | root->r.r_cg_nthreads--; |
3928 | |
3929 | __kmp_reap_thread(root->r.r_uber_thread, 1); |
3930 | |
3931 | // We canot put root thread to __kmp_thread_pool, so we have to reap it istead |
3932 | // of freeing. |
3933 | root->r.r_uber_thread = NULL__null; |
3934 | /* mark root as no longer in use */ |
3935 | root->r.r_begin = FALSE0; |
3936 | |
3937 | return n; |
3938 | } |
3939 | |
3940 | void __kmp_unregister_root_current_thread(int gtid) { |
3941 | KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: enter T#%d\n" , gtid); }; |
3942 | /* this lock should be ok, since unregister_root_current_thread is never |
3943 | called during an abort, only during a normal close. furthermore, if you |
3944 | have the forkjoin lock, you should never try to get the initz lock */ |
3945 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
3946 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { |
3947 | KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); } |
3948 | "exiting T#%d\n",if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); } |
3949 | gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: already finished, " "exiting T#%d\n", gtid); }; |
3950 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
3951 | return; |
3952 | } |
3953 | kmp_root_t *root = __kmp_root[gtid]; |
3954 | |
3955 | KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid])((__kmp_threads && __kmp_threads[gtid]) ? 0 : __kmp_debug_assert ("__kmp_threads && __kmp_threads[gtid]", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3955)); |
3956 | KMP_ASSERT(KMP_UBER_GTID(gtid))((((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3956)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3956)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) ? 0 : __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3956)); |
3957 | KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root)((root == __kmp_threads[gtid]->th.th_root) ? 0 : __kmp_debug_assert ("root == __kmp_threads[gtid]->th.th_root", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3957)); |
3958 | KMP_ASSERT(root->r.r_active == FALSE)((root->r.r_active == 0) ? 0 : __kmp_debug_assert("root->r.r_active == FALSE" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 3958)); |
3959 | |
3960 | KMP_MB(); |
3961 | |
3962 | #if OMP_45_ENABLED(50 >= 45) |
3963 | kmp_info_t *thread = __kmp_threads[gtid]; |
3964 | kmp_team_t *team = thread->th.th_team; |
3965 | kmp_task_team_t *task_team = thread->th.th_task_team; |
3966 | |
3967 | // we need to wait for the proxy tasks before finishing the thread |
3968 | if (task_team != NULL__null && task_team->tt.tt_found_proxy_tasks) { |
3969 | #if OMPT_SUPPORT1 |
3970 | // the runtime is shutting down so we won't report any events |
3971 | thread->th.ompt_thread_info.state = omp_state_undefined; |
3972 | #endif |
3973 | __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL), __null); |
3974 | } |
3975 | #endif |
3976 | |
3977 | __kmp_reset_root(gtid, root); |
3978 | |
3979 | /* free up this thread slot */ |
3980 | __kmp_gtid_set_specific(KMP_GTID_DNE(-2)); |
3981 | #ifdef KMP_TDATA_GTID1 |
3982 | __kmp_gtid = KMP_GTID_DNE(-2); |
3983 | #endif |
3984 | |
3985 | KMP_MB(); |
3986 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: T#%d unregistered\n" , gtid); } |
3987 | ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_current_thread: T#%d unregistered\n" , gtid); }; |
3988 | |
3989 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
3990 | } |
3991 | |
3992 | #if KMP_OS_WINDOWS0 |
3993 | /* __kmp_forkjoin_lock must be already held |
3994 | Unregisters a root thread that is not the current thread. Returns the number |
3995 | of __kmp_threads entries freed as a result. */ |
3996 | static int __kmp_unregister_root_other_thread(int gtid) { |
3997 | kmp_root_t *root = __kmp_root[gtid]; |
3998 | int r; |
3999 | |
4000 | KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid))if (kmp_a_debug >= 1) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: enter T#%d\n" , gtid); }; |
4001 | KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid])((__kmp_threads && __kmp_threads[gtid]) ? 0 : __kmp_debug_assert ("__kmp_threads && __kmp_threads[gtid]", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4001)); |
4002 | KMP_ASSERT(KMP_UBER_GTID(gtid))((((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4002)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4002)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) ? 0 : __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4002)); |
4003 | KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root)((root == __kmp_threads[gtid]->th.th_root) ? 0 : __kmp_debug_assert ("root == __kmp_threads[gtid]->th.th_root", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4003)); |
4004 | KMP_ASSERT(root->r.r_active == FALSE)((root->r.r_active == 0) ? 0 : __kmp_debug_assert("root->r.r_active == FALSE" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4004)); |
4005 | |
4006 | r = __kmp_reset_root(gtid, root); |
4007 | KC_TRACE(10,if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: T#%d unregistered\n" , gtid); } |
4008 | ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid))if (kmp_c_debug >= 10) { __kmp_debug_printf ("__kmp_unregister_root_other_thread: T#%d unregistered\n" , gtid); }; |
4009 | return r; |
4010 | } |
4011 | #endif |
4012 | |
4013 | #if KMP_DEBUG1 |
4014 | void __kmp_task_info() { |
4015 | |
4016 | kmp_int32 gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); |
4017 | kmp_int32 tid = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4017)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
4018 | kmp_info_t *this_thr = __kmp_threads[gtid]; |
4019 | kmp_team_t *steam = this_thr->th.th_serial_team; |
4020 | kmp_team_t *team = this_thr->th.th_team; |
4021 | |
4022 | __kmp_printf("__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p curtask=%p " |
4023 | "ptask=%p\n", |
4024 | gtid, tid, this_thr, team, this_thr->th.th_current_task, |
4025 | team->t.t_implicit_task_taskdata[tid].td_parent); |
4026 | } |
4027 | #endif // KMP_DEBUG |
4028 | |
4029 | /* TODO optimize with one big memclr, take out what isn't needed, split |
4030 | responsibility to workers as much as possible, and delay initialization of |
4031 | features as much as possible */ |
4032 | static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team, |
4033 | int tid, int gtid) { |
4034 | /* this_thr->th.th_info.ds.ds_gtid is setup in |
4035 | kmp_allocate_thread/create_worker. |
4036 | this_thr->th.th_serial_team is setup in __kmp_allocate_thread */ |
4037 | kmp_info_t *master = team->t.t_threads[0]; |
4038 | KMP_DEBUG_ASSERT(this_thr != NULL)((this_thr != __null) ? 0 : __kmp_debug_assert("this_thr != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4038)); |
4039 | KMP_DEBUG_ASSERT(this_thr->th.th_serial_team)((this_thr->th.th_serial_team) ? 0 : __kmp_debug_assert("this_thr->th.th_serial_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4039)); |
4040 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4040)); |
4041 | KMP_DEBUG_ASSERT(team->t.t_threads)((team->t.t_threads) ? 0 : __kmp_debug_assert("team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4041)); |
4042 | KMP_DEBUG_ASSERT(team->t.t_dispatch)((team->t.t_dispatch) ? 0 : __kmp_debug_assert("team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4042)); |
4043 | KMP_DEBUG_ASSERT(master)((master) ? 0 : __kmp_debug_assert("master", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4043)); |
4044 | KMP_DEBUG_ASSERT(master->th.th_root)((master->th.th_root) ? 0 : __kmp_debug_assert("master->th.th_root" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4044)); |
4045 | |
4046 | KMP_MB(); |
4047 | |
4048 | TCW_SYNC_PTR(this_thr->th.th_team, team)((this_thr->th.th_team)) = ((team)); |
4049 | |
4050 | this_thr->th.th_info.ds.ds_tid = tid; |
4051 | this_thr->th.th_set_nproc = 0; |
4052 | if (__kmp_tasking_mode != tskm_immediate_exec) |
4053 | // When tasking is possible, threads are not safe to reap until they are |
4054 | // done tasking; this will be set when tasking code is exited in wait |
4055 | this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; |
4056 | else // no tasking --> always safe to reap |
4057 | this_thr->th.th_reap_state = KMP_SAFE_TO_REAP1; |
4058 | #if OMP_40_ENABLED(50 >= 40) |
4059 | this_thr->th.th_set_proc_bind = proc_bind_default; |
4060 | #if KMP_AFFINITY_SUPPORTED1 |
4061 | this_thr->th.th_new_place = this_thr->th.th_current_place; |
4062 | #endif |
4063 | #endif |
4064 | this_thr->th.th_root = master->th.th_root; |
4065 | |
4066 | /* setup the thread's cache of the team structure */ |
4067 | this_thr->th.th_team_nproc = team->t.t_nproc; |
4068 | this_thr->th.th_team_master = master; |
4069 | this_thr->th.th_team_serialized = team->t.t_serialized; |
4070 | TCW_PTR(this_thr->th.th_sleep_loc, NULL)((this_thr->th.th_sleep_loc)) = ((__null)); |
4071 | |
4072 | KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata)((team->t.t_implicit_task_taskdata) ? 0 : __kmp_debug_assert ("team->t.t_implicit_task_taskdata", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4072)); |
4073 | |
4074 | KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); } |
4075 | tid, gtid, this_thr, this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); }; |
4076 | |
4077 | __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr, |
4078 | team, tid, TRUE(!0)); |
4079 | |
4080 | KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); } |
4081 | tid, gtid, this_thr, this_thr->th.th_current_task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n" , tid, gtid, this_thr, this_thr->th.th_current_task); }; |
4082 | // TODO: Initialize ICVs from parent; GEH - isn't that already done in |
4083 | // __kmp_initialize_team()? |
4084 | |
4085 | /* TODO no worksharing in speculative threads */ |
4086 | this_thr->th.th_dispatch = &team->t.t_dispatch[tid]; |
4087 | |
4088 | this_thr->th.th_local.this_construct = 0; |
4089 | |
4090 | if (!this_thr->th.th_pri_common) { |
4091 | this_thr->th.th_pri_common = |
4092 | (struct common_table *)__kmp_allocate(sizeof(struct common_table))___kmp_allocate((sizeof(struct common_table)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4092); |
4093 | if (__kmp_storage_map) { |
4094 | __kmp_print_storage_map_gtid( |
4095 | gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1, |
4096 | sizeof(struct common_table), "th_%d.th_pri_common\n", gtid); |
4097 | } |
4098 | this_thr->th.th_pri_head = NULL__null; |
4099 | } |
4100 | |
4101 | /* Initialize dynamic dispatch */ |
4102 | { |
4103 | volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch; |
4104 | // Use team max_nproc since this will never change for the team. |
4105 | size_t disp_size = |
4106 | sizeof(dispatch_private_info_t) * |
4107 | (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers); |
4108 | KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,if (kmp_d_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info: T#%d max_nproc: %d\n" , gtid, team->t.t_max_nproc); } |
4109 | team->t.t_max_nproc))if (kmp_d_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_info: T#%d max_nproc: %d\n" , gtid, team->t.t_max_nproc); }; |
4110 | KMP_ASSERT(dispatch)((dispatch) ? 0 : __kmp_debug_assert("dispatch", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4110)); |
4111 | KMP_DEBUG_ASSERT(team->t.t_dispatch)((team->t.t_dispatch) ? 0 : __kmp_debug_assert("team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4111)); |
4112 | KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid])((dispatch == &team->t.t_dispatch[tid]) ? 0 : __kmp_debug_assert ("dispatch == &team->t.t_dispatch[tid]", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4112)); |
4113 | |
4114 | dispatch->th_disp_index = 0; |
4115 | #if OMP_45_ENABLED(50 >= 45) |
4116 | dispatch->th_doacross_buf_idx = 0; |
4117 | #endif |
4118 | if (!dispatch->th_disp_buffer) { |
4119 | dispatch->th_disp_buffer = |
4120 | (dispatch_private_info_t *)__kmp_allocate(disp_size)___kmp_allocate((disp_size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4120); |
4121 | |
4122 | if (__kmp_storage_map) { |
4123 | __kmp_print_storage_map_gtid( |
4124 | gtid, &dispatch->th_disp_buffer[0], |
4125 | &dispatch->th_disp_buffer[team->t.t_max_nproc == 1 |
4126 | ? 1 |
4127 | : __kmp_dispatch_num_buffers], |
4128 | disp_size, "th_%d.th_dispatch.th_disp_buffer " |
4129 | "(team_%d.t_dispatch[%d].th_disp_buffer)", |
4130 | gtid, team->t.t_id, gtid); |
4131 | } |
4132 | } else { |
4133 | memset(&dispatch->th_disp_buffer[0], '\0', disp_size); |
4134 | } |
4135 | |
4136 | dispatch->th_dispatch_pr_current = 0; |
4137 | dispatch->th_dispatch_sh_current = 0; |
4138 | |
4139 | dispatch->th_deo_fcn = 0; /* ORDERED */ |
4140 | dispatch->th_dxo_fcn = 0; /* END ORDERED */ |
4141 | } |
4142 | |
4143 | this_thr->th.th_next_pool = NULL__null; |
4144 | |
4145 | if (!this_thr->th.th_task_state_memo_stack) { |
4146 | size_t i; |
4147 | this_thr->th.th_task_state_memo_stack = |
4148 | (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8))___kmp_allocate((4 * sizeof(kmp_uint8)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4148); |
4149 | this_thr->th.th_task_state_top = 0; |
4150 | this_thr->th.th_task_state_stack_sz = 4; |
4151 | for (i = 0; i < this_thr->th.th_task_state_stack_sz; |
4152 | ++i) // zero init the stack |
4153 | this_thr->th.th_task_state_memo_stack[i] = 0; |
4154 | } |
4155 | |
4156 | KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here)((!this_thr->th.th_spin_here) ? 0 : __kmp_debug_assert("!this_thr->th.th_spin_here" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4156)); |
4157 | KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0)((this_thr->th.th_next_waiting == 0) ? 0 : __kmp_debug_assert ("this_thr->th.th_next_waiting == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4157)); |
4158 | |
4159 | KMP_MB(); |
4160 | } |
4161 | |
4162 | /* allocate a new thread for the requesting team. this is only called from |
4163 | within a forkjoin critical section. we will first try to get an available |
4164 | thread from the thread pool. if none is available, we will fork a new one |
4165 | assuming we are able to create a new one. this should be assured, as the |
4166 | caller should check on this first. */ |
4167 | kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, |
4168 | int new_tid) { |
4169 | kmp_team_t *serial_team; |
4170 | kmp_info_t *new_thr; |
4171 | int new_gtid; |
4172 | |
4173 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d\n" , __kmp_get_global_thread_id()); }; |
4174 | KMP_DEBUG_ASSERT(root && team)((root && team) ? 0 : __kmp_debug_assert("root && team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4174)); |
4175 | #if !KMP_NESTED_HOT_TEAMS1 |
4176 | KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()))((((((((__kmp_get_global_thread_id())) >= 0) ? 0 : __kmp_debug_assert ("((__kmp_get_global_thread_id())) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4176)), __kmp_threads[((__kmp_get_global_thread_id()))]-> th.th_info.ds.ds_tid) == 0)) ? 0 : __kmp_debug_assert("((((((__kmp_get_global_thread_id())) >= 0) ? 0 : __kmp_debug_assert(\"((__kmp_get_global_thread_id())) >= 0\", \"/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp\", 4176)), __kmp_threads[((__kmp_get_global_thread_id()))]->th.th_info.ds.ds_tid) == 0)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4176)); |
4177 | #endif |
4178 | KMP_MB(); |
4179 | |
4180 | /* first, try to get one from the thread pool */ |
4181 | if (__kmp_thread_pool) { |
4182 | |
4183 | new_thr = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); |
4184 | __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool; |
4185 | if (new_thr == __kmp_thread_pool_insert_pt) { |
4186 | __kmp_thread_pool_insert_pt = NULL__null; |
4187 | } |
4188 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); |
4189 | // Don't touch th_active_in_pool or th_active. |
4190 | // The worker thread adjusts those flags as it sleeps/awakens. |
4191 | __kmp_thread_pool_nth--; |
4192 | |
4193 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread T#%d\n" , __kmp_get_global_thread_id(), new_thr->th.th_info.ds.ds_gtid ); } |
4194 | __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread T#%d\n" , __kmp_get_global_thread_id(), new_thr->th.th_info.ds.ds_gtid ); }; |
4195 | KMP_ASSERT(!new_thr->th.th_team)((!new_thr->th.th_team) ? 0 : __kmp_debug_assert("!new_thr->th.th_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4195)); |
4196 | KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity)((__kmp_nth < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("__kmp_nth < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4196)); |
4197 | KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0)((__kmp_thread_pool_nth >= 0) ? 0 : __kmp_debug_assert("__kmp_thread_pool_nth >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4197)); |
4198 | |
4199 | /* setup the thread structure */ |
4200 | __kmp_initialize_info(new_thr, team, new_tid, |
4201 | new_thr->th.th_info.ds.ds_gtid); |
4202 | KMP_DEBUG_ASSERT(new_thr->th.th_serial_team)((new_thr->th.th_serial_team) ? 0 : __kmp_debug_assert("new_thr->th.th_serial_team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4202)); |
4203 | |
4204 | TCW_4(__kmp_nth, __kmp_nth + 1)(__kmp_nth) = (__kmp_nth + 1); |
4205 | root->r.r_cg_nthreads++; |
4206 | |
4207 | new_thr->th.th_task_state = 0; |
4208 | new_thr->th.th_task_state_top = 0; |
4209 | new_thr->th.th_task_state_stack_sz = 4; |
4210 | |
4211 | #ifdef KMP_ADJUST_BLOCKTIME1 |
4212 | /* Adjust blocktime back to zero if necessary */ |
4213 | /* Middle initialization might not have occurred yet */ |
4214 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
4215 | if (__kmp_nth > __kmp_avail_proc) { |
4216 | __kmp_zero_bt = TRUE(!0); |
4217 | } |
4218 | } |
4219 | #endif /* KMP_ADJUST_BLOCKTIME */ |
4220 | |
4221 | #if KMP_DEBUG1 |
4222 | // If thread entered pool via __kmp_free_thread, wait_flag should != |
4223 | // KMP_BARRIER_PARENT_FLAG. |
4224 | int b; |
4225 | kmp_balign_t *balign = new_thr->th.th_bar; |
4226 | for (b = 0; b < bs_last_barrier; ++b) |
4227 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4227)); |
4228 | #endif |
4229 | |
4230 | KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread %p T#%d\n" , __kmp_get_global_thread_id(), new_thr, new_thr->th.th_info .ds.ds_gtid); } |
4231 | __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d using thread %p T#%d\n" , __kmp_get_global_thread_id(), new_thr, new_thr->th.th_info .ds.ds_gtid); }; |
4232 | |
4233 | KMP_MB(); |
4234 | return new_thr; |
4235 | } |
4236 | |
4237 | /* no, well fork a new one */ |
4238 | KMP_ASSERT(__kmp_nth == __kmp_all_nth)((__kmp_nth == __kmp_all_nth) ? 0 : __kmp_debug_assert("__kmp_nth == __kmp_all_nth" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4238)); |
4239 | KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity)((__kmp_all_nth < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("__kmp_all_nth < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4239)); |
4240 | |
4241 | #if KMP_USE_MONITOR |
4242 | // If this is the first worker thread the RTL is creating, then also |
4243 | // launch the monitor thread. We try to do this as early as possible. |
4244 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { |
4245 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); |
4246 | if (!TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { |
4247 | KF_TRACE(10, ("before __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("before __kmp_create_monitor\n" ); }; |
4248 | TCW_4(__kmp_init_monitor, 1)(__kmp_init_monitor) = (1); |
4249 | __kmp_create_monitor(&__kmp_monitor); |
4250 | KF_TRACE(10, ("after __kmp_create_monitor\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after __kmp_create_monitor\n" ); }; |
4251 | #if KMP_OS_WINDOWS0 |
4252 | // AC: wait until monitor has started. This is a fix for CQ232808. |
4253 | // The reason is that if the library is loaded/unloaded in a loop with |
4254 | // small (parallel) work in between, then there is high probability that |
4255 | // monitor thread started after the library shutdown. At shutdown it is |
4256 | // too late to cope with the problem, because when the master is in |
4257 | // DllMain (process detach) the monitor has no chances to start (it is |
4258 | // blocked), and master has no means to inform the monitor that the |
4259 | // library has gone, because all the memory which the monitor can access |
4260 | // is going to be released/reset. |
4261 | while (TCR_4(__kmp_init_monitor)(__kmp_init_monitor) < 2) { |
4262 | KMP_YIELD(TRUE){ __kmp_x86_pause(); __kmp_yield(((!0))); }; |
4263 | } |
4264 | KF_TRACE(10, ("after monitor thread has started\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("after monitor thread has started\n" ); }; |
4265 | #endif |
4266 | } |
4267 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); |
4268 | } |
4269 | #endif |
4270 | |
4271 | KMP_MB(); |
4272 | for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid])((void *)(__kmp_threads[new_gtid])) != NULL__null; ++new_gtid) { |
4273 | KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity)((new_gtid < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("new_gtid < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4273)); |
4274 | } |
4275 | |
4276 | /* allocate space for it. */ |
4277 | new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t))___kmp_allocate((sizeof(kmp_info_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4277); |
4278 | |
4279 | TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr)((__kmp_threads[new_gtid])) = ((new_thr)); |
4280 | |
4281 | if (__kmp_storage_map) { |
4282 | __kmp_print_thread_storage_map(new_thr, new_gtid); |
4283 | } |
4284 | |
4285 | // add the reserve serialized team, initialized from the team's master thread |
4286 | { |
4287 | kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team); |
4288 | KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before th_serial/serial_team\n" ); }; |
4289 | new_thr->th.th_serial_team = serial_team = |
4290 | (kmp_team_t *)__kmp_allocate_team(root, 1, 1, |
4291 | #if OMPT_SUPPORT1 |
4292 | ompt_data_none, // root parallel id |
4293 | #endif |
4294 | #if OMP_40_ENABLED(50 >= 40) |
4295 | proc_bind_default, |
4296 | #endif |
4297 | &r_icvs, 0 USE_NESTED_HOT_ARG(NULL), __null); |
4298 | } |
4299 | KMP_ASSERT(serial_team)((serial_team) ? 0 : __kmp_debug_assert("serial_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4299)); |
4300 | serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for |
4301 | // execution (it is unused for now). |
4302 | serial_team->t.t_threads[0] = new_thr; |
4303 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); } |
4304 | ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); } |
4305 | new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n" , new_thr); }; |
4306 | |
4307 | /* setup the thread structures */ |
4308 | __kmp_initialize_info(new_thr, team, new_tid, new_gtid); |
4309 | |
4310 | #if USE_FAST_MEMORY3 |
4311 | __kmp_initialize_fast_memory(new_thr); |
4312 | #endif /* USE_FAST_MEMORY */ |
4313 | |
4314 | #if KMP_USE_BGET1 |
4315 | KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL)((new_thr->th.th_local.bget_data == __null) ? 0 : __kmp_debug_assert ("new_thr->th.th_local.bget_data == __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4315)); |
4316 | __kmp_initialize_bget(new_thr); |
4317 | #endif |
4318 | |
4319 | __kmp_init_random(new_thr); // Initialize random number generator |
4320 | |
4321 | /* Initialize these only once when thread is grabbed for a team allocation */ |
4322 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); } |
4323 | ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); } |
4324 | __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n" , __kmp_get_global_thread_id(), 0, 0); }; |
4325 | |
4326 | int b; |
4327 | kmp_balign_t *balign = new_thr->th.th_bar; |
4328 | for (b = 0; b < bs_last_barrier; ++b) { |
4329 | balign[b].bb.b_go = KMP_INIT_BARRIER_STATE0; |
4330 | balign[b].bb.team = NULL__null; |
4331 | balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING0; |
4332 | balign[b].bb.use_oncore_barrier = 0; |
4333 | } |
4334 | |
4335 | new_thr->th.th_spin_here = FALSE0; |
4336 | new_thr->th.th_next_waiting = 0; |
4337 | #if KMP_OS_UNIX1 |
4338 | new_thr->th.th_blocking = false; |
4339 | #endif |
4340 | |
4341 | #if OMP_40_ENABLED(50 >= 40) && KMP_AFFINITY_SUPPORTED1 |
4342 | new_thr->th.th_current_place = KMP_PLACE_UNDEFINED(-2); |
4343 | new_thr->th.th_new_place = KMP_PLACE_UNDEFINED(-2); |
4344 | new_thr->th.th_first_place = KMP_PLACE_UNDEFINED(-2); |
4345 | new_thr->th.th_last_place = KMP_PLACE_UNDEFINED(-2); |
4346 | #endif |
4347 | |
4348 | TCW_4(new_thr->th.th_in_pool, FALSE)(new_thr->th.th_in_pool) = (0); |
4349 | new_thr->th.th_active_in_pool = FALSE0; |
4350 | TCW_4(new_thr->th.th_active, TRUE)(new_thr->th.th_active) = ((!0)); |
4351 | |
4352 | /* adjust the global counters */ |
4353 | __kmp_all_nth++; |
4354 | __kmp_nth++; |
4355 | |
4356 | root->r.r_cg_nthreads++; |
4357 | |
4358 | // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low |
4359 | // numbers of procs, and method #2 (keyed API call) for higher numbers. |
4360 | if (__kmp_adjust_gtid_mode) { |
4361 | if (__kmp_all_nth >= __kmp_tls_gtid_min) { |
4362 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 2) { |
4363 | TCW_4(__kmp_gtid_mode, 2)(__kmp_gtid_mode) = (2); |
4364 | } |
4365 | } else { |
4366 | if (TCR_4(__kmp_gtid_mode)(__kmp_gtid_mode) != 1) { |
4367 | TCW_4(__kmp_gtid_mode, 1)(__kmp_gtid_mode) = (1); |
4368 | } |
4369 | } |
4370 | } |
4371 | |
4372 | #ifdef KMP_ADJUST_BLOCKTIME1 |
4373 | /* Adjust blocktime back to zero if necessary */ |
4374 | /* Middle initialization might not have occurred yet */ |
4375 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
4376 | if (__kmp_nth > __kmp_avail_proc) { |
4377 | __kmp_zero_bt = TRUE(!0); |
4378 | } |
4379 | } |
4380 | #endif /* KMP_ADJUST_BLOCKTIME */ |
4381 | |
4382 | /* actually fork it and create the new worker thread */ |
4383 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before __kmp_create_worker: %p\n" , new_thr); } |
4384 | 10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: before __kmp_create_worker: %p\n" , new_thr); }; |
4385 | __kmp_create_worker(new_gtid, new_thr, __kmp_stksize); |
4386 | KF_TRACE(10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after __kmp_create_worker: %p\n" , new_thr); } |
4387 | ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_thread: after __kmp_create_worker: %p\n" , new_thr); }; |
4388 | |
4389 | KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d forked T#%d\n" , __kmp_get_global_thread_id(), new_gtid); } |
4390 | new_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_thread: T#%d forked T#%d\n" , __kmp_get_global_thread_id(), new_gtid); }; |
4391 | KMP_MB(); |
4392 | return new_thr; |
4393 | } |
4394 | |
4395 | /* Reinitialize team for reuse. |
4396 | The hot team code calls this case at every fork barrier, so EPCC barrier |
4397 | test are extremely sensitive to changes in it, esp. writes to the team |
4398 | struct, which cause a cache invalidation in all threads. |
4399 | IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */ |
4400 | static void __kmp_reinitialize_team(kmp_team_t *team, |
4401 | kmp_internal_control_t *new_icvs, |
4402 | ident_t *loc) { |
4403 | KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n" , team->t.t_threads[0], team); } |
4404 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n" , team->t.t_threads[0], team); }; |
4405 | KMP_DEBUG_ASSERT(team && new_icvs)((team && new_icvs) ? 0 : __kmp_debug_assert("team && new_icvs" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4405)); |
4406 | KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc)(((!(__kmp_init_parallel)) || new_icvs->nproc) ? 0 : __kmp_debug_assert ("(!(__kmp_init_parallel)) || new_icvs->nproc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4406)); |
4407 | KMP_CHECK_UPDATE(team->t.t_ident, loc)if ((team->t.t_ident) != (loc)) (team->t.t_ident) = (loc ); |
4408 | |
4409 | KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID())if ((team->t.t_id) != ((~0))) (team->t.t_id) = ((~0)); |
4410 | // Copy ICVs to the master thread's implicit taskdata |
4411 | __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE0); |
4412 | copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs); |
4413 | |
4414 | KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n" , team->t.t_threads[0], team); } |
4415 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n" , team->t.t_threads[0], team); }; |
4416 | } |
4417 | |
4418 | /* Initialize the team data structure. |
4419 | This assumes the t_threads and t_max_nproc are already set. |
4420 | Also, we don't touch the arguments */ |
4421 | static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, |
4422 | kmp_internal_control_t *new_icvs, |
4423 | ident_t *loc) { |
4424 | KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_team: enter: team=%p\n" , team); }; |
4425 | |
4426 | /* verify */ |
4427 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4427)); |
4428 | KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc)((new_nproc <= team->t.t_max_nproc) ? 0 : __kmp_debug_assert ("new_nproc <= team->t.t_max_nproc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4428)); |
4429 | KMP_DEBUG_ASSERT(team->t.t_threads)((team->t.t_threads) ? 0 : __kmp_debug_assert("team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4429)); |
4430 | KMP_MB(); |
4431 | |
4432 | team->t.t_master_tid = 0; /* not needed */ |
4433 | /* team->t.t_master_bar; not needed */ |
4434 | team->t.t_serialized = new_nproc > 1 ? 0 : 1; |
4435 | team->t.t_nproc = new_nproc; |
4436 | |
4437 | /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */ |
4438 | team->t.t_next_pool = NULL__null; |
4439 | /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess |
4440 | * up hot team */ |
4441 | |
4442 | TCW_SYNC_PTR(team->t.t_pkfn, NULL)((team->t.t_pkfn)) = ((__null)); /* not needed */ |
4443 | team->t.t_invoke = NULL__null; /* not needed */ |
4444 | |
4445 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; |
4446 | team->t.t_sched.sched = new_icvs->sched.sched; |
4447 | |
4448 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
4449 | team->t.t_fp_control_saved = FALSE0; /* not needed */ |
4450 | team->t.t_x87_fpu_control_word = 0; /* not needed */ |
4451 | team->t.t_mxcsr = 0; /* not needed */ |
4452 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
4453 | |
4454 | team->t.t_construct = 0; |
4455 | |
4456 | team->t.t_ordered.dt.t_value = 0; |
4457 | team->t.t_master_active = FALSE0; |
4458 | |
4459 | memset(&team->t.t_taskq, '\0', sizeof(kmp_taskq_t)); |
4460 | |
4461 | #ifdef KMP_DEBUG1 |
4462 | team->t.t_copypriv_data = NULL__null; /* not necessary, but nice for debugging */ |
4463 | #endif |
4464 | #if KMP_OS_WINDOWS0 |
4465 | team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */ |
4466 | #endif |
4467 | |
4468 | team->t.t_control_stack_top = NULL__null; |
4469 | |
4470 | __kmp_reinitialize_team(team, new_icvs, loc); |
4471 | |
4472 | KMP_MB(); |
4473 | KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_initialize_team: exit: team=%p\n" , team); }; |
4474 | } |
4475 | |
4476 | #if KMP_OS_LINUX1 && KMP_AFFINITY_SUPPORTED1 |
4477 | /* Sets full mask for thread and returns old mask, no changes to structures. */ |
4478 | static void |
4479 | __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) { |
4480 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { |
4481 | int status; |
4482 | if (old_mask != NULL__null) { |
4483 | status = __kmp_get_system_affinity(old_mask, TRUE)(old_mask)->get_system_affinity((!0)); |
4484 | int error = errno(*__errno_location ()); |
4485 | if (status != 0) { |
4486 | __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError)__kmp_msg_format(kmp_i18n_msg_ChangeThreadAffMaskError), KMP_ERR(error)__kmp_msg_error_code(error), |
4487 | __kmp_msg_null); |
4488 | } |
4489 | } |
4490 | __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE)(__kmp_affin_fullMask)->set_system_affinity((!0)); |
4491 | } |
4492 | } |
4493 | #endif |
4494 | |
4495 | #if OMP_40_ENABLED(50 >= 40) && KMP_AFFINITY_SUPPORTED1 |
4496 | |
4497 | // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism. |
4498 | // It calculats the worker + master thread's partition based upon the parent |
4499 | // thread's partition, and binds each worker to a thread in their partition. |
4500 | // The master thread's partition should already include its current binding. |
4501 | static void __kmp_partition_places(kmp_team_t *team, int update_master_only) { |
4502 | // Copy the master thread's place partion to the team struct |
4503 | kmp_info_t *master_th = team->t.t_threads[0]; |
4504 | KMP_DEBUG_ASSERT(master_th != NULL)((master_th != __null) ? 0 : __kmp_debug_assert("master_th != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4504)); |
4505 | kmp_proc_bind_t proc_bind = team->t.t_proc_bind; |
4506 | int first_place = master_th->th.th_first_place; |
4507 | int last_place = master_th->th.th_last_place; |
4508 | int masters_place = master_th->th.th_current_place; |
4509 | team->t.t_first_place = first_place; |
4510 | team->t.t_last_place = last_place; |
4511 | |
4512 | KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, ((((team ->t.t_threads[0]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[0]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4514)), (team->t.t_threads[0])->th.th_info.ds.ds_gtid ), team->t.t_id, masters_place, first_place, last_place); } |
4513 | "bound to place %d partition = [%d,%d]\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, ((((team ->t.t_threads[0]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[0]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4514)), (team->t.t_threads[0])->th.th_info.ds.ds_gtid ), team->t.t_id, masters_place, first_place, last_place); } |
4514 | proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, ((((team ->t.t_threads[0]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[0]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4514)), (team->t.t_threads[0])->th.th_info.ds.ds_gtid ), team->t.t_id, masters_place, first_place, last_place); } |
4515 | team->t.t_id, masters_place, first_place, last_place))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " "bound to place %d partition = [%d,%d]\n", proc_bind, ((((team ->t.t_threads[0]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[0]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4514)), (team->t.t_threads[0])->th.th_info.ds.ds_gtid ), team->t.t_id, masters_place, first_place, last_place); }; |
4516 | |
4517 | switch (proc_bind) { |
4518 | |
4519 | case proc_bind_default: |
4520 | // serial teams might have the proc_bind policy set to proc_bind_default. It |
4521 | // doesn't matter, as we don't rebind master thread for any proc_bind policy |
4522 | KMP_DEBUG_ASSERT(team->t.t_nproc == 1)((team->t.t_nproc == 1) ? 0 : __kmp_debug_assert("team->t.t_nproc == 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4522)); |
4523 | break; |
4524 | |
4525 | case proc_bind_master: { |
4526 | int f; |
4527 | int n_th = team->t.t_nproc; |
4528 | for (f = 1; f < n_th; f++) { |
4529 | kmp_info_t *th = team->t.t_threads[f]; |
4530 | KMP_DEBUG_ASSERT(th != NULL)((th != __null) ? 0 : __kmp_debug_assert("th != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4530)); |
4531 | th->th.th_first_place = first_place; |
4532 | th->th.th_last_place = last_place; |
4533 | th->th.th_new_place = masters_place; |
4534 | |
4535 | KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: master: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4537)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, masters_place, first_place, last_place ); } |
4536 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: master: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4537)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, masters_place, first_place, last_place ); } |
4537 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: master: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4537)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, masters_place, first_place, last_place ); } |
4538 | f, masters_place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: master: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4537)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, masters_place, first_place, last_place ); }; |
4539 | } |
4540 | } break; |
4541 | |
4542 | case proc_bind_close: { |
4543 | int f; |
4544 | int n_th = team->t.t_nproc; |
4545 | int n_places; |
4546 | if (first_place <= last_place) { |
4547 | n_places = last_place - first_place + 1; |
4548 | } else { |
4549 | n_places = __kmp_affinity_num_masks - first_place + last_place + 1; |
4550 | } |
4551 | if (n_th <= n_places) { |
4552 | int place = masters_place; |
4553 | for (f = 1; f < n_th; f++) { |
4554 | kmp_info_t *th = team->t.t_threads[f]; |
4555 | KMP_DEBUG_ASSERT(th != NULL)((th != __null) ? 0 : __kmp_debug_assert("th != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4555)); |
4556 | |
4557 | if (place == last_place) { |
4558 | place = first_place; |
4559 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4560 | place = 0; |
4561 | } else { |
4562 | place++; |
4563 | } |
4564 | th->th.th_first_place = first_place; |
4565 | th->th.th_last_place = last_place; |
4566 | th->th.th_new_place = place; |
4567 | |
4568 | KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4570)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, place, first_place, last_place); } |
4569 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4570)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, place, first_place, last_place); } |
4570 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4570)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, place, first_place, last_place); } |
4571 | team->t.t_id, f, place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4570)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, place, first_place, last_place); }; |
4572 | } |
4573 | } else { |
4574 | int S, rem, gap, s_count; |
4575 | S = n_th / n_places; |
4576 | s_count = 0; |
4577 | rem = n_th - (S * n_places); |
4578 | gap = rem > 0 ? n_places / rem : n_places; |
4579 | int place = masters_place; |
4580 | int gap_ct = gap; |
4581 | for (f = 0; f < n_th; f++) { |
4582 | kmp_info_t *th = team->t.t_threads[f]; |
4583 | KMP_DEBUG_ASSERT(th != NULL)((th != __null) ? 0 : __kmp_debug_assert("th != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4583)); |
4584 | |
4585 | th->th.th_first_place = first_place; |
4586 | th->th.th_last_place = last_place; |
4587 | th->th.th_new_place = place; |
4588 | s_count++; |
4589 | |
4590 | if ((s_count == S) && rem && (gap_ct == gap)) { |
4591 | // do nothing, add an extra thread to place on next iteration |
4592 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { |
4593 | // we added an extra thread to this place; move to next place |
4594 | if (place == last_place) { |
4595 | place = first_place; |
4596 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4597 | place = 0; |
4598 | } else { |
4599 | place++; |
4600 | } |
4601 | s_count = 0; |
4602 | gap_ct = 1; |
4603 | rem--; |
4604 | } else if (s_count == S) { // place full; don't add extra |
4605 | if (place == last_place) { |
4606 | place = first_place; |
4607 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4608 | place = 0; |
4609 | } else { |
4610 | place++; |
4611 | } |
4612 | gap_ct++; |
4613 | s_count = 0; |
4614 | } |
4615 | |
4616 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4619)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, first_place, last_place ); } |
4617 | ("__kmp_partition_places: close: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4619)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, first_place, last_place ); } |
4618 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4619)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, first_place, last_place ); } |
4619 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4619)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, first_place, last_place ); } |
4620 | th->th.th_new_place, first_place, last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: close: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4619)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, first_place, last_place ); }; |
4621 | } |
4622 | KMP_DEBUG_ASSERT(place == masters_place)((place == masters_place) ? 0 : __kmp_debug_assert("place == masters_place" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4622)); |
4623 | } |
4624 | } break; |
4625 | |
4626 | case proc_bind_spread: { |
4627 | int f; |
4628 | int n_th = team->t.t_nproc; |
4629 | int n_places; |
4630 | int thidx; |
4631 | if (first_place <= last_place) { |
4632 | n_places = last_place - first_place + 1; |
4633 | } else { |
4634 | n_places = __kmp_affinity_num_masks - first_place + last_place + 1; |
4635 | } |
4636 | if (n_th <= n_places) { |
4637 | int place = -1; |
4638 | |
4639 | if (n_places != static_cast<int>(__kmp_affinity_num_masks)) { |
4640 | int S = n_places / n_th; |
4641 | int s_count, rem, gap, gap_ct; |
4642 | |
4643 | place = masters_place; |
4644 | rem = n_places - n_th * S; |
4645 | gap = rem ? n_th / rem : 1; |
4646 | gap_ct = gap; |
4647 | thidx = n_th; |
4648 | if (update_master_only == 1) |
4649 | thidx = 1; |
4650 | for (f = 0; f < thidx; f++) { |
4651 | kmp_info_t *th = team->t.t_threads[f]; |
4652 | KMP_DEBUG_ASSERT(th != NULL)((th != __null) ? 0 : __kmp_debug_assert("th != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4652)); |
4653 | |
4654 | th->th.th_first_place = place; |
4655 | th->th.th_new_place = place; |
4656 | s_count = 1; |
4657 | while (s_count < S) { |
4658 | if (place == last_place) { |
4659 | place = first_place; |
4660 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4661 | place = 0; |
4662 | } else { |
4663 | place++; |
4664 | } |
4665 | s_count++; |
4666 | } |
4667 | if (rem && (gap_ct == gap)) { |
4668 | if (place == last_place) { |
4669 | place = first_place; |
4670 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4671 | place = 0; |
4672 | } else { |
4673 | place++; |
4674 | } |
4675 | rem--; |
4676 | gap_ct = 0; |
4677 | } |
4678 | th->th.th_last_place = place; |
4679 | gap_ct++; |
4680 | |
4681 | if (place == last_place) { |
4682 | place = first_place; |
4683 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4684 | place = 0; |
4685 | } else { |
4686 | place++; |
4687 | } |
4688 | |
4689 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); } |
4690 | ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); } |
4691 | "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); } |
4692 | __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); } |
4693 | f, th->th.th_new_place, th->th.th_first_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); } |
4694 | th->th.th_last_place, __kmp_affinity_num_masks))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", ((((team ->t.t_threads[f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4692)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, __kmp_affinity_num_masks); }; |
4695 | } |
4696 | } else { |
4697 | /* Having uniform space of available computation places I can create |
4698 | T partitions of round(P/T) size and put threads into the first |
4699 | place of each partition. */ |
4700 | double current = static_cast<double>(masters_place); |
4701 | double spacing = |
4702 | (static_cast<double>(n_places + 1) / static_cast<double>(n_th)); |
4703 | int first, last; |
4704 | kmp_info_t *th; |
4705 | |
4706 | thidx = n_th + 1; |
4707 | if (update_master_only == 1) |
4708 | thidx = 1; |
4709 | for (f = 0; f < thidx; f++) { |
4710 | first = static_cast<int>(current); |
4711 | last = static_cast<int>(current + spacing) - 1; |
4712 | KMP_DEBUG_ASSERT(last >= first)((last >= first) ? 0 : __kmp_debug_assert("last >= first" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4712)); |
4713 | if (first >= n_places) { |
4714 | if (masters_place) { |
4715 | first -= n_places; |
4716 | last -= n_places; |
4717 | if (first == (masters_place + 1)) { |
4718 | KMP_DEBUG_ASSERT(f == n_th)((f == n_th) ? 0 : __kmp_debug_assert("f == n_th", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4718)); |
4719 | first--; |
4720 | } |
4721 | if (last == masters_place) { |
4722 | KMP_DEBUG_ASSERT(f == (n_th - 1))((f == (n_th - 1)) ? 0 : __kmp_debug_assert("f == (n_th - 1)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4722)); |
4723 | last--; |
4724 | } |
4725 | } else { |
4726 | KMP_DEBUG_ASSERT(f == n_th)((f == n_th) ? 0 : __kmp_debug_assert("f == n_th", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4726)); |
4727 | first = 0; |
4728 | last = 0; |
4729 | } |
4730 | } |
4731 | if (last >= n_places) { |
4732 | last = (n_places - 1); |
4733 | } |
4734 | place = first; |
4735 | current += spacing; |
4736 | if (f < n_th) { |
4737 | KMP_DEBUG_ASSERT(0 <= first)((0 <= first) ? 0 : __kmp_debug_assert("0 <= first", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4737)); |
4738 | KMP_DEBUG_ASSERT(n_places > first)((n_places > first) ? 0 : __kmp_debug_assert("n_places > first" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4738)); |
4739 | KMP_DEBUG_ASSERT(0 <= last)((0 <= last) ? 0 : __kmp_debug_assert("0 <= last", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4739)); |
4740 | KMP_DEBUG_ASSERT(n_places > last)((n_places > last) ? 0 : __kmp_debug_assert("n_places > last" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4740)); |
4741 | KMP_DEBUG_ASSERT(last_place >= first_place)((last_place >= first_place) ? 0 : __kmp_debug_assert("last_place >= first_place" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4741)); |
4742 | th = team->t.t_threads[f]; |
4743 | KMP_DEBUG_ASSERT(th)((th) ? 0 : __kmp_debug_assert("th", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4743)); |
4744 | th->th.th_first_place = first; |
4745 | th->th.th_new_place = place; |
4746 | th->th.th_last_place = last; |
4747 | |
4748 | KA_TRACE(100,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); } |
4749 | ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); } |
4750 | "partition = [%d,%d], spacing = %.4f\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); } |
4751 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); } |
4752 | team->t.t_id, f, th->th.th_new_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); } |
4753 | th->th.th_first_place, th->th.th_last_place, spacing))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d], spacing = %.4f\n", ((((team->t.t_threads [f]) != __null) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4751)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place, spacing); }; |
4754 | } |
4755 | } |
4756 | } |
4757 | KMP_DEBUG_ASSERT(update_master_only || place == masters_place)((update_master_only || place == masters_place) ? 0 : __kmp_debug_assert ("update_master_only || place == masters_place", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4757)); |
4758 | } else { |
4759 | int S, rem, gap, s_count; |
4760 | S = n_th / n_places; |
4761 | s_count = 0; |
4762 | rem = n_th - (S * n_places); |
4763 | gap = rem > 0 ? n_places / rem : n_places; |
4764 | int place = masters_place; |
4765 | int gap_ct = gap; |
4766 | thidx = n_th; |
4767 | if (update_master_only == 1) |
4768 | thidx = 1; |
4769 | for (f = 0; f < thidx; f++) { |
4770 | kmp_info_t *th = team->t.t_threads[f]; |
4771 | KMP_DEBUG_ASSERT(th != NULL)((th != __null) ? 0 : __kmp_debug_assert("th != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4771)); |
4772 | |
4773 | th->th.th_first_place = place; |
4774 | th->th.th_last_place = place; |
4775 | th->th.th_new_place = place; |
4776 | s_count++; |
4777 | |
4778 | if ((s_count == S) && rem && (gap_ct == gap)) { |
4779 | // do nothing, add an extra thread to place on next iteration |
4780 | } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { |
4781 | // we added an extra thread to this place; move on to next place |
4782 | if (place == last_place) { |
4783 | place = first_place; |
4784 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4785 | place = 0; |
4786 | } else { |
4787 | place++; |
4788 | } |
4789 | s_count = 0; |
4790 | gap_ct = 1; |
4791 | rem--; |
4792 | } else if (s_count == S) { // place is full; don't add extra thread |
4793 | if (place == last_place) { |
4794 | place = first_place; |
4795 | } else if (place == (int)(__kmp_affinity_num_masks - 1)) { |
4796 | place = 0; |
4797 | } else { |
4798 | place++; |
4799 | } |
4800 | gap_ct++; |
4801 | s_count = 0; |
4802 | } |
4803 | |
4804 | KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4806)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } |
4805 | "partition = [%d,%d]\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4806)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } |
4806 | __kmp_gtid_from_thread(team->t.t_threads[f]),if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4806)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } |
4807 | team->t.t_id, f, th->th.th_new_place,if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4806)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); } |
4808 | th->th.th_first_place, th->th.th_last_place))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " "partition = [%d,%d]\n", ((((team->t.t_threads[f]) != __null ) ? 0 : __kmp_debug_assert("(team->t.t_threads[f]) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4806)), (team->t.t_threads[f])->th.th_info.ds.ds_gtid ), team->t.t_id, f, th->th.th_new_place, th->th.th_first_place , th->th.th_last_place); }; |
4809 | } |
4810 | KMP_DEBUG_ASSERT(update_master_only || place == masters_place)((update_master_only || place == masters_place) ? 0 : __kmp_debug_assert ("update_master_only || place == masters_place", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4810)); |
4811 | } |
4812 | } break; |
4813 | |
4814 | default: |
4815 | break; |
4816 | } |
4817 | |
4818 | KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_partition_places: exit T#%d\n" , team->t.t_id); }; |
4819 | } |
4820 | |
4821 | #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */ |
4822 | |
4823 | /* allocate a new team data structure to use. take one off of the free pool if |
4824 | available */ |
4825 | kmp_team_t * |
4826 | __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, |
4827 | #if OMPT_SUPPORT1 |
4828 | ompt_data_t ompt_parallel_data, |
4829 | #endif |
4830 | #if OMP_40_ENABLED(50 >= 40) |
4831 | kmp_proc_bind_t new_proc_bind, |
4832 | #endif |
4833 | kmp_internal_control_t *new_icvs, |
4834 | int argc USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { |
4835 | KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team)((void)0); |
4836 | int f; |
4837 | kmp_team_t *team; |
4838 | int use_hot_team = !root->r.r_active; |
4839 | int level = 0; |
4840 | |
4841 | KA_TRACE(20, ("__kmp_allocate_team: called\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: called\n" ); }; |
4842 | KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0)((new_nproc >= 1 && argc >= 0) ? 0 : __kmp_debug_assert ("new_nproc >= 1 && argc >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4842)); |
4843 | KMP_DEBUG_ASSERT(max_nproc >= new_nproc)((max_nproc >= new_nproc) ? 0 : __kmp_debug_assert("max_nproc >= new_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4843)); |
4844 | KMP_MB(); |
4845 | |
4846 | #if KMP_NESTED_HOT_TEAMS1 |
4847 | kmp_hot_team_ptr_t *hot_teams; |
4848 | if (master) { |
4849 | team = master->th.th_team; |
4850 | level = team->t.t_active_level; |
4851 | if (master->th.th_teams_microtask) { // in teams construct? |
4852 | if (master->th.th_teams_size.nteams > 1 && |
4853 | ( // #teams > 1 |
4854 | team->t.t_pkfn == |
4855 | (microtask_t)__kmp_teams_master || // inner fork of the teams |
4856 | master->th.th_teams_level < |
4857 | team->t.t_level)) { // or nested parallel inside the teams |
4858 | ++level; // not increment if #teams==1, or for outer fork of the teams; |
4859 | // increment otherwise |
4860 | } |
4861 | } |
4862 | hot_teams = master->th.th_hot_teams; |
4863 | if (level < __kmp_hot_teams_max_level && hot_teams && |
4864 | hot_teams[level] |
4865 | .hot_team) { // hot team has already been allocated for given level |
4866 | use_hot_team = 1; |
4867 | } else { |
4868 | use_hot_team = 0; |
4869 | } |
4870 | } |
4871 | #endif |
4872 | // Optimization to use a "hot" team |
4873 | if (use_hot_team && new_nproc > 1) { |
4874 | KMP_DEBUG_ASSERT(new_nproc == max_nproc)((new_nproc == max_nproc) ? 0 : __kmp_debug_assert("new_nproc == max_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4874)); |
4875 | #if KMP_NESTED_HOT_TEAMS1 |
4876 | team = hot_teams[level].hot_team; |
4877 | #else |
4878 | team = root->r.r_hot_team; |
4879 | #endif |
4880 | #if KMP_DEBUG1 |
4881 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
4882 | KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); } |
4883 | "task_team[1] = %p before reinit\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); } |
4884 | team->t.t_task_team[0], team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p before reinit\n", team->t.t_task_team[ 0], team->t.t_task_team[1]); }; |
4885 | } |
4886 | #endif |
4887 | |
4888 | // Has the number of threads changed? |
4889 | /* Let's assume the most common case is that the number of threads is |
4890 | unchanged, and put that case first. */ |
4891 | if (team->t.t_nproc == new_nproc) { // Check changes in number of threads |
4892 | KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team\n" ); }; |
4893 | // This case can mean that omp_set_num_threads() was called and the hot |
4894 | // team size was already reduced, so we check the special flag |
4895 | if (team->t.t_size_changed == -1) { |
4896 | team->t.t_size_changed = 1; |
4897 | } else { |
4898 | KMP_CHECK_UPDATE(team->t.t_size_changed, 0)if ((team->t.t_size_changed) != (0)) (team->t.t_size_changed ) = (0); |
4899 | } |
4900 | |
4901 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; |
4902 | kmp_r_sched_t new_sched = new_icvs->sched; |
4903 | // set master's schedule as new run-time schedule |
4904 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched)if ((team->t.t_sched.sched) != (new_sched.sched)) (team-> t.t_sched.sched) = (new_sched.sched); |
4905 | |
4906 | __kmp_reinitialize_team(team, new_icvs, |
4907 | root->r.r_uber_thread->th.th_ident); |
4908 | |
4909 | KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); } |
4910 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); }; |
4911 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); |
4912 | |
4913 | #if OMP_40_ENABLED(50 >= 40) |
4914 | #if KMP_AFFINITY_SUPPORTED1 |
4915 | if ((team->t.t_size_changed == 0) && |
4916 | (team->t.t_proc_bind == new_proc_bind)) { |
4917 | if (new_proc_bind == proc_bind_spread) { |
4918 | __kmp_partition_places( |
4919 | team, 1); // add flag to update only master for spread |
4920 | } |
4921 | KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } |
4922 | "proc_bind = %d, partition = [%d,%d]\n",if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } |
4923 | team->t.t_id, new_proc_bind, team->t.t_first_place,if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); } |
4924 | team->t.t_last_place))if (kmp_a_debug >= 200) { __kmp_debug_printf ("__kmp_allocate_team: reusing hot team #%d bindings: " "proc_bind = %d, partition = [%d,%d]\n", team->t.t_id, new_proc_bind , team->t.t_first_place, team->t.t_last_place); }; |
4925 | } else { |
4926 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); |
4927 | __kmp_partition_places(team); |
4928 | } |
4929 | #else |
4930 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); |
4931 | #endif /* KMP_AFFINITY_SUPPORTED */ |
4932 | #endif /* OMP_40_ENABLED */ |
4933 | } else if (team->t.t_nproc > new_nproc) { |
4934 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); } |
4935 | ("__kmp_allocate_team: decreasing hot team thread count to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); } |
4936 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: decreasing hot team thread count to %d\n" , new_nproc); }; |
4937 | |
4938 | team->t.t_size_changed = 1; |
4939 | #if KMP_NESTED_HOT_TEAMS1 |
4940 | if (__kmp_hot_teams_mode == 0) { |
4941 | // AC: saved number of threads should correspond to team's value in this |
4942 | // mode, can be bigger in mode 1, when hot team has threads in reserve |
4943 | KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc)((hot_teams[level].hot_team_nth == team->t.t_nproc) ? 0 : __kmp_debug_assert ("hot_teams[level].hot_team_nth == team->t.t_nproc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4943)); |
4944 | hot_teams[level].hot_team_nth = new_nproc; |
4945 | #endif // KMP_NESTED_HOT_TEAMS |
4946 | /* release the extra threads we don't need any more */ |
4947 | for (f = new_nproc; f < team->t.t_nproc; f++) { |
4948 | KMP_DEBUG_ASSERT(team->t.t_threads[f])((team->t.t_threads[f]) ? 0 : __kmp_debug_assert("team->t.t_threads[f]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4948)); |
4949 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
4950 | // When decreasing team size, threads no longer in the team should |
4951 | // unref task team. |
4952 | team->t.t_threads[f]->th.th_task_team = NULL__null; |
4953 | } |
4954 | __kmp_free_thread(team->t.t_threads[f]); |
4955 | team->t.t_threads[f] = NULL__null; |
4956 | } |
4957 | #if KMP_NESTED_HOT_TEAMS1 |
4958 | } // (__kmp_hot_teams_mode == 0) |
4959 | else { |
4960 | // When keeping extra threads in team, switch threads to wait on own |
4961 | // b_go flag |
4962 | for (f = new_nproc; f < team->t.t_nproc; ++f) { |
4963 | KMP_DEBUG_ASSERT(team->t.t_threads[f])((team->t.t_threads[f]) ? 0 : __kmp_debug_assert("team->t.t_threads[f]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4963)); |
4964 | kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar; |
4965 | for (int b = 0; b < bs_last_barrier; ++b) { |
4966 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) { |
4967 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; |
4968 | } |
4969 | KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0)if ((balign[b].bb.leaf_kids) != (0)) (balign[b].bb.leaf_kids) = (0); |
4970 | } |
4971 | } |
4972 | } |
4973 | #endif // KMP_NESTED_HOT_TEAMS |
4974 | team->t.t_nproc = new_nproc; |
4975 | // TODO???: team->t.t_max_active_levels = new_max_active_levels; |
4976 | KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched)if ((team->t.t_sched.sched) != (new_icvs->sched.sched)) (team->t.t_sched.sched) = (new_icvs->sched.sched); |
4977 | __kmp_reinitialize_team(team, new_icvs, |
4978 | root->r.r_uber_thread->th.th_ident); |
4979 | |
4980 | /* update the remaining threads */ |
4981 | for (f = 0; f < new_nproc; ++f) { |
4982 | team->t.t_threads[f]->th.th_team_nproc = new_nproc; |
4983 | } |
4984 | // restore the current task state of the master thread: should be the |
4985 | // implicit task |
4986 | KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); } |
4987 | team->t.t_threads[0], team))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n" , 0, team->t.t_threads[0], team); }; |
4988 | |
4989 | __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); |
4990 | |
4991 | #ifdef KMP_DEBUG1 |
4992 | for (f = 0; f < team->t.t_nproc; f++) { |
4993 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4995)) |
4994 | team->t.t_threads[f]->th.th_team_nproc ==((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4995)) |
4995 | team->t.t_nproc)((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 4995)); |
4996 | } |
4997 | #endif |
4998 | |
4999 | #if OMP_40_ENABLED(50 >= 40) |
5000 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); |
5001 | #if KMP_AFFINITY_SUPPORTED1 |
5002 | __kmp_partition_places(team); |
5003 | #endif |
5004 | #endif |
5005 | } else { // team->t.t_nproc < new_nproc |
5006 | #if KMP_OS_LINUX1 && KMP_AFFINITY_SUPPORTED1 |
5007 | kmp_affin_mask_t *old_mask; |
5008 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { |
5009 | KMP_CPU_ALLOC(old_mask)(old_mask = __kmp_affinity_dispatch->allocate_mask()); |
5010 | } |
5011 | #endif |
5012 | |
5013 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); } |
5014 | ("__kmp_allocate_team: increasing hot team thread count to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); } |
5015 | new_nproc))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: increasing hot team thread count to %d\n" , new_nproc); }; |
5016 | |
5017 | team->t.t_size_changed = 1; |
5018 | |
5019 | #if KMP_NESTED_HOT_TEAMS1 |
5020 | int avail_threads = hot_teams[level].hot_team_nth; |
5021 | if (new_nproc < avail_threads) |
5022 | avail_threads = new_nproc; |
5023 | kmp_info_t **other_threads = team->t.t_threads; |
5024 | for (f = team->t.t_nproc; f < avail_threads; ++f) { |
5025 | // Adjust barrier data of reserved threads (if any) of the team |
5026 | // Other data will be set in __kmp_initialize_info() below. |
5027 | int b; |
5028 | kmp_balign_t *balign = other_threads[f]->th.th_bar; |
5029 | for (b = 0; b < bs_last_barrier; ++b) { |
5030 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; |
5031 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5031)); |
5032 | #if USE_DEBUGGER0 |
5033 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; |
5034 | #endif |
5035 | } |
5036 | } |
5037 | if (hot_teams[level].hot_team_nth >= new_nproc) { |
5038 | // we have all needed threads in reserve, no need to allocate any |
5039 | // this only possible in mode 1, cannot have reserved threads in mode 0 |
5040 | KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1)((__kmp_hot_teams_mode == 1) ? 0 : __kmp_debug_assert("__kmp_hot_teams_mode == 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5040)); |
5041 | team->t.t_nproc = new_nproc; // just get reserved threads involved |
5042 | } else { |
5043 | // we may have some threads in reserve, but not enough |
5044 | team->t.t_nproc = |
5045 | hot_teams[level] |
5046 | .hot_team_nth; // get reserved threads involved if any |
5047 | hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size |
5048 | #endif // KMP_NESTED_HOT_TEAMS |
5049 | if (team->t.t_max_nproc < new_nproc) { |
5050 | /* reallocate larger arrays */ |
5051 | __kmp_reallocate_team_arrays(team, new_nproc); |
5052 | __kmp_reinitialize_team(team, new_icvs, NULL__null); |
5053 | } |
5054 | |
5055 | #if KMP_OS_LINUX1 && KMP_AFFINITY_SUPPORTED1 |
5056 | /* Temporarily set full mask for master thread before creation of |
5057 | workers. The reason is that workers inherit the affinity from master, |
5058 | so if a lot of workers are created on the single core quickly, they |
5059 | don't get a chance to set their own affinity for a long time. */ |
5060 | __kmp_set_thread_affinity_mask_full_tmp(old_mask); |
5061 | #endif |
5062 | |
5063 | /* allocate new threads for the hot team */ |
5064 | for (f = team->t.t_nproc; f < new_nproc; f++) { |
5065 | kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f); |
5066 | KMP_DEBUG_ASSERT(new_worker)((new_worker) ? 0 : __kmp_debug_assert("new_worker", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5066)); |
5067 | team->t.t_threads[f] = new_worker; |
5068 | |
5069 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
5070 | ("__kmp_allocate_team: team %d init T#%d arrived: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
5071 | "join=%llu, plain=%llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
5072 | team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
5073 | team->t.t_bar[bs_forkjoin_barrier].b_arrived,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); } |
5074 | team->t.t_bar[bs_plain_barrier].b_arrived))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init T#%d arrived: " "join=%llu, plain=%llu\n", team->t.t_id, ((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert("(f) >= 0 && (team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5072)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid ), team->t.t_id, f, team->t.t_bar[bs_forkjoin_barrier]. b_arrived, team->t.t_bar[bs_plain_barrier].b_arrived); }; |
5075 | |
5076 | { // Initialize barrier data for new threads. |
5077 | int b; |
5078 | kmp_balign_t *balign = new_worker->th.th_bar; |
5079 | for (b = 0; b < bs_last_barrier; ++b) { |
5080 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; |
5081 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5082)) |
5082 | KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5082)); |
5083 | #if USE_DEBUGGER0 |
5084 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; |
5085 | #endif |
5086 | } |
5087 | } |
5088 | } |
5089 | |
5090 | #if KMP_OS_LINUX1 && KMP_AFFINITY_SUPPORTED1 |
5091 | if (KMP_AFFINITY_CAPABLE()(__kmp_affin_mask_size > 0)) { |
5092 | /* Restore initial master thread's affinity mask */ |
5093 | __kmp_set_system_affinity(old_mask, TRUE)(old_mask)->set_system_affinity((!0)); |
5094 | KMP_CPU_FREE(old_mask)__kmp_affinity_dispatch->deallocate_mask(old_mask); |
5095 | } |
5096 | #endif |
5097 | #if KMP_NESTED_HOT_TEAMS1 |
5098 | } // end of check of t_nproc vs. new_nproc vs. hot_team_nth |
5099 | #endif // KMP_NESTED_HOT_TEAMS |
5100 | /* make sure everyone is syncronized */ |
5101 | int old_nproc = team->t.t_nproc; // save old value and use to update only |
5102 | // new threads below |
5103 | __kmp_initialize_team(team, new_nproc, new_icvs, |
5104 | root->r.r_uber_thread->th.th_ident); |
5105 | |
5106 | /* reinitialize the threads */ |
5107 | KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc)((team->t.t_nproc == new_nproc) ? 0 : __kmp_debug_assert("team->t.t_nproc == new_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5107)); |
5108 | for (f = 0; f < team->t.t_nproc; ++f) |
5109 | __kmp_initialize_info(team->t.t_threads[f], team, f, |
5110 | __kmp_gtid_from_tid(f, team)((((f) >= 0 && (team) != __null) ? 0 : __kmp_debug_assert ("(f) >= 0 && (team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5110)), team->t.t_threads[(f)]->th.th_info.ds.ds_gtid )); |
5111 | if (level) { // set th_task_state for new threads in nested hot team |
5112 | // __kmp_initialize_info() no longer zeroes th_task_state, so we should |
5113 | // only need to set the th_task_state for the new threads. th_task_state |
5114 | // for master thread will not be accurate until after this in |
5115 | // __kmp_fork_call(), so we look to the master's memo_stack to get the |
5116 | // correct value. |
5117 | for (f = old_nproc; f < team->t.t_nproc; ++f) |
5118 | team->t.t_threads[f]->th.th_task_state = |
5119 | team->t.t_threads[0]->th.th_task_state_memo_stack[level]; |
5120 | } else { // set th_task_state for new threads in non-nested hot team |
5121 | int old_state = |
5122 | team->t.t_threads[0]->th.th_task_state; // copy master's state |
5123 | for (f = old_nproc; f < team->t.t_nproc; ++f) |
5124 | team->t.t_threads[f]->th.th_task_state = old_state; |
5125 | } |
5126 | |
5127 | #ifdef KMP_DEBUG1 |
5128 | for (f = 0; f < team->t.t_nproc; ++f) { |
5129 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5131)) |
5130 | team->t.t_threads[f]->th.th_team_nproc ==((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5131)) |
5131 | team->t.t_nproc)((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5131)); |
5132 | } |
5133 | #endif |
5134 | |
5135 | #if OMP_40_ENABLED(50 >= 40) |
5136 | KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind)if ((team->t.t_proc_bind) != (new_proc_bind)) (team->t. t_proc_bind) = (new_proc_bind); |
5137 | #if KMP_AFFINITY_SUPPORTED1 |
5138 | __kmp_partition_places(team); |
5139 | #endif |
5140 | #endif |
5141 | } // Check changes in number of threads |
5142 | |
5143 | #if OMP_40_ENABLED(50 >= 40) |
5144 | kmp_info_t *master = team->t.t_threads[0]; |
5145 | if (master->th.th_teams_microtask) { |
5146 | for (f = 1; f < new_nproc; ++f) { |
5147 | // propagate teams construct specific info to workers |
5148 | kmp_info_t *thr = team->t.t_threads[f]; |
5149 | thr->th.th_teams_microtask = master->th.th_teams_microtask; |
5150 | thr->th.th_teams_level = master->th.th_teams_level; |
5151 | thr->th.th_teams_size = master->th.th_teams_size; |
5152 | } |
5153 | } |
5154 | #endif /* OMP_40_ENABLED */ |
5155 | #if KMP_NESTED_HOT_TEAMS1 |
5156 | if (level) { |
5157 | // Sync barrier state for nested hot teams, not needed for outermost hot |
5158 | // team. |
5159 | for (f = 1; f < new_nproc; ++f) { |
5160 | kmp_info_t *thr = team->t.t_threads[f]; |
5161 | int b; |
5162 | kmp_balign_t *balign = thr->th.th_bar; |
5163 | for (b = 0; b < bs_last_barrier; ++b) { |
5164 | balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; |
5165 | KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG)((balign[b].bb.wait_flag != 2) ? 0 : __kmp_debug_assert("balign[b].bb.wait_flag != 2" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5165)); |
5166 | #if USE_DEBUGGER0 |
5167 | balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; |
5168 | #endif |
5169 | } |
5170 | } |
5171 | } |
5172 | #endif // KMP_NESTED_HOT_TEAMS |
5173 | |
5174 | /* reallocate space for arguments if necessary */ |
5175 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); |
5176 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); |
5177 | // The hot team re-uses the previous task team, |
5178 | // if untouched during the previous release->gather phase. |
5179 | |
5180 | KF_TRACE(10, (" hot_team = %p\n", team))if (kmp_f_debug >= 10) { __kmp_debug_printf (" hot_team = %p\n" , team); }; |
5181 | |
5182 | #if KMP_DEBUG1 |
5183 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
5184 | KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); } |
5185 | "task_team[1] = %p after reinit\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); } |
5186 | team->t.t_task_team[0], team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: hot team task_team[0] = %p " "task_team[1] = %p after reinit\n", team->t.t_task_team[0 ], team->t.t_task_team[1]); }; |
5187 | } |
5188 | #endif |
5189 | |
5190 | #if OMPT_SUPPORT1 |
5191 | __ompt_team_assign_id(team, ompt_parallel_data); |
5192 | #endif |
5193 | |
5194 | KMP_MB(); |
5195 | |
5196 | return team; |
5197 | } |
5198 | |
5199 | /* next, let's try to take one from the team pool */ |
5200 | KMP_MB(); |
5201 | for (team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); (team);) { |
5202 | /* TODO: consider resizing undersized teams instead of reaping them, now |
5203 | that we have a resizing mechanism */ |
5204 | if (team->t.t_max_nproc >= max_nproc) { |
5205 | /* take this team from the team pool */ |
5206 | __kmp_team_pool = team->t.t_next_pool; |
5207 | |
5208 | /* setup the team for fresh use */ |
5209 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); |
5210 | |
5211 | KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); } |
5212 | "task_team[1] %p to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); } |
5213 | &team->t.t_task_team[0], &team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and " "task_team[1] %p to NULL\n", &team->t.t_task_team[0], &team->t.t_task_team[1]); }; |
5214 | team->t.t_task_team[0] = NULL__null; |
5215 | team->t.t_task_team[1] = NULL__null; |
5216 | |
5217 | /* reallocate space for arguments if necessary */ |
5218 | __kmp_alloc_argv_entries(argc, team, TRUE(!0)); |
5219 | KMP_CHECK_UPDATE(team->t.t_argc, argc)if ((team->t.t_argc) != (argc)) (team->t.t_argc) = (argc ); |
5220 | |
5221 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } |
5222 | 20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } |
5223 | team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); }; |
5224 | { // Initialize barrier data. |
5225 | int b; |
5226 | for (b = 0; b < bs_last_barrier; ++b) { |
5227 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; |
5228 | #if USE_DEBUGGER0 |
5229 | team->t.t_bar[b].b_master_arrived = 0; |
5230 | team->t.t_bar[b].b_team_arrived = 0; |
5231 | #endif |
5232 | } |
5233 | } |
5234 | |
5235 | #if OMP_40_ENABLED(50 >= 40) |
5236 | team->t.t_proc_bind = new_proc_bind; |
5237 | #endif |
5238 | |
5239 | KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: using team from pool %d.\n" , team->t.t_id); } |
5240 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: using team from pool %d.\n" , team->t.t_id); }; |
5241 | |
5242 | #if OMPT_SUPPORT1 |
5243 | __ompt_team_assign_id(team, ompt_parallel_data); |
5244 | #endif |
5245 | |
5246 | KMP_MB(); |
5247 | |
5248 | return team; |
5249 | } |
5250 | |
5251 | /* reap team if it is too small, then loop back and check the next one */ |
5252 | // not sure if this is wise, but, will be redone during the hot-teams |
5253 | // rewrite. |
5254 | /* TODO: Use technique to find the right size hot-team, don't reap them */ |
5255 | team = __kmp_reap_team(team); |
5256 | __kmp_team_pool = team; |
5257 | } |
5258 | |
5259 | /* nothing available in the pool, no matter, make a new team! */ |
5260 | KMP_MB(); |
5261 | team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t))___kmp_allocate((sizeof(kmp_team_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5261); |
5262 | |
5263 | /* and set it up */ |
5264 | team->t.t_max_nproc = max_nproc; |
5265 | /* NOTE well, for some reason allocating one big buffer and dividing it up |
5266 | seems to really hurt performance a lot on the P4, so, let's not use this */ |
5267 | __kmp_allocate_team_arrays(team, max_nproc); |
5268 | |
5269 | KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: making a new team\n" ); }; |
5270 | __kmp_initialize_team(team, new_nproc, new_icvs, NULL__null); |
5271 | |
5272 | KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); } |
5273 | "%p to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); } |
5274 | &team->t.t_task_team[0], &team->t.t_task_team[1]))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " "%p to NULL\n", &team->t.t_task_team[0], &team-> t.t_task_team[1]); }; |
5275 | team->t.t_task_team[0] = NULL__null; // to be removed, as __kmp_allocate zeroes |
5276 | // memory, no need to duplicate |
5277 | team->t.t_task_team[1] = NULL__null; // to be removed, as __kmp_allocate zeroes |
5278 | // memory, no need to duplicate |
5279 | |
5280 | if (__kmp_storage_map) { |
5281 | __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc); |
5282 | } |
5283 | |
5284 | /* allocate space for arguments */ |
5285 | __kmp_alloc_argv_entries(argc, team, FALSE0); |
5286 | team->t.t_argc = argc; |
5287 | |
5288 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } |
5289 | ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); } |
5290 | team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n" , team->t.t_id, 0, 0); }; |
5291 | { // Initialize barrier data. |
5292 | int b; |
5293 | for (b = 0; b < bs_last_barrier; ++b) { |
5294 | team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE0; |
5295 | #if USE_DEBUGGER0 |
5296 | team->t.t_bar[b].b_master_arrived = 0; |
5297 | team->t.t_bar[b].b_team_arrived = 0; |
5298 | #endif |
5299 | } |
5300 | } |
5301 | |
5302 | #if OMP_40_ENABLED(50 >= 40) |
5303 | team->t.t_proc_bind = new_proc_bind; |
5304 | #endif |
5305 | |
5306 | #if OMPT_SUPPORT1 |
5307 | __ompt_team_assign_id(team, ompt_parallel_data); |
5308 | team->t.ompt_serialized_team_info = NULL__null; |
5309 | #endif |
5310 | |
5311 | KMP_MB(); |
5312 | |
5313 | KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: done creating a new team %d.\n" , team->t.t_id); } |
5314 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_team: done creating a new team %d.\n" , team->t.t_id); }; |
5315 | |
5316 | return team; |
5317 | } |
5318 | |
5319 | /* TODO implement hot-teams at all levels */ |
5320 | /* TODO implement lazy thread release on demand (disband request) */ |
5321 | |
5322 | /* free the team. return it to the team pool. release all the threads |
5323 | * associated with it */ |
5324 | void __kmp_free_team(kmp_root_t *root, |
5325 | kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master), kmp_info_t *master) { |
5326 | int f; |
5327 | KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d freeing team %d\n" , __kmp_get_global_thread_id(), team->t.t_id); } |
5328 | team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d freeing team %d\n" , __kmp_get_global_thread_id(), team->t.t_id); }; |
5329 | |
5330 | /* verify state */ |
5331 | KMP_DEBUG_ASSERT(root)((root) ? 0 : __kmp_debug_assert("root", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5331)); |
5332 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5332)); |
5333 | KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc)((team->t.t_nproc <= team->t.t_max_nproc) ? 0 : __kmp_debug_assert ("team->t.t_nproc <= team->t.t_max_nproc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5333)); |
5334 | KMP_DEBUG_ASSERT(team->t.t_threads)((team->t.t_threads) ? 0 : __kmp_debug_assert("team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5334)); |
5335 | |
5336 | int use_hot_team = team == root->r.r_hot_team; |
5337 | #if KMP_NESTED_HOT_TEAMS1 |
5338 | int level; |
5339 | kmp_hot_team_ptr_t *hot_teams; |
5340 | if (master) { |
5341 | level = team->t.t_active_level - 1; |
5342 | if (master->th.th_teams_microtask) { // in teams construct? |
5343 | if (master->th.th_teams_size.nteams > 1) { |
5344 | ++level; // level was not increased in teams construct for |
5345 | // team_of_masters |
5346 | } |
5347 | if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && |
5348 | master->th.th_teams_level == team->t.t_level) { |
5349 | ++level; // level was not increased in teams construct for |
5350 | // team_of_workers before the parallel |
5351 | } // team->t.t_level will be increased inside parallel |
5352 | } |
5353 | hot_teams = master->th.th_hot_teams; |
5354 | if (level < __kmp_hot_teams_max_level) { |
5355 | KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team)((team == hot_teams[level].hot_team) ? 0 : __kmp_debug_assert ("team == hot_teams[level].hot_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5355)); |
5356 | use_hot_team = 1; |
5357 | } |
5358 | } |
5359 | #endif // KMP_NESTED_HOT_TEAMS |
5360 | |
5361 | /* team is done working */ |
5362 | TCW_SYNC_PTR(team->t.t_pkfn,((team->t.t_pkfn)) = ((__null)) |
5363 | NULL)((team->t.t_pkfn)) = ((__null)); // Important for Debugging Support Library. |
5364 | #if KMP_OS_WINDOWS0 |
5365 | team->t.t_copyin_counter = 0; // init counter for possible reuse |
5366 | #endif |
5367 | // Do not reset pointer to parent team to NULL for hot teams. |
5368 | |
5369 | /* if we are non-hot team, release our threads */ |
5370 | if (!use_hot_team) { |
5371 | if (__kmp_tasking_mode != tskm_immediate_exec) { |
5372 | // Wait for threads to reach reapable state |
5373 | for (f = 1; f < team->t.t_nproc; ++f) { |
5374 | KMP_DEBUG_ASSERT(team->t.t_threads[f])((team->t.t_threads[f]) ? 0 : __kmp_debug_assert("team->t.t_threads[f]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5374)); |
5375 | kmp_info_t *th = team->t.t_threads[f]; |
5376 | volatile kmp_uint32 *state = &th->th.th_reap_state; |
5377 | while (*state != KMP_SAFE_TO_REAP1) { |
5378 | #if KMP_OS_WINDOWS0 |
5379 | // On Windows a thread can be killed at any time, check this |
5380 | DWORD ecode; |
5381 | if (!__kmp_is_thread_alive(th, &ecode)) { |
5382 | *state = KMP_SAFE_TO_REAP1; // reset the flag for dead thread |
5383 | break; |
5384 | } |
5385 | #endif |
5386 | // first check if thread is sleeping |
5387 | kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th); |
5388 | if (fl.is_sleeping()) |
5389 | fl.resume(__kmp_gtid_from_thread(th)((((th) != __null) ? 0 : __kmp_debug_assert("(th) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5389)), (th)->th.th_info.ds.ds_gtid)); |
5390 | KMP_CPU_PAUSE()__kmp_x86_pause(); |
5391 | } |
5392 | } |
5393 | |
5394 | // Delete task teams |
5395 | int tt_idx; |
5396 | for (tt_idx = 0; tt_idx < 2; ++tt_idx) { |
5397 | kmp_task_team_t *task_team = team->t.t_task_team[tt_idx]; |
5398 | if (task_team != NULL__null) { |
5399 | for (f = 0; f < team->t.t_nproc; |
5400 | ++f) { // Have all threads unref task teams |
5401 | team->t.t_threads[f]->th.th_task_team = NULL__null; |
5402 | } |
5403 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } |
5404 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } |
5405 | ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); } |
5406 | __kmp_get_gtid(), task_team, team->t.t_id))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n" , __kmp_get_global_thread_id(), task_team, team->t.t_id); }; |
5407 | #if KMP_NESTED_HOT_TEAMS1 |
5408 | __kmp_free_task_team(master, task_team); |
5409 | #endif |
5410 | team->t.t_task_team[tt_idx] = NULL__null; |
5411 | } |
5412 | } |
5413 | } |
5414 | |
5415 | // Reset pointer to parent team only for non-hot teams. |
5416 | team->t.t_parent = NULL__null; |
5417 | team->t.t_level = 0; |
5418 | team->t.t_active_level = 0; |
5419 | |
5420 | /* free the worker threads */ |
5421 | for (f = 1; f < team->t.t_nproc; ++f) { |
5422 | KMP_DEBUG_ASSERT(team->t.t_threads[f])((team->t.t_threads[f]) ? 0 : __kmp_debug_assert("team->t.t_threads[f]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5422)); |
5423 | __kmp_free_thread(team->t.t_threads[f]); |
5424 | team->t.t_threads[f] = NULL__null; |
5425 | } |
5426 | |
5427 | /* put the team back in the team pool */ |
5428 | /* TODO limit size of team pool, call reap_team if pool too large */ |
5429 | team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); |
5430 | __kmp_team_pool = (volatile kmp_team_t *)team; |
5431 | } |
5432 | |
5433 | KMP_MB(); |
5434 | } |
5435 | |
5436 | /* reap the team. destroy it, reclaim all its resources and free its memory */ |
5437 | kmp_team_t *__kmp_reap_team(kmp_team_t *team) { |
5438 | kmp_team_t *next_pool = team->t.t_next_pool; |
5439 | |
5440 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5440)); |
5441 | KMP_DEBUG_ASSERT(team->t.t_dispatch)((team->t.t_dispatch) ? 0 : __kmp_debug_assert("team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5441)); |
5442 | KMP_DEBUG_ASSERT(team->t.t_disp_buffer)((team->t.t_disp_buffer) ? 0 : __kmp_debug_assert("team->t.t_disp_buffer" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5442)); |
5443 | KMP_DEBUG_ASSERT(team->t.t_threads)((team->t.t_threads) ? 0 : __kmp_debug_assert("team->t.t_threads" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5443)); |
5444 | KMP_DEBUG_ASSERT(team->t.t_argv)((team->t.t_argv) ? 0 : __kmp_debug_assert("team->t.t_argv" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5444)); |
5445 | |
5446 | /* TODO clean the threads that are a part of this? */ |
5447 | |
5448 | /* free stuff */ |
5449 | __kmp_free_team_arrays(team); |
5450 | if (team->t.t_argv != &team->t.t_inline_argv[0]) |
5451 | __kmp_free((void *)team->t.t_argv)___kmp_free(((void *)team->t.t_argv), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5451); |
5452 | __kmp_free(team)___kmp_free((team), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5452); |
5453 | |
5454 | KMP_MB(); |
5455 | return next_pool; |
5456 | } |
5457 | |
5458 | // Free the thread. Don't reap it, just place it on the pool of available |
5459 | // threads. |
5460 | // |
5461 | // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid |
5462 | // binding for the affinity mechanism to be useful. |
5463 | // |
5464 | // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid. |
5465 | // However, we want to avoid a potential performance problem by always |
5466 | // scanning through the list to find the correct point at which to insert |
5467 | // the thread (potential N**2 behavior). To do this we keep track of the |
5468 | // last place a thread struct was inserted (__kmp_thread_pool_insert_pt). |
5469 | // With single-level parallelism, threads will always be added to the tail |
5470 | // of the list, kept track of by __kmp_thread_pool_insert_pt. With nested |
5471 | // parallelism, all bets are off and we may need to scan through the entire |
5472 | // free list. |
5473 | // |
5474 | // This change also has a potentially large performance benefit, for some |
5475 | // applications. Previously, as threads were freed from the hot team, they |
5476 | // would be placed back on the free list in inverse order. If the hot team |
5477 | // grew back to it's original size, then the freed thread would be placed |
5478 | // back on the hot team in reverse order. This could cause bad cache |
5479 | // locality problems on programs where the size of the hot team regularly |
5480 | // grew and shrunk. |
5481 | // |
5482 | // Now, for single-level parallelism, the OMP tid is alway == gtid. |
5483 | void __kmp_free_thread(kmp_info_t *this_th) { |
5484 | int gtid; |
5485 | kmp_info_t **scan; |
5486 | kmp_root_t *root = this_th->th.th_root; |
5487 | |
5488 | KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n" , __kmp_get_global_thread_id(), this_th->th.th_info.ds.ds_gtid ); } |
5489 | __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n" , __kmp_get_global_thread_id(), this_th->th.th_info.ds.ds_gtid ); }; |
5490 | |
5491 | KMP_DEBUG_ASSERT(this_th)((this_th) ? 0 : __kmp_debug_assert("this_th", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5491)); |
5492 | |
5493 | // When moving thread to pool, switch thread to wait on own b_go flag, and |
5494 | // uninitialized (NULL team). |
5495 | int b; |
5496 | kmp_balign_t *balign = this_th->th.th_bar; |
5497 | for (b = 0; b < bs_last_barrier; ++b) { |
5498 | if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG2) |
5499 | balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG3; |
5500 | balign[b].bb.team = NULL__null; |
5501 | balign[b].bb.leaf_kids = 0; |
5502 | } |
5503 | this_th->th.th_task_state = 0; |
5504 | this_th->th.th_reap_state = KMP_SAFE_TO_REAP1; |
5505 | |
5506 | /* put thread back on the free pool */ |
5507 | TCW_PTR(this_th->th.th_team, NULL)((this_th->th.th_team)) = ((__null)); |
5508 | TCW_PTR(this_th->th.th_root, NULL)((this_th->th.th_root)) = ((__null)); |
5509 | TCW_PTR(this_th->th.th_dispatch, NULL)((this_th->th.th_dispatch)) = ((__null)); /* NOT NEEDED */ |
5510 | |
5511 | /* If the implicit task assigned to this thread can be used by other threads |
5512 | * -> multiple threads can share the data and try to free the task at |
5513 | * __kmp_reap_thread at exit. This duplicate use of the task data can happen |
5514 | * with higher probability when hot team is disabled but can occurs even when |
5515 | * the hot team is enabled */ |
5516 | __kmp_free_implicit_task(this_th); |
5517 | this_th->th.th_current_task = NULL__null; |
5518 | |
5519 | // If the __kmp_thread_pool_insert_pt is already past the new insert |
5520 | // point, then we need to re-scan the entire list. |
5521 | gtid = this_th->th.th_info.ds.ds_gtid; |
5522 | if (__kmp_thread_pool_insert_pt != NULL__null) { |
5523 | KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL)((__kmp_thread_pool != __null) ? 0 : __kmp_debug_assert("__kmp_thread_pool != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5523)); |
5524 | if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) { |
5525 | __kmp_thread_pool_insert_pt = NULL__null; |
5526 | } |
5527 | } |
5528 | |
5529 | // Scan down the list to find the place to insert the thread. |
5530 | // scan is the address of a link in the list, possibly the address of |
5531 | // __kmp_thread_pool itself. |
5532 | // |
5533 | // In the absence of nested parallism, the for loop will have 0 iterations. |
5534 | if (__kmp_thread_pool_insert_pt != NULL__null) { |
5535 | scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool); |
5536 | } else { |
5537 | scan = CCAST(kmp_info_t **, &__kmp_thread_pool)const_cast<kmp_info_t **>(&__kmp_thread_pool); |
5538 | } |
5539 | for (; (*scan != NULL__null) && ((*scan)->th.th_info.ds.ds_gtid < gtid); |
5540 | scan = &((*scan)->th.th_next_pool)) |
5541 | ; |
5542 | |
5543 | // Insert the new element on the list, and set __kmp_thread_pool_insert_pt |
5544 | // to its address. |
5545 | TCW_PTR(this_th->th.th_next_pool, *scan)((this_th->th.th_next_pool)) = ((*scan)); |
5546 | __kmp_thread_pool_insert_pt = *scan = this_th; |
5547 | KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||(((this_th->th.th_next_pool == __null) || (this_th->th. th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info .ds.ds_gtid)) ? 0 : __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5549)) |
5548 | (this_th->th.th_info.ds.ds_gtid <(((this_th->th.th_next_pool == __null) || (this_th->th. th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info .ds.ds_gtid)) ? 0 : __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5549)) |
5549 | this_th->th.th_next_pool->th.th_info.ds.ds_gtid))(((this_th->th.th_next_pool == __null) || (this_th->th. th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info .ds.ds_gtid)) ? 0 : __kmp_debug_assert("(this_th->th.th_next_pool == __null) || (this_th->th.th_info.ds.ds_gtid < this_th->th.th_next_pool->th.th_info.ds.ds_gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5549)); |
5550 | TCW_4(this_th->th.th_in_pool, TRUE)(this_th->th.th_in_pool) = ((!0)); |
5551 | __kmp_thread_pool_nth++; |
5552 | |
5553 | TCW_4(__kmp_nth, __kmp_nth - 1)(__kmp_nth) = (__kmp_nth - 1); |
5554 | root->r.r_cg_nthreads--; |
5555 | |
5556 | #ifdef KMP_ADJUST_BLOCKTIME1 |
5557 | /* Adjust blocktime back to user setting or default if necessary */ |
5558 | /* Middle initialization might never have occurred */ |
5559 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
5560 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)((__kmp_avail_proc > 0) ? 0 : __kmp_debug_assert("__kmp_avail_proc > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5560)); |
5561 | if (__kmp_nth <= __kmp_avail_proc) { |
5562 | __kmp_zero_bt = FALSE0; |
5563 | } |
5564 | } |
5565 | #endif /* KMP_ADJUST_BLOCKTIME */ |
5566 | |
5567 | KMP_MB(); |
5568 | } |
5569 | |
5570 | /* ------------------------------------------------------------------------ */ |
5571 | |
5572 | void *__kmp_launch_thread(kmp_info_t *this_thr) { |
5573 | int gtid = this_thr->th.th_info.ds.ds_gtid; |
5574 | /* void *stack_data;*/ |
5575 | kmp_team_t *(*volatile pteam); |
5576 | |
5577 | KMP_MB(); |
5578 | KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_launch_thread: T#%d start\n" , gtid); }; |
5579 | |
5580 | if (__kmp_env_consistency_check) { |
5581 | this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak? |
5582 | } |
5583 | |
5584 | #if OMPT_SUPPORT1 |
5585 | ompt_data_t *thread_data; |
5586 | if (ompt_enabled.enabled) { |
5587 | thread_data = &(this_thr->th.ompt_thread_info.thread_data); |
5588 | thread_data->ptr = NULL__null; |
5589 | |
5590 | this_thr->th.ompt_thread_info.state = omp_state_overhead; |
5591 | this_thr->th.ompt_thread_info.wait_id = 0; |
5592 | this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); |
5593 | if (ompt_enabled.ompt_callback_thread_begin) { |
5594 | ompt_callbacks.ompt_callback(ompt_callback_thread_begin)ompt_callback_thread_begin_callback( |
5595 | ompt_thread_worker, thread_data); |
5596 | } |
5597 | } |
5598 | #endif |
5599 | |
5600 | #if OMPT_SUPPORT1 |
5601 | if (ompt_enabled.enabled) { |
5602 | this_thr->th.ompt_thread_info.state = omp_state_idle; |
5603 | } |
5604 | #endif |
5605 | /* This is the place where threads wait for work */ |
5606 | while (!TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { |
5607 | KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid])((this_thr == __kmp_threads[gtid]) ? 0 : __kmp_debug_assert("this_thr == __kmp_threads[gtid]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5607)); |
5608 | KMP_MB(); |
5609 | |
5610 | /* wait for work to do */ |
5611 | KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d waiting for work\n" , gtid); }; |
5612 | |
5613 | /* No tid yet since not part of a team */ |
5614 | __kmp_fork_barrier(gtid, KMP_GTID_DNE(-2)); |
5615 | |
5616 | #if OMPT_SUPPORT1 |
5617 | if (ompt_enabled.enabled) { |
5618 | this_thr->th.ompt_thread_info.state = omp_state_overhead; |
5619 | } |
5620 | #endif |
5621 | |
5622 | pteam = (kmp_team_t * (*))(&this_thr->th.th_team); |
5623 | |
5624 | /* have we been allocated? */ |
5625 | if (TCR_SYNC_PTR(*pteam)((void *)(*pteam)) && !TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { |
5626 | /* we were just woken up, so run our new task */ |
5627 | if (TCR_SYNC_PTR((*pteam)->t.t_pkfn)((void *)((*pteam)->t.t_pkfn)) != NULL__null) { |
5628 | int rc; |
5629 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5631)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); } |
5630 | ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5631)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); } |
5631 | gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5631)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); } |
5632 | (*pteam)->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5631)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); }; |
5633 | |
5634 | updateHWFPControl(*pteam); |
5635 | |
5636 | #if OMPT_SUPPORT1 |
5637 | if (ompt_enabled.enabled) { |
5638 | this_thr->th.ompt_thread_info.state = omp_state_work_parallel; |
5639 | } |
5640 | #endif |
5641 | |
5642 | { |
5643 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
5644 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
5645 | rc = (*pteam)->t.t_invoke(gtid); |
5646 | } |
5647 | KMP_ASSERT(rc)((rc) ? 0 : __kmp_debug_assert("rc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5647)); |
5648 | |
5649 | KMP_MB(); |
5650 | KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5651)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); } |
5651 | gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5651)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); } |
5652 | (*pteam)->t.t_pkfn))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n" , gtid, (*pteam)->t.t_id, ((((gtid) >= 0) ? 0 : __kmp_debug_assert ("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5651)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), (* pteam)->t.t_pkfn); }; |
5653 | } |
5654 | #if OMPT_SUPPORT1 |
5655 | if (ompt_enabled.enabled) { |
5656 | /* no frame set while outside task */ |
5657 | __ompt_get_task_info_object(0)->frame.exit_frame = NULL__null; |
5658 | |
5659 | this_thr->th.ompt_thread_info.state = omp_state_overhead; |
5660 | } |
5661 | #endif |
5662 | /* join barrier after parallel region */ |
5663 | __kmp_join_barrier(gtid); |
5664 | } |
5665 | } |
5666 | TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done)((void *)((intptr_t)__kmp_global.g.g_done)); |
5667 | |
5668 | #if OMPT_SUPPORT1 |
5669 | if (ompt_enabled.ompt_callback_thread_end) { |
5670 | ompt_callbacks.ompt_callback(ompt_callback_thread_end)ompt_callback_thread_end_callback(thread_data); |
5671 | } |
5672 | #endif |
5673 | |
5674 | this_thr->th.th_task_team = NULL__null; |
5675 | /* run the destructors for the threadprivate data for this thread */ |
5676 | __kmp_common_destroy_gtid(gtid); |
5677 | |
5678 | KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_launch_thread: T#%d done\n" , gtid); }; |
5679 | KMP_MB(); |
5680 | return this_thr; |
5681 | } |
5682 | |
5683 | /* ------------------------------------------------------------------------ */ |
5684 | |
5685 | void __kmp_internal_end_dest(void *specific_gtid) { |
5686 | #if KMP_COMPILER_ICC0 |
5687 | #pragma warning(push) |
5688 | #pragma warning(disable : 810) // conversion from "void *" to "int" may lose |
5689 | // significant bits |
5690 | #endif |
5691 | // Make sure no significant bits are lost |
5692 | int gtid = (kmp_intptr_t)specific_gtid - 1; |
5693 | #if KMP_COMPILER_ICC0 |
5694 | #pragma warning(pop) |
5695 | #endif |
5696 | |
5697 | KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_internal_end_dest: T#%d\n" , gtid); }; |
5698 | /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage |
5699 | * this is because 0 is reserved for the nothing-stored case */ |
5700 | |
5701 | /* josh: One reason for setting the gtid specific data even when it is being |
5702 | destroyed by pthread is to allow gtid lookup through thread specific data |
5703 | (__kmp_gtid_get_specific). Some of the code, especially stat code, |
5704 | that gets executed in the call to __kmp_internal_end_thread, actually |
5705 | gets the gtid through the thread specific data. Setting it here seems |
5706 | rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread |
5707 | to run smoothly. |
5708 | todo: get rid of this after we remove the dependence on |
5709 | __kmp_gtid_get_specific */ |
5710 | if (gtid >= 0 && KMP_UBER_GTID(gtid)((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5710)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5710)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) |
5711 | __kmp_gtid_set_specific(gtid); |
5712 | #ifdef KMP_TDATA_GTID1 |
5713 | __kmp_gtid = gtid; |
5714 | #endif |
5715 | __kmp_internal_end_thread(gtid); |
5716 | } |
5717 | |
5718 | #if KMP_OS_UNIX1 && KMP_DYNAMIC_LIB1 |
5719 | |
5720 | // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases |
5721 | // destructors work perfectly, but in real libomp.so I have no evidence it is |
5722 | // ever called. However, -fini linker option in makefile.mk works fine. |
5723 | |
5724 | __attribute__((destructor)) void __kmp_internal_end_dtor(void) { |
5725 | __kmp_internal_end_atexit(); |
5726 | } |
5727 | |
5728 | void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); } |
5729 | |
5730 | #endif |
5731 | |
5732 | /* [Windows] josh: when the atexit handler is called, there may still be more |
5733 | than one thread alive */ |
5734 | void __kmp_internal_end_atexit(void) { |
5735 | KA_TRACE(30, ("__kmp_internal_end_atexit\n"))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_internal_end_atexit\n" ); }; |
5736 | /* [Windows] |
5737 | josh: ideally, we want to completely shutdown the library in this atexit |
5738 | handler, but stat code that depends on thread specific data for gtid fails |
5739 | because that data becomes unavailable at some point during the shutdown, so |
5740 | we call __kmp_internal_end_thread instead. We should eventually remove the |
5741 | dependency on __kmp_get_specific_gtid in the stat code and use |
5742 | __kmp_internal_end_library to cleanly shutdown the library. |
5743 | |
5744 | // TODO: Can some of this comment about GVS be removed? |
5745 | I suspect that the offending stat code is executed when the calling thread |
5746 | tries to clean up a dead root thread's data structures, resulting in GVS |
5747 | code trying to close the GVS structures for that thread, but since the stat |
5748 | code uses __kmp_get_specific_gtid to get the gtid with the assumption that |
5749 | the calling thread is cleaning up itself instead of another thread, it get |
5750 | confused. This happens because allowing a thread to unregister and cleanup |
5751 | another thread is a recent modification for addressing an issue. |
5752 | Based on the current design (20050722), a thread may end up |
5753 | trying to unregister another thread only if thread death does not trigger |
5754 | the calling of __kmp_internal_end_thread. For Linux* OS, there is the |
5755 | thread specific data destructor function to detect thread death. For |
5756 | Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there |
5757 | is nothing. Thus, the workaround is applicable only for Windows static |
5758 | stat library. */ |
5759 | __kmp_internal_end_library(-1); |
5760 | #if KMP_OS_WINDOWS0 |
5761 | __kmp_close_console(); |
5762 | #endif |
5763 | } |
5764 | |
5765 | static void __kmp_reap_thread(kmp_info_t *thread, int is_root) { |
5766 | // It is assumed __kmp_forkjoin_lock is acquired. |
5767 | |
5768 | int gtid; |
5769 | |
5770 | KMP_DEBUG_ASSERT(thread != NULL)((thread != __null) ? 0 : __kmp_debug_assert("thread != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5770)); |
5771 | |
5772 | gtid = thread->th.th_info.ds.ds_gtid; |
5773 | |
5774 | if (!is_root) { |
5775 | |
5776 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { |
5777 | /* Assume the threads are at the fork barrier here */ |
5778 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); } |
5779 | 20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); } |
5780 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n" , gtid); }; |
5781 | /* Need release fence here to prevent seg faults for tree forkjoin barrier |
5782 | * (GEH) */ |
5783 | ANNOTATE_HAPPENS_BEFORE(thread); |
5784 | kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread); |
5785 | __kmp_release_64(&flag); |
5786 | } |
5787 | |
5788 | // Terminate OS thread. |
5789 | __kmp_reap_worker(thread); |
5790 | |
5791 | // The thread was killed asynchronously. If it was actively |
5792 | // spinning in the thread pool, decrement the global count. |
5793 | // |
5794 | // There is a small timing hole here - if the worker thread was just waking |
5795 | // up after sleeping in the pool, had reset it's th_active_in_pool flag but |
5796 | // not decremented the global counter __kmp_thread_pool_active_nth yet, then |
5797 | // the global counter might not get updated. |
5798 | // |
5799 | // Currently, this can only happen as the library is unloaded, |
5800 | // so there are no harmful side effects. |
5801 | if (thread->th.th_active_in_pool) { |
5802 | thread->th.th_active_in_pool = FALSE0; |
5803 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth)(&__kmp_thread_pool_active_nth)->fetch_sub(1, std::memory_order_acq_rel ); |
5804 | KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0)((__kmp_thread_pool_active_nth >= 0) ? 0 : __kmp_debug_assert ("__kmp_thread_pool_active_nth >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5804)); |
5805 | } |
5806 | |
5807 | // Decrement # of [worker] threads in the pool. |
5808 | KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0)((__kmp_thread_pool_nth > 0) ? 0 : __kmp_debug_assert("__kmp_thread_pool_nth > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5808)); |
5809 | --__kmp_thread_pool_nth; |
5810 | } |
5811 | |
5812 | __kmp_free_implicit_task(thread); |
5813 | |
5814 | // Free the fast memory for tasking |
5815 | #if USE_FAST_MEMORY3 |
5816 | __kmp_free_fast_memory(thread); |
5817 | #endif /* USE_FAST_MEMORY */ |
5818 | |
5819 | __kmp_suspend_uninitialize_thread(thread); |
5820 | |
5821 | KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread)((__kmp_threads[gtid] == thread) ? 0 : __kmp_debug_assert("__kmp_threads[gtid] == thread" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5821)); |
5822 | TCW_SYNC_PTR(__kmp_threads[gtid], NULL)((__kmp_threads[gtid])) = ((__null)); |
5823 | |
5824 | --__kmp_all_nth; |
5825 | // __kmp_nth was decremented when thread is added to the pool. |
5826 | |
5827 | #ifdef KMP_ADJUST_BLOCKTIME1 |
5828 | /* Adjust blocktime back to user setting or default if necessary */ |
5829 | /* Middle initialization might never have occurred */ |
5830 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
5831 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)((__kmp_avail_proc > 0) ? 0 : __kmp_debug_assert("__kmp_avail_proc > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5831)); |
5832 | if (__kmp_nth <= __kmp_avail_proc) { |
5833 | __kmp_zero_bt = FALSE0; |
5834 | } |
5835 | } |
5836 | #endif /* KMP_ADJUST_BLOCKTIME */ |
5837 | |
5838 | /* free the memory being used */ |
5839 | if (__kmp_env_consistency_check) { |
5840 | if (thread->th.th_cons) { |
5841 | __kmp_free_cons_stack(thread->th.th_cons); |
5842 | thread->th.th_cons = NULL__null; |
5843 | } |
5844 | } |
5845 | |
5846 | if (thread->th.th_pri_common != NULL__null) { |
5847 | __kmp_free(thread->th.th_pri_common)___kmp_free((thread->th.th_pri_common), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5847); |
5848 | thread->th.th_pri_common = NULL__null; |
5849 | } |
5850 | |
5851 | if (thread->th.th_task_state_memo_stack != NULL__null) { |
5852 | __kmp_free(thread->th.th_task_state_memo_stack)___kmp_free((thread->th.th_task_state_memo_stack), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5852); |
5853 | thread->th.th_task_state_memo_stack = NULL__null; |
5854 | } |
5855 | |
5856 | #if KMP_USE_BGET1 |
5857 | if (thread->th.th_local.bget_data != NULL__null) { |
5858 | __kmp_finalize_bget(thread); |
5859 | } |
5860 | #endif |
5861 | |
5862 | #if KMP_AFFINITY_SUPPORTED1 |
5863 | if (thread->th.th_affin_mask != NULL__null) { |
5864 | KMP_CPU_FREE(thread->th.th_affin_mask)__kmp_affinity_dispatch->deallocate_mask(thread->th.th_affin_mask ); |
5865 | thread->th.th_affin_mask = NULL__null; |
5866 | } |
5867 | #endif /* KMP_AFFINITY_SUPPORTED */ |
5868 | |
5869 | #if KMP_USE_HIER_SCHED0 |
5870 | if (thread->th.th_hier_bar_data != NULL__null) { |
5871 | __kmp_free(thread->th.th_hier_bar_data)___kmp_free((thread->th.th_hier_bar_data), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5871); |
5872 | thread->th.th_hier_bar_data = NULL__null; |
5873 | } |
5874 | #endif |
5875 | |
5876 | __kmp_reap_team(thread->th.th_serial_team); |
5877 | thread->th.th_serial_team = NULL__null; |
5878 | __kmp_free(thread)___kmp_free((thread), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5878); |
5879 | |
5880 | KMP_MB(); |
5881 | |
5882 | } // __kmp_reap_thread |
5883 | |
5884 | static void __kmp_internal_end(void) { |
5885 | int i; |
5886 | |
5887 | /* First, unregister the library */ |
5888 | __kmp_unregister_library(); |
5889 | |
5890 | #if KMP_OS_WINDOWS0 |
5891 | /* In Win static library, we can't tell when a root actually dies, so we |
5892 | reclaim the data structures for any root threads that have died but not |
5893 | unregistered themselves, in order to shut down cleanly. |
5894 | In Win dynamic library we also can't tell when a thread dies. */ |
5895 | __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of |
5896 | // dead roots |
5897 | #endif |
5898 | |
5899 | for (i = 0; i < __kmp_threads_capacity; i++) |
5900 | if (__kmp_root[i]) |
5901 | if (__kmp_root[i]->r.r_active) |
5902 | break; |
5903 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
5904 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); |
5905 | |
5906 | if (i < __kmp_threads_capacity) { |
5907 | #if KMP_USE_MONITOR |
5908 | // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor?? |
5909 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
5910 | |
5911 | // Need to check that monitor was initialized before reaping it. If we are |
5912 | // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then |
5913 | // __kmp_monitor will appear to contain valid data, but it is only valid in |
5914 | // the parent process, not the child. |
5915 | // New behavior (201008): instead of keying off of the flag |
5916 | // __kmp_init_parallel, the monitor thread creation is keyed off |
5917 | // of the new flag __kmp_init_monitor. |
5918 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); |
5919 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { |
5920 | __kmp_reap_monitor(&__kmp_monitor); |
5921 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); |
5922 | } |
5923 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); |
5924 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; |
5925 | #endif // KMP_USE_MONITOR |
5926 | } else { |
5927 | /* TODO move this to cleanup code */ |
5928 | #ifdef KMP_DEBUG1 |
5929 | /* make sure that everything has properly ended */ |
5930 | for (i = 0; i < __kmp_threads_capacity; i++) { |
5931 | if (__kmp_root[i]) { |
5932 | // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC: |
5933 | // there can be uber threads alive here |
5934 | KMP_ASSERT(!__kmp_root[i]->r.r_active)((!__kmp_root[i]->r.r_active) ? 0 : __kmp_debug_assert("!__kmp_root[i]->r.r_active" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5934)); // TODO: can they be active? |
5935 | } |
5936 | } |
5937 | #endif |
5938 | |
5939 | KMP_MB(); |
5940 | |
5941 | // Reap the worker threads. |
5942 | // This is valid for now, but be careful if threads are reaped sooner. |
5943 | while (__kmp_thread_pool != NULL__null) { // Loop thru all the thread in the pool. |
5944 | // Get the next thread from the pool. |
5945 | kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); |
5946 | __kmp_thread_pool = thread->th.th_next_pool; |
5947 | // Reap it. |
5948 | KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP)((thread->th.th_reap_state == 1) ? 0 : __kmp_debug_assert( "thread->th.th_reap_state == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 5948)); |
5949 | thread->th.th_next_pool = NULL__null; |
5950 | thread->th.th_in_pool = FALSE0; |
5951 | __kmp_reap_thread(thread, 0); |
5952 | } |
5953 | __kmp_thread_pool_insert_pt = NULL__null; |
5954 | |
5955 | // Reap teams. |
5956 | while (__kmp_team_pool != NULL__null) { // Loop thru all the teams in the pool. |
5957 | // Get the next team from the pool. |
5958 | kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool)const_cast<kmp_team_t *>(__kmp_team_pool); |
5959 | __kmp_team_pool = team->t.t_next_pool; |
5960 | // Reap it. |
5961 | team->t.t_next_pool = NULL__null; |
5962 | __kmp_reap_team(team); |
5963 | } |
5964 | |
5965 | __kmp_reap_task_teams(); |
5966 | |
5967 | #if KMP_OS_UNIX1 |
5968 | // Threads that are not reaped should not access any resources since they |
5969 | // are going to be deallocated soon, so the shutdown sequence should wait |
5970 | // until all threads either exit the final spin-waiting loop or begin |
5971 | // sleeping after the given blocktime. |
5972 | for (i = 0; i < __kmp_threads_capacity; i++) { |
5973 | kmp_info_t *thr = __kmp_threads[i]; |
5974 | while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking)(&thr->th.th_blocking)->load(std::memory_order_acquire )) |
5975 | KMP_CPU_PAUSE()__kmp_x86_pause(); |
5976 | } |
5977 | #endif |
5978 | |
5979 | for (i = 0; i < __kmp_threads_capacity; ++i) { |
5980 | // TBD: Add some checking... |
5981 | // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL ); |
5982 | } |
5983 | |
5984 | /* Make sure all threadprivate destructors get run by joining with all |
5985 | worker threads before resetting this flag */ |
5986 | TCW_SYNC_4(__kmp_init_common, FALSE)(__kmp_init_common) = (0); |
5987 | |
5988 | KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: all workers reaped\n" ); }; |
5989 | KMP_MB(); |
5990 | |
5991 | #if KMP_USE_MONITOR |
5992 | // See note above: One of the possible fixes for CQ138434 / CQ140126 |
5993 | // |
5994 | // FIXME: push both code fragments down and CSE them? |
5995 | // push them into __kmp_cleanup() ? |
5996 | __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); |
5997 | if (TCR_4(__kmp_init_monitor)(__kmp_init_monitor)) { |
5998 | __kmp_reap_monitor(&__kmp_monitor); |
5999 | TCW_4(__kmp_init_monitor, 0)(__kmp_init_monitor) = (0); |
6000 | } |
6001 | __kmp_release_bootstrap_lock(&__kmp_monitor_lock); |
6002 | KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end: monitor reaped\n" ); }; |
6003 | #endif |
6004 | } /* else !__kmp_global.t_active */ |
6005 | TCW_4(__kmp_init_gtid, FALSE)(__kmp_init_gtid) = (0); |
6006 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
6007 | |
6008 | __kmp_cleanup(); |
6009 | #if OMPT_SUPPORT1 |
6010 | ompt_fini(); |
6011 | #endif |
6012 | } |
6013 | |
6014 | void __kmp_internal_end_library(int gtid_req) { |
6015 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ |
6016 | /* this shouldn't be a race condition because __kmp_internal_end() is the |
6017 | only place to clear __kmp_serial_init */ |
6018 | /* we'll check this later too, after we get the lock */ |
6019 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks |
6020 | // redundaant, because the next check will work in any case. |
6021 | if (__kmp_global.g.g_abort) { |
6022 | KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"))if (kmp_a_debug >= 11) { __kmp_debug_printf ("__kmp_internal_end_library: abort, exiting\n" ); }; |
6023 | /* TODO abort? */ |
6024 | return; |
6025 | } |
6026 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { |
6027 | KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: already finished\n" ); }; |
6028 | return; |
6029 | } |
6030 | |
6031 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
6032 | |
6033 | /* find out who we are and what we should do */ |
6034 | { |
6035 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); |
6036 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: enter T#%d (%d)\n" , gtid, gtid_req); } |
6037 | 10, ("__kmp_internal_end_library: enter T#%d (%d)\n", gtid, gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: enter T#%d (%d)\n" , gtid, gtid_req); }; |
6038 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { |
6039 | KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n"); } |
6040 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: !__kmp_init_runtime, system " "already shutdown\n"); }; |
6041 | return; |
6042 | } else if (gtid == KMP_GTID_MONITOR(-4)) { |
6043 | KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: monitor thread, gtid not " "registered, or system shutdown\n"); } |
6044 | "registered, or system shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: monitor thread, gtid not " "registered, or system shutdown\n"); }; |
6045 | return; |
6046 | } else if (gtid == KMP_GTID_DNE(-2)) { |
6047 | KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n"); } |
6048 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: gtid not registered or system " "shutdown\n"); }; |
6049 | /* we don't know who we are, but we may still shutdown the library */ |
6050 | } else if (KMP_UBER_GTID(gtid)((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6050)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6050)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) { |
6051 | /* unregister ourselves as an uber thread. gtid is no longer valid */ |
6052 | if (__kmp_root[gtid]->r.r_active) { |
6053 | __kmp_global.g.g_abort = -1; |
6054 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); |
6055 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); } |
6056 | ("__kmp_internal_end_library: root still active, abort T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); } |
6057 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: root still active, abort T#%d\n" , gtid); }; |
6058 | return; |
6059 | } else { |
6060 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } |
6061 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); } |
6062 | ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: unregistering sibling T#%d\n" , gtid); }; |
6063 | __kmp_unregister_root_current_thread(gtid); |
6064 | } |
6065 | } else { |
6066 | /* worker threads may call this function through the atexit handler, if they |
6067 | * call exit() */ |
6068 | /* For now, skip the usual subsequent processing and just dump the debug buffer. |
6069 | TODO: do a thorough shutdown instead */ |
6070 | #ifdef DUMP_DEBUG_ON_EXIT |
6071 | if (__kmp_debug_buf) |
6072 | __kmp_dump_debug_buffer(); |
6073 | #endif |
6074 | return; |
6075 | } |
6076 | } |
6077 | /* synchronize the termination process */ |
6078 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
6079 | |
6080 | /* have we already finished */ |
6081 | if (__kmp_global.g.g_abort) { |
6082 | KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: abort, exiting\n" ); }; |
6083 | /* TODO abort? */ |
6084 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6085 | return; |
6086 | } |
6087 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { |
6088 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6089 | return; |
6090 | } |
6091 | |
6092 | /* We need this lock to enforce mutex between this reading of |
6093 | __kmp_threads_capacity and the writing by __kmp_register_root. |
6094 | Alternatively, we can use a counter of roots that is atomically updated by |
6095 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and |
6096 | __kmp_internal_end_*. */ |
6097 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
6098 | |
6099 | /* now we can safely conduct the actual termination */ |
6100 | __kmp_internal_end(); |
6101 | |
6102 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
6103 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6104 | |
6105 | KA_TRACE(10, ("__kmp_internal_end_library: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_library: exit\n" ); }; |
6106 | |
6107 | #ifdef DUMP_DEBUG_ON_EXIT |
6108 | if (__kmp_debug_buf) |
6109 | __kmp_dump_debug_buffer(); |
6110 | #endif |
6111 | |
6112 | #if KMP_OS_WINDOWS0 |
6113 | __kmp_close_console(); |
6114 | #endif |
6115 | |
6116 | __kmp_fini_allocator(); |
6117 | |
6118 | } // __kmp_internal_end_library |
6119 | |
6120 | void __kmp_internal_end_thread(int gtid_req) { |
6121 | int i; |
6122 | |
6123 | /* if we have already cleaned up, don't try again, it wouldn't be pretty */ |
6124 | /* this shouldn't be a race condition because __kmp_internal_end() is the |
6125 | * only place to clear __kmp_serial_init */ |
6126 | /* we'll check this later too, after we get the lock */ |
6127 | // 2009-09-06: We do not set g_abort without setting g_done. This check looks |
6128 | // redundant, because the next check will work in any case. |
6129 | if (__kmp_global.g.g_abort) { |
6130 | KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"))if (kmp_a_debug >= 11) { __kmp_debug_printf ("__kmp_internal_end_thread: abort, exiting\n" ); }; |
6131 | /* TODO abort? */ |
6132 | return; |
6133 | } |
6134 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { |
6135 | KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: already finished\n" ); }; |
6136 | return; |
6137 | } |
6138 | |
6139 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
6140 | |
6141 | /* find out who we are and what we should do */ |
6142 | { |
6143 | int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); |
6144 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: enter T#%d (%d)\n" , gtid, gtid_req); } |
6145 | ("__kmp_internal_end_thread: enter T#%d (%d)\n", gtid, gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: enter T#%d (%d)\n" , gtid, gtid_req); }; |
6146 | if (gtid == KMP_GTID_SHUTDOWN(-3)) { |
6147 | KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n"); } |
6148 | "already shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: !__kmp_init_runtime, system " "already shutdown\n"); }; |
6149 | return; |
6150 | } else if (gtid == KMP_GTID_MONITOR(-4)) { |
6151 | KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: monitor thread, gtid not " "registered, or system shutdown\n"); } |
6152 | "registered, or system shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: monitor thread, gtid not " "registered, or system shutdown\n"); }; |
6153 | return; |
6154 | } else if (gtid == KMP_GTID_DNE(-2)) { |
6155 | KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n"); } |
6156 | "shutdown\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: gtid not registered or system " "shutdown\n"); }; |
6157 | return; |
6158 | /* we don't know who we are */ |
6159 | } else if (KMP_UBER_GTID(gtid)((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6159)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6159)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) { |
6160 | /* unregister ourselves as an uber thread. gtid is no longer valid */ |
6161 | if (__kmp_root[gtid]->r.r_active) { |
6162 | __kmp_global.g.g_abort = -1; |
6163 | TCW_SYNC_4(__kmp_global.g.g_done, TRUE)(__kmp_global.g.g_done) = ((!0)); |
6164 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); } |
6165 | ("__kmp_internal_end_thread: root still active, abort T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); } |
6166 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: root still active, abort T#%d\n" , gtid); }; |
6167 | return; |
6168 | } else { |
6169 | KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: unregistering sibling T#%d\n" , gtid); } |
6170 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: unregistering sibling T#%d\n" , gtid); }; |
6171 | __kmp_unregister_root_current_thread(gtid); |
6172 | } |
6173 | } else { |
6174 | /* just a worker thread, let's leave */ |
6175 | KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread T#%d\n" , gtid); }; |
6176 | |
6177 | if (gtid >= 0) { |
6178 | __kmp_threads[gtid]->th.th_task_team = NULL__null; |
6179 | } |
6180 | |
6181 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); } |
6182 | ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); } |
6183 | gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n" , gtid); }; |
6184 | return; |
6185 | } |
6186 | } |
6187 | #if defined KMP_DYNAMIC_LIB1 |
6188 | // AC: lets not shutdown the Linux* OS dynamic library at the exit of uber |
6189 | // thread, because we will better shutdown later in the library destructor. |
6190 | // The reason of this change is performance problem when non-openmp thread in |
6191 | // a loop forks and joins many openmp threads. We can save a lot of time |
6192 | // keeping worker threads alive until the program shutdown. |
6193 | // OM: Removed Linux* OS restriction to fix the crash on OS X* (DPD200239966) |
6194 | // and Windows(DPD200287443) that occurs when using critical sections from |
6195 | // foreign threads. |
6196 | KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: exiting T#%d\n" , gtid_req); }; |
6197 | return; |
6198 | #endif |
6199 | /* synchronize the termination process */ |
6200 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
6201 | |
6202 | /* have we already finished */ |
6203 | if (__kmp_global.g.g_abort) { |
6204 | KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: abort, exiting\n" ); }; |
6205 | /* TODO abort? */ |
6206 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6207 | return; |
6208 | } |
6209 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done) || !__kmp_init_serial) { |
6210 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6211 | return; |
6212 | } |
6213 | |
6214 | /* We need this lock to enforce mutex between this reading of |
6215 | __kmp_threads_capacity and the writing by __kmp_register_root. |
6216 | Alternatively, we can use a counter of roots that is atomically updated by |
6217 | __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and |
6218 | __kmp_internal_end_*. */ |
6219 | |
6220 | /* should we finish the run-time? are all siblings done? */ |
6221 | __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); |
6222 | |
6223 | for (i = 0; i < __kmp_threads_capacity; ++i) { |
6224 | if (KMP_UBER_GTID(i)((((i) >= (-6)) ? 0 : __kmp_debug_assert("(i) >= (-6)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6224)), (((i) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(i) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6224)), (i) >= 0 && __kmp_root[(i)] && __kmp_threads [(i)] && (__kmp_threads[(i)] == __kmp_root[(i)]->r .r_uber_thread))) { |
6225 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } |
6226 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); } |
6227 | ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n" , i); }; |
6228 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
6229 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6230 | return; |
6231 | } |
6232 | } |
6233 | |
6234 | /* now we can safely conduct the actual termination */ |
6235 | |
6236 | __kmp_internal_end(); |
6237 | |
6238 | __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); |
6239 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6240 | |
6241 | KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_internal_end_thread: exit T#%d\n" , gtid_req); }; |
6242 | |
6243 | #ifdef DUMP_DEBUG_ON_EXIT |
6244 | if (__kmp_debug_buf) |
6245 | __kmp_dump_debug_buffer(); |
6246 | #endif |
6247 | } // __kmp_internal_end_thread |
6248 | |
6249 | // ----------------------------------------------------------------------------- |
6250 | // Library registration stuff. |
6251 | |
6252 | static long __kmp_registration_flag = 0; |
6253 | // Random value used to indicate library initialization. |
6254 | static char *__kmp_registration_str = NULL__null; |
6255 | // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>. |
6256 | |
6257 | static inline char *__kmp_reg_status_name() { |
6258 | /* On RHEL 3u5 if linked statically, getpid() returns different values in |
6259 | each thread. If registration and unregistration go in different threads |
6260 | (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env |
6261 | env var can not be found, because the name will contain different pid. */ |
6262 | return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid()); |
6263 | } // __kmp_reg_status_get |
6264 | |
6265 | void __kmp_register_library_startup(void) { |
6266 | |
6267 | char *name = __kmp_reg_status_name(); // Name of the environment variable. |
6268 | int done = 0; |
6269 | union { |
6270 | double dtime; |
6271 | long ltime; |
6272 | } time; |
6273 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
6274 | __kmp_initialize_system_tick(); |
6275 | #endif |
6276 | __kmp_read_system_time(&time.dtime); |
6277 | __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL); |
6278 | __kmp_registration_str = |
6279 | __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag, |
6280 | __kmp_registration_flag, KMP_LIBRARY_FILE"libomp.so.5"); |
6281 | |
6282 | KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,if (kmp_a_debug >= 50) { __kmp_debug_printf ("__kmp_register_library_startup: %s=\"%s\"\n" , name, __kmp_registration_str); } |
6283 | __kmp_registration_str))if (kmp_a_debug >= 50) { __kmp_debug_printf ("__kmp_register_library_startup: %s=\"%s\"\n" , name, __kmp_registration_str); }; |
6284 | |
6285 | while (!done) { |
6286 | |
6287 | char *value = NULL__null; // Actual value of the environment variable. |
6288 | |
6289 | // Set environment variable, but do not overwrite if it is exist. |
6290 | __kmp_env_set(name, __kmp_registration_str, 0); |
6291 | // Check the variable is written. |
6292 | value = __kmp_env_get(name); |
6293 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { |
6294 | |
6295 | done = 1; // Ok, environment variable set successfully, exit the loop. |
6296 | |
6297 | } else { |
6298 | |
6299 | // Oops. Write failed. Another copy of OpenMP RTL is in memory. |
6300 | // Check whether it alive or dead. |
6301 | int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead. |
6302 | char *tail = value; |
6303 | char *flag_addr_str = NULL__null; |
6304 | char *flag_val_str = NULL__null; |
6305 | char const *file_name = NULL__null; |
6306 | __kmp_str_split(tail, '-', &flag_addr_str, &tail); |
6307 | __kmp_str_split(tail, '-', &flag_val_str, &tail); |
6308 | file_name = tail; |
6309 | if (tail != NULL__null) { |
6310 | long *flag_addr = 0; |
6311 | long flag_val = 0; |
6312 | KMP_SSCANFsscanf(flag_addr_str, "%p", &flag_addr); |
6313 | KMP_SSCANFsscanf(flag_val_str, "%lx", &flag_val); |
6314 | if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) { |
6315 | // First, check whether environment-encoded address is mapped into |
6316 | // addr space. |
6317 | // If so, dereference it to see if it still has the right value. |
6318 | if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) { |
6319 | neighbor = 1; |
6320 | } else { |
6321 | // If not, then we know the other copy of the library is no longer |
6322 | // running. |
6323 | neighbor = 2; |
6324 | } |
6325 | } |
6326 | } |
6327 | switch (neighbor) { |
6328 | case 0: // Cannot parse environment variable -- neighbor status unknown. |
6329 | // Assume it is the incompatible format of future version of the |
6330 | // library. Assume the other library is alive. |
6331 | // WARN( ... ); // TODO: Issue a warning. |
6332 | file_name = "unknown library"; |
6333 | // Attention! Falling to the next case. That's intentional. |
6334 | case 1: { // Neighbor is alive. |
6335 | // Check it is allowed. |
6336 | char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK"); |
6337 | if (!__kmp_str_match_true(duplicate_ok)) { |
6338 | // That's not allowed. Issue fatal error. |
6339 | __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name)__kmp_msg_format(kmp_i18n_msg_DuplicateLibrary, "libomp.so.5" , file_name), |
6340 | KMP_HNT(DuplicateLibrary)__kmp_msg_format(kmp_i18n_hnt_DuplicateLibrary), __kmp_msg_null); |
6341 | } |
6342 | KMP_INTERNAL_FREE(duplicate_ok)free(duplicate_ok); |
6343 | __kmp_duplicate_library_ok = 1; |
6344 | done = 1; // Exit the loop. |
6345 | } break; |
6346 | case 2: { // Neighbor is dead. |
6347 | // Clear the variable and try to register library again. |
6348 | __kmp_env_unset(name); |
6349 | } break; |
6350 | default: { KMP_DEBUG_ASSERT(0)((0) ? 0 : __kmp_debug_assert("0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6350)); } break; |
6351 | } |
6352 | } |
6353 | KMP_INTERNAL_FREE((void *)value)free((void *)value); |
6354 | } |
6355 | KMP_INTERNAL_FREE((void *)name)free((void *)name); |
6356 | |
6357 | } // func __kmp_register_library_startup |
6358 | |
6359 | void __kmp_unregister_library(void) { |
6360 | |
6361 | char *name = __kmp_reg_status_name(); |
6362 | char *value = __kmp_env_get(name); |
6363 | |
6364 | KMP_DEBUG_ASSERT(__kmp_registration_flag != 0)((__kmp_registration_flag != 0) ? 0 : __kmp_debug_assert("__kmp_registration_flag != 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6364)); |
6365 | KMP_DEBUG_ASSERT(__kmp_registration_str != NULL)((__kmp_registration_str != __null) ? 0 : __kmp_debug_assert( "__kmp_registration_str != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6365)); |
6366 | if (value != NULL__null && strcmp(value, __kmp_registration_str) == 0) { |
6367 | // Ok, this is our variable. Delete it. |
6368 | __kmp_env_unset(name); |
6369 | } |
6370 | |
6371 | KMP_INTERNAL_FREE(__kmp_registration_str)free(__kmp_registration_str); |
6372 | KMP_INTERNAL_FREE(value)free(value); |
6373 | KMP_INTERNAL_FREE(name)free(name); |
6374 | |
6375 | __kmp_registration_flag = 0; |
6376 | __kmp_registration_str = NULL__null; |
6377 | |
6378 | } // __kmp_unregister_library |
6379 | |
6380 | // End of Library registration stuff. |
6381 | // ----------------------------------------------------------------------------- |
6382 | |
6383 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
6384 | |
6385 | static void __kmp_check_mic_type() { |
6386 | kmp_cpuid_t cpuid_state = {0}; |
6387 | kmp_cpuid_t *cs_p = &cpuid_state; |
6388 | __kmp_x86_cpuid(1, 0, cs_p); |
6389 | // We don't support mic1 at the moment |
6390 | if ((cs_p->eax & 0xff0) == 0xB10) { |
6391 | __kmp_mic_type = mic2; |
6392 | } else if ((cs_p->eax & 0xf0ff0) == 0x50670) { |
6393 | __kmp_mic_type = mic3; |
6394 | } else { |
6395 | __kmp_mic_type = non_mic; |
6396 | } |
6397 | } |
6398 | |
6399 | #endif /* KMP_MIC_SUPPORTED */ |
6400 | |
6401 | static void __kmp_do_serial_initialize(void) { |
6402 | int i, gtid; |
6403 | int size; |
6404 | |
6405 | KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: enter\n" ); }; |
6406 | |
6407 | KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4)((sizeof(kmp_int32) == 4) ? 0 : __kmp_debug_assert("sizeof(kmp_int32) == 4" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6407)); |
6408 | KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4)((sizeof(kmp_uint32) == 4) ? 0 : __kmp_debug_assert("sizeof(kmp_uint32) == 4" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6408)); |
6409 | KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8)((sizeof(kmp_int64) == 8) ? 0 : __kmp_debug_assert("sizeof(kmp_int64) == 8" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6409)); |
6410 | KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8)((sizeof(kmp_uint64) == 8) ? 0 : __kmp_debug_assert("sizeof(kmp_uint64) == 8" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6410)); |
6411 | KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *))((sizeof(kmp_intptr_t) == sizeof(void *)) ? 0 : __kmp_debug_assert ("sizeof(kmp_intptr_t) == sizeof(void *)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6411)); |
6412 | |
6413 | #if OMPT_SUPPORT1 |
6414 | ompt_pre_init(); |
6415 | #endif |
6416 | |
6417 | __kmp_validate_locks(); |
6418 | |
6419 | /* Initialize internal memory allocator */ |
6420 | __kmp_init_allocator(); |
6421 | |
6422 | /* Register the library startup via an environment variable and check to see |
6423 | whether another copy of the library is already registered. */ |
6424 | |
6425 | __kmp_register_library_startup(); |
6426 | |
6427 | /* TODO reinitialization of library */ |
6428 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { |
6429 | KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: reinitialization of library\n" ); }; |
6430 | } |
6431 | |
6432 | __kmp_global.g.g_abort = 0; |
6433 | TCW_SYNC_4(__kmp_global.g.g_done, FALSE)(__kmp_global.g.g_done) = (0); |
6434 | |
6435 | /* initialize the locks */ |
6436 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
6437 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 |
6438 | __kmp_init_speculative_stats(); |
6439 | #endif |
6440 | #endif |
6441 | #if KMP_STATS_ENABLED0 |
6442 | __kmp_stats_init(); |
6443 | #endif |
6444 | __kmp_init_lock(&__kmp_global_lock); |
6445 | __kmp_init_queuing_lock(&__kmp_dispatch_lock); |
6446 | __kmp_init_lock(&__kmp_debug_lock); |
6447 | __kmp_init_atomic_lock(&__kmp_atomic_lock); |
6448 | __kmp_init_atomic_lock(&__kmp_atomic_lock_1i); |
6449 | __kmp_init_atomic_lock(&__kmp_atomic_lock_2i); |
6450 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4i); |
6451 | __kmp_init_atomic_lock(&__kmp_atomic_lock_4r); |
6452 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8i); |
6453 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8r); |
6454 | __kmp_init_atomic_lock(&__kmp_atomic_lock_8c); |
6455 | __kmp_init_atomic_lock(&__kmp_atomic_lock_10r); |
6456 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16r); |
6457 | __kmp_init_atomic_lock(&__kmp_atomic_lock_16c); |
6458 | __kmp_init_atomic_lock(&__kmp_atomic_lock_20c); |
6459 | __kmp_init_atomic_lock(&__kmp_atomic_lock_32c); |
6460 | __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock); |
6461 | __kmp_init_bootstrap_lock(&__kmp_exit_lock); |
6462 | #if KMP_USE_MONITOR |
6463 | __kmp_init_bootstrap_lock(&__kmp_monitor_lock); |
6464 | #endif |
6465 | __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock); |
6466 | |
6467 | /* conduct initialization and initial setup of configuration */ |
6468 | |
6469 | __kmp_runtime_initialize(); |
6470 | |
6471 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
6472 | __kmp_check_mic_type(); |
6473 | #endif |
6474 | |
6475 | // Some global variable initialization moved here from kmp_env_initialize() |
6476 | #ifdef KMP_DEBUG1 |
6477 | kmp_diag = 0; |
6478 | #endif |
6479 | __kmp_abort_delay = 0; |
6480 | |
6481 | // From __kmp_init_dflt_team_nth() |
6482 | /* assume the entire machine will be used */ |
6483 | __kmp_dflt_team_nth_ub = __kmp_xproc; |
6484 | if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH1) { |
6485 | __kmp_dflt_team_nth_ub = KMP_MIN_NTH1; |
6486 | } |
6487 | if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) { |
6488 | __kmp_dflt_team_nth_ub = __kmp_sys_max_nth; |
6489 | } |
6490 | __kmp_max_nth = __kmp_sys_max_nth; |
6491 | __kmp_cg_max_nth = __kmp_sys_max_nth; |
6492 | __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default |
6493 | if (__kmp_teams_max_nth > __kmp_sys_max_nth) { |
6494 | __kmp_teams_max_nth = __kmp_sys_max_nth; |
6495 | } |
6496 | |
6497 | // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME" |
6498 | // part |
6499 | __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME(200); |
6500 | #if KMP_USE_MONITOR |
6501 | __kmp_monitor_wakeups = |
6502 | KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); |
6503 | __kmp_bt_intervals = |
6504 | KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); |
6505 | #endif |
6506 | // From "KMP_LIBRARY" part of __kmp_env_initialize() |
6507 | __kmp_library = library_throughput; |
6508 | // From KMP_SCHEDULE initialization |
6509 | __kmp_static = kmp_sch_static_balanced; |
6510 | // AC: do not use analytical here, because it is non-monotonous |
6511 | //__kmp_guided = kmp_sch_guided_iterative_chunked; |
6512 | //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no |
6513 | // need to repeat assignment |
6514 | // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch |
6515 | // bit control and barrier method control parts |
6516 | #if KMP_FAST_REDUCTION_BARRIER1 |
6517 | #define kmp_reduction_barrier_gather_bb ((int)1) |
6518 | #define kmp_reduction_barrier_release_bb ((int)1) |
6519 | #define kmp_reduction_barrier_gather_pat bp_hyper_bar |
6520 | #define kmp_reduction_barrier_release_pat bp_hyper_bar |
6521 | #endif // KMP_FAST_REDUCTION_BARRIER |
6522 | for (i = bs_plain_barrier; i < bs_last_barrier; i++) { |
6523 | __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt; |
6524 | __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt; |
6525 | __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt; |
6526 | __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt; |
6527 | #if KMP_FAST_REDUCTION_BARRIER1 |
6528 | if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only ( |
6529 | // lin_64 ): hyper,1 |
6530 | __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb; |
6531 | __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb; |
6532 | __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat; |
6533 | __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat; |
6534 | } |
6535 | #endif // KMP_FAST_REDUCTION_BARRIER |
6536 | } |
6537 | #if KMP_FAST_REDUCTION_BARRIER1 |
6538 | #undef kmp_reduction_barrier_release_pat |
6539 | #undef kmp_reduction_barrier_gather_pat |
6540 | #undef kmp_reduction_barrier_release_bb |
6541 | #undef kmp_reduction_barrier_gather_bb |
6542 | #endif // KMP_FAST_REDUCTION_BARRIER |
6543 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
6544 | if (__kmp_mic_type == mic2) { // KNC |
6545 | // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC |
6546 | __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather |
6547 | __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] = |
6548 | 1; // forkjoin release |
6549 | __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; |
6550 | __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; |
6551 | } |
6552 | #if KMP_FAST_REDUCTION_BARRIER1 |
6553 | if (__kmp_mic_type == mic2) { // KNC |
6554 | __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar; |
6555 | __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar; |
6556 | } |
6557 | #endif // KMP_FAST_REDUCTION_BARRIER |
6558 | #endif // KMP_MIC_SUPPORTED |
6559 | |
6560 | // From KMP_CHECKS initialization |
6561 | #ifdef KMP_DEBUG1 |
6562 | __kmp_env_checks = TRUE(!0); /* development versions have the extra checks */ |
6563 | #else |
6564 | __kmp_env_checks = FALSE0; /* port versions do not have the extra checks */ |
6565 | #endif |
6566 | |
6567 | // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization |
6568 | __kmp_foreign_tp = TRUE(!0); |
6569 | |
6570 | __kmp_global.g.g_dynamic = FALSE0; |
6571 | __kmp_global.g.g_dynamic_mode = dynamic_default; |
6572 | |
6573 | __kmp_env_initialize(NULL__null); |
6574 | |
6575 | // Print all messages in message catalog for testing purposes. |
6576 | #ifdef KMP_DEBUG1 |
6577 | char const *val = __kmp_env_get("KMP_DUMP_CATALOG"); |
6578 | if (__kmp_str_match_true(val)) { |
6579 | kmp_str_buf_t buffer; |
6580 | __kmp_str_buf_init(&buffer){ (&buffer)->str = (&buffer)->bulk; (&buffer )->size = sizeof((&buffer)->bulk); (&buffer)-> used = 0; (&buffer)->bulk[0] = 0; }; |
6581 | __kmp_i18n_dump_catalog(&buffer); |
6582 | __kmp_printf("%s", buffer.str); |
6583 | __kmp_str_buf_free(&buffer); |
6584 | } |
6585 | __kmp_env_free(&val); |
6586 | #endif |
6587 | |
6588 | __kmp_threads_capacity = |
6589 | __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub); |
6590 | // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part |
6591 | __kmp_tp_capacity = __kmp_default_tp_capacity( |
6592 | __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified); |
6593 | |
6594 | // If the library is shut down properly, both pools must be NULL. Just in |
6595 | // case, set them to NULL -- some memory may leak, but subsequent code will |
6596 | // work even if pools are not freed. |
6597 | KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL)((__kmp_thread_pool == __null) ? 0 : __kmp_debug_assert("__kmp_thread_pool == __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6597)); |
6598 | KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL)((__kmp_thread_pool_insert_pt == __null) ? 0 : __kmp_debug_assert ("__kmp_thread_pool_insert_pt == __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6598)); |
6599 | KMP_DEBUG_ASSERT(__kmp_team_pool == NULL)((__kmp_team_pool == __null) ? 0 : __kmp_debug_assert("__kmp_team_pool == __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6599)); |
6600 | __kmp_thread_pool = NULL__null; |
6601 | __kmp_thread_pool_insert_pt = NULL__null; |
6602 | __kmp_team_pool = NULL__null; |
6603 | |
6604 | /* Allocate all of the variable sized records */ |
6605 | /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are |
6606 | * expandable */ |
6607 | /* Since allocation is cache-aligned, just add extra padding at the end */ |
6608 | size = |
6609 | (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity + |
6610 | CACHE_LINE64; |
6611 | __kmp_threads = (kmp_info_t **)__kmp_allocate(size)___kmp_allocate((size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6611); |
6612 | __kmp_root = (kmp_root_t **)((char *)__kmp_threads + |
6613 | sizeof(kmp_info_t *) * __kmp_threads_capacity); |
6614 | |
6615 | /* init thread counts */ |
6616 | KMP_DEBUG_ASSERT(__kmp_all_nth ==((__kmp_all_nth == 0) ? 0 : __kmp_debug_assert("__kmp_all_nth == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6617)) |
6617 | 0)((__kmp_all_nth == 0) ? 0 : __kmp_debug_assert("__kmp_all_nth == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6617)); // Asserts fail if the library is reinitializing and |
6618 | KMP_DEBUG_ASSERT(__kmp_nth == 0)((__kmp_nth == 0) ? 0 : __kmp_debug_assert("__kmp_nth == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6618)); // something was wrong in termination. |
6619 | __kmp_all_nth = 0; |
6620 | __kmp_nth = 0; |
6621 | |
6622 | /* setup the uber master thread and hierarchy */ |
6623 | gtid = __kmp_register_root(TRUE(!0)); |
6624 | KA_TRACE(10, ("__kmp_do_serial_initialize T#%d\n", gtid))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize T#%d\n" , gtid); }; |
6625 | KMP_ASSERT(KMP_UBER_GTID(gtid))((((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6625)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6625)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) ? 0 : __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6625)); |
6626 | KMP_ASSERT(KMP_INITIAL_GTID(gtid))((((gtid) == 0)) ? 0 : __kmp_debug_assert("KMP_INITIAL_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6626)); |
6627 | |
6628 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
6629 | |
6630 | __kmp_common_initialize(); |
6631 | |
6632 | #if KMP_OS_UNIX1 |
6633 | /* invoke the child fork handler */ |
6634 | __kmp_register_atfork(); |
6635 | #endif |
6636 | |
6637 | #if !defined KMP_DYNAMIC_LIB1 |
6638 | { |
6639 | /* Invoke the exit handler when the program finishes, only for static |
6640 | library. For dynamic library, we already have _fini and DllMain. */ |
6641 | int rc = atexit(__kmp_internal_end_atexit); |
6642 | if (rc != 0) { |
6643 | __kmp_fatal(KMP_MSG(FunctionError, "atexit()")__kmp_msg_format(kmp_i18n_msg_FunctionError, "atexit()"), KMP_ERR(rc)__kmp_msg_error_code(rc), |
6644 | __kmp_msg_null); |
6645 | } |
6646 | } |
6647 | #endif |
6648 | |
6649 | #if KMP_HANDLE_SIGNALS(1 || 0) |
6650 | #if KMP_OS_UNIX1 |
6651 | /* NOTE: make sure that this is called before the user installs their own |
6652 | signal handlers so that the user handlers are called first. this way they |
6653 | can return false, not call our handler, avoid terminating the library, and |
6654 | continue execution where they left off. */ |
6655 | __kmp_install_signals(FALSE0); |
6656 | #endif /* KMP_OS_UNIX */ |
6657 | #if KMP_OS_WINDOWS0 |
6658 | __kmp_install_signals(TRUE(!0)); |
6659 | #endif /* KMP_OS_WINDOWS */ |
6660 | #endif |
6661 | |
6662 | /* we have finished the serial initialization */ |
6663 | __kmp_init_counter++; |
6664 | |
6665 | __kmp_init_serial = TRUE(!0); |
6666 | |
6667 | if (__kmp_settings) { |
6668 | __kmp_env_print(); |
6669 | } |
6670 | |
6671 | #if OMP_40_ENABLED(50 >= 40) |
6672 | if (__kmp_display_env || __kmp_display_env_verbose) { |
6673 | __kmp_env_print_2(); |
6674 | } |
6675 | #endif // OMP_40_ENABLED |
6676 | |
6677 | #if OMPT_SUPPORT1 |
6678 | ompt_post_init(); |
6679 | #endif |
6680 | |
6681 | KMP_MB(); |
6682 | |
6683 | KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_serial_initialize: exit\n" ); }; |
6684 | } |
6685 | |
6686 | void __kmp_serial_initialize(void) { |
6687 | if (__kmp_init_serial) { |
6688 | return; |
6689 | } |
6690 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
6691 | if (__kmp_init_serial) { |
6692 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6693 | return; |
6694 | } |
6695 | __kmp_do_serial_initialize(); |
6696 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6697 | } |
6698 | |
6699 | static void __kmp_do_middle_initialize(void) { |
6700 | int i, j; |
6701 | int prev_dflt_team_nth; |
6702 | |
6703 | if (!__kmp_init_serial) { |
6704 | __kmp_do_serial_initialize(); |
6705 | } |
6706 | |
6707 | KA_TRACE(10, ("__kmp_middle_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_middle_initialize: enter\n" ); }; |
6708 | |
6709 | // Save the previous value for the __kmp_dflt_team_nth so that |
6710 | // we can avoid some reinitialization if it hasn't changed. |
6711 | prev_dflt_team_nth = __kmp_dflt_team_nth; |
6712 | |
6713 | #if KMP_AFFINITY_SUPPORTED1 |
6714 | // __kmp_affinity_initialize() will try to set __kmp_ncores to the |
6715 | // number of cores on the machine. |
6716 | __kmp_affinity_initialize(); |
6717 | |
6718 | // Run through the __kmp_threads array and set the affinity mask |
6719 | // for each root thread that is currently registered with the RTL. |
6720 | for (i = 0; i < __kmp_threads_capacity; i++) { |
6721 | if (TCR_PTR(__kmp_threads[i])((void *)(__kmp_threads[i])) != NULL__null) { |
6722 | __kmp_affinity_set_init_mask(i, TRUE(!0)); |
6723 | } |
6724 | } |
6725 | #endif /* KMP_AFFINITY_SUPPORTED */ |
6726 | |
6727 | KMP_ASSERT(__kmp_xproc > 0)((__kmp_xproc > 0) ? 0 : __kmp_debug_assert("__kmp_xproc > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6727)); |
6728 | if (__kmp_avail_proc == 0) { |
6729 | __kmp_avail_proc = __kmp_xproc; |
6730 | } |
6731 | |
6732 | // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3), |
6733 | // correct them now |
6734 | j = 0; |
6735 | while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) { |
6736 | __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub = |
6737 | __kmp_avail_proc; |
6738 | j++; |
6739 | } |
6740 | |
6741 | if (__kmp_dflt_team_nth == 0) { |
6742 | #ifdef KMP_DFLT_NTH_CORES |
6743 | // Default #threads = #cores |
6744 | __kmp_dflt_team_nth = __kmp_ncores; |
6745 | KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); } |
6746 | "__kmp_ncores (%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); } |
6747 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_ncores (%d)\n", __kmp_dflt_team_nth); }; |
6748 | #else |
6749 | // Default #threads = #available OS procs |
6750 | __kmp_dflt_team_nth = __kmp_avail_proc; |
6751 | KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); } |
6752 | "__kmp_avail_proc(%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); } |
6753 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " "__kmp_avail_proc(%d)\n", __kmp_dflt_team_nth); }; |
6754 | #endif /* KMP_DFLT_NTH_CORES */ |
6755 | } |
6756 | |
6757 | if (__kmp_dflt_team_nth < KMP_MIN_NTH1) { |
6758 | __kmp_dflt_team_nth = KMP_MIN_NTH1; |
6759 | } |
6760 | if (__kmp_dflt_team_nth > __kmp_sys_max_nth) { |
6761 | __kmp_dflt_team_nth = __kmp_sys_max_nth; |
6762 | } |
6763 | |
6764 | // There's no harm in continuing if the following check fails, |
6765 | // but it indicates an error in the previous logic. |
6766 | KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub)((__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub) ? 0 : __kmp_debug_assert ("__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6766)); |
6767 | |
6768 | if (__kmp_dflt_team_nth != prev_dflt_team_nth) { |
6769 | // Run through the __kmp_threads array and set the num threads icv for each |
6770 | // root thread that is currently registered with the RTL (which has not |
6771 | // already explicitly set its nthreads-var with a call to |
6772 | // omp_set_num_threads()). |
6773 | for (i = 0; i < __kmp_threads_capacity; i++) { |
6774 | kmp_info_t *thread = __kmp_threads[i]; |
6775 | if (thread == NULL__null) |
6776 | continue; |
6777 | if (thread->th.th_current_task->td_icvs.nproc != 0) |
6778 | continue; |
6779 | |
6780 | set__nproc(__kmp_threads[i], __kmp_dflt_team_nth)(((__kmp_threads[i])->th.th_current_task->td_icvs.nproc ) = (__kmp_dflt_team_nth)); |
6781 | } |
6782 | } |
6783 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } |
6784 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } |
6785 | ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); } |
6786 | __kmp_dflt_team_nth))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n" , __kmp_dflt_team_nth); }; |
6787 | |
6788 | #ifdef KMP_ADJUST_BLOCKTIME1 |
6789 | /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */ |
6790 | if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { |
6791 | KMP_DEBUG_ASSERT(__kmp_avail_proc > 0)((__kmp_avail_proc > 0) ? 0 : __kmp_debug_assert("__kmp_avail_proc > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6791)); |
6792 | if (__kmp_nth > __kmp_avail_proc) { |
6793 | __kmp_zero_bt = TRUE(!0); |
6794 | } |
6795 | } |
6796 | #endif /* KMP_ADJUST_BLOCKTIME */ |
6797 | |
6798 | /* we have finished middle initialization */ |
6799 | TCW_SYNC_4(__kmp_init_middle, TRUE)(__kmp_init_middle) = ((!0)); |
6800 | |
6801 | KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_do_middle_initialize: exit\n" ); }; |
6802 | } |
6803 | |
6804 | void __kmp_middle_initialize(void) { |
6805 | if (__kmp_init_middle) { |
6806 | return; |
6807 | } |
6808 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
6809 | if (__kmp_init_middle) { |
6810 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6811 | return; |
6812 | } |
6813 | __kmp_do_middle_initialize(); |
6814 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6815 | } |
6816 | |
6817 | void __kmp_parallel_initialize(void) { |
6818 | int gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); // this might be a new root |
6819 | |
6820 | /* synchronize parallel initialization (for sibling) */ |
6821 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) |
6822 | return; |
6823 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
6824 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { |
6825 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6826 | return; |
6827 | } |
6828 | |
6829 | /* TODO reinitialization after we have already shut down */ |
6830 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { |
6831 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } |
6832 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); } |
6833 | ("__kmp_parallel_initialize: attempt to init while shutting down\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: attempt to init while shutting down\n" ); }; |
6834 | __kmp_infinite_loop(); |
6835 | } |
6836 | |
6837 | /* jc: The lock __kmp_initz_lock is already held, so calling |
6838 | __kmp_serial_initialize would cause a deadlock. So we call |
6839 | __kmp_do_serial_initialize directly. */ |
6840 | if (!__kmp_init_middle) { |
6841 | __kmp_do_middle_initialize(); |
6842 | } |
6843 | |
6844 | /* begin initialization */ |
6845 | KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: enter\n" ); }; |
6846 | KMP_ASSERT(KMP_UBER_GTID(gtid))((((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6846)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6846)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) ? 0 : __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6846)); |
6847 | |
6848 | #if KMP_ARCH_X860 || KMP_ARCH_X86_641 |
6849 | // Save the FP control regs. |
6850 | // Worker threads will set theirs to these values at thread startup. |
6851 | __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); |
6852 | __kmp_store_mxcsr(&__kmp_init_mxcsr); |
6853 | __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK0xffffffc0; |
6854 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
6855 | |
6856 | #if KMP_OS_UNIX1 |
6857 | #if KMP_HANDLE_SIGNALS(1 || 0) |
6858 | /* must be after __kmp_serial_initialize */ |
6859 | __kmp_install_signals(TRUE(!0)); |
6860 | #endif |
6861 | #endif |
6862 | |
6863 | __kmp_suspend_initialize(); |
6864 | |
6865 | #if defined(USE_LOAD_BALANCE1) |
6866 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { |
6867 | __kmp_global.g.g_dynamic_mode = dynamic_load_balance; |
6868 | } |
6869 | #else |
6870 | if (__kmp_global.g.g_dynamic_mode == dynamic_default) { |
6871 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; |
6872 | } |
6873 | #endif |
6874 | |
6875 | if (__kmp_version) { |
6876 | __kmp_print_version_2(); |
6877 | } |
6878 | |
6879 | /* we have finished parallel initialization */ |
6880 | TCW_SYNC_4(__kmp_init_parallel, TRUE)(__kmp_init_parallel) = ((!0)); |
6881 | |
6882 | KMP_MB(); |
6883 | KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_parallel_initialize: exit\n" ); }; |
6884 | |
6885 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
6886 | } |
6887 | |
6888 | /* ------------------------------------------------------------------------ */ |
6889 | |
6890 | void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr, |
6891 | kmp_team_t *team) { |
6892 | kmp_disp_t *dispatch; |
6893 | |
6894 | KMP_MB(); |
6895 | |
6896 | /* none of the threads have encountered any constructs, yet. */ |
6897 | this_thr->th.th_local.this_construct = 0; |
6898 | #if KMP_CACHE_MANAGE |
6899 | KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived); |
6900 | #endif /* KMP_CACHE_MANAGE */ |
6901 | dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch)((void *)(this_thr->th.th_dispatch)); |
6902 | KMP_DEBUG_ASSERT(dispatch)((dispatch) ? 0 : __kmp_debug_assert("dispatch", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6902)); |
6903 | KMP_DEBUG_ASSERT(team->t.t_dispatch)((team->t.t_dispatch) ? 0 : __kmp_debug_assert("team->t.t_dispatch" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6903)); |
6904 | // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[ |
6905 | // this_thr->th.th_info.ds.ds_tid ] ); |
6906 | |
6907 | dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */ |
6908 | #if OMP_45_ENABLED(50 >= 45) |
6909 | dispatch->th_doacross_buf_idx = |
6910 | 0; /* reset the doacross dispatch buffer counter */ |
6911 | #endif |
6912 | if (__kmp_env_consistency_check) |
6913 | __kmp_push_parallel(gtid, team->t.t_ident); |
6914 | |
6915 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
6916 | } |
6917 | |
6918 | void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr, |
6919 | kmp_team_t *team) { |
6920 | if (__kmp_env_consistency_check) |
6921 | __kmp_pop_parallel(gtid, team->t.t_ident); |
6922 | |
6923 | __kmp_finish_implicit_task(this_thr); |
6924 | } |
6925 | |
6926 | int __kmp_invoke_task_func(int gtid) { |
6927 | int rc; |
6928 | int tid = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6928)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
6929 | kmp_info_t *this_thr = __kmp_threads[gtid]; |
6930 | kmp_team_t *team = this_thr->th.th_team; |
6931 | |
6932 | __kmp_run_before_invoked_task(gtid, tid, this_thr, team); |
6933 | #if USE_ITT_BUILD1 |
6934 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { |
6935 | __kmp_itt_stack_callee_enter( |
6936 | (__itt_caller) |
6937 | team->t.t_stack_id); // inform ittnotify about entering user's code |
6938 | } |
6939 | #endif /* USE_ITT_BUILD */ |
6940 | #if INCLUDE_SSC_MARKS(1 && 1) |
6941 | SSC_MARK_INVOKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd695) : "%ebx"); |
6942 | #endif |
6943 | |
6944 | #if OMPT_SUPPORT1 |
6945 | void *dummy; |
6946 | void **exit_runtime_p; |
6947 | ompt_data_t *my_task_data; |
6948 | ompt_data_t *my_parallel_data; |
6949 | int ompt_team_size; |
6950 | |
6951 | if (ompt_enabled.enabled) { |
6952 | exit_runtime_p = &( |
6953 | team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame); |
6954 | } else { |
6955 | exit_runtime_p = &dummy; |
6956 | } |
6957 | |
6958 | my_task_data = |
6959 | &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data); |
6960 | my_parallel_data = &(team->t.ompt_team_info.parallel_data); |
6961 | if (ompt_enabled.ompt_callback_implicit_task) { |
6962 | ompt_team_size = team->t.t_nproc; |
6963 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
6964 | ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size, |
6965 | __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6965)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)); |
6966 | OMPT_CUR_TASK_INFO(this_thr)(&(this_thr->th.th_current_task->ompt_task_info))->thread_num = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 6966)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); |
6967 | } |
6968 | #endif |
6969 | |
6970 | { |
6971 | KMP_TIME_PARTITIONED_BLOCK(OMP_parallel)((void)0); |
6972 | KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK)((void)0); |
6973 | rc = |
6974 | __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn)((void *)(team->t.t_pkfn)), gtid, |
6975 | tid, (int)team->t.t_argc, (void **)team->t.t_argv |
6976 | #if OMPT_SUPPORT1 |
6977 | , |
6978 | exit_runtime_p |
6979 | #endif |
6980 | ); |
6981 | #if OMPT_SUPPORT1 |
6982 | *exit_runtime_p = NULL__null; |
6983 | #endif |
6984 | } |
6985 | |
6986 | #if USE_ITT_BUILD1 |
6987 | if (__itt_stack_caller_create_ptr__kmp_itt_stack_caller_create_ptr__3_0) { |
6988 | __kmp_itt_stack_callee_leave( |
6989 | (__itt_caller) |
6990 | team->t.t_stack_id); // inform ittnotify about leaving user's code |
6991 | } |
6992 | #endif /* USE_ITT_BUILD */ |
6993 | __kmp_run_after_invoked_task(gtid, tid, this_thr, team); |
6994 | |
6995 | return rc; |
6996 | } |
6997 | |
6998 | #if OMP_40_ENABLED(50 >= 40) |
6999 | void __kmp_teams_master(int gtid) { |
7000 | // This routine is called by all master threads in teams construct |
7001 | kmp_info_t *thr = __kmp_threads[gtid]; |
7002 | kmp_team_t *team = thr->th.th_team; |
7003 | ident_t *loc = team->t.t_ident; |
7004 | thr->th.th_set_nproc = thr->th.th_teams_size.nth; |
7005 | KMP_DEBUG_ASSERT(thr->th.th_teams_microtask)((thr->th.th_teams_microtask) ? 0 : __kmp_debug_assert("thr->th.th_teams_microtask" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7005)); |
7006 | KMP_DEBUG_ASSERT(thr->th.th_set_nproc)((thr->th.th_set_nproc) ? 0 : __kmp_debug_assert("thr->th.th_set_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7006)); |
7007 | KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n" , gtid, ((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7008)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), thr ->th.th_teams_microtask); } |
7008 | __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n" , gtid, ((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7008)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid), thr ->th.th_teams_microtask); }; |
7009 | // Launch league of teams now, but not let workers execute |
7010 | // (they hang on fork barrier until next parallel) |
7011 | #if INCLUDE_SSC_MARKS(1 && 1) |
7012 | SSC_MARK_FORKING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd693) : "%ebx"); |
7013 | #endif |
7014 | __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc, |
7015 | (microtask_t)thr->th.th_teams_microtask, // "wrapped" task |
7016 | VOLATILE_CAST(launch_t)(launch_t) __kmp_invoke_task_func, NULL__null); |
7017 | #if INCLUDE_SSC_MARKS(1 && 1) |
7018 | SSC_MARK_JOINING()__asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0xd694) : "%ebx"); |
7019 | #endif |
7020 | |
7021 | // AC: last parameter "1" eliminates join barrier which won't work because |
7022 | // worker threads are in a fork barrier waiting for more parallel regions |
7023 | __kmp_join_call(loc, gtid |
7024 | #if OMPT_SUPPORT1 |
7025 | , |
7026 | fork_context_intel |
7027 | #endif |
7028 | , |
7029 | 1); |
7030 | } |
7031 | |
7032 | int __kmp_invoke_teams_master(int gtid) { |
7033 | kmp_info_t *this_thr = __kmp_threads[gtid]; |
7034 | kmp_team_t *team = this_thr->th.th_team; |
7035 | #if KMP_DEBUG1 |
7036 | if (!__kmp_threads[gtid]->th.th_team->t.t_serialized) |
7037 | KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==(((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == ( void *)__kmp_teams_master) ? 0 : __kmp_debug_assert("(void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7038)) |
7038 | (void *)__kmp_teams_master)(((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == ( void *)__kmp_teams_master) ? 0 : __kmp_debug_assert("(void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == (void *)__kmp_teams_master" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7038)); |
7039 | #endif |
7040 | __kmp_run_before_invoked_task(gtid, 0, this_thr, team); |
7041 | __kmp_teams_master(gtid); |
7042 | __kmp_run_after_invoked_task(gtid, 0, this_thr, team); |
7043 | return 1; |
7044 | } |
7045 | #endif /* OMP_40_ENABLED */ |
7046 | |
7047 | /* this sets the requested number of threads for the next parallel region |
7048 | encountered by this team. since this should be enclosed in the forkjoin |
7049 | critical section it should avoid race conditions with assymmetrical nested |
7050 | parallelism */ |
7051 | |
7052 | void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) { |
7053 | kmp_info_t *thr = __kmp_threads[gtid]; |
7054 | |
7055 | if (num_threads > 0) |
7056 | thr->th.th_set_nproc = num_threads; |
7057 | } |
7058 | |
7059 | #if OMP_40_ENABLED(50 >= 40) |
7060 | |
7061 | /* this sets the requested number of teams for the teams region and/or |
7062 | the number of threads for the next parallel region encountered */ |
7063 | void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams, |
7064 | int num_threads) { |
7065 | kmp_info_t *thr = __kmp_threads[gtid]; |
7066 | KMP_DEBUG_ASSERT(num_teams >= 0)((num_teams >= 0) ? 0 : __kmp_debug_assert("num_teams >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7066)); |
7067 | KMP_DEBUG_ASSERT(num_threads >= 0)((num_threads >= 0) ? 0 : __kmp_debug_assert("num_threads >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7067)); |
7068 | |
7069 | if (num_teams == 0) |
7070 | num_teams = 1; // default number of teams is 1. |
7071 | if (num_teams > __kmp_teams_max_nth) { // if too many teams requested? |
7072 | if (!__kmp_reserve_warn) { |
7073 | __kmp_reserve_warn = 1; |
7074 | __kmp_msg(kmp_ms_warning, |
7075 | KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_teams, __kmp_teams_max_nth ), |
7076 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); |
7077 | } |
7078 | num_teams = __kmp_teams_max_nth; |
7079 | } |
7080 | // Set number of teams (number of threads in the outer "parallel" of the |
7081 | // teams) |
7082 | thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; |
7083 | |
7084 | // Remember the number of threads for inner parallel regions |
7085 | if (num_threads == 0) { |
7086 | if (!TCR_4(__kmp_init_middle)(__kmp_init_middle)) |
7087 | __kmp_middle_initialize(); // get __kmp_avail_proc calculated |
7088 | num_threads = __kmp_avail_proc / num_teams; |
7089 | if (num_teams * num_threads > __kmp_teams_max_nth) { |
7090 | // adjust num_threads w/o warning as it is not user setting |
7091 | num_threads = __kmp_teams_max_nth / num_teams; |
7092 | } |
7093 | } else { |
7094 | if (num_teams * num_threads > __kmp_teams_max_nth) { |
7095 | int new_threads = __kmp_teams_max_nth / num_teams; |
7096 | if (!__kmp_reserve_warn) { // user asked for too many threads |
7097 | __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT |
7098 | __kmp_msg(kmp_ms_warning, |
7099 | KMP_MSG(CantFormThrTeam, num_threads, new_threads)__kmp_msg_format(kmp_i18n_msg_CantFormThrTeam, num_threads, new_threads ), |
7100 | KMP_HNT(Unset_ALL_THREADS)__kmp_msg_format(kmp_i18n_hnt_Unset_ALL_THREADS), __kmp_msg_null); |
7101 | } |
7102 | num_threads = new_threads; |
7103 | } |
7104 | } |
7105 | thr->th.th_teams_size.nth = num_threads; |
7106 | } |
7107 | |
7108 | // Set the proc_bind var to use in the following parallel region. |
7109 | void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) { |
7110 | kmp_info_t *thr = __kmp_threads[gtid]; |
7111 | thr->th.th_set_proc_bind = proc_bind; |
7112 | } |
7113 | |
7114 | #endif /* OMP_40_ENABLED */ |
7115 | |
7116 | /* Launch the worker threads into the microtask. */ |
7117 | |
7118 | void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) { |
7119 | kmp_info_t *this_thr = __kmp_threads[gtid]; |
7120 | |
7121 | #ifdef KMP_DEBUG1 |
7122 | int f; |
7123 | #endif /* KMP_DEBUG */ |
7124 | |
7125 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7125)); |
7126 | KMP_DEBUG_ASSERT(this_thr->th.th_team == team)((this_thr->th.th_team == team) ? 0 : __kmp_debug_assert("this_thr->th.th_team == team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7126)); |
7127 | KMP_ASSERT(KMP_MASTER_GTID(gtid))((((((((gtid)) >= 0) ? 0 : __kmp_debug_assert("((gtid)) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7127)), __kmp_threads[((gtid))]->th.th_info.ds.ds_tid) == 0)) ? 0 : __kmp_debug_assert("KMP_MASTER_GTID(gtid)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7127)); |
7128 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
7129 | |
7130 | team->t.t_construct = 0; /* no single directives seen yet */ |
7131 | team->t.t_ordered.dt.t_value = |
7132 | 0; /* thread 0 enters the ordered section first */ |
7133 | |
7134 | /* Reset the identifiers on the dispatch buffer */ |
7135 | KMP_DEBUG_ASSERT(team->t.t_disp_buffer)((team->t.t_disp_buffer) ? 0 : __kmp_debug_assert("team->t.t_disp_buffer" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7135)); |
7136 | if (team->t.t_max_nproc > 1) { |
7137 | int i; |
7138 | for (i = 0; i < __kmp_dispatch_num_buffers; ++i) { |
7139 | team->t.t_disp_buffer[i].buffer_index = i; |
7140 | #if OMP_45_ENABLED(50 >= 45) |
7141 | team->t.t_disp_buffer[i].doacross_buf_idx = i; |
7142 | #endif |
7143 | } |
7144 | } else { |
7145 | team->t.t_disp_buffer[0].buffer_index = 0; |
7146 | #if OMP_45_ENABLED(50 >= 45) |
7147 | team->t.t_disp_buffer[0].doacross_buf_idx = 0; |
7148 | #endif |
7149 | } |
7150 | |
7151 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
7152 | KMP_ASSERT(this_thr->th.th_team == team)((this_thr->th.th_team == team) ? 0 : __kmp_debug_assert("this_thr->th.th_team == team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7152)); |
7153 | |
7154 | #ifdef KMP_DEBUG1 |
7155 | for (f = 0; f < team->t.t_nproc; f++) { |
7156 | KMP_DEBUG_ASSERT(team->t.t_threads[f] &&((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7157)) |
7157 | team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc)((team->t.t_threads[f] && team->t.t_threads[f]-> th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert ("team->t.t_threads[f] && team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7157)); |
7158 | } |
7159 | #endif /* KMP_DEBUG */ |
7160 | |
7161 | /* release the worker threads so they may begin working */ |
7162 | __kmp_fork_barrier(gtid, 0); |
7163 | } |
7164 | |
7165 | void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) { |
7166 | kmp_info_t *this_thr = __kmp_threads[gtid]; |
7167 | |
7168 | KMP_DEBUG_ASSERT(team)((team) ? 0 : __kmp_debug_assert("team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7168)); |
7169 | KMP_DEBUG_ASSERT(this_thr->th.th_team == team)((this_thr->th.th_team == team) ? 0 : __kmp_debug_assert("this_thr->th.th_team == team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7169)); |
7170 | KMP_ASSERT(KMP_MASTER_GTID(gtid))((((((((gtid)) >= 0) ? 0 : __kmp_debug_assert("((gtid)) >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7170)), __kmp_threads[((gtid))]->th.th_info.ds.ds_tid) == 0)) ? 0 : __kmp_debug_assert("KMP_MASTER_GTID(gtid)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7170)); |
7171 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
7172 | |
7173 | /* Join barrier after fork */ |
7174 | |
7175 | #ifdef KMP_DEBUG1 |
7176 | if (__kmp_threads[gtid] && |
7177 | __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) { |
7178 | __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid, |
7179 | __kmp_threads[gtid]); |
7180 | __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, " |
7181 | "team->t.t_nproc=%d\n", |
7182 | gtid, __kmp_threads[gtid]->th.th_team_nproc, team, |
7183 | team->t.t_nproc); |
7184 | __kmp_print_structure(); |
7185 | } |
7186 | KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&((__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert("__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7187)) |
7187 | __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc)((__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc) ? 0 : __kmp_debug_assert("__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7187)); |
7188 | #endif /* KMP_DEBUG */ |
7189 | |
7190 | __kmp_join_barrier(gtid); /* wait for everyone */ |
7191 | #if OMPT_SUPPORT1 |
7192 | if (ompt_enabled.enabled && |
7193 | this_thr->th.ompt_thread_info.state == omp_state_wait_barrier_implicit) { |
7194 | int ds_tid = this_thr->th.th_info.ds.ds_tid; |
7195 | ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr)(&(this_thr->th.th_current_task->ompt_task_info.task_data )); |
7196 | this_thr->th.ompt_thread_info.state = omp_state_overhead; |
7197 | #if OMPT_OPTIONAL1 |
7198 | void *codeptr = NULL__null; |
7199 | if (KMP_MASTER_TID(ds_tid)((ds_tid) == 0) && |
7200 | (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback || |
7201 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback)) |
7202 | codeptr = OMPT_CUR_TEAM_INFO(this_thr)(&(this_thr->th.th_team->t.ompt_team_info))->master_return_address; |
7203 | |
7204 | if (ompt_enabled.ompt_callback_sync_region_wait) { |
7205 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( |
7206 | ompt_sync_region_barrier, ompt_scope_end, NULL__null, task_data, codeptr); |
7207 | } |
7208 | if (ompt_enabled.ompt_callback_sync_region) { |
7209 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( |
7210 | ompt_sync_region_barrier, ompt_scope_end, NULL__null, task_data, codeptr); |
7211 | } |
7212 | #endif |
7213 | if (!KMP_MASTER_TID(ds_tid)((ds_tid) == 0) && ompt_enabled.ompt_callback_implicit_task) { |
7214 | ompt_callbacks.ompt_callback(ompt_callback_implicit_task)ompt_callback_implicit_task_callback( |
7215 | ompt_scope_end, NULL__null, task_data, 0, ds_tid); |
7216 | } |
7217 | } |
7218 | #endif |
7219 | |
7220 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
7221 | KMP_ASSERT(this_thr->th.th_team == team)((this_thr->th.th_team == team) ? 0 : __kmp_debug_assert("this_thr->th.th_team == team" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7221)); |
7222 | } |
7223 | |
7224 | /* ------------------------------------------------------------------------ */ |
7225 | |
7226 | #ifdef USE_LOAD_BALANCE1 |
7227 | |
7228 | // Return the worker threads actively spinning in the hot team, if we |
7229 | // are at the outermost level of parallelism. Otherwise, return 0. |
7230 | static int __kmp_active_hot_team_nproc(kmp_root_t *root) { |
7231 | int i; |
7232 | int retval; |
7233 | kmp_team_t *hot_team; |
7234 | |
7235 | if (root->r.r_active) { |
7236 | return 0; |
7237 | } |
7238 | hot_team = root->r.r_hot_team; |
7239 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME(2147483647)) { |
7240 | return hot_team->t.t_nproc - 1; // Don't count master thread |
7241 | } |
7242 | |
7243 | // Skip the master thread - it is accounted for elsewhere. |
7244 | retval = 0; |
7245 | for (i = 1; i < hot_team->t.t_nproc; i++) { |
7246 | if (hot_team->t.t_threads[i]->th.th_active) { |
7247 | retval++; |
7248 | } |
7249 | } |
7250 | return retval; |
7251 | } |
7252 | |
7253 | // Perform an automatic adjustment to the number of |
7254 | // threads used by the next parallel region. |
7255 | static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) { |
7256 | int retval; |
7257 | int pool_active; |
7258 | int hot_team_active; |
7259 | int team_curr_active; |
7260 | int system_active; |
7261 | |
7262 | KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n" , root, set_nproc); } |
7263 | set_nproc))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n" , root, set_nproc); }; |
7264 | KMP_DEBUG_ASSERT(root)((root) ? 0 : __kmp_debug_assert("root", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7264)); |
7265 | KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]((root->r.r_root_team->t.t_threads[0] ->th.th_current_task ->td_icvs.dynamic == (!0)) ? 0 : __kmp_debug_assert("root->r.r_root_team->t.t_threads[0] ->th.th_current_task->td_icvs.dynamic == (!0)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7266)) |
7266 | ->th.th_current_task->td_icvs.dynamic == TRUE)((root->r.r_root_team->t.t_threads[0] ->th.th_current_task ->td_icvs.dynamic == (!0)) ? 0 : __kmp_debug_assert("root->r.r_root_team->t.t_threads[0] ->th.th_current_task->td_icvs.dynamic == (!0)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7266)); |
7267 | KMP_DEBUG_ASSERT(set_nproc > 1)((set_nproc > 1) ? 0 : __kmp_debug_assert("set_nproc > 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7267)); |
7268 | |
7269 | if (set_nproc == 1) { |
7270 | KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: serial execution.\n" ); }; |
7271 | return 1; |
7272 | } |
7273 | |
7274 | // Threads that are active in the thread pool, active in the hot team for this |
7275 | // particular root (if we are at the outer par level), and the currently |
7276 | // executing thread (to become the master) are available to add to the new |
7277 | // team, but are currently contributing to the system load, and must be |
7278 | // accounted for. |
7279 | pool_active = __kmp_thread_pool_active_nth; |
7280 | hot_team_active = __kmp_active_hot_team_nproc(root); |
7281 | team_curr_active = pool_active + hot_team_active + 1; |
7282 | |
7283 | // Check the system load. |
7284 | system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active); |
7285 | KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); } |
7286 | "hot team active = %d\n",if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); } |
7287 | system_active, pool_active, hot_team_active))if (kmp_b_debug >= 30) { __kmp_debug_printf ("__kmp_load_balance_nproc: system active = %d pool active = %d " "hot team active = %d\n", system_active, pool_active, hot_team_active ); }; |
7288 | |
7289 | if (system_active < 0) { |
7290 | // There was an error reading the necessary info from /proc, so use the |
7291 | // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode |
7292 | // = dynamic_thread_limit, we shouldn't wind up getting back here. |
7293 | __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; |
7294 | KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_CantLoadBalUsing , "KMP_DYNAMIC_MODE=thread limit"), __kmp_msg_null); |
7295 | |
7296 | // Make this call behave like the thread limit algorithm. |
7297 | retval = __kmp_avail_proc - __kmp_nth + |
7298 | (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); |
7299 | if (retval > set_nproc) { |
7300 | retval = set_nproc; |
7301 | } |
7302 | if (retval < KMP_MIN_NTH1) { |
7303 | retval = KMP_MIN_NTH1; |
7304 | } |
7305 | |
7306 | KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n" , retval); } |
7307 | retval))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n" , retval); }; |
7308 | return retval; |
7309 | } |
7310 | |
7311 | // There is a slight delay in the load balance algorithm in detecting new |
7312 | // running procs. The real system load at this instant should be at least as |
7313 | // large as the #active omp thread that are available to add to the team. |
7314 | if (system_active < team_curr_active) { |
7315 | system_active = team_curr_active; |
7316 | } |
7317 | retval = __kmp_avail_proc - system_active + team_curr_active; |
7318 | if (retval > set_nproc) { |
7319 | retval = set_nproc; |
7320 | } |
7321 | if (retval < KMP_MIN_NTH1) { |
7322 | retval = KMP_MIN_NTH1; |
7323 | } |
7324 | |
7325 | KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval))if (kmp_b_debug >= 20) { __kmp_debug_printf ("__kmp_load_balance_nproc: exit. retval:%d\n" , retval); }; |
7326 | return retval; |
7327 | } // __kmp_load_balance_nproc() |
7328 | |
7329 | #endif /* USE_LOAD_BALANCE */ |
7330 | |
7331 | /* ------------------------------------------------------------------------ */ |
7332 | |
7333 | /* NOTE: this is called with the __kmp_init_lock held */ |
7334 | void __kmp_cleanup(void) { |
7335 | int f; |
7336 | |
7337 | KA_TRACE(10, ("__kmp_cleanup: enter\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: enter\n" ); }; |
7338 | |
7339 | if (TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { |
7340 | #if KMP_HANDLE_SIGNALS(1 || 0) |
7341 | __kmp_remove_signals(); |
7342 | #endif |
7343 | TCW_4(__kmp_init_parallel, FALSE)(__kmp_init_parallel) = (0); |
7344 | } |
7345 | |
7346 | if (TCR_4(__kmp_init_middle)(__kmp_init_middle)) { |
7347 | #if KMP_AFFINITY_SUPPORTED1 |
7348 | __kmp_affinity_uninitialize(); |
7349 | #endif /* KMP_AFFINITY_SUPPORTED */ |
7350 | __kmp_cleanup_hierarchy(); |
7351 | TCW_4(__kmp_init_middle, FALSE)(__kmp_init_middle) = (0); |
7352 | } |
7353 | |
7354 | KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: go serial cleanup\n" ); }; |
7355 | |
7356 | if (__kmp_init_serial) { |
7357 | __kmp_runtime_destroy(); |
7358 | __kmp_init_serial = FALSE0; |
7359 | } |
7360 | |
7361 | __kmp_cleanup_threadprivate_caches(); |
7362 | |
7363 | for (f = 0; f < __kmp_threads_capacity; f++) { |
7364 | if (__kmp_root[f] != NULL__null) { |
7365 | __kmp_free(__kmp_root[f])___kmp_free((__kmp_root[f]), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7365); |
7366 | __kmp_root[f] = NULL__null; |
7367 | } |
7368 | } |
7369 | __kmp_free(__kmp_threads)___kmp_free((__kmp_threads), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7369); |
7370 | // __kmp_threads and __kmp_root were allocated at once, as single block, so |
7371 | // there is no need in freeing __kmp_root. |
7372 | __kmp_threads = NULL__null; |
7373 | __kmp_root = NULL__null; |
7374 | __kmp_threads_capacity = 0; |
7375 | |
7376 | #if KMP_USE_DYNAMIC_LOCK1 |
7377 | __kmp_cleanup_indirect_user_locks(); |
7378 | #else |
7379 | __kmp_cleanup_user_locks(); |
7380 | #endif |
7381 | |
7382 | #if KMP_AFFINITY_SUPPORTED1 |
7383 | KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file))free(const_cast<char *>(__kmp_cpuinfo_file)); |
7384 | __kmp_cpuinfo_file = NULL__null; |
7385 | #endif /* KMP_AFFINITY_SUPPORTED */ |
7386 | |
7387 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
7388 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 |
7389 | __kmp_print_speculative_stats(); |
7390 | #endif |
7391 | #endif |
7392 | KMP_INTERNAL_FREE(__kmp_nested_nth.nth)free(__kmp_nested_nth.nth); |
7393 | __kmp_nested_nth.nth = NULL__null; |
7394 | __kmp_nested_nth.size = 0; |
7395 | __kmp_nested_nth.used = 0; |
7396 | KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types)free(__kmp_nested_proc_bind.bind_types); |
7397 | __kmp_nested_proc_bind.bind_types = NULL__null; |
7398 | __kmp_nested_proc_bind.size = 0; |
7399 | __kmp_nested_proc_bind.used = 0; |
7400 | |
7401 | __kmp_i18n_catclose(); |
7402 | |
7403 | #if KMP_USE_HIER_SCHED0 |
7404 | __kmp_hier_scheds.deallocate(); |
7405 | #endif |
7406 | |
7407 | #if KMP_STATS_ENABLED0 |
7408 | __kmp_stats_fini(); |
7409 | #endif |
7410 | |
7411 | KA_TRACE(10, ("__kmp_cleanup: exit\n"))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_cleanup: exit\n" ); }; |
7412 | } |
7413 | |
7414 | /* ------------------------------------------------------------------------ */ |
7415 | |
7416 | int __kmp_ignore_mppbeg(void) { |
7417 | char *env; |
7418 | |
7419 | if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL__null) { |
7420 | if (__kmp_str_match_false(env)) |
7421 | return FALSE0; |
7422 | } |
7423 | // By default __kmpc_begin() is no-op. |
7424 | return TRUE(!0); |
7425 | } |
7426 | |
7427 | int __kmp_ignore_mppend(void) { |
7428 | char *env; |
7429 | |
7430 | if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL__null) { |
7431 | if (__kmp_str_match_false(env)) |
7432 | return FALSE0; |
7433 | } |
7434 | // By default __kmpc_end() is no-op. |
7435 | return TRUE(!0); |
7436 | } |
7437 | |
7438 | void __kmp_internal_begin(void) { |
7439 | int gtid; |
7440 | kmp_root_t *root; |
7441 | |
7442 | /* this is a very important step as it will register new sibling threads |
7443 | and assign these new uber threads a new gtid */ |
7444 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); |
7445 | root = __kmp_threads[gtid]->th.th_root; |
7446 | KMP_ASSERT(KMP_UBER_GTID(gtid))((((((gtid) >= (-6)) ? 0 : __kmp_debug_assert("(gtid) >= (-6)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7446)), (((gtid) < __kmp_threads_capacity) ? 0 : __kmp_debug_assert ("(gtid) < __kmp_threads_capacity", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7446)), (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && (__kmp_threads[(gtid)] == __kmp_root [(gtid)]->r.r_uber_thread))) ? 0 : __kmp_debug_assert("KMP_UBER_GTID(gtid)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7446)); |
7447 | |
7448 | if (root->r.r_begin) |
7449 | return; |
7450 | __kmp_acquire_lock(&root->r.r_begin_lock, gtid); |
7451 | if (root->r.r_begin) { |
7452 | __kmp_release_lock(&root->r.r_begin_lock, gtid); |
7453 | return; |
7454 | } |
7455 | |
7456 | root->r.r_begin = TRUE(!0); |
7457 | |
7458 | __kmp_release_lock(&root->r.r_begin_lock, gtid); |
7459 | } |
7460 | |
7461 | /* ------------------------------------------------------------------------ */ |
7462 | |
7463 | void __kmp_user_set_library(enum library_type arg) { |
7464 | int gtid; |
7465 | kmp_root_t *root; |
7466 | kmp_info_t *thread; |
7467 | |
7468 | /* first, make sure we are initialized so we can get our gtid */ |
7469 | |
7470 | gtid = __kmp_entry_gtid()__kmp_get_global_thread_id_reg(); |
7471 | thread = __kmp_threads[gtid]; |
7472 | |
7473 | root = thread->th.th_root; |
7474 | |
7475 | KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n" , gtid, arg, library_serial); } |
7476 | library_serial))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n" , gtid, arg, library_serial); }; |
7477 | if (root->r.r_in_parallel) { /* Must be called in serial section of top-level |
7478 | thread */ |
7479 | KMP_WARNING(SetLibraryIncorrectCall)__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_SetLibraryIncorrectCall ), __kmp_msg_null); |
7480 | return; |
7481 | } |
7482 | |
7483 | switch (arg) { |
7484 | case library_serial: |
7485 | thread->th.th_set_nproc = 0; |
7486 | set__nproc(thread, 1)(((thread)->th.th_current_task->td_icvs.nproc) = (1)); |
7487 | break; |
7488 | case library_turnaround: |
7489 | thread->th.th_set_nproc = 0; |
7490 | set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)) |
7491 | : __kmp_dflt_team_nth_ub)(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)); |
7492 | break; |
7493 | case library_throughput: |
7494 | thread->th.th_set_nproc = 0; |
7495 | set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)) |
7496 | : __kmp_dflt_team_nth_ub)(((thread)->th.th_current_task->td_icvs.nproc) = (__kmp_dflt_team_nth ? __kmp_dflt_team_nth : __kmp_dflt_team_nth_ub)); |
7497 | break; |
7498 | default: |
7499 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); |
7500 | } |
7501 | |
7502 | __kmp_aux_set_library(arg); |
7503 | } |
7504 | |
7505 | void __kmp_aux_set_stacksize(size_t arg) { |
7506 | if (!__kmp_init_serial) |
7507 | __kmp_serial_initialize(); |
7508 | |
7509 | #if KMP_OS_DARWIN0 |
7510 | if (arg & (0x1000 - 1)) { |
7511 | arg &= ~(0x1000 - 1); |
7512 | if (arg + 0x1000) /* check for overflow if we round up */ |
7513 | arg += 0x1000; |
7514 | } |
7515 | #endif |
7516 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); |
7517 | |
7518 | /* only change the default stacksize before the first parallel region */ |
7519 | if (!TCR_4(__kmp_init_parallel)(__kmp_init_parallel)) { |
7520 | size_t value = arg; /* argument is in bytes */ |
7521 | |
7522 | if (value < __kmp_sys_min_stksize) |
7523 | value = __kmp_sys_min_stksize; |
7524 | else if (value > KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) ))) |
7525 | value = KMP_MAX_STKSIZE(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1) )); |
7526 | |
7527 | __kmp_stksize = value; |
7528 | |
7529 | __kmp_env_stksize = TRUE(!0); /* was KMP_STACKSIZE specified? */ |
7530 | } |
7531 | |
7532 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); |
7533 | } |
7534 | |
7535 | /* set the behaviour of the runtime library */ |
7536 | /* TODO this can cause some odd behaviour with sibling parallelism... */ |
7537 | void __kmp_aux_set_library(enum library_type arg) { |
7538 | __kmp_library = arg; |
7539 | |
7540 | switch (__kmp_library) { |
7541 | case library_serial: { |
7542 | KMP_INFORM(LibraryIsSerial)__kmp_msg(kmp_ms_inform, __kmp_msg_format(kmp_i18n_msg_LibraryIsSerial ), __kmp_msg_null); |
7543 | (void)__kmp_change_library(TRUE(!0)); |
7544 | } break; |
7545 | case library_turnaround: |
7546 | (void)__kmp_change_library(TRUE(!0)); |
7547 | break; |
7548 | case library_throughput: |
7549 | (void)__kmp_change_library(FALSE0); |
7550 | break; |
7551 | default: |
7552 | KMP_FATAL(UnknownLibraryType, arg)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_UnknownLibraryType, arg), __kmp_msg_null); |
7553 | } |
7554 | } |
7555 | |
7556 | /* ------------------------------------------------------------------------ */ |
7557 | |
7558 | void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) { |
7559 | int blocktime = arg; /* argument is in milliseconds */ |
7560 | #if KMP_USE_MONITOR |
7561 | int bt_intervals; |
7562 | #endif |
7563 | int bt_set; |
7564 | |
7565 | __kmp_save_internal_controls(thread); |
7566 | |
7567 | /* Normalize and set blocktime for the teams */ |
7568 | if (blocktime < KMP_MIN_BLOCKTIME(0)) |
7569 | blocktime = KMP_MIN_BLOCKTIME(0); |
7570 | else if (blocktime > KMP_MAX_BLOCKTIME(2147483647)) |
7571 | blocktime = KMP_MAX_BLOCKTIME(2147483647); |
7572 | |
7573 | set__blocktime_team(thread->th.th_team, tid, blocktime)(((thread->th.th_team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.blocktime) = (blocktime)); |
7574 | set__blocktime_team(thread->th.th_serial_team, 0, blocktime)(((thread->th.th_serial_team)->t.t_threads[(0)]->th. th_current_task->td_icvs.blocktime) = (blocktime)); |
7575 | |
7576 | #if KMP_USE_MONITOR |
7577 | /* Calculate and set blocktime intervals for the teams */ |
7578 | bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups); |
7579 | |
7580 | set__bt_intervals_team(thread->th.th_team, tid, bt_intervals); |
7581 | set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals); |
7582 | #endif |
7583 | |
7584 | /* Set whether blocktime has been set to "TRUE" */ |
7585 | bt_set = TRUE(!0); |
7586 | |
7587 | set__bt_set_team(thread->th.th_team, tid, bt_set)(((thread->th.th_team)->t.t_threads[(tid)]->th.th_current_task ->td_icvs.bt_set) = (bt_set)); |
7588 | set__bt_set_team(thread->th.th_serial_team, 0, bt_set)(((thread->th.th_serial_team)->t.t_threads[(0)]->th. th_current_task->td_icvs.bt_set) = (bt_set)); |
7589 | #if KMP_USE_MONITOR |
7590 | KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", ((((tid) >= 0 && (thread->th.th_team) != __null) ? 0 : __kmp_debug_assert( "(tid) >= 0 && (thread->th.th_team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7592)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime , bt_intervals, __kmp_monitor_wakeups); } |
7591 | "bt_intervals=%d, monitor_updates=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", ((((tid) >= 0 && (thread->th.th_team) != __null) ? 0 : __kmp_debug_assert( "(tid) >= 0 && (thread->th.th_team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7592)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime , bt_intervals, __kmp_monitor_wakeups); } |
7592 | __kmp_gtid_from_tid(tid, thread->th.th_team),if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", ((((tid) >= 0 && (thread->th.th_team) != __null) ? 0 : __kmp_debug_assert( "(tid) >= 0 && (thread->th.th_team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7592)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime , bt_intervals, __kmp_monitor_wakeups); } |
7593 | thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", ((((tid) >= 0 && (thread->th.th_team) != __null) ? 0 : __kmp_debug_assert( "(tid) >= 0 && (thread->th.th_team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7592)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime , bt_intervals, __kmp_monitor_wakeups); } |
7594 | __kmp_monitor_wakeups))if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " "bt_intervals=%d, monitor_updates=%d\n", ((((tid) >= 0 && (thread->th.th_team) != __null) ? 0 : __kmp_debug_assert( "(tid) >= 0 && (thread->th.th_team) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7592)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime , bt_intervals, __kmp_monitor_wakeups); }; |
7595 | #else |
7596 | KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , ((((tid) >= 0 && (thread->th.th_team) != __null ) ? 0 : __kmp_debug_assert("(tid) >= 0 && (thread->th.th_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7597)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime ); } |
7597 | __kmp_gtid_from_tid(tid, thread->th.th_team),if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , ((((tid) >= 0 && (thread->th.th_team) != __null ) ? 0 : __kmp_debug_assert("(tid) >= 0 && (thread->th.th_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7597)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime ); } |
7598 | thread->th.th_team->t.t_id, tid, blocktime))if (kmp_f_debug >= 10) { __kmp_debug_printf ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n" , ((((tid) >= 0 && (thread->th.th_team) != __null ) ? 0 : __kmp_debug_assert("(tid) >= 0 && (thread->th.th_team) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7597)), thread->th.th_team->t.t_threads[(tid)]->th .th_info.ds.ds_gtid), thread->th.th_team->t.t_id, tid, blocktime ); }; |
7599 | #endif |
7600 | } |
7601 | |
7602 | void __kmp_aux_set_defaults(char const *str, int len) { |
7603 | if (!__kmp_init_serial) { |
7604 | __kmp_serial_initialize(); |
7605 | } |
7606 | __kmp_env_initialize(str); |
7607 | |
7608 | if (__kmp_settings |
7609 | #if OMP_40_ENABLED(50 >= 40) |
7610 | || __kmp_display_env || __kmp_display_env_verbose |
7611 | #endif // OMP_40_ENABLED |
7612 | ) { |
7613 | __kmp_env_print(); |
7614 | } |
7615 | } // __kmp_aux_set_defaults |
7616 | |
7617 | /* ------------------------------------------------------------------------ */ |
7618 | /* internal fast reduction routines */ |
7619 | |
7620 | PACKED_REDUCTION_METHOD_T |
7621 | __kmp_determine_reduction_method( |
7622 | ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, |
7623 | void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), |
7624 | kmp_critical_name *lck) { |
7625 | |
7626 | // Default reduction method: critical construct ( lck != NULL, like in current |
7627 | // PAROPT ) |
7628 | // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method |
7629 | // can be selected by RTL |
7630 | // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method |
7631 | // can be selected by RTL |
7632 | // Finally, it's up to OpenMP RTL to make a decision on which method to select |
7633 | // among generated by PAROPT. |
7634 | |
7635 | PACKED_REDUCTION_METHOD_T retval; |
7636 | |
7637 | int team_size; |
7638 | |
7639 | KMP_DEBUG_ASSERT(loc)((loc) ? 0 : __kmp_debug_assert("loc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7639)); // it would be nice to test ( loc != 0 ) |
7640 | KMP_DEBUG_ASSERT(lck)((lck) ? 0 : __kmp_debug_assert("lck", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7640)); // it would be nice to test ( lck != 0 ) |
7641 | |
7642 | #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \ |
7643 | ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE0x10)) == (KMP_IDENT_ATOMIC_REDUCE0x10)) |
7644 | #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func)) |
7645 | |
7646 | retval = critical_reduce_block; |
7647 | |
7648 | // another choice of getting a team size (with 1 dynamic deference) is slower |
7649 | team_size = __kmp_get_team_num_threads(global_tid)(__kmp_threads[(global_tid)]->th.th_team->t.t_nproc); |
7650 | if (team_size == 1) { |
7651 | |
7652 | retval = empty_reduce_block; |
7653 | |
7654 | } else { |
7655 | |
7656 | int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; |
7657 | int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; |
7658 | |
7659 | #if KMP_ARCH_X86_641 || KMP_ARCH_PPC64(0 || 0) || KMP_ARCH_AARCH640 || KMP_ARCH_MIPS640 |
7660 | |
7661 | #if KMP_OS_LINUX1 || KMP_OS_FREEBSD0 || KMP_OS_NETBSD0 || KMP_OS_WINDOWS0 || \ |
7662 | KMP_OS_DARWIN0 |
7663 | |
7664 | int teamsize_cutoff = 4; |
7665 | |
7666 | #if KMP_MIC_SUPPORTED((0 || 1) && (1 || 0)) |
7667 | if (__kmp_mic_type != non_mic) { |
7668 | teamsize_cutoff = 8; |
7669 | } |
7670 | #endif |
7671 | if (tree_available) { |
7672 | if (team_size <= teamsize_cutoff) { |
7673 | if (atomic_available) { |
7674 | retval = atomic_reduce_block; |
7675 | } |
7676 | } else { |
7677 | retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); |
7678 | } |
7679 | } else if (atomic_available) { |
7680 | retval = atomic_reduce_block; |
7681 | } |
7682 | #else |
7683 | #error "Unknown or unsupported OS" |
7684 | #endif // KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS || |
7685 | // KMP_OS_DARWIN |
7686 | |
7687 | #elif KMP_ARCH_X860 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS0 |
7688 | |
7689 | #if KMP_OS_LINUX1 || KMP_OS_WINDOWS0 |
7690 | |
7691 | // basic tuning |
7692 | |
7693 | if (atomic_available) { |
7694 | if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ??? |
7695 | retval = atomic_reduce_block; |
7696 | } |
7697 | } // otherwise: use critical section |
7698 | |
7699 | #elif KMP_OS_DARWIN0 |
7700 | |
7701 | if (atomic_available && (num_vars <= 3)) { |
7702 | retval = atomic_reduce_block; |
7703 | } else if (tree_available) { |
7704 | if ((reduce_size > (9 * sizeof(kmp_real64))) && |
7705 | (reduce_size < (2000 * sizeof(kmp_real64)))) { |
7706 | retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER(((tree_reduce_block) | (bs_plain_barrier))); |
7707 | } |
7708 | } // otherwise: use critical section |
7709 | |
7710 | #else |
7711 | #error "Unknown or unsupported OS" |
7712 | #endif |
7713 | |
7714 | #else |
7715 | #error "Unknown or unsupported architecture" |
7716 | #endif |
7717 | } |
7718 | |
7719 | // KMP_FORCE_REDUCTION |
7720 | |
7721 | // If the team is serialized (team_size == 1), ignore the forced reduction |
7722 | // method and stay with the unsynchronized method (empty_reduce_block) |
7723 | if (__kmp_force_reduction_method != reduction_method_not_defined && |
7724 | team_size != 1) { |
7725 | |
7726 | PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block; |
7727 | |
7728 | int atomic_available, tree_available; |
7729 | |
7730 | switch ((forced_retval = __kmp_force_reduction_method)) { |
7731 | case critical_reduce_block: |
7732 | KMP_ASSERT(lck)((lck) ? 0 : __kmp_debug_assert("lck", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7732)); // lck should be != 0 |
7733 | break; |
7734 | |
7735 | case atomic_reduce_block: |
7736 | atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; |
7737 | if (!atomic_available) { |
7738 | KMP_WARNING(RedMethodNotSupported, "atomic")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "atomic"), __kmp_msg_null); |
7739 | forced_retval = critical_reduce_block; |
7740 | } |
7741 | break; |
7742 | |
7743 | case tree_reduce_block: |
7744 | tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; |
7745 | if (!tree_available) { |
7746 | KMP_WARNING(RedMethodNotSupported, "tree")__kmp_msg(kmp_ms_warning, __kmp_msg_format(kmp_i18n_msg_RedMethodNotSupported , "tree"), __kmp_msg_null); |
7747 | forced_retval = critical_reduce_block; |
7748 | } else { |
7749 | #if KMP_FAST_REDUCTION_BARRIER1 |
7750 | forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER(((tree_reduce_block) | (bs_reduction_barrier))); |
7751 | #endif |
7752 | } |
7753 | break; |
7754 | |
7755 | default: |
7756 | KMP_ASSERT(0)((0) ? 0 : __kmp_debug_assert("0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_runtime.cpp" , 7756)); // "unsupported method specified" |
7757 | } |
7758 | |
7759 | retval = forced_retval; |
7760 | } |
7761 | |
7762 | KA_TRACE(10, ("reduction method selected=%08x\n", retval))if (kmp_a_debug >= 10) { __kmp_debug_printf ("reduction method selected=%08x\n" , retval); }; |
7763 | |
7764 | #undef FAST_REDUCTION_TREE_METHOD_GENERATED |
7765 | #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED |
7766 | |
7767 | return (retval); |
7768 | } |
7769 | |
7770 | // this function is for testing set/get/determine reduce method |
7771 | kmp_int32 __kmp_get_reduce_method(void) { |
7772 | return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8); |
7773 | } |