File: | projects/openmp/runtime/src/kmp_tasking.cpp |
Warning: | line 824, column 36 Dereference of null pointer |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* | |||
2 | * kmp_tasking.cpp -- OpenMP 3.0 tasking support. | |||
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_i18n.h" | |||
16 | #include "kmp_itt.h" | |||
17 | #include "kmp_stats.h" | |||
18 | #include "kmp_wait_release.h" | |||
19 | ||||
20 | #if OMPT_SUPPORT1 | |||
21 | #include "ompt-specific.h" | |||
22 | #endif | |||
23 | ||||
24 | #include "tsan_annotations.h" | |||
25 | ||||
26 | /* forward declaration */ | |||
27 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
28 | kmp_info_t *this_thr); | |||
29 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
30 | kmp_thread_data_t *thread_data); | |||
31 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
32 | kmp_task_team_t *task_team); | |||
33 | ||||
34 | #ifdef OMP_45_ENABLED(50 >= 45) | |||
35 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask); | |||
36 | #endif | |||
37 | ||||
38 | #ifdef BUILD_TIED_TASK_STACK | |||
39 | ||||
40 | // __kmp_trace_task_stack: print the tied tasks from the task stack in order | |||
41 | // from top do bottom | |||
42 | // | |||
43 | // gtid: global thread identifier for thread containing stack | |||
44 | // thread_data: thread data for task team thread containing stack | |||
45 | // threshold: value above which the trace statement triggers | |||
46 | // location: string identifying call site of this function (for trace) | |||
47 | static void __kmp_trace_task_stack(kmp_int32 gtid, | |||
48 | kmp_thread_data_t *thread_data, | |||
49 | int threshold, char *location) { | |||
50 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
51 | kmp_taskdata_t **stack_top = task_stack->ts_top; | |||
52 | kmp_int32 entries = task_stack->ts_entries; | |||
53 | kmp_taskdata_t *tied_task; | |||
54 | ||||
55 | KA_TRACE(if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
56 | threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
57 | ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, "if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
58 | "first_block = %p, stack_top = %p \n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
59 | location, gtid, entries, task_stack->ts_first_block, stack_top))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); }; | |||
60 | ||||
61 | KMP_DEBUG_ASSERT(stack_top != NULL)((stack_top != __null) ? 0 : __kmp_debug_assert("stack_top != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 61)); | |||
62 | KMP_DEBUG_ASSERT(entries > 0)((entries > 0) ? 0 : __kmp_debug_assert("entries > 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 62)); | |||
63 | ||||
64 | while (entries != 0) { | |||
65 | KMP_DEBUG_ASSERT(stack_top != &task_stack->ts_first_block.sb_block[0])((stack_top != &task_stack->ts_first_block.sb_block[0] ) ? 0 : __kmp_debug_assert("stack_top != &task_stack->ts_first_block.sb_block[0]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 65)); | |||
66 | // fix up ts_top if we need to pop from previous block | |||
67 | if (entries & TASK_STACK_INDEX_MASK == 0) { | |||
68 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(stack_top); | |||
69 | ||||
70 | stack_block = stack_block->sb_prev; | |||
71 | stack_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
72 | } | |||
73 | ||||
74 | // finish bookkeeping | |||
75 | stack_top--; | |||
76 | entries--; | |||
77 | ||||
78 | tied_task = *stack_top; | |||
79 | ||||
80 | KMP_DEBUG_ASSERT(tied_task != NULL)((tied_task != __null) ? 0 : __kmp_debug_assert("tied_task != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 80)); | |||
81 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 81)); | |||
82 | ||||
83 | KA_TRACE(threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
84 | ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, "if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
85 | "stack_top=%p, tied_task=%p\n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
86 | location, gtid, entries, stack_top, tied_task))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); }; | |||
87 | } | |||
88 | KMP_DEBUG_ASSERT(stack_top == &task_stack->ts_first_block.sb_block[0])((stack_top == &task_stack->ts_first_block.sb_block[0] ) ? 0 : __kmp_debug_assert("stack_top == &task_stack->ts_first_block.sb_block[0]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 88)); | |||
89 | ||||
90 | KA_TRACE(threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); } | |||
91 | ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); } | |||
92 | location, gtid))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); }; | |||
93 | } | |||
94 | ||||
95 | // __kmp_init_task_stack: initialize the task stack for the first time | |||
96 | // after a thread_data structure is created. | |||
97 | // It should not be necessary to do this again (assuming the stack works). | |||
98 | // | |||
99 | // gtid: global thread identifier of calling thread | |||
100 | // thread_data: thread data for task team thread containing stack | |||
101 | static void __kmp_init_task_stack(kmp_int32 gtid, | |||
102 | kmp_thread_data_t *thread_data) { | |||
103 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
104 | kmp_stack_block_t *first_block; | |||
105 | ||||
106 | // set up the first block of the stack | |||
107 | first_block = &task_stack->ts_first_block; | |||
108 | task_stack->ts_top = (kmp_taskdata_t **)first_block; | |||
109 | memset((void *)first_block, '\0', | |||
110 | TASK_STACK_BLOCK_SIZE * sizeof(kmp_taskdata_t *)); | |||
111 | ||||
112 | // initialize the stack to be empty | |||
113 | task_stack->ts_entries = TASK_STACK_EMPTY; | |||
114 | first_block->sb_next = NULL__null; | |||
115 | first_block->sb_prev = NULL__null; | |||
116 | } | |||
117 | ||||
118 | // __kmp_free_task_stack: free the task stack when thread_data is destroyed. | |||
119 | // | |||
120 | // gtid: global thread identifier for calling thread | |||
121 | // thread_data: thread info for thread containing stack | |||
122 | static void __kmp_free_task_stack(kmp_int32 gtid, | |||
123 | kmp_thread_data_t *thread_data) { | |||
124 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
125 | kmp_stack_block_t *stack_block = &task_stack->ts_first_block; | |||
126 | ||||
127 | KMP_DEBUG_ASSERT(task_stack->ts_entries == TASK_STACK_EMPTY)((task_stack->ts_entries == TASK_STACK_EMPTY) ? 0 : __kmp_debug_assert ("task_stack->ts_entries == TASK_STACK_EMPTY", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 127)); | |||
128 | // free from the second block of the stack | |||
129 | while (stack_block != NULL__null) { | |||
130 | kmp_stack_block_t *next_block = (stack_block) ? stack_block->sb_next : NULL__null; | |||
131 | ||||
132 | stack_block->sb_next = NULL__null; | |||
133 | stack_block->sb_prev = NULL__null; | |||
134 | if (stack_block != &task_stack->ts_first_block) { | |||
135 | __kmp_thread_free(thread,___kmp_thread_free((thread), (stack_block), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 136) | |||
136 | stack_block)___kmp_thread_free((thread), (stack_block), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 136); // free the block, if not the first | |||
137 | } | |||
138 | stack_block = next_block; | |||
139 | } | |||
140 | // initialize the stack to be empty | |||
141 | task_stack->ts_entries = 0; | |||
142 | task_stack->ts_top = NULL__null; | |||
143 | } | |||
144 | ||||
145 | // __kmp_push_task_stack: Push the tied task onto the task stack. | |||
146 | // Grow the stack if necessary by allocating another block. | |||
147 | // | |||
148 | // gtid: global thread identifier for calling thread | |||
149 | // thread: thread info for thread containing stack | |||
150 | // tied_task: the task to push on the stack | |||
151 | static void __kmp_push_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
152 | kmp_taskdata_t *tied_task) { | |||
153 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
154 | kmp_thread_data_t *thread_data = | |||
155 | &thread->th.th_task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 155)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
156 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
157 | ||||
158 | if (tied_task->td_flags.team_serial || tied_task->td_flags.tasking_ser) { | |||
159 | return; // Don't push anything on stack if team or team tasks are serialized | |||
160 | } | |||
161 | ||||
162 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 162)); | |||
163 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)((task_stack->ts_top != __null) ? 0 : __kmp_debug_assert("task_stack->ts_top != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 163)); | |||
164 | ||||
165 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); } | |||
166 | ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); } | |||
167 | gtid, thread, tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); }; | |||
168 | // Store entry | |||
169 | *(task_stack->ts_top) = tied_task; | |||
170 | ||||
171 | // Do bookkeeping for next push | |||
172 | task_stack->ts_top++; | |||
173 | task_stack->ts_entries++; | |||
174 | ||||
175 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
176 | // Find beginning of this task block | |||
177 | kmp_stack_block_t *stack_block = | |||
178 | (kmp_stack_block_t *)(task_stack->ts_top - TASK_STACK_BLOCK_SIZE); | |||
179 | ||||
180 | // Check if we already have a block | |||
181 | if (stack_block->sb_next != | |||
182 | NULL__null) { // reset ts_top to beginning of next block | |||
183 | task_stack->ts_top = &stack_block->sb_next->sb_block[0]; | |||
184 | } else { // Alloc new block and link it up | |||
185 | kmp_stack_block_t *new_block = (kmp_stack_block_t *)__kmp_thread_calloc( | |||
186 | thread, sizeof(kmp_stack_block_t)); | |||
187 | ||||
188 | task_stack->ts_top = &new_block->sb_block[0]; | |||
189 | stack_block->sb_next = new_block; | |||
190 | new_block->sb_prev = stack_block; | |||
191 | new_block->sb_next = NULL__null; | |||
192 | ||||
193 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
194 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
195 | ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
196 | gtid, tied_task, new_block))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); }; | |||
197 | } | |||
198 | } | |||
199 | KA_TRACE(20, ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); } | |||
200 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
201 | } | |||
202 | ||||
203 | // __kmp_pop_task_stack: Pop the tied task from the task stack. Don't return | |||
204 | // the task, just check to make sure it matches the ending task passed in. | |||
205 | // | |||
206 | // gtid: global thread identifier for the calling thread | |||
207 | // thread: thread info structure containing stack | |||
208 | // tied_task: the task popped off the stack | |||
209 | // ending_task: the task that is ending (should match popped task) | |||
210 | static void __kmp_pop_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
211 | kmp_taskdata_t *ending_task) { | |||
212 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
213 | kmp_thread_data_t *thread_data = | |||
214 | &thread->th.th_task_team->tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 214)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
215 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
216 | kmp_taskdata_t *tied_task; | |||
217 | ||||
218 | if (ending_task->td_flags.team_serial || ending_task->td_flags.tasking_ser) { | |||
219 | // Don't pop anything from stack if team or team tasks are serialized | |||
220 | return; | |||
221 | } | |||
222 | ||||
223 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)((task_stack->ts_top != __null) ? 0 : __kmp_debug_assert("task_stack->ts_top != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 223)); | |||
224 | KMP_DEBUG_ASSERT(task_stack->ts_entries > 0)((task_stack->ts_entries > 0) ? 0 : __kmp_debug_assert( "task_stack->ts_entries > 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 224)); | |||
225 | ||||
226 | KA_TRACE(20, ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n" , gtid, thread); } | |||
227 | thread))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n" , gtid, thread); }; | |||
228 | ||||
229 | // fix up ts_top if we need to pop from previous block | |||
230 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
231 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(task_stack->ts_top); | |||
232 | ||||
233 | stack_block = stack_block->sb_prev; | |||
234 | task_stack->ts_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
235 | } | |||
236 | ||||
237 | // finish bookkeeping | |||
238 | task_stack->ts_top--; | |||
239 | task_stack->ts_entries--; | |||
240 | ||||
241 | tied_task = *(task_stack->ts_top); | |||
242 | ||||
243 | KMP_DEBUG_ASSERT(tied_task != NULL)((tied_task != __null) ? 0 : __kmp_debug_assert("tied_task != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 243)); | |||
244 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 244)); | |||
245 | KMP_DEBUG_ASSERT(tied_task == ending_task)((tied_task == ending_task) ? 0 : __kmp_debug_assert("tied_task == ending_task" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 245)); // If we built the stack correctly | |||
246 | ||||
247 | KA_TRACE(20, ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); } | |||
248 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
249 | return; | |||
250 | } | |||
251 | #endif /* BUILD_TIED_TASK_STACK */ | |||
252 | ||||
253 | // __kmp_push_task: Add a task to the thread's deque | |||
254 | static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) { | |||
255 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
256 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
257 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
258 | 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_tasking.cpp" , 258)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); | |||
259 | kmp_thread_data_t *thread_data; | |||
260 | ||||
261 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d trying to push task %p.\n" , gtid, taskdata); } | |||
262 | ("__kmp_push_task: T#%d trying to push task %p.\n", gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d trying to push task %p.\n" , gtid, taskdata); }; | |||
263 | ||||
264 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
265 | // untied task needs to increment counter so that the task structure is not | |||
266 | // freed prematurely | |||
267 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
268 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
269 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
270 | ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
271 | gtid, counter, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); }; | |||
272 | } | |||
273 | ||||
274 | // The first check avoids building task_team thread data if serialized | |||
275 | if (taskdata->td_flags.task_serial) { | |||
276 | KA_TRACE(20, ("__kmp_push_task: T#%d team serialized; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
277 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
278 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
279 | return TASK_NOT_PUSHED1; | |||
280 | } | |||
281 | ||||
282 | // Now that serialized tasks have returned, we can assume that we are not in | |||
283 | // immediate exec mode | |||
284 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 284)); | |||
285 | if (!KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
286 | __kmp_enable_tasking(task_team, thread); | |||
287 | } | |||
288 | KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE)(((task_team->tt.tt_found_tasks) == (!0)) ? 0 : __kmp_debug_assert ("(task_team->tt.tt_found_tasks) == (!0)", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 288)); | |||
289 | KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL)((((void *)(task_team->tt.tt_threads_data)) != __null) ? 0 : __kmp_debug_assert("((void *)(task_team->tt.tt_threads_data)) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 289)); | |||
290 | ||||
291 | // Find tasking deque specific to encountering thread | |||
292 | thread_data = &task_team->tt.tt_threads_data[tid]; | |||
293 | ||||
294 | // No lock needed since only owner can allocate | |||
295 | if (thread_data->td.td_deque == NULL__null) { | |||
296 | __kmp_alloc_task_deque(thread, thread_data); | |||
297 | } | |||
298 | ||||
299 | // Check if deque is full | |||
300 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
301 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
302 | KA_TRACE(20, ("__kmp_push_task: T#%d deque is full; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
303 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
304 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
305 | return TASK_NOT_PUSHED1; | |||
306 | } | |||
307 | ||||
308 | // Lock the deque for the task push operation | |||
309 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
310 | ||||
311 | #if OMP_45_ENABLED(50 >= 45) | |||
312 | // Need to recheck as we can get a proxy task from a thread outside of OpenMP | |||
313 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
314 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
315 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
316 | KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
317 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
318 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
319 | return TASK_NOT_PUSHED1; | |||
320 | } | |||
321 | #else | |||
322 | // Must have room since no thread can add tasks but calling thread | |||
323 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <(((thread_data->td.td_deque_ntasks) < ((thread_data-> td).td_deque_size)) ? 0 : __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 324)) | |||
324 | TASK_DEQUE_SIZE(thread_data->td))(((thread_data->td.td_deque_ntasks) < ((thread_data-> td).td_deque_size)) ? 0 : __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 324)); | |||
325 | #endif | |||
326 | ||||
327 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = | |||
328 | taskdata; // Push taskdata | |||
329 | // Wrap index. | |||
330 | thread_data->td.td_deque_tail = | |||
331 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
332 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
333 | TCR_4(thread_data->td.td_deque_ntasks) + 1)(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1); // Adjust task count | |||
334 | ||||
335 | KA_TRACE(20, ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
336 | "task=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
337 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
338 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); }; | |||
339 | ||||
340 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
341 | ||||
342 | return TASK_SUCCESSFULLY_PUSHED0; | |||
343 | } | |||
344 | ||||
345 | // __kmp_pop_current_task_from_thread: set up current task from called thread | |||
346 | // when team ends | |||
347 | // | |||
348 | // this_thr: thread structure to set current_task in. | |||
349 | void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) { | |||
350 | KF_TRACE(10, ("__kmp_pop_current_task_from_thread(enter): T#%d "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
351 | "this_thread=%p, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
352 | "curtask_parent=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
353 | 0, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
354 | this_thr->th.th_current_task->td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); }; | |||
355 | ||||
356 | this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent; | |||
357 | ||||
358 | KF_TRACE(10, ("__kmp_pop_current_task_from_thread(exit): T#%d "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
359 | "this_thread=%p, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
360 | "curtask_parent=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
361 | 0, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
362 | this_thr->th.th_current_task->td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); }; | |||
363 | } | |||
364 | ||||
365 | // __kmp_push_current_task_to_thread: set up current task in called thread for a | |||
366 | // new team | |||
367 | // | |||
368 | // this_thr: thread structure to set up | |||
369 | // team: team for implicit task data | |||
370 | // tid: thread within team to set up | |||
371 | void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team, | |||
372 | int tid) { | |||
373 | // current task of the thread is a parent of the new just created implicit | |||
374 | // tasks of new team | |||
375 | KF_TRACE(10, ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
376 | "curtask=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
377 | "parent_task=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
378 | tid, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
379 | team->t.t_implicit_task_taskdata[tid].td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); }; | |||
380 | ||||
381 | 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_tasking.cpp" , 381)); | |||
382 | ||||
383 | if (tid == 0) { | |||
384 | if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) { | |||
385 | team->t.t_implicit_task_taskdata[0].td_parent = | |||
386 | this_thr->th.th_current_task; | |||
387 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0]; | |||
388 | } | |||
389 | } else { | |||
390 | team->t.t_implicit_task_taskdata[tid].td_parent = | |||
391 | team->t.t_implicit_task_taskdata[0].td_parent; | |||
392 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid]; | |||
393 | } | |||
394 | ||||
395 | KF_TRACE(10, ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
396 | "curtask=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
397 | "parent_task=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
398 | tid, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
399 | team->t.t_implicit_task_taskdata[tid].td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); }; | |||
400 | } | |||
401 | ||||
402 | // __kmp_task_start: bookkeeping for a task starting execution | |||
403 | // | |||
404 | // GTID: global thread id of calling thread | |||
405 | // task: task starting execution | |||
406 | // current_task: task suspending | |||
407 | static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task, | |||
408 | kmp_taskdata_t *current_task) { | |||
409 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
410 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
411 | ||||
412 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); } | |||
413 | ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); } | |||
414 | gtid, taskdata, current_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); }; | |||
415 | ||||
416 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 416)); | |||
417 | ||||
418 | // mark currently executing task as suspended | |||
419 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
420 | // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 ); | |||
421 | current_task->td_flags.executing = 0; | |||
422 | ||||
423 | // Add task to stack if tied | |||
424 | #ifdef BUILD_TIED_TASK_STACK | |||
425 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
426 | __kmp_push_task_stack(gtid, thread, taskdata); | |||
427 | } | |||
428 | #endif /* BUILD_TIED_TASK_STACK */ | |||
429 | ||||
430 | // mark starting task as executing and as current task | |||
431 | thread->th.th_current_task = taskdata; | |||
432 | ||||
433 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||((taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 434)) | |||
434 | taskdata->td_flags.tiedness == TASK_UNTIED)((taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 434)); | |||
435 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||((taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 436)) | |||
436 | taskdata->td_flags.tiedness == TASK_UNTIED)((taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 436)); | |||
437 | taskdata->td_flags.started = 1; | |||
438 | taskdata->td_flags.executing = 1; | |||
439 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 439)); | |||
440 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 440)); | |||
441 | ||||
442 | // GEH TODO: shouldn't we pass some sort of location identifier here? | |||
443 | // APT: yes, we will pass location here. | |||
444 | // need to store current thread state (in a thread or taskdata structure) | |||
445 | // before setting work_state, otherwise wrong state is set after end of task | |||
446 | ||||
447 | KA_TRACE(10, ("__kmp_task_start(exit): T#%d task=%p\n", gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(exit): T#%d task=%p\n" , gtid, taskdata); }; | |||
448 | ||||
449 | return; | |||
450 | } | |||
451 | ||||
452 | #if OMPT_SUPPORT1 | |||
453 | //------------------------------------------------------------------------------ | |||
454 | // __ompt_task_init: | |||
455 | // Initialize OMPT fields maintained by a task. This will only be called after | |||
456 | // ompt_start_tool, so we already know whether ompt is enabled or not. | |||
457 | ||||
458 | static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) { | |||
459 | // The calls to __ompt_task_init already have the ompt_enabled condition. | |||
460 | task->ompt_task_info.task_data.value = 0; | |||
461 | task->ompt_task_info.frame.exit_frame = NULL__null; | |||
462 | task->ompt_task_info.frame.enter_frame = NULL__null; | |||
463 | #if OMP_40_ENABLED(50 >= 40) | |||
464 | task->ompt_task_info.ndeps = 0; | |||
465 | task->ompt_task_info.deps = NULL__null; | |||
466 | #endif /* OMP_40_ENABLED */ | |||
467 | } | |||
468 | ||||
469 | // __ompt_task_start: | |||
470 | // Build and trigger task-begin event | |||
471 | static inline void __ompt_task_start(kmp_task_t *task, | |||
472 | kmp_taskdata_t *current_task, | |||
473 | kmp_int32 gtid) { | |||
474 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
475 | ompt_task_status_t status = ompt_task_others; | |||
476 | if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) { | |||
477 | status = ompt_task_yield; | |||
478 | __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0; | |||
479 | } | |||
480 | /* let OMPT know that we're about to run this task */ | |||
481 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
482 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
483 | &(current_task->ompt_task_info.task_data), status, | |||
484 | &(taskdata->ompt_task_info.task_data)); | |||
485 | } | |||
486 | taskdata->ompt_task_info.scheduling_parent = current_task; | |||
487 | } | |||
488 | ||||
489 | // __ompt_task_finish: | |||
490 | // Build and trigger final task-schedule event | |||
491 | static inline void | |||
492 | __ompt_task_finish(kmp_task_t *task, kmp_taskdata_t *resumed_task, | |||
493 | ompt_task_status_t status = ompt_task_complete) { | |||
494 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
495 | if (__kmp_omp_cancellation && taskdata->td_taskgroup && | |||
496 | taskdata->td_taskgroup->cancel_request == cancel_taskgroup) { | |||
497 | status = ompt_task_cancel; | |||
498 | } | |||
499 | ||||
500 | /* let OMPT know that we're returning to the callee task */ | |||
501 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
502 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
503 | &(taskdata->ompt_task_info.task_data), status, | |||
504 | &((resumed_task ? resumed_task | |||
505 | : (taskdata->ompt_task_info.scheduling_parent | |||
506 | ? taskdata->ompt_task_info.scheduling_parent | |||
507 | : taskdata->td_parent)) | |||
508 | ->ompt_task_info.task_data)); | |||
509 | } | |||
510 | } | |||
511 | #endif | |||
512 | ||||
513 | template <bool ompt> | |||
514 | static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid, | |||
515 | kmp_task_t *task, | |||
516 | void *frame_address, | |||
517 | void *return_address) { | |||
518 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
519 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
520 | ||||
521 | KA_TRACE(10, ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p " "current_task=%p\n", gtid, loc_ref, taskdata, current_task); } | |||
522 | "current_task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p " "current_task=%p\n", gtid, loc_ref, taskdata, current_task); } | |||
523 | gtid, loc_ref, taskdata, current_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p " "current_task=%p\n", gtid, loc_ref, taskdata, current_task); }; | |||
524 | ||||
525 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
526 | // untied task needs to increment counter so that the task structure is not | |||
527 | // freed prematurely | |||
528 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
529 | KA_TRACE(20, ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) " "incremented for task %p\n", gtid, counter, taskdata); } | |||
530 | "incremented for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) " "incremented for task %p\n", gtid, counter, taskdata); } | |||
531 | gtid, counter, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) " "incremented for task %p\n", gtid, counter, taskdata); }; | |||
532 | } | |||
533 | ||||
534 | taskdata->td_flags.task_serial = | |||
535 | 1; // Execute this task immediately, not deferred. | |||
536 | __kmp_task_start(gtid, task, current_task); | |||
537 | ||||
538 | #if OMPT_SUPPORT1 | |||
539 | if (ompt) { | |||
540 | if (current_task->ompt_task_info.frame.enter_frame == NULL__null) { | |||
541 | current_task->ompt_task_info.frame.enter_frame = | |||
542 | taskdata->ompt_task_info.frame.exit_frame = frame_address; | |||
543 | } | |||
544 | if (ompt_enabled.ompt_callback_task_create) { | |||
545 | ompt_task_info_t *parent_info = &(current_task->ompt_task_info); | |||
546 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
547 | &(parent_info->task_data), &(parent_info->frame), | |||
548 | &(taskdata->ompt_task_info.task_data), | |||
549 | ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(taskdata)((taskdata->td_flags.task_serial || taskdata->td_flags. tasking_ser) ? ompt_task_undeferred : 0x0) | ((!(taskdata-> td_flags.tiedness)) ? ompt_task_untied : 0x0) | (taskdata-> td_flags.final ? ompt_task_final : 0x0) | (taskdata->td_flags .merged_if0 ? ompt_task_mergeable : 0x0), 0, | |||
550 | return_address); | |||
551 | } | |||
552 | __ompt_task_start(task, current_task, gtid); | |||
553 | } | |||
554 | #endif // OMPT_SUPPORT | |||
555 | ||||
556 | KA_TRACE(10, ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n", gtid,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n" , gtid, loc_ref, taskdata); } | |||
557 | loc_ref, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n" , gtid, loc_ref, taskdata); }; | |||
558 | } | |||
559 | ||||
560 | #if OMPT_SUPPORT1 | |||
561 | OMPT_NOINLINE__attribute__((noinline)) | |||
562 | static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
563 | kmp_task_t *task, | |||
564 | void *frame_address, | |||
565 | void *return_address) { | |||
566 | __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address, | |||
567 | return_address); | |||
568 | } | |||
569 | #endif // OMPT_SUPPORT | |||
570 | ||||
571 | // __kmpc_omp_task_begin_if0: report that a given serialized task has started | |||
572 | // execution | |||
573 | // | |||
574 | // loc_ref: source location information; points to beginning of task block. | |||
575 | // gtid: global thread number. | |||
576 | // task: task thunk for the started task. | |||
577 | void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
578 | kmp_task_t *task) { | |||
579 | #if OMPT_SUPPORT1 | |||
580 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
581 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
582 | __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task, | |||
583 | OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1), | |||
584 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
585 | return; | |||
586 | } | |||
587 | #endif | |||
588 | __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL__null, NULL__null); | |||
589 | } | |||
590 | ||||
591 | #ifdef TASK_UNUSED | |||
592 | // __kmpc_omp_task_begin: report that a given task has started execution | |||
593 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
594 | void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) { | |||
595 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
596 | ||||
597 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1), current_task ); } | |||
598 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1), current_task ); } | |||
599 | ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1), current_task ); } | |||
600 | gtid, loc_ref, KMP_TASK_TO_TASKDATA(task), current_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1), current_task ); }; | |||
601 | ||||
602 | __kmp_task_start(gtid, task, current_task); | |||
603 | ||||
604 | KA_TRACE(10, ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n", gtid,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); } | |||
605 | loc_ref, KMP_TASK_TO_TASKDATA(task)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); }; | |||
606 | return; | |||
607 | } | |||
608 | #endif // TASK_UNUSED | |||
609 | ||||
610 | // __kmp_free_task: free the current task space and the space for shareds | |||
611 | // | |||
612 | // gtid: Global thread ID of calling thread | |||
613 | // taskdata: task to free | |||
614 | // thread: thread data structure of caller | |||
615 | static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata, | |||
616 | kmp_info_t *thread) { | |||
617 | KA_TRACE(30, ("__kmp_free_task: T#%d freeing data from task %p\n", gtid,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_free_task: T#%d freeing data from task %p\n" , gtid, taskdata); } | |||
618 | taskdata))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_free_task: T#%d freeing data from task %p\n" , gtid, taskdata); }; | |||
619 | ||||
620 | // Check to make sure all flags and counters have the correct values | |||
621 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 621)); | |||
622 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0)((taskdata->td_flags.executing == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.executing == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 622)); | |||
623 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1)((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 623)); | |||
624 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 624)); | |||
625 | KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||((taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1) ? 0 : __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 626)) | |||
626 | taskdata->td_flags.task_serial == 1)((taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1) ? 0 : __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 626)); | |||
627 | KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0)((taskdata->td_incomplete_child_tasks == 0) ? 0 : __kmp_debug_assert ("taskdata->td_incomplete_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 627)); | |||
628 | ||||
629 | taskdata->td_flags.freed = 1; | |||
630 | ANNOTATE_HAPPENS_BEFORE(taskdata); | |||
631 | // deallocate the taskdata and shared variable blocks associated with this task | |||
632 | #if USE_FAST_MEMORY3 | |||
633 | __kmp_fast_free(thread, taskdata)___kmp_fast_free((thread), (taskdata), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 633); | |||
634 | #else /* ! USE_FAST_MEMORY */ | |||
635 | __kmp_thread_free(thread, taskdata)___kmp_thread_free((thread), (taskdata), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 635); | |||
636 | #endif | |||
637 | ||||
638 | KA_TRACE(20, ("__kmp_free_task: T#%d freed task %p\n", gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task: T#%d freed task %p\n" , gtid, taskdata); }; | |||
639 | } | |||
640 | ||||
641 | // __kmp_free_task_and_ancestors: free the current task and ancestors without | |||
642 | // children | |||
643 | // | |||
644 | // gtid: Global thread ID of calling thread | |||
645 | // taskdata: task to free | |||
646 | // thread: thread data structure of caller | |||
647 | static void __kmp_free_task_and_ancestors(kmp_int32 gtid, | |||
648 | kmp_taskdata_t *taskdata, | |||
649 | kmp_info_t *thread) { | |||
650 | #if OMP_45_ENABLED(50 >= 45) | |||
651 | // Proxy tasks must always be allowed to free their parents | |||
652 | // because they can be run in background even in serial mode. | |||
653 | kmp_int32 team_serial = | |||
654 | (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) && | |||
655 | !taskdata->td_flags.proxy; | |||
656 | #else | |||
657 | kmp_int32 team_serial = | |||
658 | taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser; | |||
659 | #endif | |||
660 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 660)); | |||
661 | ||||
662 | kmp_int32 children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks)(&taskdata->td_allocated_child_tasks)->fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
663 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 663)); | |||
664 | ||||
665 | // Now, go up the ancestor tree to see if any ancestors can now be freed. | |||
666 | while (children == 0) { | |||
667 | kmp_taskdata_t *parent_taskdata = taskdata->td_parent; | |||
668 | ||||
669 | KA_TRACE(20, ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete " "and freeing itself\n", gtid, taskdata); } | |||
670 | "and freeing itself\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete " "and freeing itself\n", gtid, taskdata); } | |||
671 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete " "and freeing itself\n", gtid, taskdata); }; | |||
672 | ||||
673 | // --- Deallocate my ancestor task --- | |||
674 | __kmp_free_task(gtid, taskdata, thread); | |||
675 | ||||
676 | taskdata = parent_taskdata; | |||
677 | ||||
678 | // Stop checking ancestors at implicit task instead of walking up ancestor | |||
679 | // tree to avoid premature deallocation of ancestors. | |||
680 | if (team_serial || taskdata->td_flags.tasktype == TASK_IMPLICIT0) | |||
681 | return; | |||
682 | ||||
683 | // Predecrement simulated by "- 1" calculation | |||
684 | children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks)(&taskdata->td_allocated_child_tasks)->fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
685 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 685)); | |||
686 | } | |||
687 | ||||
688 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; " "not freeing it yet\n", gtid, taskdata, children); } | |||
689 | 20, ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; " "not freeing it yet\n", gtid, taskdata, children); } | |||
690 | "not freeing it yet\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; " "not freeing it yet\n", gtid, taskdata, children); } | |||
691 | gtid, taskdata, children))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; " "not freeing it yet\n", gtid, taskdata, children); }; | |||
692 | } | |||
693 | ||||
694 | // __kmp_task_finish: bookkeeping to do when a task finishes execution | |||
695 | // | |||
696 | // gtid: global thread ID for calling thread | |||
697 | // task: task to be finished | |||
698 | // resumed_task: task to be resumed. (may be NULL if task is serialized) | |||
699 | template <bool ompt> | |||
700 | static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task, | |||
701 | kmp_taskdata_t *resumed_task) { | |||
702 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
703 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
704 | kmp_task_team_t *task_team = | |||
705 | thread->th.th_task_team; // might be NULL for serial teams... | |||
706 | kmp_int32 children = 0; | |||
707 | ||||
708 | KA_TRACE(10, ("__kmp_task_finish(enter): T#%d finishing task %p and resuming "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(enter): T#%d finishing task %p and resuming " "task %p\n", gtid, taskdata, resumed_task); } | |||
709 | "task %p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(enter): T#%d finishing task %p and resuming " "task %p\n", gtid, taskdata, resumed_task); } | |||
710 | gtid, taskdata, resumed_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(enter): T#%d finishing task %p and resuming " "task %p\n", gtid, taskdata, resumed_task); }; | |||
711 | ||||
712 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 712)); | |||
713 | ||||
714 | // Pop task from stack if tied | |||
715 | #ifdef BUILD_TIED_TASK_STACK | |||
716 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
717 | __kmp_pop_task_stack(gtid, thread, taskdata); | |||
718 | } | |||
719 | #endif /* BUILD_TIED_TASK_STACK */ | |||
720 | ||||
721 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
722 | // untied task needs to check the counter so that the task structure is not | |||
723 | // freed prematurely | |||
724 | kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_sub(1, std::memory_order_acq_rel ) - 1; | |||
725 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n" , gtid, counter, taskdata); } | |||
726 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n" , gtid, counter, taskdata); } | |||
727 | ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n" , gtid, counter, taskdata); } | |||
728 | gtid, counter, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n" , gtid, counter, taskdata); }; | |||
729 | if (counter > 0) { | |||
730 | // untied task is not done, to be continued possibly by other thread, do | |||
731 | // not free it now | |||
732 | if (resumed_task == NULL__null) { | |||
733 | KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial)((taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("taskdata->td_flags.task_serial", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 733)); | |||
734 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
735 | // task is the parent | |||
736 | } | |||
737 | thread->th.th_current_task = resumed_task; // restore current_task | |||
738 | resumed_task->td_flags.executing = 1; // resume previous task | |||
739 | KA_TRACE(10, ("__kmp_task_finish(exit): T#%d partially done task %p, "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d partially done task %p, " "resuming task %p\n", gtid, taskdata, resumed_task); } | |||
740 | "resuming task %p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d partially done task %p, " "resuming task %p\n", gtid, taskdata, resumed_task); } | |||
741 | gtid, taskdata, resumed_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d partially done task %p, " "resuming task %p\n", gtid, taskdata, resumed_task); }; | |||
742 | return; | |||
743 | } | |||
744 | } | |||
745 | #if OMPT_SUPPORT1 | |||
746 | if (ompt) | |||
747 | __ompt_task_finish(task, resumed_task); | |||
748 | #endif | |||
749 | ||||
750 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 750)); | |||
751 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
752 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1)((taskdata->td_flags.started == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.started == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 752)); | |||
753 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 753)); | |||
754 | ||||
755 | // Only need to keep track of count if team parallel and tasking not | |||
756 | // serialized | |||
757 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) { | |||
758 | // Predecrement simulated by "- 1" calculation | |||
759 | children = | |||
760 | KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks)(&taskdata->td_parent->td_incomplete_child_tasks)-> fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
761 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 761)); | |||
762 | #if OMP_40_ENABLED(50 >= 40) | |||
763 | if (taskdata->td_taskgroup) | |||
764 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
765 | #if OMP_45_ENABLED(50 >= 45) | |||
766 | } | |||
767 | // if we found proxy tasks there could exist a dependency chain | |||
768 | // with the proxy task as origin | |||
769 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) || | |||
770 | (task_team && task_team->tt.tt_found_proxy_tasks)) { | |||
771 | #endif | |||
772 | __kmp_release_deps(gtid, taskdata); | |||
773 | #endif | |||
774 | } | |||
775 | ||||
776 | // td_flags.executing must be marked as 0 after __kmp_release_deps has been | |||
777 | // called. Othertwise, if a task is executed immediately from the release_deps | |||
778 | // code, the flag will be reset to 1 again by this same function | |||
779 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1)((taskdata->td_flags.executing == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.executing == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 779)); | |||
780 | taskdata->td_flags.executing = 0; // suspend the finishing task | |||
781 | ||||
782 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n" , gtid, taskdata, children); } | |||
783 | 20, ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n" , gtid, taskdata, children); } | |||
784 | gtid, taskdata, children))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n" , gtid, taskdata, children); }; | |||
785 | ||||
786 | #if OMP_40_ENABLED(50 >= 40) | |||
787 | /* If the tasks' destructor thunk flag has been set, we need to invoke the | |||
788 | destructor thunk that has been generated by the compiler. The code is | |||
789 | placed here, since at this point other tasks might have been released | |||
790 | hence overlapping the destructor invokations with some other work in the | |||
791 | released tasks. The OpenMP spec is not specific on when the destructors | |||
792 | are invoked, so we should be free to choose. */ | |||
793 | if (taskdata->td_flags.destructors_thunk) { | |||
794 | kmp_routine_entry_t destr_thunk = task->data1.destructors; | |||
795 | KMP_ASSERT(destr_thunk)((destr_thunk) ? 0 : __kmp_debug_assert("destr_thunk", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 795)); | |||
796 | destr_thunk(gtid, task); | |||
797 | } | |||
798 | #endif // OMP_40_ENABLED | |||
799 | ||||
800 | // bookkeeping for resuming task: | |||
801 | // GEH - note tasking_ser => task_serial | |||
802 | KMP_DEBUG_ASSERT((((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 804)) | |||
803 | (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==(((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 804)) | |||
804 | taskdata->td_flags.task_serial)(((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 804)); | |||
805 | if (taskdata->td_flags.task_serial) { | |||
806 | if (resumed_task == NULL__null) { | |||
807 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
808 | // task is the parent | |||
809 | } | |||
810 | } else { | |||
811 | KMP_DEBUG_ASSERT(resumed_task !=((resumed_task != __null) ? 0 : __kmp_debug_assert("resumed_task != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 812)) | |||
812 | NULL)((resumed_task != __null) ? 0 : __kmp_debug_assert("resumed_task != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 812)); // verify that resumed task is passed as arguemnt | |||
813 | } | |||
814 | ||||
815 | // Free this task and then ancestor tasks if they have no children. | |||
816 | // Restore th_current_task first as suggested by John: | |||
817 | // johnmc: if an asynchronous inquiry peers into the runtime system | |||
818 | // it doesn't see the freed task as the current task. | |||
819 | thread->th.th_current_task = resumed_task; | |||
820 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
821 | ||||
822 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
823 | // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 ); | |||
824 | resumed_task->td_flags.executing = 1; // resume previous task | |||
| ||||
825 | ||||
826 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n" , gtid, taskdata, resumed_task); } | |||
827 | 10, ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n" , gtid, taskdata, resumed_task); } | |||
828 | gtid, taskdata, resumed_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n" , gtid, taskdata, resumed_task); }; | |||
829 | ||||
830 | return; | |||
831 | } | |||
832 | ||||
833 | template <bool ompt> | |||
834 | static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref, | |||
835 | kmp_int32 gtid, | |||
836 | kmp_task_t *task) { | |||
837 | KA_TRACE(10, ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); } | |||
838 | gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); }; | |||
839 | // this routine will provide task to resume | |||
840 | __kmp_task_finish<ompt>(gtid, task, NULL__null); | |||
841 | ||||
842 | KA_TRACE(10, ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); } | |||
843 | gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); }; | |||
844 | ||||
845 | #if OMPT_SUPPORT1 | |||
846 | if (ompt) { | |||
847 | omp_frame_t *ompt_frame; | |||
848 | __ompt_get_task_info_internal(0, NULL__null, NULL__null, &ompt_frame, NULL__null, NULL__null); | |||
849 | ompt_frame->enter_frame = NULL__null; | |||
850 | } | |||
851 | #endif | |||
852 | ||||
853 | return; | |||
854 | } | |||
855 | ||||
856 | #if OMPT_SUPPORT1 | |||
857 | OMPT_NOINLINE__attribute__((noinline)) | |||
858 | void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
859 | kmp_task_t *task) { | |||
860 | __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task); | |||
861 | } | |||
862 | #endif // OMPT_SUPPORT | |||
863 | ||||
864 | // __kmpc_omp_task_complete_if0: report that a task has completed execution | |||
865 | // | |||
866 | // loc_ref: source location information; points to end of task block. | |||
867 | // gtid: global thread number. | |||
868 | // task: task thunk for the completed task. | |||
869 | void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
870 | kmp_task_t *task) { | |||
871 | #if OMPT_SUPPORT1 | |||
872 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| ||||
873 | __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task); | |||
874 | return; | |||
875 | } | |||
876 | #endif | |||
877 | __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task); | |||
878 | } | |||
879 | ||||
880 | #ifdef TASK_UNUSED | |||
881 | // __kmpc_omp_task_complete: report that a task has completed execution | |||
882 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
883 | void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, | |||
884 | kmp_task_t *task) { | |||
885 | KA_TRACE(10, ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n", gtid,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); } | |||
886 | loc_ref, KMP_TASK_TO_TASKDATA(task)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); }; | |||
887 | ||||
888 | __kmp_task_finish<false>(gtid, task, | |||
889 | NULL__null); // Not sure how to find task to resume | |||
890 | ||||
891 | KA_TRACE(10, ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n", gtid,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); } | |||
892 | loc_ref, KMP_TASK_TO_TASKDATA(task)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n" , gtid, loc_ref, (((kmp_taskdata_t *)task) - 1)); }; | |||
893 | return; | |||
894 | } | |||
895 | #endif // TASK_UNUSED | |||
896 | ||||
897 | // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit | |||
898 | // task for a given thread | |||
899 | // | |||
900 | // loc_ref: reference to source location of parallel region | |||
901 | // this_thr: thread data structure corresponding to implicit task | |||
902 | // team: team for this_thr | |||
903 | // tid: thread id of given thread within team | |||
904 | // set_curr_task: TRUE if need to push current task to thread | |||
905 | // NOTE: Routine does not set up the implicit task ICVS. This is assumed to | |||
906 | // have already been done elsewhere. | |||
907 | // TODO: Get better loc_ref. Value passed in may be NULL | |||
908 | void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, | |||
909 | kmp_team_t *team, int tid, int set_curr_task) { | |||
910 | kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid]; | |||
911 | ||||
912 | KF_TRACE(if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n" , tid, team, task, set_curr_task ? "TRUE" : "FALSE"); } | |||
913 | 10,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n" , tid, team, task, set_curr_task ? "TRUE" : "FALSE"); } | |||
914 | ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n" , tid, team, task, set_curr_task ? "TRUE" : "FALSE"); } | |||
915 | tid, team, task, set_curr_task ? "TRUE" : "FALSE"))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n" , tid, team, task, set_curr_task ? "TRUE" : "FALSE"); }; | |||
916 | ||||
917 | task->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
918 | task->td_team = team; | |||
919 | // task->td_parent = NULL; // fix for CQ230101 (broken parent task info | |||
920 | // in debugger) | |||
921 | task->td_ident = loc_ref; | |||
922 | task->td_taskwait_ident = NULL__null; | |||
923 | task->td_taskwait_counter = 0; | |||
924 | task->td_taskwait_thread = 0; | |||
925 | ||||
926 | task->td_flags.tiedness = TASK_TIED1; | |||
927 | task->td_flags.tasktype = TASK_IMPLICIT0; | |||
928 | #if OMP_45_ENABLED(50 >= 45) | |||
929 | task->td_flags.proxy = TASK_FULL0; | |||
930 | #endif | |||
931 | ||||
932 | // All implicit tasks are executed immediately, not deferred | |||
933 | task->td_flags.task_serial = 1; | |||
934 | task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
935 | task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
936 | ||||
937 | task->td_flags.started = 1; | |||
938 | task->td_flags.executing = 1; | |||
939 | task->td_flags.complete = 0; | |||
940 | task->td_flags.freed = 0; | |||
941 | ||||
942 | #if OMP_40_ENABLED(50 >= 40) | |||
943 | task->td_depnode = NULL__null; | |||
944 | #endif | |||
945 | task->td_last_tied = task; | |||
946 | ||||
947 | if (set_curr_task) { // only do this init first time thread is created | |||
948 | KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0)(&task->td_incomplete_child_tasks)->store(0, std::memory_order_release ); | |||
949 | // Not used: don't need to deallocate implicit task | |||
950 | KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0)(&task->td_allocated_child_tasks)->store(0, std::memory_order_release ); | |||
951 | #if OMP_40_ENABLED(50 >= 40) | |||
952 | task->td_taskgroup = NULL__null; // An implicit task does not have taskgroup | |||
953 | task->td_dephash = NULL__null; | |||
954 | #endif | |||
955 | __kmp_push_current_task_to_thread(this_thr, team, tid); | |||
956 | } else { | |||
957 | KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0)((task->td_incomplete_child_tasks == 0) ? 0 : __kmp_debug_assert ("task->td_incomplete_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 957)); | |||
958 | KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0)((task->td_allocated_child_tasks == 0) ? 0 : __kmp_debug_assert ("task->td_allocated_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 958)); | |||
959 | } | |||
960 | ||||
961 | #if OMPT_SUPPORT1 | |||
962 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
963 | __ompt_task_init(task, tid); | |||
964 | #endif | |||
965 | ||||
966 | KF_TRACE(10, ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n", tid,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n" , tid, team, task); } | |||
967 | team, task))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n" , tid, team, task); }; | |||
968 | } | |||
969 | ||||
970 | // __kmp_finish_implicit_task: Release resources associated to implicit tasks | |||
971 | // at the end of parallel regions. Some resources are kept for reuse in the next | |||
972 | // parallel region. | |||
973 | // | |||
974 | // thread: thread data structure corresponding to implicit task | |||
975 | void __kmp_finish_implicit_task(kmp_info_t *thread) { | |||
976 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
977 | if (task->td_dephash) | |||
978 | __kmp_dephash_free_entries(thread, task->td_dephash); | |||
979 | } | |||
980 | ||||
981 | // __kmp_free_implicit_task: Release resources associated to implicit tasks | |||
982 | // when these are destroyed regions | |||
983 | // | |||
984 | // thread: thread data structure corresponding to implicit task | |||
985 | void __kmp_free_implicit_task(kmp_info_t *thread) { | |||
986 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
987 | if (task && task->td_dephash) { | |||
988 | __kmp_dephash_free(thread, task->td_dephash); | |||
989 | task->td_dephash = NULL__null; | |||
990 | } | |||
991 | } | |||
992 | ||||
993 | // Round up a size to a power of two specified by val: Used to insert padding | |||
994 | // between structures co-allocated using a single malloc() call | |||
995 | static size_t __kmp_round_up_to_val(size_t size, size_t val) { | |||
996 | if (size & (val - 1)) { | |||
997 | size &= ~(val - 1); | |||
998 | if (size <= KMP_SIZE_T_MAX(0xFFFFFFFFFFFFFFFF) - val) { | |||
999 | size += val; // Round up if there is no overflow. | |||
1000 | } | |||
1001 | } | |||
1002 | return size; | |||
1003 | } // __kmp_round_up_to_va | |||
1004 | ||||
1005 | // __kmp_task_alloc: Allocate the taskdata and task data structures for a task | |||
1006 | // | |||
1007 | // loc_ref: source location information | |||
1008 | // gtid: global thread number. | |||
1009 | // flags: include tiedness & task type (explicit vs. implicit) of the ''new'' | |||
1010 | // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine. | |||
1011 | // sizeof_kmp_task_t: Size in bytes of kmp_task_t data structure including | |||
1012 | // private vars accessed in task. | |||
1013 | // sizeof_shareds: Size in bytes of array of pointers to shared vars accessed | |||
1014 | // in task. | |||
1015 | // task_entry: Pointer to task code entry point generated by compiler. | |||
1016 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
1017 | kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
1018 | kmp_tasking_flags_t *flags, | |||
1019 | size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
1020 | kmp_routine_entry_t task_entry) { | |||
1021 | kmp_task_t *task; | |||
1022 | kmp_taskdata_t *taskdata; | |||
1023 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
1024 | kmp_team_t *team = thread->th.th_team; | |||
1025 | kmp_taskdata_t *parent_task = thread->th.th_current_task; | |||
1026 | size_t shareds_offset; | |||
1027 | ||||
1028 | KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , *((kmp_int32 *)flags), sizeof_kmp_task_t, sizeof_shareds, task_entry ); } | |||
1029 | "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , *((kmp_int32 *)flags), sizeof_kmp_task_t, sizeof_shareds, task_entry ); } | |||
1030 | gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , *((kmp_int32 *)flags), sizeof_kmp_task_t, sizeof_shareds, task_entry ); } | |||
1031 | sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , *((kmp_int32 *)flags), sizeof_kmp_task_t, sizeof_shareds, task_entry ); }; | |||
1032 | ||||
1033 | if (parent_task->td_flags.final) { | |||
1034 | if (flags->merged_if0) { | |||
1035 | } | |||
1036 | flags->final = 1; | |||
1037 | } | |||
1038 | if (flags->tiedness == TASK_UNTIED0 && !team->t.t_serialized) { | |||
1039 | // Untied task encountered causes the TSC algorithm to check entire deque of | |||
1040 | // the victim thread. If no untied task encountered, then checking the head | |||
1041 | // of the deque should be enough. | |||
1042 | KMP_CHECK_UPDATE(thread->th.th_task_team->tt.tt_untied_task_encountered, 1)if ((thread->th.th_task_team->tt.tt_untied_task_encountered ) != (1)) (thread->th.th_task_team->tt.tt_untied_task_encountered ) = (1); | |||
1043 | } | |||
1044 | ||||
1045 | #if OMP_45_ENABLED(50 >= 45) | |||
1046 | if (flags->proxy == TASK_PROXY1) { | |||
1047 | flags->tiedness = TASK_UNTIED0; | |||
1048 | flags->merged_if0 = 1; | |||
1049 | ||||
1050 | /* are we running in a sequential parallel or tskm_immediate_exec... we need | |||
1051 | tasking support enabled */ | |||
1052 | if ((thread->th.th_task_team) == NULL__null) { | |||
1053 | /* This should only happen if the team is serialized | |||
1054 | setup a task team and propagate it to the thread */ | |||
1055 | KMP_DEBUG_ASSERT(team->t.t_serialized)((team->t.t_serialized) ? 0 : __kmp_debug_assert("team->t.t_serialized" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1055)); | |||
1056 | KA_TRACE(30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d creating task team in __kmp_task_alloc for proxy task\n" , gtid); } | |||
1057 | ("T#%d creating task team in __kmp_task_alloc for proxy task\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d creating task team in __kmp_task_alloc for proxy task\n" , gtid); } | |||
1058 | gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d creating task team in __kmp_task_alloc for proxy task\n" , gtid); }; | |||
1059 | __kmp_task_team_setup( | |||
1060 | thread, team, | |||
1061 | 1); // 1 indicates setup the current team regardless of nthreads | |||
1062 | thread->th.th_task_team = team->t.t_task_team[thread->th.th_task_state]; | |||
1063 | } | |||
1064 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
1065 | ||||
1066 | /* tasking must be enabled now as the task might not be pushed */ | |||
1067 | if (!KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
1068 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
1069 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
1070 | ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n", gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); }; | |||
1071 | __kmp_enable_tasking(task_team, thread); | |||
1072 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
1073 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
1074 | // No lock needed since only owner can allocate | |||
1075 | if (thread_data->td.td_deque == NULL__null) { | |||
1076 | __kmp_alloc_task_deque(thread, thread_data); | |||
1077 | } | |||
1078 | } | |||
1079 | ||||
1080 | if (task_team->tt.tt_found_proxy_tasks == FALSE0) | |||
1081 | TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE)(task_team->tt.tt_found_proxy_tasks) = ((!0)); | |||
1082 | } | |||
1083 | #endif | |||
1084 | ||||
1085 | // Calculate shared structure offset including padding after kmp_task_t struct | |||
1086 | // to align pointers in shared struct | |||
1087 | shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t; | |||
1088 | shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *)); | |||
1089 | ||||
1090 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
1091 | KA_TRACE(30, ("__kmp_task_alloc: T#%d First malloc size: %ld\n", gtid,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d First malloc size: %ld\n" , gtid, shareds_offset); } | |||
1092 | shareds_offset))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d First malloc size: %ld\n" , gtid, shareds_offset); }; | |||
1093 | KA_TRACE(30, ("__kmp_task_alloc: T#%d Second malloc size: %ld\n", gtid,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d Second malloc size: %ld\n" , gtid, sizeof_shareds); } | |||
1094 | sizeof_shareds))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d Second malloc size: %ld\n" , gtid, sizeof_shareds); }; | |||
1095 | ||||
1096 | // Avoid double allocation here by combining shareds with taskdata | |||
1097 | #if USE_FAST_MEMORY3 | |||
1098 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, shareds_offset +___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1099) | |||
1099 | sizeof_shareds)___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1099); | |||
1100 | #else /* ! USE_FAST_MEMORY */ | |||
1101 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, shareds_offset +___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1102) | |||
1102 | sizeof_shareds)___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1102); | |||
1103 | #endif /* USE_FAST_MEMORY */ | |||
1104 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
1105 | ||||
1106 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
1107 | ||||
1108 | // Make sure task & taskdata are aligned appropriately | |||
1109 | #if KMP_ARCH_X860 || KMP_ARCH_PPC64(0 || 0) || !KMP_HAVE_QUAD0 | |||
1110 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0)(((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0 ) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1110)); | |||
1111 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0)(((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1111)); | |||
1112 | #else | |||
1113 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0)(((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1113)); | |||
1114 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0)(((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1114)); | |||
1115 | #endif | |||
1116 | if (sizeof_shareds > 0) { | |||
1117 | // Avoid double allocation here by combining shareds with taskdata | |||
1118 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
1119 | // Make sure shareds struct is aligned to pointer size | |||
1120 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1121)) | |||
1121 | 0)(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1121)); | |||
1122 | } else { | |||
1123 | task->shareds = NULL__null; | |||
1124 | } | |||
1125 | task->routine = task_entry; | |||
1126 | task->part_id = 0; // AC: Always start with 0 part id | |||
1127 | ||||
1128 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
1129 | taskdata->td_team = team; | |||
1130 | taskdata->td_alloc_thread = thread; | |||
1131 | taskdata->td_parent = parent_task; | |||
1132 | taskdata->td_level = parent_task->td_level + 1; // increment nesting level | |||
1133 | KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0)(&taskdata->td_untied_count)->store(0, std::memory_order_relaxed ); | |||
1134 | taskdata->td_ident = loc_ref; | |||
1135 | taskdata->td_taskwait_ident = NULL__null; | |||
1136 | taskdata->td_taskwait_counter = 0; | |||
1137 | taskdata->td_taskwait_thread = 0; | |||
1138 | KMP_DEBUG_ASSERT(taskdata->td_parent != NULL)((taskdata->td_parent != __null) ? 0 : __kmp_debug_assert( "taskdata->td_parent != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1138)); | |||
1139 | #if OMP_45_ENABLED(50 >= 45) | |||
1140 | // avoid copying icvs for proxy tasks | |||
1141 | if (flags->proxy == TASK_FULL0) | |||
1142 | #endif | |||
1143 | copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs); | |||
1144 | ||||
1145 | taskdata->td_flags.tiedness = flags->tiedness; | |||
1146 | taskdata->td_flags.final = flags->final; | |||
1147 | taskdata->td_flags.merged_if0 = flags->merged_if0; | |||
1148 | #if OMP_40_ENABLED(50 >= 40) | |||
1149 | taskdata->td_flags.destructors_thunk = flags->destructors_thunk; | |||
1150 | #endif // OMP_40_ENABLED | |||
1151 | #if OMP_45_ENABLED(50 >= 45) | |||
1152 | taskdata->td_flags.proxy = flags->proxy; | |||
1153 | taskdata->td_task_team = thread->th.th_task_team; | |||
1154 | taskdata->td_size_alloc = shareds_offset + sizeof_shareds; | |||
1155 | #endif | |||
1156 | taskdata->td_flags.tasktype = TASK_EXPLICIT1; | |||
1157 | ||||
1158 | // GEH - TODO: fix this to copy parent task's value of tasking_ser flag | |||
1159 | taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
1160 | ||||
1161 | // GEH - TODO: fix this to copy parent task's value of team_serial flag | |||
1162 | taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
1163 | ||||
1164 | // GEH - Note we serialize the task if the team is serialized to make sure | |||
1165 | // implicit parallel region tasks are not left until program termination to | |||
1166 | // execute. Also, it helps locality to execute immediately. | |||
1167 | ||||
1168 | taskdata->td_flags.task_serial = | |||
1169 | (parent_task->td_flags.final || taskdata->td_flags.team_serial || | |||
1170 | taskdata->td_flags.tasking_ser); | |||
1171 | ||||
1172 | taskdata->td_flags.started = 0; | |||
1173 | taskdata->td_flags.executing = 0; | |||
1174 | taskdata->td_flags.complete = 0; | |||
1175 | taskdata->td_flags.freed = 0; | |||
1176 | ||||
1177 | taskdata->td_flags.native = flags->native; | |||
1178 | ||||
1179 | KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0)(&taskdata->td_incomplete_child_tasks)->store(0, std ::memory_order_relaxed); | |||
1180 | // start at one because counts current task and children | |||
1181 | KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1)(&taskdata->td_allocated_child_tasks)->store(1, std ::memory_order_relaxed); | |||
1182 | #if OMP_40_ENABLED(50 >= 40) | |||
1183 | taskdata->td_taskgroup = | |||
1184 | parent_task->td_taskgroup; // task inherits taskgroup from the parent task | |||
1185 | taskdata->td_dephash = NULL__null; | |||
1186 | taskdata->td_depnode = NULL__null; | |||
1187 | #endif | |||
1188 | if (flags->tiedness == TASK_UNTIED0) | |||
1189 | taskdata->td_last_tied = NULL__null; // will be set when the task is scheduled | |||
1190 | else | |||
1191 | taskdata->td_last_tied = taskdata; | |||
1192 | ||||
1193 | #if OMPT_SUPPORT1 | |||
1194 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
1195 | __ompt_task_init(taskdata, gtid); | |||
1196 | #endif | |||
1197 | // Only need to keep track of child task counts if team parallel and tasking not | |||
1198 | // serialized or if it is a proxy task | |||
1199 | #if OMP_45_ENABLED(50 >= 45) | |||
1200 | if (flags->proxy == TASK_PROXY1 || | |||
1201 | !(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) | |||
1202 | #else | |||
1203 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) | |||
1204 | #endif | |||
1205 | { | |||
1206 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
1207 | #if OMP_40_ENABLED(50 >= 40) | |||
1208 | if (parent_task->td_taskgroup) | |||
1209 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
1210 | #endif | |||
1211 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
1212 | // implicit not deallocated | |||
1213 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) { | |||
1214 | KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks)(&taskdata->td_parent->td_allocated_child_tasks)-> fetch_add(1, std::memory_order_acq_rel); | |||
1215 | } | |||
1216 | } | |||
1217 | ||||
1218 | KA_TRACE(20, ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n" , gtid, taskdata, taskdata->td_parent); } | |||
1219 | gtid, taskdata, taskdata->td_parent))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n" , gtid, taskdata, taskdata->td_parent); }; | |||
1220 | ANNOTATE_HAPPENS_BEFORE(task); | |||
1221 | ||||
1222 | return task; | |||
1223 | } | |||
1224 | ||||
1225 | kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
1226 | kmp_int32 flags, size_t sizeof_kmp_task_t, | |||
1227 | size_t sizeof_shareds, | |||
1228 | kmp_routine_entry_t task_entry) { | |||
1229 | kmp_task_t *retval; | |||
1230 | kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; | |||
1231 | ||||
1232 | input_flags->native = FALSE0; | |||
1233 | // __kmp_task_alloc() sets up all other runtime flags | |||
1234 | ||||
1235 | #if OMP_45_ENABLED(50 >= 45) | |||
1236 | KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
1237 | "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
1238 | gtid, loc_ref, input_flags->tiedness ? "tied " : "untied",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
1239 | input_flags->proxy ? "proxy" : "", sizeof_kmp_task_t,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
1240 | sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); }; | |||
1241 | #else | |||
1242 | KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
1243 | "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
1244 | gtid, loc_ref, input_flags->tiedness ? "tied " : "untied",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
1245 | sizeof_kmp_task_t, sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); }; | |||
1246 | #endif | |||
1247 | ||||
1248 | retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t, | |||
1249 | sizeof_shareds, task_entry); | |||
1250 | ||||
1251 | KA_TRACE(20, ("__kmpc_omp_task_alloc(exit): T#%d retval %p\n", gtid, retval))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_omp_task_alloc(exit): T#%d retval %p\n" , gtid, retval); }; | |||
1252 | ||||
1253 | return retval; | |||
1254 | } | |||
1255 | ||||
1256 | // __kmp_invoke_task: invoke the specified task | |||
1257 | // | |||
1258 | // gtid: global thread ID of caller | |||
1259 | // task: the task to invoke | |||
1260 | // current_task: the task to resume after task invokation | |||
1261 | static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task, | |||
1262 | kmp_taskdata_t *current_task) { | |||
1263 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
1264 | kmp_uint64 cur_time; | |||
1265 | #if OMP_40_ENABLED(50 >= 40) | |||
1266 | int discard = 0 /* false */; | |||
1267 | #endif | |||
1268 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n" , gtid, taskdata, current_task); } | |||
1269 | 30, ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n" , gtid, taskdata, current_task); } | |||
1270 | gtid, taskdata, current_task))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n" , gtid, taskdata, current_task); }; | |||
1271 | KMP_DEBUG_ASSERT(task)((task) ? 0 : __kmp_debug_assert("task", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1271)); | |||
1272 | #if OMP_45_ENABLED(50 >= 45) | |||
1273 | if (taskdata->td_flags.proxy == TASK_PROXY1 && | |||
1274 | taskdata->td_flags.complete == 1) { | |||
1275 | // This is a proxy task that was already completed but it needs to run | |||
1276 | // its bottom-half finish | |||
1277 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); } | |||
1278 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); } | |||
1279 | ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); } | |||
1280 | gtid, taskdata))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); }; | |||
1281 | ||||
1282 | __kmp_bottom_half_finish_proxy(gtid, task); | |||
1283 | ||||
1284 | KA_TRACE(30, ("__kmp_invoke_task(exit): T#%d completed bottom finish for "if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed bottom finish for " "proxy task %p, resuming task %p\n", gtid, taskdata, current_task ); } | |||
1285 | "proxy task %p, resuming task %p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed bottom finish for " "proxy task %p, resuming task %p\n", gtid, taskdata, current_task ); } | |||
1286 | gtid, taskdata, current_task))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed bottom finish for " "proxy task %p, resuming task %p\n", gtid, taskdata, current_task ); }; | |||
1287 | ||||
1288 | return; | |||
1289 | } | |||
1290 | #endif | |||
1291 | ||||
1292 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
1293 | if (__kmp_forkjoin_frames_mode == 3) { | |||
1294 | // Get the current time stamp to measure task execution time to correct | |||
1295 | // barrier imbalance time | |||
1296 | cur_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | |||
1297 | } | |||
1298 | #endif | |||
1299 | ||||
1300 | #if OMPT_SUPPORT1 | |||
1301 | // For untied tasks, the first task executed only calls __kmpc_omp_task and | |||
1302 | // does not execute code. | |||
1303 | ompt_thread_info_t oldInfo; | |||
1304 | kmp_info_t *thread; | |||
1305 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1306 | // Store the threads states and restore them after the task | |||
1307 | thread = __kmp_threads[gtid]; | |||
1308 | oldInfo = thread->th.ompt_thread_info; | |||
1309 | thread->th.ompt_thread_info.wait_id = 0; | |||
1310 | thread->th.ompt_thread_info.state = (thread->th.th_team_serialized) | |||
1311 | ? omp_state_work_serial | |||
1312 | : omp_state_work_parallel; | |||
1313 | taskdata->ompt_task_info.frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
1314 | } | |||
1315 | #endif | |||
1316 | ||||
1317 | #if OMP_45_ENABLED(50 >= 45) | |||
1318 | // Proxy tasks are not handled by the runtime | |||
1319 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
1320 | #endif | |||
1321 | ANNOTATE_HAPPENS_AFTER(task); | |||
1322 | __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded | |||
1323 | #if OMP_45_ENABLED(50 >= 45) | |||
1324 | } | |||
1325 | #endif | |||
1326 | ||||
1327 | #if OMP_40_ENABLED(50 >= 40) | |||
1328 | // TODO: cancel tasks if the parallel region has also been cancelled | |||
1329 | // TODO: check if this sequence can be hoisted above __kmp_task_start | |||
1330 | // if cancellation has been enabled for this run ... | |||
1331 | if (__kmp_omp_cancellation) { | |||
1332 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
1333 | kmp_team_t *this_team = this_thr->th.th_team; | |||
1334 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
1335 | if ((taskgroup && taskgroup->cancel_request) || | |||
1336 | (this_team->t.t_cancel_request == cancel_parallel)) { | |||
1337 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
1338 | ompt_data_t *task_data; | |||
1339 | if (UNLIKELY(ompt_enabled.ompt_callback_cancel)__builtin_expect(!!(ompt_enabled.ompt_callback_cancel), 0)) { | |||
1340 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, NULL__null, NULL__null); | |||
1341 | ompt_callbacks.ompt_callback(ompt_callback_cancel)ompt_callback_cancel_callback( | |||
1342 | task_data, | |||
1343 | ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup | |||
1344 | : ompt_cancel_parallel) | | |||
1345 | ompt_cancel_discarded_task, | |||
1346 | NULL__null); | |||
1347 | } | |||
1348 | #endif | |||
1349 | KMP_COUNT_BLOCK(TASK_cancelled)((void)0); | |||
1350 | // this task belongs to a task group and we need to cancel it | |||
1351 | discard = 1 /* true */; | |||
1352 | } | |||
1353 | } | |||
1354 | ||||
1355 | // Invoke the task routine and pass in relevant data. | |||
1356 | // Thunks generated by gcc take a different argument list. | |||
1357 | if (!discard) { | |||
1358 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
1359 | taskdata->td_last_tied = current_task->td_last_tied; | |||
1360 | KMP_DEBUG_ASSERT(taskdata->td_last_tied)((taskdata->td_last_tied) ? 0 : __kmp_debug_assert("taskdata->td_last_tied" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1360)); | |||
1361 | } | |||
1362 | #if KMP_STATS_ENABLED0 | |||
1363 | KMP_COUNT_BLOCK(TASK_executed)((void)0); | |||
1364 | switch (KMP_GET_THREAD_STATE()((void)0)) { | |||
1365 | case FORK_JOIN_BARRIER: | |||
1366 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar)((void)0); | |||
1367 | break; | |||
1368 | case PLAIN_BARRIER: | |||
1369 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar)((void)0); | |||
1370 | break; | |||
1371 | case TASKYIELD: | |||
1372 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield)((void)0); | |||
1373 | break; | |||
1374 | case TASKWAIT: | |||
1375 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait)((void)0); | |||
1376 | break; | |||
1377 | case TASKGROUP: | |||
1378 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup)((void)0); | |||
1379 | break; | |||
1380 | default: | |||
1381 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate)((void)0); | |||
1382 | break; | |||
1383 | } | |||
1384 | #endif // KMP_STATS_ENABLED | |||
1385 | #endif // OMP_40_ENABLED | |||
1386 | ||||
1387 | // OMPT task begin | |||
1388 | #if OMPT_SUPPORT1 | |||
1389 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
1390 | __ompt_task_start(task, current_task, gtid); | |||
1391 | #endif | |||
1392 | ||||
1393 | #ifdef KMP_GOMP_COMPAT | |||
1394 | if (taskdata->td_flags.native) { | |||
1395 | ((void (*)(void *))(*(task->routine)))(task->shareds); | |||
1396 | } else | |||
1397 | #endif /* KMP_GOMP_COMPAT */ | |||
1398 | { | |||
1399 | (*(task->routine))(gtid, task); | |||
1400 | } | |||
1401 | KMP_POP_PARTITIONED_TIMER()((void)0); | |||
1402 | ||||
1403 | #if OMP_40_ENABLED(50 >= 40) | |||
1404 | } | |||
1405 | #endif // OMP_40_ENABLED | |||
1406 | ||||
1407 | ||||
1408 | #if OMP_45_ENABLED(50 >= 45) | |||
1409 | // Proxy tasks are not handled by the runtime | |||
1410 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
1411 | #endif | |||
1412 | ANNOTATE_HAPPENS_BEFORE(taskdata->td_parent); | |||
1413 | #if OMPT_SUPPORT1 | |||
1414 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1415 | thread->th.ompt_thread_info = oldInfo; | |||
1416 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
1417 | taskdata->ompt_task_info.frame.exit_frame = NULL__null; | |||
1418 | } | |||
1419 | __kmp_task_finish<true>(gtid, task, current_task); | |||
1420 | } else | |||
1421 | #endif | |||
1422 | __kmp_task_finish<false>(gtid, task, current_task); | |||
1423 | #if OMP_45_ENABLED(50 >= 45) | |||
1424 | } | |||
1425 | #endif | |||
1426 | ||||
1427 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
1428 | // Barrier imbalance - correct arrive time after the task finished | |||
1429 | if (__kmp_forkjoin_frames_mode == 3) { | |||
1430 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
1431 | if (this_thr->th.th_bar_arrive_time) { | |||
1432 | this_thr->th.th_bar_arrive_time += (__itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0() - cur_time); | |||
1433 | } | |||
1434 | } | |||
1435 | #endif | |||
1436 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n" , gtid, taskdata, current_task); } | |||
1437 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n" , gtid, taskdata, current_task); } | |||
1438 | ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n" , gtid, taskdata, current_task); } | |||
1439 | gtid, taskdata, current_task))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n" , gtid, taskdata, current_task); }; | |||
1440 | return; | |||
1441 | } | |||
1442 | ||||
1443 | // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution | |||
1444 | // | |||
1445 | // loc_ref: location of original task pragma (ignored) | |||
1446 | // gtid: Global Thread ID of encountering thread | |||
1447 | // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task'' | |||
1448 | // Returns: | |||
1449 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
1450 | // be resumed later. | |||
1451 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
1452 | // resumed later. | |||
1453 | kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, | |||
1454 | kmp_task_t *new_task) { | |||
1455 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
1456 | ||||
1457 | KA_TRACE(10, ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n", gtid,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); } | |||
1458 | loc_ref, new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); }; | |||
1459 | ||||
1460 | #if OMPT_SUPPORT1 | |||
1461 | kmp_taskdata_t *parent; | |||
1462 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1463 | parent = new_taskdata->td_parent; | |||
1464 | if (ompt_enabled.ompt_callback_task_create) { | |||
1465 | ompt_data_t task_data = ompt_data_none; | |||
1466 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
1467 | parent ? &(parent->ompt_task_info.task_data) : &task_data, | |||
1468 | parent ? &(parent->ompt_task_info.frame) : NULL__null, | |||
1469 | &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0, | |||
1470 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
1471 | } | |||
1472 | } | |||
1473 | #endif | |||
1474 | ||||
1475 | /* Should we execute the new task or queue it? For now, let's just always try | |||
1476 | to queue it. If the queue fills up, then we'll execute it. */ | |||
1477 | ||||
1478 | if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
1479 | { // Execute this task immediately | |||
1480 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
1481 | new_taskdata->td_flags.task_serial = 1; | |||
1482 | __kmp_invoke_task(gtid, new_task, current_task); | |||
1483 | } | |||
1484 | ||||
1485 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: " "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n", gtid, loc_ref , new_taskdata); } | |||
1486 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: " "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n", gtid, loc_ref , new_taskdata); } | |||
1487 | ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: " "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n", gtid, loc_ref , new_taskdata); } | |||
1488 | "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: " "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n", gtid, loc_ref , new_taskdata); } | |||
1489 | gtid, loc_ref, new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: " "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n", gtid, loc_ref , new_taskdata); }; | |||
1490 | ||||
1491 | ANNOTATE_HAPPENS_BEFORE(new_task); | |||
1492 | #if OMPT_SUPPORT1 | |||
1493 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1494 | parent->ompt_task_info.frame.enter_frame = NULL__null; | |||
1495 | } | |||
1496 | #endif | |||
1497 | return TASK_CURRENT_NOT_QUEUED0; | |||
1498 | } | |||
1499 | ||||
1500 | // __kmp_omp_task: Schedule a non-thread-switchable task for execution | |||
1501 | // | |||
1502 | // gtid: Global Thread ID of encountering thread | |||
1503 | // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc() | |||
1504 | // serialize_immediate: if TRUE then if the task is executed immediately its | |||
1505 | // execution will be serialized | |||
1506 | // Returns: | |||
1507 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
1508 | // be resumed later. | |||
1509 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
1510 | // resumed later. | |||
1511 | kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, | |||
1512 | bool serialize_immediate) { | |||
1513 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
1514 | ||||
1515 | /* Should we execute the new task or queue it? For now, let's just always try to | |||
1516 | queue it. If the queue fills up, then we'll execute it. */ | |||
1517 | #if OMP_45_ENABLED(50 >= 45) | |||
1518 | if (new_taskdata->td_flags.proxy == TASK_PROXY1 || | |||
1519 | __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
1520 | #else | |||
1521 | if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
1522 | #endif | |||
1523 | { // Execute this task immediately | |||
1524 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
1525 | if (serialize_immediate) | |||
1526 | new_taskdata->td_flags.task_serial = 1; | |||
1527 | __kmp_invoke_task(gtid, new_task, current_task); | |||
1528 | } | |||
1529 | ||||
1530 | ANNOTATE_HAPPENS_BEFORE(new_task); | |||
1531 | return TASK_CURRENT_NOT_QUEUED0; | |||
1532 | } | |||
1533 | ||||
1534 | // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a | |||
1535 | // non-thread-switchable task from the parent thread only! | |||
1536 | // | |||
1537 | // loc_ref: location of original task pragma (ignored) | |||
1538 | // gtid: Global Thread ID of encountering thread | |||
1539 | // new_task: non-thread-switchable task thunk allocated by | |||
1540 | // __kmp_omp_task_alloc() | |||
1541 | // Returns: | |||
1542 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
1543 | // be resumed later. | |||
1544 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
1545 | // resumed later. | |||
1546 | kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, | |||
1547 | kmp_task_t *new_task) { | |||
1548 | kmp_int32 res; | |||
1549 | KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK)((void)0); | |||
1550 | ||||
1551 | #if KMP_DEBUG1 || OMPT_SUPPORT1 | |||
1552 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
1553 | #endif | |||
1554 | KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); } | |||
1555 | new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); }; | |||
1556 | ||||
1557 | #if OMPT_SUPPORT1 | |||
1558 | kmp_taskdata_t *parent = NULL__null; | |||
1559 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1560 | if (!new_taskdata->td_flags.started) { | |||
1561 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
1562 | parent = new_taskdata->td_parent; | |||
1563 | if (!parent->ompt_task_info.frame.enter_frame) { | |||
1564 | parent->ompt_task_info.frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); | |||
1565 | } | |||
1566 | if (ompt_enabled.ompt_callback_task_create) { | |||
1567 | ompt_data_t task_data = ompt_data_none; | |||
1568 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
1569 | parent ? &(parent->ompt_task_info.task_data) : &task_data, | |||
1570 | parent ? &(parent->ompt_task_info.frame) : NULL__null, | |||
1571 | &(new_taskdata->ompt_task_info.task_data), | |||
1572 | ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata)((new_taskdata->td_flags.task_serial || new_taskdata->td_flags .tasking_ser) ? ompt_task_undeferred : 0x0) | ((!(new_taskdata ->td_flags.tiedness)) ? ompt_task_untied : 0x0) | (new_taskdata ->td_flags.final ? ompt_task_final : 0x0) | (new_taskdata-> td_flags.merged_if0 ? ompt_task_mergeable : 0x0), 0, | |||
1573 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
1574 | } | |||
1575 | } else { | |||
1576 | // We are scheduling the continuation of an UNTIED task. | |||
1577 | // Scheduling back to the parent task. | |||
1578 | __ompt_task_finish(new_task, | |||
1579 | new_taskdata->ompt_task_info.scheduling_parent, | |||
1580 | ompt_task_others); | |||
1581 | new_taskdata->ompt_task_info.frame.exit_frame = NULL__null; | |||
1582 | } | |||
1583 | } | |||
1584 | #endif | |||
1585 | ||||
1586 | res = __kmp_omp_task(gtid, new_task, true); | |||
1587 | ||||
1588 | KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); } | |||
1589 | "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); } | |||
1590 | gtid, loc_ref, new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); }; | |||
1591 | #if OMPT_SUPPORT1 | |||
1592 | if (UNLIKELY(ompt_enabled.enabled && parent != NULL)__builtin_expect(!!(ompt_enabled.enabled && parent != __null), 0)) { | |||
1593 | parent->ompt_task_info.frame.enter_frame = NULL__null; | |||
1594 | } | |||
1595 | #endif | |||
1596 | return res; | |||
1597 | } | |||
1598 | ||||
1599 | // __kmp_omp_taskloop_task: Wrapper around __kmp_omp_task to schedule | |||
1600 | // a taskloop task with the correct OMPT return address | |||
1601 | // | |||
1602 | // loc_ref: location of original task pragma (ignored) | |||
1603 | // gtid: Global Thread ID of encountering thread | |||
1604 | // new_task: non-thread-switchable task thunk allocated by | |||
1605 | // __kmp_omp_task_alloc() | |||
1606 | // codeptr_ra: return address for OMPT callback | |||
1607 | // Returns: | |||
1608 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
1609 | // be resumed later. | |||
1610 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
1611 | // resumed later. | |||
1612 | kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid, | |||
1613 | kmp_task_t *new_task, void *codeptr_ra) { | |||
1614 | kmp_int32 res; | |||
1615 | KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK)((void)0); | |||
1616 | ||||
1617 | #if KMP_DEBUG1 || OMPT_SUPPORT1 | |||
1618 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
1619 | #endif | |||
1620 | KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); } | |||
1621 | new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n" , gtid, loc_ref, new_taskdata); }; | |||
1622 | ||||
1623 | #if OMPT_SUPPORT1 | |||
1624 | kmp_taskdata_t *parent = NULL__null; | |||
1625 | if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)__builtin_expect(!!(ompt_enabled.enabled && !new_taskdata ->td_flags.started), 0)) { | |||
1626 | parent = new_taskdata->td_parent; | |||
1627 | if (!parent->ompt_task_info.frame.enter_frame) | |||
1628 | parent->ompt_task_info.frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); | |||
1629 | if (ompt_enabled.ompt_callback_task_create) { | |||
1630 | ompt_data_t task_data = ompt_data_none; | |||
1631 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
1632 | parent ? &(parent->ompt_task_info.task_data) : &task_data, | |||
1633 | parent ? &(parent->ompt_task_info.frame) : NULL__null, | |||
1634 | &(new_taskdata->ompt_task_info.task_data), | |||
1635 | ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata)((new_taskdata->td_flags.task_serial || new_taskdata->td_flags .tasking_ser) ? ompt_task_undeferred : 0x0) | ((!(new_taskdata ->td_flags.tiedness)) ? ompt_task_untied : 0x0) | (new_taskdata ->td_flags.final ? ompt_task_final : 0x0) | (new_taskdata-> td_flags.merged_if0 ? ompt_task_mergeable : 0x0), 0, | |||
1636 | codeptr_ra); | |||
1637 | } | |||
1638 | } | |||
1639 | #endif | |||
1640 | ||||
1641 | res = __kmp_omp_task(gtid, new_task, true); | |||
1642 | ||||
1643 | KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); } | |||
1644 | "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); } | |||
1645 | gtid, loc_ref, new_taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task(exit): T#%d returning " "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n", gtid, loc_ref, new_taskdata ); }; | |||
1646 | #if OMPT_SUPPORT1 | |||
1647 | if (UNLIKELY(ompt_enabled.enabled && parent != NULL)__builtin_expect(!!(ompt_enabled.enabled && parent != __null), 0)) { | |||
1648 | parent->ompt_task_info.frame.enter_frame = NULL__null; | |||
1649 | } | |||
1650 | #endif | |||
1651 | return res; | |||
1652 | } | |||
1653 | ||||
1654 | template <bool ompt> | |||
1655 | static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid, | |||
1656 | void *frame_address, | |||
1657 | void *return_address) { | |||
1658 | kmp_taskdata_t *taskdata; | |||
1659 | kmp_info_t *thread; | |||
1660 | int thread_finished = FALSE0; | |||
1661 | KMP_SET_THREAD_STATE_BLOCK(TASKWAIT)((void)0); | |||
1662 | ||||
1663 | KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n" , gtid, loc_ref); }; | |||
1664 | ||||
1665 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
1666 | thread = __kmp_threads[gtid]; | |||
1667 | taskdata = thread->th.th_current_task; | |||
1668 | ||||
1669 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
1670 | ompt_data_t *my_task_data; | |||
1671 | ompt_data_t *my_parallel_data; | |||
1672 | ||||
1673 | if (ompt) { | |||
1674 | my_task_data = &(taskdata->ompt_task_info.task_data); | |||
1675 | my_parallel_data = OMPT_CUR_TEAM_DATA(thread)(&(thread->th.th_team->t.ompt_team_info.parallel_data )); | |||
1676 | ||||
1677 | taskdata->ompt_task_info.frame.enter_frame = frame_address; | |||
1678 | ||||
1679 | if (ompt_enabled.ompt_callback_sync_region) { | |||
1680 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
1681 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
1682 | my_task_data, return_address); | |||
1683 | } | |||
1684 | ||||
1685 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
1686 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
1687 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
1688 | my_task_data, return_address); | |||
1689 | } | |||
1690 | } | |||
1691 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
1692 | ||||
1693 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
1694 | // taskwait. | |||
1695 | #if USE_ITT_BUILD1 | |||
1696 | // Note: These values are used by ITT events as well. | |||
1697 | #endif /* USE_ITT_BUILD */ | |||
1698 | taskdata->td_taskwait_counter += 1; | |||
1699 | taskdata->td_taskwait_ident = loc_ref; | |||
1700 | taskdata->td_taskwait_thread = gtid + 1; | |||
1701 | ||||
1702 | #if USE_ITT_BUILD1 | |||
1703 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
1704 | if (itt_sync_obj != NULL__null) | |||
1705 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
1706 | #endif /* USE_ITT_BUILD */ | |||
1707 | ||||
1708 | bool must_wait = | |||
1709 | !taskdata->td_flags.team_serial && !taskdata->td_flags.final; | |||
1710 | ||||
1711 | #if OMP_45_ENABLED(50 >= 45) | |||
1712 | must_wait = must_wait || (thread->th.th_task_team != NULL__null && | |||
1713 | thread->th.th_task_team->tt.tt_found_proxy_tasks); | |||
1714 | #endif | |||
1715 | if (must_wait) { | |||
1716 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)) | |||
1717 | &(taskdata->td_incomplete_child_tasks))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)), | |||
1718 | 0U); | |||
1719 | while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) != 0) { | |||
1720 | flag.execute_tasks(thread, gtid, FALSE0, | |||
1721 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
1722 | __kmp_task_stealing_constraint); | |||
1723 | } | |||
1724 | } | |||
1725 | #if USE_ITT_BUILD1 | |||
1726 | if (itt_sync_obj != NULL__null) | |||
1727 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
1728 | #endif /* USE_ITT_BUILD */ | |||
1729 | ||||
1730 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
1731 | // negated. | |||
1732 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
1733 | ||||
1734 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
1735 | if (ompt) { | |||
1736 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
1737 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
1738 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
1739 | my_task_data, return_address); | |||
1740 | } | |||
1741 | if (ompt_enabled.ompt_callback_sync_region) { | |||
1742 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
1743 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
1744 | my_task_data, return_address); | |||
1745 | } | |||
1746 | taskdata->ompt_task_info.frame.enter_frame = NULL__null; | |||
1747 | } | |||
1748 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
1749 | ||||
1750 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
1751 | } | |||
1752 | ||||
1753 | KA_TRACE(10, ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); } | |||
1754 | "returning TASK_CURRENT_NOT_QUEUED\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); } | |||
1755 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); }; | |||
1756 | ||||
1757 | return TASK_CURRENT_NOT_QUEUED0; | |||
1758 | } | |||
1759 | ||||
1760 | #if OMPT_SUPPORT1 | |||
1761 | OMPT_NOINLINE__attribute__((noinline)) | |||
1762 | static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
1763 | void *frame_address, | |||
1764 | void *return_address) { | |||
1765 | return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address, | |||
1766 | return_address); | |||
1767 | } | |||
1768 | #endif // OMPT_SUPPORT | |||
1769 | ||||
1770 | // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are | |||
1771 | // complete | |||
1772 | kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) { | |||
1773 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
1774 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
1775 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
1776 | return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1), | |||
1777 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
1778 | } | |||
1779 | #endif | |||
1780 | return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL__null, NULL__null); | |||
1781 | } | |||
1782 | ||||
1783 | // __kmpc_omp_taskyield: switch to a different task | |||
1784 | kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) { | |||
1785 | kmp_taskdata_t *taskdata; | |||
1786 | kmp_info_t *thread; | |||
1787 | int thread_finished = FALSE0; | |||
1788 | ||||
1789 | KMP_COUNT_BLOCK(OMP_TASKYIELD)((void)0); | |||
1790 | KMP_SET_THREAD_STATE_BLOCK(TASKYIELD)((void)0); | |||
1791 | ||||
1792 | KA_TRACE(10, ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n" , gtid, loc_ref, end_part); } | |||
1793 | gtid, loc_ref, end_part))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n" , gtid, loc_ref, end_part); }; | |||
1794 | ||||
1795 | if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) { | |||
1796 | thread = __kmp_threads[gtid]; | |||
1797 | taskdata = thread->th.th_current_task; | |||
1798 | // Should we model this as a task wait or not? | |||
1799 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
1800 | // taskwait. | |||
1801 | #if USE_ITT_BUILD1 | |||
1802 | // Note: These values are used by ITT events as well. | |||
1803 | #endif /* USE_ITT_BUILD */ | |||
1804 | taskdata->td_taskwait_counter += 1; | |||
1805 | taskdata->td_taskwait_ident = loc_ref; | |||
1806 | taskdata->td_taskwait_thread = gtid + 1; | |||
1807 | ||||
1808 | #if USE_ITT_BUILD1 | |||
1809 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
1810 | if (itt_sync_obj != NULL__null) | |||
1811 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
1812 | #endif /* USE_ITT_BUILD */ | |||
1813 | if (!taskdata->td_flags.team_serial) { | |||
1814 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
1815 | if (task_team != NULL__null) { | |||
1816 | if (KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
1817 | #if OMPT_SUPPORT1 | |||
1818 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
1819 | thread->th.ompt_thread_info.ompt_task_yielded = 1; | |||
1820 | #endif | |||
1821 | __kmp_execute_tasks_32( | |||
1822 | thread, gtid, NULL__null, FALSE0, | |||
1823 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
1824 | __kmp_task_stealing_constraint); | |||
1825 | #if OMPT_SUPPORT1 | |||
1826 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
1827 | thread->th.ompt_thread_info.ompt_task_yielded = 0; | |||
1828 | #endif | |||
1829 | } | |||
1830 | } | |||
1831 | } | |||
1832 | #if USE_ITT_BUILD1 | |||
1833 | if (itt_sync_obj != NULL__null) | |||
1834 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
1835 | #endif /* USE_ITT_BUILD */ | |||
1836 | ||||
1837 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
1838 | // negated. | |||
1839 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
1840 | } | |||
1841 | ||||
1842 | KA_TRACE(10, ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); } | |||
1843 | "returning TASK_CURRENT_NOT_QUEUED\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); } | |||
1844 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, " "returning TASK_CURRENT_NOT_QUEUED\n", gtid, taskdata); }; | |||
1845 | ||||
1846 | return TASK_CURRENT_NOT_QUEUED0; | |||
1847 | } | |||
1848 | ||||
1849 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
1850 | #if OMP_45_ENABLED(50 >= 45) | |||
1851 | // Task Reduction implementation | |||
1852 | ||||
1853 | typedef struct kmp_task_red_flags { | |||
1854 | unsigned lazy_priv : 1; // hint: (1) use lazy allocation (big objects) | |||
1855 | unsigned reserved31 : 31; | |||
1856 | } kmp_task_red_flags_t; | |||
1857 | ||||
1858 | // internal structure for reduction data item related info | |||
1859 | typedef struct kmp_task_red_data { | |||
1860 | void *reduce_shar; // shared reduction item | |||
1861 | size_t reduce_size; // size of data item | |||
1862 | void *reduce_priv; // thread specific data | |||
1863 | void *reduce_pend; // end of private data for comparison op | |||
1864 | void *reduce_init; // data initialization routine | |||
1865 | void *reduce_fini; // data finalization routine | |||
1866 | void *reduce_comb; // data combiner routine | |||
1867 | kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
1868 | } kmp_task_red_data_t; | |||
1869 | ||||
1870 | // structure sent us by compiler - one per reduction item | |||
1871 | typedef struct kmp_task_red_input { | |||
1872 | void *reduce_shar; // shared reduction item | |||
1873 | size_t reduce_size; // size of data item | |||
1874 | void *reduce_init; // data initialization routine | |||
1875 | void *reduce_fini; // data finalization routine | |||
1876 | void *reduce_comb; // data combiner routine | |||
1877 | kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
1878 | } kmp_task_red_input_t; | |||
1879 | ||||
1880 | /*! | |||
1881 | @ingroup TASKING | |||
1882 | @param gtid Global thread ID | |||
1883 | @param num Number of data items to reduce | |||
1884 | @param data Array of data for reduction | |||
1885 | @return The taskgroup identifier | |||
1886 | ||||
1887 | Initialize task reduction for the taskgroup. | |||
1888 | */ | |||
1889 | void *__kmpc_task_reduction_init(int gtid, int num, void *data) { | |||
1890 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
1891 | kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup; | |||
1892 | kmp_int32 nth = thread->th.th_team_nproc; | |||
1893 | kmp_task_red_input_t *input = (kmp_task_red_input_t *)data; | |||
1894 | kmp_task_red_data_t *arr; | |||
1895 | ||||
1896 | // check input data just in case | |||
1897 | KMP_ASSERT(tg != NULL)((tg != __null) ? 0 : __kmp_debug_assert("tg != NULL", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1897)); | |||
1898 | KMP_ASSERT(data != NULL)((data != __null) ? 0 : __kmp_debug_assert("data != NULL", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1898)); | |||
1899 | KMP_ASSERT(num > 0)((num > 0) ? 0 : __kmp_debug_assert("num > 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1899)); | |||
1900 | if (nth == 1) { | |||
1901 | KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n" , gtid, tg); } | |||
1902 | gtid, tg))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n" , gtid, tg); }; | |||
1903 | return (void *)tg; | |||
1904 | } | |||
1905 | KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n" , gtid, tg, num); } | |||
1906 | gtid, tg, num))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n" , gtid, tg, num); }; | |||
1907 | arr = (kmp_task_red_data_t *)__kmp_thread_malloc(___kmp_thread_malloc((thread), (num * sizeof(kmp_task_red_data_t )), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1908) | |||
1908 | thread, num * sizeof(kmp_task_red_data_t))___kmp_thread_malloc((thread), (num * sizeof(kmp_task_red_data_t )), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1908); | |||
1909 | for (int i = 0; i < num; ++i) { | |||
1910 | void (*f_init)(void *) = (void (*)(void *))(input[i].reduce_init); | |||
1911 | size_t size = input[i].reduce_size - 1; | |||
1912 | // round the size up to cache line per thread-specific item | |||
1913 | size += CACHE_LINE64 - size % CACHE_LINE64; | |||
1914 | KMP_ASSERT(input[i].reduce_comb != NULL)((input[i].reduce_comb != __null) ? 0 : __kmp_debug_assert("input[i].reduce_comb != NULL" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1914)); // combiner is mandatory | |||
1915 | arr[i].reduce_shar = input[i].reduce_shar; | |||
1916 | arr[i].reduce_size = size; | |||
1917 | arr[i].reduce_init = input[i].reduce_init; | |||
1918 | arr[i].reduce_fini = input[i].reduce_fini; | |||
1919 | arr[i].reduce_comb = input[i].reduce_comb; | |||
1920 | arr[i].flags = input[i].flags; | |||
1921 | if (!input[i].flags.lazy_priv) { | |||
1922 | // allocate cache-line aligned block and fill it with zeros | |||
1923 | arr[i].reduce_priv = __kmp_allocate(nth * size)___kmp_allocate((nth * size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1923); | |||
1924 | arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size; | |||
1925 | if (f_init != NULL__null) { | |||
1926 | // initialize thread-specific items | |||
1927 | for (int j = 0; j < nth; ++j) { | |||
1928 | f_init((char *)(arr[i].reduce_priv) + j * size); | |||
1929 | } | |||
1930 | } | |||
1931 | } else { | |||
1932 | // only allocate space for pointers now, | |||
1933 | // objects will be lazily allocated/initialized once requested | |||
1934 | arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *))___kmp_allocate((nth * sizeof(void *)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1934); | |||
1935 | } | |||
1936 | } | |||
1937 | tg->reduce_data = (void *)arr; | |||
1938 | tg->reduce_num_data = num; | |||
1939 | return (void *)tg; | |||
1940 | } | |||
1941 | ||||
1942 | /*! | |||
1943 | @ingroup TASKING | |||
1944 | @param gtid Global thread ID | |||
1945 | @param tskgrp The taskgroup ID (optional) | |||
1946 | @param data Shared location of the item | |||
1947 | @return The pointer to per-thread data | |||
1948 | ||||
1949 | Get thread-specific location of data item | |||
1950 | */ | |||
1951 | void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) { | |||
1952 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
1953 | kmp_int32 nth = thread->th.th_team_nproc; | |||
1954 | if (nth == 1) | |||
1955 | return data; // nothing to do | |||
1956 | ||||
1957 | kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp; | |||
1958 | if (tg == NULL__null) | |||
1959 | tg = thread->th.th_current_task->td_taskgroup; | |||
1960 | KMP_ASSERT(tg != NULL)((tg != __null) ? 0 : __kmp_debug_assert("tg != NULL", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1960)); | |||
1961 | kmp_task_red_data_t *arr = (kmp_task_red_data_t *)(tg->reduce_data); | |||
1962 | kmp_int32 num = tg->reduce_num_data; | |||
1963 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
1964 | ||||
1965 | KMP_ASSERT(data != NULL)((data != __null) ? 0 : __kmp_debug_assert("data != NULL", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1965)); | |||
1966 | while (tg != NULL__null) { | |||
1967 | for (int i = 0; i < num; ++i) { | |||
1968 | if (!arr[i].flags.lazy_priv) { | |||
1969 | if (data == arr[i].reduce_shar || | |||
1970 | (data >= arr[i].reduce_priv && data < arr[i].reduce_pend)) | |||
1971 | return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size; | |||
1972 | } else { | |||
1973 | // check shared location first | |||
1974 | void **p_priv = (void **)(arr[i].reduce_priv); | |||
1975 | if (data == arr[i].reduce_shar) | |||
1976 | goto found; | |||
1977 | // check if we get some thread specific location as parameter | |||
1978 | for (int j = 0; j < nth; ++j) | |||
1979 | if (data == p_priv[j]) | |||
1980 | goto found; | |||
1981 | continue; // not found, continue search | |||
1982 | found: | |||
1983 | if (p_priv[tid] == NULL__null) { | |||
1984 | // allocate thread specific object lazily | |||
1985 | void (*f_init)(void *) = (void (*)(void *))(arr[i].reduce_init); | |||
1986 | p_priv[tid] = __kmp_allocate(arr[i].reduce_size)___kmp_allocate((arr[i].reduce_size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1986); | |||
1987 | if (f_init != NULL__null) { | |||
1988 | f_init(p_priv[tid]); | |||
1989 | } | |||
1990 | } | |||
1991 | return p_priv[tid]; | |||
1992 | } | |||
1993 | } | |||
1994 | tg = tg->parent; | |||
1995 | arr = (kmp_task_red_data_t *)(tg->reduce_data); | |||
1996 | num = tg->reduce_num_data; | |||
1997 | } | |||
1998 | KMP_ASSERT2(0, "Unknown task reduction item")((0) ? 0 : __kmp_debug_assert(("Unknown task reduction item") , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 1998)); | |||
1999 | return NULL__null; // ERROR, this line never executed | |||
2000 | } | |||
2001 | ||||
2002 | // Finalize task reduction. | |||
2003 | // Called from __kmpc_end_taskgroup() | |||
2004 | static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) { | |||
2005 | kmp_int32 nth = th->th.th_team_nproc; | |||
2006 | KMP_DEBUG_ASSERT(nth > 1)((nth > 1) ? 0 : __kmp_debug_assert("nth > 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2006)); // should not be called if nth == 1 | |||
2007 | kmp_task_red_data_t *arr = (kmp_task_red_data_t *)tg->reduce_data; | |||
2008 | kmp_int32 num = tg->reduce_num_data; | |||
2009 | for (int i = 0; i < num; ++i) { | |||
2010 | void *sh_data = arr[i].reduce_shar; | |||
2011 | void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini); | |||
2012 | void (*f_comb)(void *, void *) = | |||
2013 | (void (*)(void *, void *))(arr[i].reduce_comb); | |||
2014 | if (!arr[i].flags.lazy_priv) { | |||
2015 | void *pr_data = arr[i].reduce_priv; | |||
2016 | size_t size = arr[i].reduce_size; | |||
2017 | for (int j = 0; j < nth; ++j) { | |||
2018 | void *priv_data = (char *)pr_data + j * size; | |||
2019 | f_comb(sh_data, priv_data); // combine results | |||
2020 | if (f_fini) | |||
2021 | f_fini(priv_data); // finalize if needed | |||
2022 | } | |||
2023 | } else { | |||
2024 | void **pr_data = (void **)(arr[i].reduce_priv); | |||
2025 | for (int j = 0; j < nth; ++j) { | |||
2026 | if (pr_data[j] != NULL__null) { | |||
2027 | f_comb(sh_data, pr_data[j]); // combine results | |||
2028 | if (f_fini) | |||
2029 | f_fini(pr_data[j]); // finalize if needed | |||
2030 | __kmp_free(pr_data[j])___kmp_free((pr_data[j]), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2030); | |||
2031 | } | |||
2032 | } | |||
2033 | } | |||
2034 | __kmp_free(arr[i].reduce_priv)___kmp_free((arr[i].reduce_priv), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2034); | |||
2035 | } | |||
2036 | __kmp_thread_free(th, arr)___kmp_thread_free((th), (arr), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2036); | |||
2037 | tg->reduce_data = NULL__null; | |||
2038 | tg->reduce_num_data = 0; | |||
2039 | } | |||
2040 | #endif | |||
2041 | ||||
2042 | #if OMP_40_ENABLED(50 >= 40) | |||
2043 | // __kmpc_taskgroup: Start a new taskgroup | |||
2044 | void __kmpc_taskgroup(ident_t *loc, int gtid) { | |||
2045 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
2046 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
2047 | kmp_taskgroup_t *tg_new = | |||
2048 | (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t))___kmp_thread_malloc((thread), (sizeof(kmp_taskgroup_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2048); | |||
2049 | KA_TRACE(10, ("__kmpc_taskgroup: T#%d loc=%p group=%p\n", gtid, loc, tg_new))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_taskgroup: T#%d loc=%p group=%p\n" , gtid, loc, tg_new); }; | |||
2050 | KMP_ATOMIC_ST_RLX(&tg_new->count, 0)(&tg_new->count)->store(0, std::memory_order_relaxed ); | |||
2051 | KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq)(&tg_new->cancel_request)->store(cancel_noreq, std:: memory_order_relaxed); | |||
2052 | tg_new->parent = taskdata->td_taskgroup; | |||
2053 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
2054 | #if OMP_45_ENABLED(50 >= 45) | |||
2055 | tg_new->reduce_data = NULL__null; | |||
2056 | tg_new->reduce_num_data = 0; | |||
2057 | #endif | |||
2058 | taskdata->td_taskgroup = tg_new; | |||
2059 | ||||
2060 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
2061 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
2062 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
2063 | if (!codeptr) | |||
2064 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
2065 | kmp_team_t *team = thread->th.th_team; | |||
2066 | ompt_data_t my_task_data = taskdata->ompt_task_info.task_data; | |||
2067 | // FIXME: I think this is wrong for lwt! | |||
2068 | ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
2069 | ||||
2070 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
2071 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
2072 | &(my_task_data), codeptr); | |||
2073 | } | |||
2074 | #endif | |||
2075 | } | |||
2076 | ||||
2077 | // __kmpc_end_taskgroup: Wait until all tasks generated by the current task | |||
2078 | // and its descendants are complete | |||
2079 | void __kmpc_end_taskgroup(ident_t *loc, int gtid) { | |||
2080 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
2081 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
2082 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
2083 | int thread_finished = FALSE0; | |||
2084 | ||||
2085 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
2086 | kmp_team_t *team; | |||
2087 | ompt_data_t my_task_data; | |||
2088 | ompt_data_t my_parallel_data; | |||
2089 | void *codeptr; | |||
2090 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
2091 | team = thread->th.th_team; | |||
2092 | my_task_data = taskdata->ompt_task_info.task_data; | |||
2093 | // FIXME: I think this is wrong for lwt! | |||
2094 | my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
2095 | codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
2096 | if (!codeptr) | |||
2097 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
2098 | } | |||
2099 | #endif | |||
2100 | ||||
2101 | KA_TRACE(10, ("__kmpc_end_taskgroup(enter): T#%d loc=%p\n", gtid, loc))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_end_taskgroup(enter): T#%d loc=%p\n" , gtid, loc); }; | |||
2102 | KMP_DEBUG_ASSERT(taskgroup != NULL)((taskgroup != __null) ? 0 : __kmp_debug_assert("taskgroup != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2102)); | |||
2103 | KMP_SET_THREAD_STATE_BLOCK(TASKGROUP)((void)0); | |||
2104 | ||||
2105 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
2106 | // mark task as waiting not on a barrier | |||
2107 | taskdata->td_taskwait_counter += 1; | |||
2108 | taskdata->td_taskwait_ident = loc; | |||
2109 | taskdata->td_taskwait_thread = gtid + 1; | |||
2110 | #if USE_ITT_BUILD1 | |||
2111 | // For ITT the taskgroup wait is similar to taskwait until we need to | |||
2112 | // distinguish them | |||
2113 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
2114 | if (itt_sync_obj != NULL__null) | |||
2115 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
2116 | #endif /* USE_ITT_BUILD */ | |||
2117 | ||||
2118 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
2119 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
2120 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
2121 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
2122 | &(my_task_data), codeptr); | |||
2123 | } | |||
2124 | #endif | |||
2125 | ||||
2126 | #if OMP_45_ENABLED(50 >= 45) | |||
2127 | if (!taskdata->td_flags.team_serial || | |||
2128 | (thread->th.th_task_team != NULL__null && | |||
2129 | thread->th.th_task_team->tt.tt_found_proxy_tasks)) | |||
2130 | #else | |||
2131 | if (!taskdata->td_flags.team_serial) | |||
2132 | #endif | |||
2133 | { | |||
2134 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskgroup->count)), | |||
2135 | 0U); | |||
2136 | while (KMP_ATOMIC_LD_ACQ(&taskgroup->count)(&taskgroup->count)->load(std::memory_order_acquire ) != 0) { | |||
2137 | flag.execute_tasks(thread, gtid, FALSE0, | |||
2138 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
2139 | __kmp_task_stealing_constraint); | |||
2140 | } | |||
2141 | } | |||
2142 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting | |||
2143 | ||||
2144 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
2145 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
2146 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
2147 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
2148 | &(my_task_data), codeptr); | |||
2149 | } | |||
2150 | #endif | |||
2151 | ||||
2152 | #if USE_ITT_BUILD1 | |||
2153 | if (itt_sync_obj != NULL__null) | |||
2154 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
2155 | #endif /* USE_ITT_BUILD */ | |||
2156 | } | |||
2157 | KMP_DEBUG_ASSERT(taskgroup->count == 0)((taskgroup->count == 0) ? 0 : __kmp_debug_assert("taskgroup->count == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2157)); | |||
2158 | ||||
2159 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
2160 | #if OMP_45_ENABLED(50 >= 45) | |||
2161 | if (taskgroup->reduce_data != NULL__null) // need to reduce? | |||
2162 | __kmp_task_reduction_fini(thread, taskgroup); | |||
2163 | #endif | |||
2164 | // Restore parent taskgroup for the current task | |||
2165 | taskdata->td_taskgroup = taskgroup->parent; | |||
2166 | __kmp_thread_free(thread, taskgroup)___kmp_thread_free((thread), (taskgroup), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2166); | |||
2167 | ||||
2168 | KA_TRACE(10, ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n" , gtid, taskdata); } | |||
2169 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n" , gtid, taskdata); }; | |||
2170 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
2171 | ||||
2172 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
2173 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
2174 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
2175 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
2176 | &(my_task_data), codeptr); | |||
2177 | } | |||
2178 | #endif | |||
2179 | } | |||
2180 | #endif | |||
2181 | ||||
2182 | // __kmp_remove_my_task: remove a task from my own deque | |||
2183 | static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid, | |||
2184 | kmp_task_team_t *task_team, | |||
2185 | kmp_int32 is_constrained) { | |||
2186 | kmp_task_t *task; | |||
2187 | kmp_taskdata_t *taskdata; | |||
2188 | kmp_thread_data_t *thread_data; | |||
2189 | kmp_uint32 tail; | |||
2190 | ||||
2191 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2191)); | |||
2192 | KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=((task_team->tt.tt_threads_data != __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2193)) | |||
2193 | NULL)((task_team->tt.tt_threads_data != __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2193)); // Caller should check this condition | |||
2194 | ||||
2195 | thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2195)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
2196 | ||||
2197 | KA_TRACE(10, ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n" , gtid, thread_data->td.td_deque_ntasks, thread_data->td .td_deque_head, thread_data->td.td_deque_tail); } | |||
2198 | gtid, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n" , gtid, thread_data->td.td_deque_ntasks, thread_data->td .td_deque_head, thread_data->td.td_deque_tail); } | |||
2199 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n" , gtid, thread_data->td.td_deque_ntasks, thread_data->td .td_deque_head, thread_data->td.td_deque_tail); }; | |||
2200 | ||||
2201 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
2202 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2203 | ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2204 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2205 | gtid, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2206 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); }; | |||
2207 | return NULL__null; | |||
2208 | } | |||
2209 | ||||
2210 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2211 | ||||
2212 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
2213 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2214 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2215 | ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2216 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2217 | gtid, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2218 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); }; | |||
2219 | return NULL__null; | |||
2220 | } | |||
2221 | ||||
2222 | tail = (thread_data->td.td_deque_tail - 1) & | |||
2223 | TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); // Wrap index. | |||
2224 | taskdata = thread_data->td.td_deque[tail]; | |||
2225 | ||||
2226 | if (is_constrained && (taskdata->td_flags.tiedness == TASK_TIED1)) { | |||
2227 | // we need to check if the candidate obeys task scheduling constraint (TSC) | |||
2228 | // only descendant of all deferred tied tasks can be scheduled, checking | |||
2229 | // the last one is enough, as it in turn is the descendant of all others | |||
2230 | kmp_taskdata_t *current = thread->th.th_current_task->td_last_tied; | |||
2231 | KMP_DEBUG_ASSERT(current != NULL)((current != __null) ? 0 : __kmp_debug_assert("current != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2231)); | |||
2232 | // check if last tied task is not suspended on barrier | |||
2233 | if (current->td_flags.tasktype == TASK_EXPLICIT1 || | |||
2234 | current->td_taskwait_thread > 0) { // <= 0 on barrier | |||
2235 | kmp_int32 level = current->td_level; | |||
2236 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
2237 | while (parent != current && parent->td_level > level) { | |||
2238 | parent = parent->td_parent; // check generation up to the level of the | |||
2239 | // current task | |||
2240 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2240)); | |||
2241 | } | |||
2242 | if (parent != current) { | |||
2243 | // The TSC does not allow to steal victim task | |||
2244 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2245 | KA_TRACE(10, ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2246 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2247 | gtid, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2248 | thread_data->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
2249 | thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); }; | |||
2250 | return NULL__null; | |||
2251 | } | |||
2252 | } | |||
2253 | } | |||
2254 | ||||
2255 | thread_data->td.td_deque_tail = tail; | |||
2256 | TCW_4(thread_data->td.td_deque_ntasks, thread_data->td.td_deque_ntasks - 1)(thread_data->td.td_deque_ntasks) = (thread_data->td.td_deque_ntasks - 1); | |||
2257 | ||||
2258 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2259 | ||||
2260 | KA_TRACE(10, ("__kmp_remove_my_task(exit #2): T#%d task %p removed: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
2261 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
2262 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
2263 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); }; | |||
2264 | ||||
2265 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
2266 | return task; | |||
2267 | } | |||
2268 | ||||
2269 | // __kmp_steal_task: remove a task from another thread's deque | |||
2270 | // Assume that calling thread has already checked existence of | |||
2271 | // task_team thread_data before calling this routine. | |||
2272 | static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid, | |||
2273 | kmp_task_team_t *task_team, | |||
2274 | std::atomic<kmp_int32> *unfinished_threads, | |||
2275 | int *thread_finished, | |||
2276 | kmp_int32 is_constrained) { | |||
2277 | kmp_task_t *task; | |||
2278 | kmp_taskdata_t *taskdata; | |||
2279 | kmp_taskdata_t *current; | |||
2280 | kmp_thread_data_t *victim_td, *threads_data; | |||
2281 | kmp_int32 level, target; | |||
2282 | kmp_int32 victim_tid; | |||
2283 | ||||
2284 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2284)); | |||
2285 | ||||
2286 | threads_data = task_team->tt.tt_threads_data; | |||
2287 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2287)); // Caller should check this condition | |||
2288 | ||||
2289 | victim_tid = victim_thr->th.th_info.ds.ds_tid; | |||
2290 | victim_td = &threads_data[victim_tid]; | |||
2291 | ||||
2292 | KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2294)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2293 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2294)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2294 | gtid, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2294)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2295 | victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2294)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2296 | victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2294)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
2297 | ||||
2298 | if (TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks) == 0) { | |||
2299 | KA_TRACE(10, ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2301)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2300 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2301)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2301 | gtid, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2301)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2302 | victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2301)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
2303 | victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2301)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
2304 | return NULL__null; | |||
2305 | } | |||
2306 | ||||
2307 | __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
2308 | ||||
2309 | int ntasks = TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks); | |||
2310 | // Check again after we acquire the lock | |||
2311 | if (ntasks == 0) { | |||
2312 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
2313 | KA_TRACE(10, ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2315)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2314 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2315)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2315 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2315)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2316 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2315)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
2317 | return NULL__null; | |||
2318 | } | |||
2319 | ||||
2320 | KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL)((victim_td->td.td_deque != __null) ? 0 : __kmp_debug_assert ("victim_td->td.td_deque != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2320)); | |||
2321 | ||||
2322 | taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head]; | |||
2323 | if (is_constrained && (taskdata->td_flags.tiedness == TASK_TIED1)) { | |||
2324 | // we need to check if the candidate obeys task scheduling constraint (TSC) | |||
2325 | // only descendant of all deferred tied tasks can be scheduled, checking | |||
2326 | // the last one is enough, as it in turn is the descendant of all others | |||
2327 | current = __kmp_threads[gtid]->th.th_current_task->td_last_tied; | |||
2328 | KMP_DEBUG_ASSERT(current != NULL)((current != __null) ? 0 : __kmp_debug_assert("current != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2328)); | |||
2329 | // check if last tied task is not suspended on barrier | |||
2330 | if (current->td_flags.tasktype == TASK_EXPLICIT1 || | |||
2331 | current->td_taskwait_thread > 0) { // <= 0 on barrier | |||
2332 | level = current->td_level; | |||
2333 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
2334 | while (parent != current && parent->td_level > level) { | |||
2335 | parent = parent->td_parent; // check generation up to the level of the | |||
2336 | // current task | |||
2337 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2337)); | |||
2338 | } | |||
2339 | if (parent != current) { | |||
2340 | if (!task_team->tt.tt_untied_task_encountered) { | |||
2341 | // The TSC does not allow to steal victim task | |||
2342 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
2343 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2346)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2344 | ("__kmp_steal_task(exit #3): T#%d could not steal from "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2346)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2345 | "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2346)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2346 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2346)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2347 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2346)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
2348 | return NULL__null; | |||
2349 | } | |||
2350 | taskdata = NULL__null; // will check other tasks in victim's deque | |||
2351 | } | |||
2352 | } | |||
2353 | } | |||
2354 | if (taskdata != NULL__null) { | |||
2355 | // Bump head pointer and Wrap. | |||
2356 | victim_td->td.td_deque_head = | |||
2357 | (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
2358 | } else { | |||
2359 | int i; | |||
2360 | // walk through victim's deque trying to steal any task | |||
2361 | target = victim_td->td.td_deque_head; | |||
2362 | for (i = 1; i < ntasks; ++i) { | |||
2363 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
2364 | taskdata = victim_td->td.td_deque[target]; | |||
2365 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
2366 | // check if the candidate obeys the TSC | |||
2367 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
2368 | // check generation up to the level of the current task | |||
2369 | while (parent != current && parent->td_level > level) { | |||
2370 | parent = parent->td_parent; | |||
2371 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2371)); | |||
2372 | } | |||
2373 | if (parent != current) { | |||
2374 | // The TSC does not allow to steal the candidate | |||
2375 | taskdata = NULL__null; | |||
2376 | continue; | |||
2377 | } else { | |||
2378 | // found victim tied task | |||
2379 | break; | |||
2380 | } | |||
2381 | } else { | |||
2382 | // found victim untied task | |||
2383 | break; | |||
2384 | } | |||
2385 | } | |||
2386 | if (taskdata == NULL__null) { | |||
2387 | // No appropriate candidate to steal found | |||
2388 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
2389 | KA_TRACE(10, ("__kmp_steal_task(exit #4): T#%d could not steal from "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2391)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2390 | "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2391)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2391 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2391)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2392 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2391)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
2393 | return NULL__null; | |||
2394 | } | |||
2395 | int prev = target; | |||
2396 | for (i = i + 1; i < ntasks; ++i) { | |||
2397 | // shift remaining tasks in the deque left by 1 | |||
2398 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
2399 | victim_td->td.td_deque[prev] = victim_td->td.td_deque[target]; | |||
2400 | prev = target; | |||
2401 | } | |||
2402 | KMP_DEBUG_ASSERT(victim_td->td.td_deque_tail ==((victim_td->td.td_deque_tail == ((target + 1) & ((victim_td ->td).td_deque_size - 1))) ? 0 : __kmp_debug_assert("victim_td->td.td_deque_tail == ((target + 1) & ((victim_td->td).td_deque_size - 1))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2403)) | |||
2403 | ((target + 1) & TASK_DEQUE_MASK(victim_td->td)))((victim_td->td.td_deque_tail == ((target + 1) & ((victim_td ->td).td_deque_size - 1))) ? 0 : __kmp_debug_assert("victim_td->td.td_deque_tail == ((target + 1) & ((victim_td->td).td_deque_size - 1))" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2403)); | |||
2404 | victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped)) | |||
2405 | } | |||
2406 | if (*thread_finished) { | |||
2407 | // We need to un-mark this victim as a finished victim. This must be done | |||
2408 | // before releasing the lock, or else other threads (starting with the | |||
2409 | // master victim) might be prematurely released from the barrier!!! | |||
2410 | kmp_int32 count; | |||
2411 | ||||
2412 | count = KMP_ATOMIC_INC(unfinished_threads)(unfinished_threads)->fetch_add(1, std::memory_order_acq_rel ); | |||
2413 | ||||
2414 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n" , gtid, count + 1, task_team); } | |||
2415 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n" , gtid, count + 1, task_team); } | |||
2416 | ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n" , gtid, count + 1, task_team); } | |||
2417 | gtid, count + 1, task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n" , gtid, count + 1, task_team); }; | |||
2418 | ||||
2419 | *thread_finished = FALSE0; | |||
2420 | } | |||
2421 | TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1)(victim_td->td.td_deque_ntasks) = (ntasks - 1); | |||
2422 | ||||
2423 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
2424 | ||||
2425 | KMP_COUNT_BLOCK(TASK_stolen)((void)0); | |||
2426 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2429)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2427 | ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2429)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2428 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2429)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2429 | gtid, taskdata, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2429)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
2430 | ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2429)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
2431 | ||||
2432 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
2433 | return task; | |||
2434 | } | |||
2435 | ||||
2436 | // __kmp_execute_tasks_template: Choose and execute tasks until either the | |||
2437 | // condition is statisfied (return true) or there are none left (return false). | |||
2438 | // | |||
2439 | // final_spin is TRUE if this is the spin at the release barrier. | |||
2440 | // thread_finished indicates whether the thread is finished executing all | |||
2441 | // the tasks it has on its deque, and is at the release barrier. | |||
2442 | // spinner is the location on which to spin. | |||
2443 | // spinner == NULL means only execute a single task and return. | |||
2444 | // checker is the value to check to terminate the spin. | |||
2445 | template <class C> | |||
2446 | static inline int __kmp_execute_tasks_template( | |||
2447 | kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin, | |||
2448 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
2449 | kmp_int32 is_constrained) { | |||
2450 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
2451 | kmp_thread_data_t *threads_data; | |||
2452 | kmp_task_t *task; | |||
2453 | kmp_info_t *other_thread; | |||
2454 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
2455 | std::atomic<kmp_int32> *unfinished_threads; | |||
2456 | kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0, | |||
2457 | tid = thread->th.th_info.ds.ds_tid; | |||
2458 | ||||
2459 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2459)); | |||
2460 | KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid])((thread == __kmp_threads[gtid]) ? 0 : __kmp_debug_assert("thread == __kmp_threads[gtid]" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2460)); | |||
2461 | ||||
2462 | if (task_team == NULL__null) | |||
2463 | return FALSE0; | |||
2464 | ||||
2465 | KA_TRACE(15, ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d "if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d " "*thread_finished=%d\n", gtid, final_spin, *thread_finished) ; } | |||
2466 | "*thread_finished=%d\n",if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d " "*thread_finished=%d\n", gtid, final_spin, *thread_finished) ; } | |||
2467 | gtid, final_spin, *thread_finished))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d " "*thread_finished=%d\n", gtid, final_spin, *thread_finished) ; }; | |||
2468 | ||||
2469 | thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; | |||
2470 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
2471 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2471)); | |||
2472 | ||||
2473 | nthreads = task_team->tt.tt_nproc; | |||
2474 | unfinished_threads = &(task_team->tt.tt_unfinished_threads); | |||
2475 | #if OMP_45_ENABLED(50 >= 45) | |||
2476 | KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks)((nthreads > 1 || task_team->tt.tt_found_proxy_tasks) ? 0 : __kmp_debug_assert("nthreads > 1 || task_team->tt.tt_found_proxy_tasks" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2476)); | |||
2477 | #else | |||
2478 | KMP_DEBUG_ASSERT(nthreads > 1)((nthreads > 1) ? 0 : __kmp_debug_assert("nthreads > 1" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2478)); | |||
2479 | #endif | |||
2480 | KMP_DEBUG_ASSERT(*unfinished_threads >= 0)((*unfinished_threads >= 0) ? 0 : __kmp_debug_assert("*unfinished_threads >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2480)); | |||
2481 | ||||
2482 | while (1) { // Outer loop keeps trying to find tasks in case of single thread | |||
2483 | // getting tasks from target constructs | |||
2484 | while (1) { // Inner loop to find a task and execute it | |||
2485 | task = NULL__null; | |||
2486 | if (use_own_tasks) { // check on own queue first | |||
2487 | task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained); | |||
2488 | } | |||
2489 | if ((task == NULL__null) && (nthreads > 1)) { // Steal a task | |||
2490 | int asleep = 1; | |||
2491 | use_own_tasks = 0; | |||
2492 | // Try to steal from the last place I stole from successfully. | |||
2493 | if (victim_tid == -2) { // haven't stolen anything yet | |||
2494 | victim_tid = threads_data[tid].td.td_deque_last_stolen; | |||
2495 | if (victim_tid != | |||
2496 | -1) // if we have a last stolen from victim, get the thread | |||
2497 | other_thread = threads_data[victim_tid].td.td_thr; | |||
2498 | } | |||
2499 | if (victim_tid != -1) { // found last victim | |||
2500 | asleep = 0; | |||
2501 | } else if (!new_victim) { // no recent steals and we haven't already | |||
2502 | // used a new victim; select a random thread | |||
2503 | do { // Find a different thread to steal work from. | |||
2504 | // Pick a random thread. Initial plan was to cycle through all the | |||
2505 | // threads, and only return if we tried to steal from every thread, | |||
2506 | // and failed. Arch says that's not such a great idea. | |||
2507 | victim_tid = __kmp_get_random(thread) % (nthreads - 1); | |||
2508 | if (victim_tid >= tid) { | |||
2509 | ++victim_tid; // Adjusts random distribution to exclude self | |||
2510 | } | |||
2511 | // Found a potential victim | |||
2512 | other_thread = threads_data[victim_tid].td.td_thr; | |||
2513 | // There is a slight chance that __kmp_enable_tasking() did not wake | |||
2514 | // up all threads waiting at the barrier. If victim is sleeping, | |||
2515 | // then wake it up. Since we were going to pay the cache miss | |||
2516 | // penalty for referencing another thread's kmp_info_t struct | |||
2517 | // anyway, | |||
2518 | // the check shouldn't cost too much performance at this point. In | |||
2519 | // extra barrier mode, tasks do not sleep at the separate tasking | |||
2520 | // barrier, so this isn't a problem. | |||
2521 | asleep = 0; | |||
2522 | if ((__kmp_tasking_mode == tskm_task_teams) && | |||
2523 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) && | |||
2524 | (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc))((void *)(const_cast<void *>(other_thread->th.th_sleep_loc ))) != | |||
2525 | NULL__null)) { | |||
2526 | asleep = 1; | |||
2527 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(other_thread)((((other_thread) != __null) ? 0 : __kmp_debug_assert("(other_thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2527)), (other_thread)->th.th_info.ds.ds_gtid), | |||
2528 | other_thread->th.th_sleep_loc); | |||
2529 | // A sleeping thread should not have any tasks on it's queue. | |||
2530 | // There is a slight possibility that it resumes, steals a task | |||
2531 | // from another thread, which spawns more tasks, all in the time | |||
2532 | // that it takes this thread to check => don't write an assertion | |||
2533 | // that the victim's queue is empty. Try stealing from a | |||
2534 | // different thread. | |||
2535 | } | |||
2536 | } while (asleep); | |||
2537 | } | |||
2538 | ||||
2539 | if (!asleep) { | |||
2540 | // We have a victim to try to steal from | |||
2541 | task = __kmp_steal_task(other_thread, gtid, task_team, | |||
2542 | unfinished_threads, thread_finished, | |||
2543 | is_constrained); | |||
2544 | } | |||
2545 | if (task != NULL__null) { // set last stolen to victim | |||
2546 | if (threads_data[tid].td.td_deque_last_stolen != victim_tid) { | |||
2547 | threads_data[tid].td.td_deque_last_stolen = victim_tid; | |||
2548 | // The pre-refactored code did not try more than 1 successful new | |||
2549 | // vicitm, unless the last one generated more local tasks; | |||
2550 | // new_victim keeps track of this | |||
2551 | new_victim = 1; | |||
2552 | } | |||
2553 | } else { // No tasks found; unset last_stolen | |||
2554 | KMP_CHECK_UPDATE(threads_data[tid].td.td_deque_last_stolen, -1)if ((threads_data[tid].td.td_deque_last_stolen) != (-1)) (threads_data [tid].td.td_deque_last_stolen) = (-1); | |||
2555 | victim_tid = -2; // no successful victim found | |||
2556 | } | |||
2557 | } | |||
2558 | ||||
2559 | if (task == NULL__null) // break out of tasking loop | |||
2560 | break; | |||
2561 | ||||
2562 | // Found a task; execute it | |||
2563 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
2564 | if (__itt_sync_create_ptr__kmp_itt_sync_create_ptr__3_0 || KMP_ITT_DEBUG0) { | |||
2565 | if (itt_sync_obj == NULL__null) { // we are at fork barrier where we could not | |||
2566 | // get the object reliably | |||
2567 | itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier); | |||
2568 | } | |||
2569 | __kmp_itt_task_starting(itt_sync_obj); | |||
2570 | } | |||
2571 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ | |||
2572 | __kmp_invoke_task(gtid, task, current_task); | |||
2573 | #if USE_ITT_BUILD1 | |||
2574 | if (itt_sync_obj != NULL__null) | |||
2575 | __kmp_itt_task_finished(itt_sync_obj); | |||
2576 | #endif /* USE_ITT_BUILD */ | |||
2577 | // If this thread is only partway through the barrier and the condition is | |||
2578 | // met, then return now, so that the barrier gather/release pattern can | |||
2579 | // proceed. If this thread is in the last spin loop in the barrier, | |||
2580 | // waiting to be released, we know that the termination condition will not | |||
2581 | // be satisified, so don't waste any cycles checking it. | |||
2582 | if (flag == NULL__null || (!final_spin && flag->done_check())) { | |||
2583 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2584 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2585 | ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2586 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
2587 | return TRUE(!0); | |||
2588 | } | |||
2589 | if (thread->th.th_task_team == NULL__null) { | |||
2590 | break; | |||
2591 | } | |||
2592 | // Yield before executing next task | |||
2593 | KMP_YIELD(__kmp_library == library_throughput){ __kmp_x86_pause(); __kmp_yield((__kmp_library == library_throughput )); }; | |||
2594 | // If execution of a stolen task results in more tasks being placed on our | |||
2595 | // run queue, reset use_own_tasks | |||
2596 | if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks)(threads_data[tid].td.td_deque_ntasks) != 0) { | |||
2597 | KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d stolen task spawned "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d stolen task spawned " "other tasks, restart\n", gtid); } | |||
2598 | "other tasks, restart\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d stolen task spawned " "other tasks, restart\n", gtid); } | |||
2599 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d stolen task spawned " "other tasks, restart\n", gtid); }; | |||
2600 | use_own_tasks = 1; | |||
2601 | new_victim = 0; | |||
2602 | } | |||
2603 | } | |||
2604 | ||||
2605 | // The task source has been exhausted. If in final spin loop of barrier, check | |||
2606 | // if termination condition is satisfied. | |||
2607 | #if OMP_45_ENABLED(50 >= 45) | |||
2608 | // The work queue may be empty but there might be proxy tasks still | |||
2609 | // executing | |||
2610 | if (final_spin && | |||
2611 | KMP_ATOMIC_LD_ACQ(¤t_task->td_incomplete_child_tasks)(¤t_task->td_incomplete_child_tasks)->load(std ::memory_order_acquire) == 0) | |||
2612 | #else | |||
2613 | if (final_spin) | |||
2614 | #endif | |||
2615 | { | |||
2616 | // First, decrement the #unfinished threads, if that has not already been | |||
2617 | // done. This decrement might be to the spin location, and result in the | |||
2618 | // termination condition being satisfied. | |||
2619 | if (!*thread_finished) { | |||
2620 | kmp_int32 count; | |||
2621 | ||||
2622 | count = KMP_ATOMIC_DEC(unfinished_threads)(unfinished_threads)->fetch_sub(1, std::memory_order_acq_rel ) - 1; | |||
2623 | KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d dec "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d dec " "unfinished_threads to %d task_team=%p\n", gtid, count, task_team ); } | |||
2624 | "unfinished_threads to %d task_team=%p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d dec " "unfinished_threads to %d task_team=%p\n", gtid, count, task_team ); } | |||
2625 | gtid, count, task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d dec " "unfinished_threads to %d task_team=%p\n", gtid, count, task_team ); }; | |||
2626 | *thread_finished = TRUE(!0); | |||
2627 | } | |||
2628 | ||||
2629 | // It is now unsafe to reference thread->th.th_team !!! | |||
2630 | // Decrementing task_team->tt.tt_unfinished_threads can allow the master | |||
2631 | // thread to pass through the barrier, where it might reset each thread's | |||
2632 | // th.th_team field for the next parallel region. If we can steal more | |||
2633 | // work, we know that this has not happened yet. | |||
2634 | if (flag != NULL__null && flag->done_check()) { | |||
2635 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2636 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2637 | ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
2638 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
2639 | return TRUE(!0); | |||
2640 | } | |||
2641 | } | |||
2642 | ||||
2643 | // If this thread's task team is NULL, master has recognized that there are | |||
2644 | // no more tasks; bail out | |||
2645 | if (thread->th.th_task_team == NULL__null) { | |||
2646 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d no more tasks\n" , gtid); } | |||
2647 | ("__kmp_execute_tasks_template: T#%d no more tasks\n", gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d no more tasks\n" , gtid); }; | |||
2648 | return FALSE0; | |||
2649 | } | |||
2650 | ||||
2651 | #if OMP_45_ENABLED(50 >= 45) | |||
2652 | // We could be getting tasks from target constructs; if this is the only | |||
2653 | // thread, keep trying to execute tasks from own queue | |||
2654 | if (nthreads == 1) | |||
2655 | use_own_tasks = 1; | |||
2656 | else | |||
2657 | #endif | |||
2658 | { | |||
2659 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d can't find work\n" , gtid); } | |||
2660 | ("__kmp_execute_tasks_template: T#%d can't find work\n", gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d can't find work\n" , gtid); }; | |||
2661 | return FALSE0; | |||
2662 | } | |||
2663 | } | |||
2664 | } | |||
2665 | ||||
2666 | int __kmp_execute_tasks_32( | |||
2667 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32 *flag, int final_spin, | |||
2668 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
2669 | kmp_int32 is_constrained) { | |||
2670 | return __kmp_execute_tasks_template( | |||
2671 | thread, gtid, flag, final_spin, | |||
2672 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
2673 | } | |||
2674 | ||||
2675 | int __kmp_execute_tasks_64( | |||
2676 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64 *flag, int final_spin, | |||
2677 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
2678 | kmp_int32 is_constrained) { | |||
2679 | return __kmp_execute_tasks_template( | |||
2680 | thread, gtid, flag, final_spin, | |||
2681 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
2682 | } | |||
2683 | ||||
2684 | int __kmp_execute_tasks_oncore( | |||
2685 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin, | |||
2686 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
2687 | kmp_int32 is_constrained) { | |||
2688 | return __kmp_execute_tasks_template( | |||
2689 | thread, gtid, flag, final_spin, | |||
2690 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
2691 | } | |||
2692 | ||||
2693 | // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the | |||
2694 | // next barrier so they can assist in executing enqueued tasks. | |||
2695 | // First thread in allocates the task team atomically. | |||
2696 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
2697 | kmp_info_t *this_thr) { | |||
2698 | kmp_thread_data_t *threads_data; | |||
2699 | int nthreads, i, is_init_thread; | |||
2700 | ||||
2701 | KA_TRACE(10, ("__kmp_enable_tasking(enter): T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(enter): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2702)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
2702 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(enter): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2702)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
2703 | ||||
2704 | KMP_DEBUG_ASSERT(task_team != NULL)((task_team != __null) ? 0 : __kmp_debug_assert("task_team != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2704)); | |||
2705 | KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL)((this_thr->th.th_team != __null) ? 0 : __kmp_debug_assert ("this_thr->th.th_team != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2705)); | |||
2706 | ||||
2707 | nthreads = task_team->tt.tt_nproc; | |||
2708 | KMP_DEBUG_ASSERT(nthreads > 0)((nthreads > 0) ? 0 : __kmp_debug_assert("nthreads > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2708)); | |||
2709 | KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc)((nthreads == this_thr->th.th_team->t.t_nproc) ? 0 : __kmp_debug_assert ("nthreads == this_thr->th.th_team->t.t_nproc", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2709)); | |||
2710 | ||||
2711 | // Allocate or increase the size of threads_data if necessary | |||
2712 | is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team); | |||
2713 | ||||
2714 | if (!is_init_thread) { | |||
2715 | // Some other thread already set up the array. | |||
2716 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2719)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
2717 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2719)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
2718 | ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2719)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
2719 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2719)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
2720 | return; | |||
2721 | } | |||
2722 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
2723 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2723)); | |||
2724 | ||||
2725 | if ((__kmp_tasking_mode == tskm_task_teams) && | |||
2726 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647))) { | |||
2727 | // Release any threads sleeping at the barrier, so that they can steal | |||
2728 | // tasks and execute them. In extra barrier mode, tasks do not sleep | |||
2729 | // at the separate tasking barrier, so this isn't a problem. | |||
2730 | for (i = 0; i < nthreads; i++) { | |||
2731 | volatile void *sleep_loc; | |||
2732 | kmp_info_t *thread = threads_data[i].td.td_thr; | |||
2733 | ||||
2734 | if (i == this_thr->th.th_info.ds.ds_tid) { | |||
2735 | continue; | |||
2736 | } | |||
2737 | // Since we haven't locked the thread's suspend mutex lock at this | |||
2738 | // point, there is a small window where a thread might be putting | |||
2739 | // itself to sleep, but hasn't set the th_sleep_loc field yet. | |||
2740 | // To work around this, __kmp_execute_tasks_template() periodically checks | |||
2741 | // see if other threads are sleeping (using the same random mechanism that | |||
2742 | // is used for task stealing) and awakens them if they are. | |||
2743 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
2744 | NULL__null) { | |||
2745 | KF_TRACE(50, ("__kmp_enable_tasking: T#%d waking up thread T#%d\n",if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2746)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2747)), (thread)->th.th_info.ds.ds_gtid)); } | |||
2746 | __kmp_gtid_from_thread(this_thr),if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2746)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2747)), (thread)->th.th_info.ds.ds_gtid)); } | |||
2747 | __kmp_gtid_from_thread(thread)))if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2746)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2747)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
2748 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2748)), (thread)->th.th_info.ds.ds_gtid), sleep_loc); | |||
2749 | } else { | |||
2750 | KF_TRACE(50, ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n",if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2751)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2752)), (thread)->th.th_info.ds.ds_gtid)); } | |||
2751 | __kmp_gtid_from_thread(this_thr),if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2751)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2752)), (thread)->th.th_info.ds.ds_gtid)); } | |||
2752 | __kmp_gtid_from_thread(thread)))if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2751)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2752)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
2753 | } | |||
2754 | } | |||
2755 | } | |||
2756 | ||||
2757 | KA_TRACE(10, ("__kmp_enable_tasking(exit): T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2758)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
2758 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2758)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
2759 | } | |||
2760 | ||||
2761 | /* // TODO: Check the comment consistency | |||
2762 | * Utility routines for "task teams". A task team (kmp_task_t) is kind of | |||
2763 | * like a shadow of the kmp_team_t data struct, with a different lifetime. | |||
2764 | * After a child * thread checks into a barrier and calls __kmp_release() from | |||
2765 | * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no | |||
2766 | * longer assume that the kmp_team_t structure is intact (at any moment, the | |||
2767 | * master thread may exit the barrier code and free the team data structure, | |||
2768 | * and return the threads to the thread pool). | |||
2769 | * | |||
2770 | * This does not work with the the tasking code, as the thread is still | |||
2771 | * expected to participate in the execution of any tasks that may have been | |||
2772 | * spawned my a member of the team, and the thread still needs access to all | |||
2773 | * to each thread in the team, so that it can steal work from it. | |||
2774 | * | |||
2775 | * Enter the existence of the kmp_task_team_t struct. It employs a reference | |||
2776 | * counting mechanims, and is allocated by the master thread before calling | |||
2777 | * __kmp_<barrier_kind>_release, and then is release by the last thread to | |||
2778 | * exit __kmp_<barrier_kind>_release at the next barrier. I.e. the lifetimes | |||
2779 | * of the kmp_task_team_t structs for consecutive barriers can overlap | |||
2780 | * (and will, unless the master thread is the last thread to exit the barrier | |||
2781 | * release phase, which is not typical). | |||
2782 | * | |||
2783 | * The existence of such a struct is useful outside the context of tasking, | |||
2784 | * but for now, I'm trying to keep it specific to the OMP_30_ENABLED macro, | |||
2785 | * so that any performance differences show up when comparing the 2.5 vs. 3.0 | |||
2786 | * libraries. | |||
2787 | * | |||
2788 | * We currently use the existence of the threads array as an indicator that | |||
2789 | * tasks were spawned since the last barrier. If the structure is to be | |||
2790 | * useful outside the context of tasking, then this will have to change, but | |||
2791 | * not settting the field minimizes the performance impact of tasking on | |||
2792 | * barriers, when no explicit tasks were spawned (pushed, actually). | |||
2793 | */ | |||
2794 | ||||
2795 | static kmp_task_team_t *__kmp_free_task_teams = | |||
2796 | NULL__null; // Free list for task_team data structures | |||
2797 | // Lock for task team data structures | |||
2798 | kmp_bootstrap_lock_t __kmp_task_team_lock = | |||
2799 | KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock){ { { true } , &((__kmp_task_team_lock)), __null, { 0U }, { 0U }, { 0 }, { -1 } } }; | |||
2800 | ||||
2801 | // __kmp_alloc_task_deque: | |||
2802 | // Allocates a task deque for a particular thread, and initialize the necessary | |||
2803 | // data structures relating to the deque. This only happens once per thread | |||
2804 | // per task team since task teams are recycled. No lock is needed during | |||
2805 | // allocation since each thread allocates its own deque. | |||
2806 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
2807 | kmp_thread_data_t *thread_data) { | |||
2808 | __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2809 | KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL)((thread_data->td.td_deque == __null) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque == __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2809)); | |||
2810 | ||||
2811 | // Initialize last stolen task field to "none" | |||
2812 | thread_data->td.td_deque_last_stolen = -1; | |||
2813 | ||||
2814 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0)(((thread_data->td.td_deque_ntasks) == 0) ? 0 : __kmp_debug_assert ("(thread_data->td.td_deque_ntasks) == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2814)); | |||
2815 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0)((thread_data->td.td_deque_head == 0) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque_head == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2815)); | |||
2816 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0)((thread_data->td.td_deque_tail == 0) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque_tail == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2816)); | |||
2817 | ||||
2818 | KE_TRACE(if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2821)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
2819 | 10,if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2821)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
2820 | ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2821)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
2821 | __kmp_gtid_from_thread(thread), INITIAL_TASK_DEQUE_SIZE, thread_data))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2821)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); }; | |||
2822 | // Allocate space for task deque, and zero the deque | |||
2823 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
2824 | // kmp_reap_task_team( ). | |||
2825 | thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2826) | |||
2826 | INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *))___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2826); | |||
2827 | thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE(1 << 8); | |||
2828 | } | |||
2829 | ||||
2830 | // __kmp_realloc_task_deque: | |||
2831 | // Re-allocates a task deque for a particular thread, copies the content from | |||
2832 | // the old deque and adjusts the necessary data structures relating to the | |||
2833 | // deque. This operation must be done with a the deque_lock being held | |||
2834 | static void __kmp_realloc_task_deque(kmp_info_t *thread, | |||
2835 | kmp_thread_data_t *thread_data) { | |||
2836 | kmp_int32 size = TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size); | |||
2837 | kmp_int32 new_size = 2 * size; | |||
2838 | ||||
2839 | KE_TRACE(10, ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2841)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); } | |||
2840 | "%d] for thread_data %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2841)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); } | |||
2841 | __kmp_gtid_from_thread(thread), size, new_size, thread_data))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2841)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); }; | |||
2842 | ||||
2843 | kmp_taskdata_t **new_deque = | |||
2844 | (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *))___kmp_allocate((new_size * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2844); | |||
2845 | ||||
2846 | int i, j; | |||
2847 | for (i = thread_data->td.td_deque_head, j = 0; j < size; | |||
2848 | i = (i + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1), j++) | |||
2849 | new_deque[j] = thread_data->td.td_deque[i]; | |||
2850 | ||||
2851 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2851); | |||
2852 | ||||
2853 | thread_data->td.td_deque_head = 0; | |||
2854 | thread_data->td.td_deque_tail = size; | |||
2855 | thread_data->td.td_deque = new_deque; | |||
2856 | thread_data->td.td_deque_size = new_size; | |||
2857 | } | |||
2858 | ||||
2859 | // __kmp_free_task_deque: | |||
2860 | // Deallocates a task deque for a particular thread. Happens at library | |||
2861 | // deallocation so don't need to reset all thread data fields. | |||
2862 | static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) { | |||
2863 | if (thread_data->td.td_deque != NULL__null) { | |||
2864 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2865 | TCW_4(thread_data->td.td_deque_ntasks, 0)(thread_data->td.td_deque_ntasks) = (0); | |||
2866 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2866); | |||
2867 | thread_data->td.td_deque = NULL__null; | |||
2868 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
2869 | } | |||
2870 | ||||
2871 | #ifdef BUILD_TIED_TASK_STACK | |||
2872 | // GEH: Figure out what to do here for td_susp_tied_tasks | |||
2873 | if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) { | |||
2874 | __kmp_free_task_stack(__kmp_thread_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2874)), __kmp_threads[(gtid)]), thread_data); | |||
2875 | } | |||
2876 | #endif // BUILD_TIED_TASK_STACK | |||
2877 | } | |||
2878 | ||||
2879 | // __kmp_realloc_task_threads_data: | |||
2880 | // Allocates a threads_data array for a task team, either by allocating an | |||
2881 | // initial array or enlarging an existing array. Only the first thread to get | |||
2882 | // the lock allocs or enlarges the array and re-initializes the array eleemnts. | |||
2883 | // That thread returns "TRUE", the rest return "FALSE". | |||
2884 | // Assumes that the new array size is given by task_team -> tt.tt_nproc. | |||
2885 | // The current size is given by task_team -> tt.tt_max_threads. | |||
2886 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
2887 | kmp_task_team_t *task_team) { | |||
2888 | kmp_thread_data_t **threads_data_p; | |||
2889 | kmp_int32 nthreads, maxthreads; | |||
2890 | int is_init_thread = FALSE0; | |||
2891 | ||||
2892 | if (TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
2893 | // Already reallocated and initialized. | |||
2894 | return FALSE0; | |||
2895 | } | |||
2896 | ||||
2897 | threads_data_p = &task_team->tt.tt_threads_data; | |||
2898 | nthreads = task_team->tt.tt_nproc; | |||
2899 | maxthreads = task_team->tt.tt_max_threads; | |||
2900 | ||||
2901 | // All threads must lock when they encounter the first task of the implicit | |||
2902 | // task region to make sure threads_data fields are (re)initialized before | |||
2903 | // used. | |||
2904 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
2905 | ||||
2906 | if (!TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
2907 | // first thread to enable tasking | |||
2908 | kmp_team_t *team = thread->th.th_team; | |||
2909 | int i; | |||
2910 | ||||
2911 | is_init_thread = TRUE(!0); | |||
2912 | if (maxthreads < nthreads) { | |||
2913 | ||||
2914 | if (*threads_data_p != NULL__null) { | |||
2915 | kmp_thread_data_t *old_data = *threads_data_p; | |||
2916 | kmp_thread_data_t *new_data = NULL__null; | |||
2917 | ||||
2918 | KE_TRACE(if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2922)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
2919 | 10,if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2922)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
2920 | ("__kmp_realloc_task_threads_data: T#%d reallocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2922)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
2921 | "threads data for task_team %p, new_size = %d, old_size = %d\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2922)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
2922 | __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2922)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); }; | |||
2923 | // Reallocate threads_data to have more elements than current array | |||
2924 | // Cannot use __kmp_thread_realloc() because threads not around for | |||
2925 | // kmp_reap_task_team( ). Note all new array entries are initialized | |||
2926 | // to zero by __kmp_allocate(). | |||
2927 | new_data = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2928) | |||
2928 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2928); | |||
2929 | // copy old data to new data | |||
2930 | KMP_MEMCPY_S((void *)new_data, nthreads * sizeof(kmp_thread_data_t),memcpy((void *)new_data, (void *)old_data, maxthreads * sizeof (kmp_thread_data_t)) | |||
2931 | (void *)old_data, maxthreads * sizeof(kmp_thread_data_t))memcpy((void *)new_data, (void *)old_data, maxthreads * sizeof (kmp_thread_data_t)); | |||
2932 | ||||
2933 | #ifdef BUILD_TIED_TASK_STACK | |||
2934 | // GEH: Figure out if this is the right thing to do | |||
2935 | for (i = maxthreads; i < nthreads; i++) { | |||
2936 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
2937 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2937)), (thread)->th.th_info.ds.ds_gtid), thread_data); | |||
2938 | } | |||
2939 | #endif // BUILD_TIED_TASK_STACK | |||
2940 | // Install the new data and free the old data | |||
2941 | (*threads_data_p) = new_data; | |||
2942 | __kmp_free(old_data)___kmp_free((old_data), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2942); | |||
2943 | } else { | |||
2944 | KE_TRACE(10, ("__kmp_realloc_task_threads_data: T#%d allocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2946)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); } | |||
2945 | "threads data for task_team %p, size = %d\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2946)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); } | |||
2946 | __kmp_gtid_from_thread(thread), task_team, nthreads))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2946)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); }; | |||
2947 | // Make the initial allocate for threads_data array, and zero entries | |||
2948 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
2949 | // kmp_reap_task_team( ). | |||
2950 | ANNOTATE_IGNORE_WRITES_BEGIN(); | |||
2951 | *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2952) | |||
2952 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2952); | |||
2953 | ANNOTATE_IGNORE_WRITES_END(); | |||
2954 | #ifdef BUILD_TIED_TASK_STACK | |||
2955 | // GEH: Figure out if this is the right thing to do | |||
2956 | for (i = 0; i < nthreads; i++) { | |||
2957 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
2958 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2958)), (thread)->th.th_info.ds.ds_gtid), thread_data); | |||
2959 | } | |||
2960 | #endif // BUILD_TIED_TASK_STACK | |||
2961 | } | |||
2962 | task_team->tt.tt_max_threads = nthreads; | |||
2963 | } else { | |||
2964 | // If array has (more than) enough elements, go ahead and use it | |||
2965 | KMP_DEBUG_ASSERT(*threads_data_p != NULL)((*threads_data_p != __null) ? 0 : __kmp_debug_assert("*threads_data_p != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2965)); | |||
2966 | } | |||
2967 | ||||
2968 | // initialize threads_data pointers back to thread_info structures | |||
2969 | for (i = 0; i < nthreads; i++) { | |||
2970 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
2971 | thread_data->td.td_thr = team->t.t_threads[i]; | |||
2972 | ||||
2973 | if (thread_data->td.td_deque_last_stolen >= nthreads) { | |||
2974 | // The last stolen field survives across teams / barrier, and the number | |||
2975 | // of threads may have changed. It's possible (likely?) that a new | |||
2976 | // parallel region will exhibit the same behavior as previous region. | |||
2977 | thread_data->td.td_deque_last_stolen = -1; | |||
2978 | } | |||
2979 | } | |||
2980 | ||||
2981 | KMP_MB(); | |||
2982 | TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE)(task_team->tt.tt_found_tasks) = ((!0)); | |||
2983 | } | |||
2984 | ||||
2985 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
2986 | return is_init_thread; | |||
2987 | } | |||
2988 | ||||
2989 | // __kmp_free_task_threads_data: | |||
2990 | // Deallocates a threads_data array for a task team, including any attached | |||
2991 | // tasking deques. Only occurs at library shutdown. | |||
2992 | static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) { | |||
2993 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
2994 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
2995 | int i; | |||
2996 | for (i = 0; i < task_team->tt.tt_max_threads; i++) { | |||
2997 | __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]); | |||
2998 | } | |||
2999 | __kmp_free(task_team->tt.tt_threads_data)___kmp_free((task_team->tt.tt_threads_data), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 2999); | |||
3000 | task_team->tt.tt_threads_data = NULL__null; | |||
3001 | } | |||
3002 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
3003 | } | |||
3004 | ||||
3005 | // __kmp_allocate_task_team: | |||
3006 | // Allocates a task team associated with a specific team, taking it from | |||
3007 | // the global task team free list if possible. Also initializes data | |||
3008 | // structures. | |||
3009 | static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread, | |||
3010 | kmp_team_t *team) { | |||
3011 | kmp_task_team_t *task_team = NULL__null; | |||
3012 | int nthreads; | |||
3013 | ||||
3014 | KA_TRACE(20, ("__kmp_allocate_task_team: T#%d entering; team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d entering; team = %p\n" , (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3015)), (thread)->th.th_info.ds.ds_gtid) : -1), team); } | |||
3015 | (thread ? __kmp_gtid_from_thread(thread) : -1), team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d entering; team = %p\n" , (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3015)), (thread)->th.th_info.ds.ds_gtid) : -1), team); }; | |||
3016 | ||||
3017 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
3018 | // Take a task team from the task team pool | |||
3019 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
3020 | if (__kmp_free_task_teams != NULL__null) { | |||
3021 | task_team = __kmp_free_task_teams; | |||
3022 | TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next)((__kmp_free_task_teams)) = ((task_team->tt.tt_next)); | |||
3023 | task_team->tt.tt_next = NULL__null; | |||
3024 | } | |||
3025 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
3026 | } | |||
3027 | ||||
3028 | if (task_team == NULL__null) { | |||
3029 | KE_TRACE(10, ("__kmp_allocate_task_team: T#%d allocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3031)), (thread)->th.th_info.ds.ds_gtid), team); } | |||
3030 | "task team for team %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3031)), (thread)->th.th_info.ds.ds_gtid), team); } | |||
3031 | __kmp_gtid_from_thread(thread), team))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3031)), (thread)->th.th_info.ds.ds_gtid), team); }; | |||
3032 | // Allocate a new task team if one is not available. | |||
3033 | // Cannot use __kmp_thread_malloc() because threads not around for | |||
3034 | // kmp_reap_task_team( ). | |||
3035 | task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t))___kmp_allocate((sizeof(kmp_task_team_t)), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3035); | |||
3036 | __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
3037 | // AC: __kmp_allocate zeroes returned memory | |||
3038 | // task_team -> tt.tt_threads_data = NULL; | |||
3039 | // task_team -> tt.tt_max_threads = 0; | |||
3040 | // task_team -> tt.tt_next = NULL; | |||
3041 | } | |||
3042 | ||||
3043 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
3044 | #if OMP_45_ENABLED(50 >= 45) | |||
3045 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
3046 | #endif | |||
3047 | task_team->tt.tt_nproc = nthreads = team->t.t_nproc; | |||
3048 | ||||
3049 | KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads)(&task_team->tt.tt_unfinished_threads)->store(nthreads , std::memory_order_release); | |||
3050 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
3051 | ||||
3052 | KA_TRACE(20, ("__kmp_allocate_task_team: T#%d exiting; task_team = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3054)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
3053 | "unfinished_threads init'd to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3054)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
3054 | (thread ? __kmp_gtid_from_thread(thread) : -1), task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3054)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
3055 | KMP_ATOMIC_LD_RLX(&task_team->tt.tt_unfinished_threads)))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3054)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); }; | |||
3056 | return task_team; | |||
3057 | } | |||
3058 | ||||
3059 | // __kmp_free_task_team: | |||
3060 | // Frees the task team associated with a specific thread, and adds it | |||
3061 | // to the global task team free list. | |||
3062 | void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) { | |||
3063 | KA_TRACE(20, ("__kmp_free_task_team: T#%d task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_team: T#%d task_team = %p\n" , thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3064)), (thread)->th.th_info.ds.ds_gtid) : -1, task_team ); } | |||
3064 | thread ? __kmp_gtid_from_thread(thread) : -1, task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_team: T#%d task_team = %p\n" , thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3064)), (thread)->th.th_info.ds.ds_gtid) : -1, task_team ); }; | |||
3065 | ||||
3066 | // Put task team back on free list | |||
3067 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
3068 | ||||
3069 | KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL)((task_team->tt.tt_next == __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_next == __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3069)); | |||
3070 | task_team->tt.tt_next = __kmp_free_task_teams; | |||
3071 | TCW_PTR(__kmp_free_task_teams, task_team)((__kmp_free_task_teams)) = ((task_team)); | |||
3072 | ||||
3073 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
3074 | } | |||
3075 | ||||
3076 | // __kmp_reap_task_teams: | |||
3077 | // Free all the task teams on the task team free list. | |||
3078 | // Should only be done during library shutdown. | |||
3079 | // Cannot do anything that needs a thread structure or gtid since they are | |||
3080 | // already gone. | |||
3081 | void __kmp_reap_task_teams(void) { | |||
3082 | kmp_task_team_t *task_team; | |||
3083 | ||||
3084 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
3085 | // Free all task_teams on the free list | |||
3086 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
3087 | while ((task_team = __kmp_free_task_teams) != NULL__null) { | |||
3088 | __kmp_free_task_teams = task_team->tt.tt_next; | |||
3089 | task_team->tt.tt_next = NULL__null; | |||
3090 | ||||
3091 | // Free threads_data if necessary | |||
3092 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
3093 | __kmp_free_task_threads_data(task_team); | |||
3094 | } | |||
3095 | __kmp_free(task_team)___kmp_free((task_team), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3095); | |||
3096 | } | |||
3097 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
3098 | } | |||
3099 | } | |||
3100 | ||||
3101 | // __kmp_wait_to_unref_task_teams: | |||
3102 | // Some threads could still be in the fork barrier release code, possibly | |||
3103 | // trying to steal tasks. Wait for each thread to unreference its task team. | |||
3104 | void __kmp_wait_to_unref_task_teams(void) { | |||
3105 | kmp_info_t *thread; | |||
3106 | kmp_uint32 spins; | |||
3107 | int done; | |||
3108 | ||||
3109 | KMP_INIT_YIELD(spins){ (spins) = __kmp_yield_init; }; | |||
3110 | ||||
3111 | for (;;) { | |||
3112 | done = TRUE(!0); | |||
3113 | ||||
3114 | // TODO: GEH - this may be is wrong because some sync would be necessary | |||
3115 | // in case threads are added to the pool during the traversal. Need to | |||
3116 | // verify that lock for thread pool is held when calling this routine. | |||
3117 | for (thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); thread != NULL__null; | |||
3118 | thread = thread->th.th_next_pool) { | |||
3119 | #if KMP_OS_WINDOWS0 | |||
3120 | DWORD exit_val; | |||
3121 | #endif | |||
3122 | if (TCR_PTR(thread->th.th_task_team)((void *)(thread->th.th_task_team)) == NULL__null) { | |||
3123 | KA_TRACE(10, ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3124)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3124 | __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3124)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
3125 | continue; | |||
3126 | } | |||
3127 | #if KMP_OS_WINDOWS0 | |||
3128 | // TODO: GEH - add this check for Linux* OS / OS X* as well? | |||
3129 | if (!__kmp_is_thread_alive(thread, &exit_val)) { | |||
3130 | thread->th.th_task_team = NULL__null; | |||
3131 | continue; | |||
3132 | } | |||
3133 | #endif | |||
3134 | ||||
3135 | done = FALSE0; // Because th_task_team pointer is not NULL for this thread | |||
3136 | ||||
3137 | KA_TRACE(10, ("__kmp_wait_to_unref_task_team: Waiting for T#%d to "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3139)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3138 | "unreference task_team\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3139)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3139 | __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3139)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
3140 | ||||
3141 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | |||
3142 | volatile void *sleep_loc; | |||
3143 | // If the thread is sleeping, awaken it. | |||
3144 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
3145 | NULL__null) { | |||
3146 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3147 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3148 | ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid)); } | |||
3149 | __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3149)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
3150 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3150)), (thread)->th.th_info.ds.ds_gtid), sleep_loc); | |||
3151 | } | |||
3152 | } | |||
3153 | } | |||
3154 | if (done) { | |||
3155 | break; | |||
3156 | } | |||
3157 | ||||
3158 | // If we are oversubscribed, or have waited a bit (and library mode is | |||
3159 | // throughput), yield. Pause is in the following code. | |||
3160 | KMP_YIELD(TCR_4(__kmp_nth) > __kmp_avail_proc){ __kmp_x86_pause(); __kmp_yield(((__kmp_nth) > __kmp_avail_proc )); }; | |||
3161 | KMP_YIELD_SPIN(spins){ __kmp_x86_pause(); (spins) -= 2; if (!(spins)) { __kmp_yield (1); (spins) = __kmp_yield_next; } }; // Yields only if KMP_LIBRARY=throughput | |||
3162 | } | |||
3163 | } | |||
3164 | ||||
3165 | // __kmp_task_team_setup: Create a task_team for the current team, but use | |||
3166 | // an already created, unused one if it already exists. | |||
3167 | void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) { | |||
3168 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3168)); | |||
3169 | ||||
3170 | // If this task_team hasn't been created yet, allocate it. It will be used in | |||
3171 | // the region after the next. | |||
3172 | // If it exists, it is the current task team and shouldn't be touched yet as | |||
3173 | // it may still be in use. | |||
3174 | if (team->t.t_task_team[this_thr->th.th_task_state] == NULL__null && | |||
3175 | (always || team->t.t_nproc > 1)) { | |||
3176 | team->t.t_task_team[this_thr->th.th_task_state] = | |||
3177 | __kmp_allocate_task_team(this_thr, team); | |||
3178 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created new task_team %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
3179 | "for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
3180 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
3181 | team->t.t_task_team[this_thr->th.th_task_state],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
3182 | ((team != NULL) ? team->t.t_id : -1),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
3183 | this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3180)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); }; | |||
3184 | } | |||
3185 | ||||
3186 | // After threads exit the release, they will call sync, and then point to this | |||
3187 | // other task_team; make sure it is allocated and properly initialized. As | |||
3188 | // threads spin in the barrier release phase, they will continue to use the | |||
3189 | // previous task_team struct(above), until they receive the signal to stop | |||
3190 | // checking for tasks (they can't safely reference the kmp_team_t struct, | |||
3191 | // which could be reallocated by the master thread). No task teams are formed | |||
3192 | // for serialized teams. | |||
3193 | if (team->t.t_nproc > 1) { | |||
3194 | int other_team = 1 - this_thr->th.th_task_state; | |||
3195 | if (team->t.t_task_team[other_team] == NULL__null) { // setup other team as well | |||
3196 | team->t.t_task_team[other_team] = | |||
3197 | __kmp_allocate_task_team(this_thr, team); | |||
3198 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created second new "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3199 | "task_team %p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3200 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3201 | team->t.t_task_team[other_team],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3202 | ((team != NULL) ? team->t.t_id : -1), other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); }; | |||
3203 | } else { // Leave the old task team struct in place for the upcoming region; | |||
3204 | // adjust as needed | |||
3205 | kmp_task_team_t *task_team = team->t.t_task_team[other_team]; | |||
3206 | if (!task_team->tt.tt_active || | |||
3207 | team->t.t_nproc != task_team->tt.tt_nproc) { | |||
3208 | TCW_4(task_team->tt.tt_nproc, team->t.t_nproc)(task_team->tt.tt_nproc) = (team->t.t_nproc); | |||
3209 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
3210 | #if OMP_45_ENABLED(50 >= 45) | |||
3211 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
3212 | #endif | |||
3213 | KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads,(&task_team->tt.tt_unfinished_threads)->store(team-> t.t_nproc, std::memory_order_release) | |||
3214 | team->t.t_nproc)(&task_team->tt.tt_unfinished_threads)->store(team-> t.t_nproc, std::memory_order_release); | |||
3215 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
3216 | } | |||
3217 | // if team size has changed, the first thread to enable tasking will | |||
3218 | // realloc threads_data if necessary | |||
3219 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d reset next task_team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3220 | "%p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3221 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3222 | team->t.t_task_team[other_team],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
3223 | ((team != NULL) ? team->t.t_id : -1), other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); }; | |||
3224 | } | |||
3225 | } | |||
3226 | } | |||
3227 | ||||
3228 | // __kmp_task_team_sync: Propagation of task team data from team to threads | |||
3229 | // which happens just after the release phase of a team barrier. This may be | |||
3230 | // called by any thread, but only for teams with # threads > 1. | |||
3231 | void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) { | |||
3232 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3232)); | |||
3233 | ||||
3234 | // Toggle the th_task_state field, to switch which task_team this thread | |||
3235 | // refers to | |||
3236 | this_thr->th.th_task_state = 1 - this_thr->th.th_task_state; | |||
3237 | // It is now safe to propagate the task team pointer from the team struct to | |||
3238 | // the current thread. | |||
3239 | TCW_PTR(this_thr->th.th_task_team,((this_thr->th.th_task_team)) = ((team->t.t_task_team[this_thr ->th.th_task_state])) | |||
3240 | team->t.t_task_team[this_thr->th.th_task_state])((this_thr->th.th_task_team)) = ((team->t.t_task_team[this_thr ->th.th_task_state])); | |||
3241 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3244)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
3242 | ("__kmp_task_team_sync: Thread T#%d task team switched to task_team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3244)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
3243 | "%p from Team #%d (parity=%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3244)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
3244 | __kmp_gtid_from_thread(this_thr), this_thr->th.th_task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3244)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
3245 | ((team != NULL) ? team->t.t_id : -1), this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3244)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); }; | |||
3246 | } | |||
3247 | ||||
3248 | // __kmp_task_team_wait: Master thread waits for outstanding tasks after the | |||
3249 | // barrier gather phase. Only called by master thread if #threads in team > 1 or | |||
3250 | // if proxy tasks were created. | |||
3251 | // | |||
3252 | // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off | |||
3253 | // by passing in 0 optionally as the last argument. When wait is zero, master | |||
3254 | // thread does not wait for unfinished_threads to reach 0. | |||
3255 | void __kmp_task_team_wait( | |||
3256 | kmp_info_t *this_thr, | |||
3257 | kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, int wait) { | |||
3258 | kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state]; | |||
3259 | ||||
3260 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3260)); | |||
3261 | KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team)((task_team == this_thr->th.th_task_team) ? 0 : __kmp_debug_assert ("task_team == this_thr->th.th_task_team", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3261)); | |||
3262 | ||||
3263 | if ((task_team != NULL__null) && KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
3264 | if (wait) { | |||
3265 | KA_TRACE(20, ("__kmp_task_team_wait: Master T#%d waiting for all tasks "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3267)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3266 | "(for unfinished_threads to reach 0) on task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3267)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3267 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3267)), (this_thr)->th.th_info.ds.ds_gtid), task_team); }; | |||
3268 | // Worker threads may have dropped through to release phase, but could | |||
3269 | // still be executing tasks. Wait here for tasks to complete. To avoid | |||
3270 | // memory contention, only master thread checks termination condition. | |||
3271 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads) | |||
3272 | &task_team->tt.tt_unfinished_threads)reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads), | |||
3273 | 0U); | |||
3274 | flag.wait(this_thr, TRUE(!0) USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj); | |||
3275 | } | |||
3276 | // Deactivate the old task team, so that the worker threads will stop | |||
3277 | // referencing it while spinning. | |||
3278 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3279 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3280 | ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3281 | "setting active to false, setting local and team's pointer to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
3282 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)), (this_thr)->th.th_info.ds.ds_gtid), task_team); }; | |||
3283 | #if OMP_45_ENABLED(50 >= 45) | |||
3284 | KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||((task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)) ? 0 : __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3285)) | |||
3285 | task_team->tt.tt_found_proxy_tasks == TRUE)((task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)) ? 0 : __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3285)); | |||
3286 | TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
3287 | #else | |||
3288 | KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1)((task_team->tt.tt_nproc > 1) ? 0 : __kmp_debug_assert( "task_team->tt.tt_nproc > 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3288)); | |||
3289 | #endif | |||
3290 | KMP_CHECK_UPDATE(task_team->tt.tt_untied_task_encountered, 0)if ((task_team->tt.tt_untied_task_encountered) != (0)) (task_team ->tt.tt_untied_task_encountered) = (0); | |||
3291 | TCW_SYNC_4(task_team->tt.tt_active, FALSE)(task_team->tt.tt_active) = (0); | |||
3292 | KMP_MB(); | |||
3293 | ||||
3294 | TCW_PTR(this_thr->th.th_task_team, NULL)((this_thr->th.th_task_team)) = ((__null)); | |||
3295 | } | |||
3296 | } | |||
3297 | ||||
3298 | // __kmp_tasking_barrier: | |||
3299 | // This routine may only called when __kmp_tasking_mode == tskm_extra_barrier. | |||
3300 | // Internal function to execute all tasks prior to a regular barrier or a join | |||
3301 | // barrier. It is a full barrier itself, which unfortunately turns regular | |||
3302 | // barriers into double barriers and join barriers into 1 1/2 barriers. | |||
3303 | void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) { | |||
3304 | std::atomic<kmp_uint32> *spin = RCAST(reinterpret_cast<std::atomic<kmp_uint32> *>(& team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads ) | |||
3305 | std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads ) | |||
3306 | &team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads)reinterpret_cast<std::atomic<kmp_uint32> *>(& team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads ); | |||
3307 | int flag = FALSE0; | |||
3308 | KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier)((__kmp_tasking_mode == tskm_extra_barrier) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode == tskm_extra_barrier", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3308)); | |||
3309 | ||||
3310 | #if USE_ITT_BUILD1 | |||
3311 | KMP_FSYNC_SPIN_INIT(spin, NULL)int sync_iters = 0; if (__kmp_itt_fsync_prepare_ptr__3_0) { if (spin == __null) { spin = __null; } } __asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0x4376) : "%ebx"); | |||
3312 | #endif /* USE_ITT_BUILD */ | |||
3313 | kmp_flag_32 spin_flag(spin, 0U); | |||
3314 | while (!spin_flag.execute_tasks(thread, gtid, TRUE(!0), | |||
3315 | &flag USE_ITT_BUILD_ARG(NULL), __null, 0)) { | |||
3316 | #if USE_ITT_BUILD1 | |||
3317 | // TODO: What about itt_sync_obj?? | |||
3318 | KMP_FSYNC_SPIN_PREPARE(RCAST(void *, spin))do { if (__kmp_itt_fsync_prepare_ptr__3_0 && sync_iters < __kmp_itt_prepare_delay) { ++sync_iters; if (sync_iters >= __kmp_itt_prepare_delay) { (!__kmp_itt_fsync_prepare_ptr__3_0 ) ? (void)0 : __kmp_itt_fsync_prepare_ptr__3_0((void *)((void *)reinterpret_cast<void *>(spin))); } } } while (0); | |||
3319 | #endif /* USE_ITT_BUILD */ | |||
3320 | ||||
3321 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
3322 | if (__kmp_global.g.g_abort) | |||
3323 | __kmp_abort_thread(); | |||
3324 | break; | |||
3325 | } | |||
3326 | KMP_YIELD(TRUE){ __kmp_x86_pause(); __kmp_yield(((!0))); }; // GH: We always yield here | |||
3327 | } | |||
3328 | #if USE_ITT_BUILD1 | |||
3329 | KMP_FSYNC_SPIN_ACQUIRED(RCAST(void *, spin))do { __asm__ __volatile__("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(0x4377) : "%ebx"); if (sync_iters >= __kmp_itt_prepare_delay ) { (!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)((void *)reinterpret_cast<void *>(spin))); } } while (0); | |||
3330 | #endif /* USE_ITT_BUILD */ | |||
3331 | } | |||
3332 | ||||
3333 | #if OMP_45_ENABLED(50 >= 45) | |||
3334 | ||||
3335 | // __kmp_give_task puts a task into a given thread queue if: | |||
3336 | // - the queue for that thread was created | |||
3337 | // - there's space in that queue | |||
3338 | // Because of this, __kmp_push_task needs to check if there's space after | |||
3339 | // getting the lock | |||
3340 | static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task, | |||
3341 | kmp_int32 pass) { | |||
3342 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
3343 | kmp_task_team_t *task_team = taskdata->td_task_team; | |||
3344 | ||||
3345 | KA_TRACE(20, ("__kmp_give_task: trying to give task %p to thread %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_give_task: trying to give task %p to thread %d.\n" , taskdata, tid); } | |||
3346 | taskdata, tid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_give_task: trying to give task %p to thread %d.\n" , taskdata, tid); }; | |||
3347 | ||||
3348 | // If task_team is NULL something went really bad... | |||
3349 | KMP_DEBUG_ASSERT(task_team != NULL)((task_team != __null) ? 0 : __kmp_debug_assert("task_team != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3349)); | |||
3350 | ||||
3351 | bool result = false; | |||
3352 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
3353 | ||||
3354 | if (thread_data->td.td_deque == NULL__null) { | |||
3355 | // There's no queue in this thread, go find another one | |||
3356 | // We're guaranteed that at least one thread has a queue | |||
3357 | KA_TRACE(30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: thread %d has no queue while giving task %p.\n" , tid, taskdata); } | |||
3358 | ("__kmp_give_task: thread %d has no queue while giving task %p.\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: thread %d has no queue while giving task %p.\n" , tid, taskdata); } | |||
3359 | tid, taskdata))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: thread %d has no queue while giving task %p.\n" , tid, taskdata); }; | |||
3360 | return result; | |||
3361 | } | |||
3362 | ||||
3363 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
3364 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
3365 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to thread %d.\n" , taskdata, tid); } | |||
3366 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to thread %d.\n" , taskdata, tid); } | |||
3367 | ("__kmp_give_task: queue is full while giving task %p to thread %d.\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to thread %d.\n" , taskdata, tid); } | |||
3368 | taskdata, tid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to thread %d.\n" , taskdata, tid); }; | |||
3369 | ||||
3370 | // if this deque is bigger than the pass ratio give a chance to another | |||
3371 | // thread | |||
3372 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
3373 | return result; | |||
3374 | ||||
3375 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
3376 | __kmp_realloc_task_deque(thread, thread_data); | |||
3377 | ||||
3378 | } else { | |||
3379 | ||||
3380 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
3381 | ||||
3382 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
3383 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
3384 | KA_TRACE(30, ("__kmp_give_task: queue is full while giving task %p to "if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to " "thread %d.\n", taskdata, tid); } | |||
3385 | "thread %d.\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to " "thread %d.\n", taskdata, tid); } | |||
3386 | taskdata, tid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: queue is full while giving task %p to " "thread %d.\n", taskdata, tid); }; | |||
3387 | ||||
3388 | // if this deque is bigger than the pass ratio give a chance to another | |||
3389 | // thread | |||
3390 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
3391 | goto release_and_exit; | |||
3392 | ||||
3393 | __kmp_realloc_task_deque(thread, thread_data); | |||
3394 | } | |||
3395 | } | |||
3396 | ||||
3397 | // lock is held here, and there is space in the deque | |||
3398 | ||||
3399 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata; | |||
3400 | // Wrap index. | |||
3401 | thread_data->td.td_deque_tail = | |||
3402 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
3403 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
3404 | TCR_4(thread_data->td.td_deque_ntasks) + 1)(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1); | |||
3405 | ||||
3406 | result = true; | |||
3407 | KA_TRACE(30, ("__kmp_give_task: successfully gave task %p to thread %d.\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: successfully gave task %p to thread %d.\n" , taskdata, tid); } | |||
3408 | taskdata, tid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: successfully gave task %p to thread %d.\n" , taskdata, tid); }; | |||
3409 | ||||
3410 | release_and_exit: | |||
3411 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
3412 | ||||
3413 | return result; | |||
3414 | } | |||
3415 | ||||
3416 | /* The finish of the proxy tasks is divided in two pieces: | |||
3417 | - the top half is the one that can be done from a thread outside the team | |||
3418 | - the bottom half must be run from a them within the team | |||
3419 | ||||
3420 | In order to run the bottom half the task gets queued back into one of the | |||
3421 | threads of the team. Once the td_incomplete_child_task counter of the parent | |||
3422 | is decremented the threads can leave the barriers. So, the bottom half needs | |||
3423 | to be queued before the counter is decremented. The top half is therefore | |||
3424 | divided in two parts: | |||
3425 | - things that can be run before queuing the bottom half | |||
3426 | - things that must be run after queuing the bottom half | |||
3427 | ||||
3428 | This creates a second race as the bottom half can free the task before the | |||
3429 | second top half is executed. To avoid this we use the | |||
3430 | td_incomplete_child_task of the proxy task to synchronize the top and bottom | |||
3431 | half. */ | |||
3432 | static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
3433 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3433)); | |||
3434 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3434)); | |||
3435 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3435)); | |||
3436 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3436)); | |||
3437 | ||||
3438 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
3439 | ||||
3440 | if (taskdata->td_taskgroup) | |||
3441 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
3442 | ||||
3443 | // Create an imaginary children for this task so the bottom half cannot | |||
3444 | // release the task before we have completed the second top half | |||
3445 | KMP_ATOMIC_INC(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->fetch_add(1 , std::memory_order_acq_rel); | |||
3446 | } | |||
3447 | ||||
3448 | static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
3449 | kmp_int32 children = 0; | |||
3450 | ||||
3451 | // Predecrement simulated by "- 1" calculation | |||
3452 | children = | |||
3453 | KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks)(&taskdata->td_parent->td_incomplete_child_tasks)-> fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
3454 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3454)); | |||
3455 | ||||
3456 | // Remove the imaginary children | |||
3457 | KMP_ATOMIC_DEC(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->fetch_sub(1 , std::memory_order_acq_rel); | |||
3458 | } | |||
3459 | ||||
3460 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) { | |||
3461 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
3462 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
3463 | ||||
3464 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3464)); | |||
3465 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3466)) | |||
3466 | 1)((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3466)); // top half must run before bottom half | |||
3467 | ||||
3468 | // We need to wait to make sure the top half is finished | |||
3469 | // Spinning here should be ok as this should happen quickly | |||
3470 | while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) > 0) | |||
3471 | ; | |||
3472 | ||||
3473 | __kmp_release_deps(gtid, taskdata); | |||
3474 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
3475 | } | |||
3476 | ||||
3477 | /*! | |||
3478 | @ingroup TASKING | |||
3479 | @param gtid Global Thread ID of encountering thread | |||
3480 | @param ptask Task which execution is completed | |||
3481 | ||||
3482 | Execute the completation of a proxy task from a thread of that is part of the | |||
3483 | team. Run first and bottom halves directly. | |||
3484 | */ | |||
3485 | void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) { | |||
3486 | KMP_DEBUG_ASSERT(ptask != NULL)((ptask != __null) ? 0 : __kmp_debug_assert("ptask != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3486)); | |||
3487 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
3488 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
3489 | 10, ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
3490 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
3491 | ||||
3492 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3492)); | |||
3493 | ||||
3494 | __kmp_first_top_half_finish_proxy(taskdata); | |||
3495 | __kmp_second_top_half_finish_proxy(taskdata); | |||
3496 | __kmp_bottom_half_finish_proxy(gtid, ptask); | |||
3497 | ||||
3498 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
3499 | ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
3500 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
3501 | } | |||
3502 | ||||
3503 | /*! | |||
3504 | @ingroup TASKING | |||
3505 | @param ptask Task which execution is completed | |||
3506 | ||||
3507 | Execute the completation of a proxy task from a thread that could not belong to | |||
3508 | the team. | |||
3509 | */ | |||
3510 | void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) { | |||
3511 | KMP_DEBUG_ASSERT(ptask != NULL)((ptask != __null) ? 0 : __kmp_debug_assert("ptask != __null" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3511)); | |||
3512 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
3513 | ||||
3514 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
3515 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
3516 | ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
3517 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); }; | |||
3518 | ||||
3519 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3519)); | |||
3520 | ||||
3521 | __kmp_first_top_half_finish_proxy(taskdata); | |||
3522 | ||||
3523 | // Enqueue task to complete bottom half completion from a thread within the | |||
3524 | // corresponding team | |||
3525 | kmp_team_t *team = taskdata->td_team; | |||
3526 | kmp_int32 nthreads = team->t.t_nproc; | |||
3527 | kmp_info_t *thread; | |||
3528 | ||||
3529 | // This should be similar to start_k = __kmp_get_random( thread ) % nthreads | |||
3530 | // but we cannot use __kmp_get_random here | |||
3531 | kmp_int32 start_k = 0; | |||
3532 | kmp_int32 pass = 1; | |||
3533 | kmp_int32 k = start_k; | |||
3534 | ||||
3535 | do { | |||
3536 | // For now we're just linearly trying to find a thread | |||
3537 | thread = team->t.t_threads[k]; | |||
3538 | k = (k + 1) % nthreads; | |||
3539 | ||||
3540 | // we did a full pass through all the threads | |||
3541 | if (k == start_k) | |||
3542 | pass = pass << 1; | |||
3543 | ||||
3544 | } while (!__kmp_give_task(thread, k, ptask, pass)); | |||
3545 | ||||
3546 | __kmp_second_top_half_finish_proxy(taskdata); | |||
3547 | ||||
3548 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
3549 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
3550 | ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
3551 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); }; | |||
3552 | } | |||
3553 | ||||
3554 | // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task | |||
3555 | // for taskloop | |||
3556 | // | |||
3557 | // thread: allocating thread | |||
3558 | // task_src: pointer to source task to be duplicated | |||
3559 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
3560 | kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) { | |||
3561 | kmp_task_t *task; | |||
3562 | kmp_taskdata_t *taskdata; | |||
3563 | kmp_taskdata_t *taskdata_src; | |||
3564 | kmp_taskdata_t *parent_task = thread->th.th_current_task; | |||
3565 | size_t shareds_offset; | |||
3566 | size_t task_size; | |||
3567 | ||||
3568 | KA_TRACE(10, ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n", thread,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n" , thread, task_src); } | |||
3569 | task_src))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n" , thread, task_src); }; | |||
3570 | taskdata_src = KMP_TASK_TO_TASKDATA(task_src)(((kmp_taskdata_t *)task_src) - 1); | |||
3571 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==((taskdata_src->td_flags.proxy == 0) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3572)) | |||
3572 | TASK_FULL)((taskdata_src->td_flags.proxy == 0) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3572)); // it should not be proxy task | |||
3573 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT)((taskdata_src->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3573)); | |||
3574 | task_size = taskdata_src->td_size_alloc; | |||
3575 | ||||
3576 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
3577 | KA_TRACE(30, ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n", thread,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n" , thread, task_size); } | |||
3578 | task_size))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n" , thread, task_size); }; | |||
3579 | #if USE_FAST_MEMORY3 | |||
3580 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size)___kmp_fast_allocate((thread), (task_size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3580); | |||
3581 | #else | |||
3582 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size)___kmp_thread_malloc((thread), (task_size), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3582); | |||
3583 | #endif /* USE_FAST_MEMORY */ | |||
3584 | KMP_MEMCPYmemcpy(taskdata, taskdata_src, task_size); | |||
3585 | ||||
3586 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
3587 | ||||
3588 | // Initialize new task (only specific fields not affected by memcpy) | |||
3589 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
3590 | if (task->shareds != NULL__null) { // need setup shareds pointer | |||
3591 | shareds_offset = (char *)task_src->shareds - (char *)taskdata_src; | |||
3592 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
3593 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3594)) | |||
3594 | 0)(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3594)); | |||
3595 | } | |||
3596 | taskdata->td_alloc_thread = thread; | |||
3597 | taskdata->td_parent = parent_task; | |||
3598 | taskdata->td_taskgroup = | |||
3599 | parent_task | |||
3600 | ->td_taskgroup; // task inherits the taskgroup from the parent task | |||
3601 | ||||
3602 | // Only need to keep track of child task counts if team parallel and tasking | |||
3603 | // not serialized | |||
3604 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) { | |||
3605 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
3606 | if (parent_task->td_taskgroup) | |||
3607 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
3608 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
3609 | // implicit not deallocated | |||
3610 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) | |||
3611 | KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks)(&taskdata->td_parent->td_allocated_child_tasks)-> fetch_add(1, std::memory_order_acq_rel); | |||
3612 | } | |||
3613 | ||||
3614 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n" , thread, taskdata, taskdata->td_parent); } | |||
3615 | ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n" , thread, taskdata, taskdata->td_parent); } | |||
3616 | thread, taskdata, taskdata->td_parent))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n" , thread, taskdata, taskdata->td_parent); }; | |||
3617 | #if OMPT_SUPPORT1 | |||
3618 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
3619 | __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid); | |||
3620 | #endif | |||
3621 | return task; | |||
3622 | } | |||
3623 | ||||
3624 | // Routine optionally generated by the compiler for setting the lastprivate flag | |||
3625 | // and calling needed constructors for private/firstprivate objects | |||
3626 | // (used to form taskloop tasks from pattern task) | |||
3627 | // Parameters: dest task, src task, lastprivate flag. | |||
3628 | typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); | |||
3629 | ||||
3630 | KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8)static_assert(sizeof(long) == 4 || sizeof(long) == 8, "Build condition error" ); | |||
3631 | ||||
3632 | // class to encapsulate manipulating loop bounds in a taskloop task. | |||
3633 | // this abstracts away the Intel vs GOMP taskloop interface for setting/getting | |||
3634 | // the loop bound variables. | |||
3635 | class kmp_taskloop_bounds_t { | |||
3636 | kmp_task_t *task; | |||
3637 | const kmp_taskdata_t *taskdata; | |||
3638 | size_t lower_offset; | |||
3639 | size_t upper_offset; | |||
3640 | ||||
3641 | public: | |||
3642 | kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub) | |||
3643 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1)), | |||
3644 | lower_offset((char *)lb - (char *)task), | |||
3645 | upper_offset((char *)ub - (char *)task) { | |||
3646 | KMP_DEBUG_ASSERT((char *)lb > (char *)_task)(((char *)lb > (char *)_task) ? 0 : __kmp_debug_assert("(char *)lb > (char *)_task" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3646)); | |||
3647 | KMP_DEBUG_ASSERT((char *)ub > (char *)_task)(((char *)ub > (char *)_task) ? 0 : __kmp_debug_assert("(char *)ub > (char *)_task" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3647)); | |||
3648 | } | |||
3649 | kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds) | |||
3650 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)(((kmp_taskdata_t *)_task) - 1)), | |||
3651 | lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {} | |||
3652 | size_t get_lower_offset() const { return lower_offset; } | |||
3653 | size_t get_upper_offset() const { return upper_offset; } | |||
3654 | kmp_uint64 get_lb() const { | |||
3655 | kmp_int64 retval; | |||
3656 | #if defined(KMP_GOMP_COMPAT) | |||
3657 | // Intel task just returns the lower bound normally | |||
3658 | if (!taskdata->td_flags.native) { | |||
3659 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
3660 | } else { | |||
3661 | // GOMP task has to take into account the sizeof(long) | |||
3662 | if (taskdata->td_size_loop_bounds == 4) { | |||
3663 | kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds); | |||
3664 | retval = (kmp_int64)*lb; | |||
3665 | } else { | |||
3666 | kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds); | |||
3667 | retval = (kmp_int64)*lb; | |||
3668 | } | |||
3669 | } | |||
3670 | #else | |||
3671 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
3672 | #endif // defined(KMP_GOMP_COMPAT) | |||
3673 | return retval; | |||
3674 | } | |||
3675 | kmp_uint64 get_ub() const { | |||
3676 | kmp_int64 retval; | |||
3677 | #if defined(KMP_GOMP_COMPAT) | |||
3678 | // Intel task just returns the upper bound normally | |||
3679 | if (!taskdata->td_flags.native) { | |||
3680 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
3681 | } else { | |||
3682 | // GOMP task has to take into account the sizeof(long) | |||
3683 | if (taskdata->td_size_loop_bounds == 4) { | |||
3684 | kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds) + 1; | |||
3685 | retval = (kmp_int64)*ub; | |||
3686 | } else { | |||
3687 | kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds) + 1; | |||
3688 | retval = (kmp_int64)*ub; | |||
3689 | } | |||
3690 | } | |||
3691 | #else | |||
3692 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
3693 | #endif // defined(KMP_GOMP_COMPAT) | |||
3694 | return retval; | |||
3695 | } | |||
3696 | void set_lb(kmp_uint64 lb) { | |||
3697 | #if defined(KMP_GOMP_COMPAT) | |||
3698 | // Intel task just sets the lower bound normally | |||
3699 | if (!taskdata->td_flags.native) { | |||
3700 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
3701 | } else { | |||
3702 | // GOMP task has to take into account the sizeof(long) | |||
3703 | if (taskdata->td_size_loop_bounds == 4) { | |||
3704 | kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds); | |||
3705 | *lower = (kmp_uint32)lb; | |||
3706 | } else { | |||
3707 | kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds); | |||
3708 | *lower = (kmp_uint64)lb; | |||
3709 | } | |||
3710 | } | |||
3711 | #else | |||
3712 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
3713 | #endif // defined(KMP_GOMP_COMPAT) | |||
3714 | } | |||
3715 | void set_ub(kmp_uint64 ub) { | |||
3716 | #if defined(KMP_GOMP_COMPAT) | |||
3717 | // Intel task just sets the upper bound normally | |||
3718 | if (!taskdata->td_flags.native) { | |||
3719 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
3720 | } else { | |||
3721 | // GOMP task has to take into account the sizeof(long) | |||
3722 | if (taskdata->td_size_loop_bounds == 4) { | |||
3723 | kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds) + 1; | |||
3724 | *upper = (kmp_uint32)ub; | |||
3725 | } else { | |||
3726 | kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds) + 1; | |||
3727 | *upper = (kmp_uint64)ub; | |||
3728 | } | |||
3729 | } | |||
3730 | #else | |||
3731 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
3732 | #endif // defined(KMP_GOMP_COMPAT) | |||
3733 | } | |||
3734 | }; | |||
3735 | ||||
3736 | // __kmp_taskloop_linear: Start tasks of the taskloop linearly | |||
3737 | // | |||
3738 | // loc Source location information | |||
3739 | // gtid Global thread ID | |||
3740 | // task Pattern task, exposes the loop iteration range | |||
3741 | // lb Pointer to loop lower bound in task structure | |||
3742 | // ub Pointer to loop upper bound in task structure | |||
3743 | // st Loop stride | |||
3744 | // ub_glob Global upper bound (used for lastprivate check) | |||
3745 | // num_tasks Number of tasks to execute | |||
3746 | // grainsize Number of loop iterations per task | |||
3747 | // extras Number of chunks with grainsize+1 iterations | |||
3748 | // tc Iterations count | |||
3749 | // task_dup Tasks duplication routine | |||
3750 | // codeptr_ra Return address for OMPT events | |||
3751 | void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task, | |||
3752 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
3753 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
3754 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
3755 | kmp_uint64 tc, | |||
3756 | #if OMPT_SUPPORT1 | |||
3757 | void *codeptr_ra, | |||
3758 | #endif | |||
3759 | void *task_dup) { | |||
3760 | KMP_COUNT_BLOCK(OMP_TASKLOOP)((void)0); | |||
3761 | KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling)((void)0); | |||
3762 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
3763 | // compiler provides global bounds here | |||
3764 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
3765 | kmp_uint64 lower = task_bounds.get_lb(); | |||
3766 | kmp_uint64 upper = task_bounds.get_ub(); | |||
3767 | kmp_uint64 i; | |||
3768 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
3769 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
3770 | kmp_task_t *next_task; | |||
3771 | kmp_int32 lastpriv = 0; | |||
3772 | ||||
3773 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3773)); | |||
3774 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3774)); | |||
3775 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3775)); | |||
3776 | KA_TRACE(20, ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
3777 | "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
3778 | gtid, num_tasks, grainsize, extras, lower, upper, ub_glob, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
3779 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); }; | |||
3780 | ||||
3781 | // Launch num_tasks tasks, assign grainsize iterations each task | |||
3782 | for (i = 0; i < num_tasks; ++i) { | |||
3783 | kmp_uint64 chunk_minus_1; | |||
3784 | if (extras == 0) { | |||
3785 | chunk_minus_1 = grainsize - 1; | |||
3786 | } else { | |||
3787 | chunk_minus_1 = grainsize; | |||
3788 | --extras; // first extras iterations get bigger chunk (grainsize+1) | |||
3789 | } | |||
3790 | upper = lower + st * chunk_minus_1; | |||
3791 | if (i == num_tasks - 1) { | |||
3792 | // schedule the last task, set lastprivate flag if needed | |||
3793 | if (st == 1) { // most common case | |||
3794 | KMP_DEBUG_ASSERT(upper == *ub)((upper == *ub) ? 0 : __kmp_debug_assert("upper == *ub", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3794)); | |||
3795 | if (upper == ub_glob) | |||
3796 | lastpriv = 1; | |||
3797 | } else if (st > 0) { // positive loop stride | |||
3798 | KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper)(((kmp_uint64)st > *ub - upper) ? 0 : __kmp_debug_assert("(kmp_uint64)st > *ub - upper" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3798)); | |||
3799 | if ((kmp_uint64)st > ub_glob - upper) | |||
3800 | lastpriv = 1; | |||
3801 | } else { // negative loop stride | |||
3802 | KMP_DEBUG_ASSERT(upper + st < *ub)((upper + st < *ub) ? 0 : __kmp_debug_assert("upper + st < *ub" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3802)); | |||
3803 | if (upper - ub_glob < (kmp_uint64)(-st)) | |||
3804 | lastpriv = 1; | |||
3805 | } | |||
3806 | } | |||
3807 | next_task = __kmp_task_dup_alloc(thread, task); // allocate new task | |||
3808 | kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task)(((kmp_taskdata_t *)next_task) - 1); | |||
3809 | kmp_taskloop_bounds_t next_task_bounds = | |||
3810 | kmp_taskloop_bounds_t(next_task, task_bounds); | |||
3811 | ||||
3812 | // adjust task-specific bounds | |||
3813 | next_task_bounds.set_lb(lower); | |||
3814 | if (next_taskdata->td_flags.native) { | |||
3815 | next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1)); | |||
3816 | } else { | |||
3817 | next_task_bounds.set_ub(upper); | |||
3818 | } | |||
3819 | if (ptask_dup != NULL__null) // set lastprivate flag, construct fistprivates, etc. | |||
3820 | ptask_dup(next_task, task, lastpriv); | |||
3821 | KA_TRACE(40,if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); } | |||
3822 | ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, "if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); } | |||
3823 | "upper %lld stride %lld, (offsets %p %p)\n",if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); } | |||
3824 | gtid, i, next_task, lower, upper, st,if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); } | |||
3825 | next_task_bounds.get_lower_offset(),if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); } | |||
3826 | next_task_bounds.get_upper_offset()))if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, " "upper %lld stride %lld, (offsets %p %p)\n", gtid, i, next_task , lower, upper, st, next_task_bounds.get_lower_offset(), next_task_bounds .get_upper_offset()); }; | |||
3827 | #if OMPT_SUPPORT1 | |||
3828 | __kmp_omp_taskloop_task(NULL__null, gtid, next_task, | |||
3829 | codeptr_ra); // schedule new task | |||
3830 | #else | |||
3831 | __kmp_omp_task(gtid, next_task, true); // schedule new task | |||
3832 | #endif | |||
3833 | lower = upper + st; // adjust lower bound for the next iteration | |||
3834 | } | |||
3835 | // free the pattern task and exit | |||
3836 | __kmp_task_start(gtid, task, current_task); // make internal bookkeeping | |||
3837 | // do not execute the pattern task, just do internal bookkeeping | |||
3838 | __kmp_task_finish<false>(gtid, task, current_task); | |||
3839 | } | |||
3840 | ||||
3841 | // Structure to keep taskloop parameters for auxiliary task | |||
3842 | // kept in the shareds of the task structure. | |||
3843 | typedef struct __taskloop_params { | |||
3844 | kmp_task_t *task; | |||
3845 | kmp_uint64 *lb; | |||
3846 | kmp_uint64 *ub; | |||
3847 | void *task_dup; | |||
3848 | kmp_int64 st; | |||
3849 | kmp_uint64 ub_glob; | |||
3850 | kmp_uint64 num_tasks; | |||
3851 | kmp_uint64 grainsize; | |||
3852 | kmp_uint64 extras; | |||
3853 | kmp_uint64 tc; | |||
3854 | kmp_uint64 num_t_min; | |||
3855 | #if OMPT_SUPPORT1 | |||
3856 | void *codeptr_ra; | |||
3857 | #endif | |||
3858 | } __taskloop_params_t; | |||
3859 | ||||
3860 | void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *, | |||
3861 | kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64, | |||
3862 | kmp_uint64, kmp_uint64, kmp_uint64, kmp_uint64, | |||
3863 | #if OMPT_SUPPORT1 | |||
3864 | void *, | |||
3865 | #endif | |||
3866 | void *); | |||
3867 | ||||
3868 | // Execute part of the the taskloop submitted as a task. | |||
3869 | int __kmp_taskloop_task(int gtid, void *ptask) { | |||
3870 | __taskloop_params_t *p = | |||
3871 | (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds; | |||
3872 | kmp_task_t *task = p->task; | |||
3873 | kmp_uint64 *lb = p->lb; | |||
3874 | kmp_uint64 *ub = p->ub; | |||
3875 | void *task_dup = p->task_dup; | |||
3876 | // p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
3877 | kmp_int64 st = p->st; | |||
3878 | kmp_uint64 ub_glob = p->ub_glob; | |||
3879 | kmp_uint64 num_tasks = p->num_tasks; | |||
3880 | kmp_uint64 grainsize = p->grainsize; | |||
3881 | kmp_uint64 extras = p->extras; | |||
3882 | kmp_uint64 tc = p->tc; | |||
3883 | kmp_uint64 num_t_min = p->num_t_min; | |||
3884 | #if OMPT_SUPPORT1 | |||
3885 | void *codeptr_ra = p->codeptr_ra; | |||
3886 | #endif | |||
3887 | #if KMP_DEBUG1 | |||
3888 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
3889 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3889)); | |||
3890 | KA_TRACE(20, ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3891 | " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3892 | gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3893 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); }; | |||
3894 | #endif | |||
3895 | KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min)((num_tasks * 2 + 1 > num_t_min) ? 0 : __kmp_debug_assert( "num_tasks * 2 + 1 > num_t_min", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3895)); | |||
3896 | if (num_tasks > num_t_min) | |||
3897 | __kmp_taskloop_recur(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
3898 | grainsize, extras, tc, num_t_min, | |||
3899 | #if OMPT_SUPPORT1 | |||
3900 | codeptr_ra, | |||
3901 | #endif | |||
3902 | task_dup); | |||
3903 | else | |||
3904 | __kmp_taskloop_linear(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
3905 | grainsize, extras, tc, | |||
3906 | #if OMPT_SUPPORT1 | |||
3907 | codeptr_ra, | |||
3908 | #endif | |||
3909 | task_dup); | |||
3910 | ||||
3911 | KA_TRACE(40, ("__kmp_taskloop_task(exit): T#%d\n", gtid))if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_task(exit): T#%d\n" , gtid); }; | |||
3912 | return 0; | |||
3913 | } | |||
3914 | ||||
3915 | // Schedule part of the the taskloop as a task, | |||
3916 | // execute the rest of the the taskloop. | |||
3917 | // | |||
3918 | // loc Source location information | |||
3919 | // gtid Global thread ID | |||
3920 | // task Pattern task, exposes the loop iteration range | |||
3921 | // lb Pointer to loop lower bound in task structure | |||
3922 | // ub Pointer to loop upper bound in task structure | |||
3923 | // st Loop stride | |||
3924 | // ub_glob Global upper bound (used for lastprivate check) | |||
3925 | // num_tasks Number of tasks to execute | |||
3926 | // grainsize Number of loop iterations per task | |||
3927 | // extras Number of chunks with grainsize+1 iterations | |||
3928 | // tc Iterations count | |||
3929 | // num_t_min Threashold to launch tasks recursively | |||
3930 | // task_dup Tasks duplication routine | |||
3931 | // codeptr_ra Return address for OMPT events | |||
3932 | void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task, | |||
3933 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
3934 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
3935 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
3936 | kmp_uint64 tc, kmp_uint64 num_t_min, | |||
3937 | #if OMPT_SUPPORT1 | |||
3938 | void *codeptr_ra, | |||
3939 | #endif | |||
3940 | void *task_dup) { | |||
3941 | #if KMP_DEBUG1 | |||
3942 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
3943 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3943)); | |||
3944 | KMP_DEBUG_ASSERT(num_tasks > num_t_min)((num_tasks > num_t_min) ? 0 : __kmp_debug_assert("num_tasks > num_t_min" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3944)); | |||
3945 | KA_TRACE(20, ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3946 | " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3947 | gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
3948 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); }; | |||
3949 | #endif | |||
3950 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
3951 | kmp_uint64 lower = *lb; | |||
3952 | kmp_uint64 upper = *ub; | |||
3953 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
3954 | // kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
3955 | kmp_task_t *next_task; | |||
3956 | kmp_int32 lastpriv = 0; | |||
3957 | size_t lower_offset = | |||
3958 | (char *)lb - (char *)task; // remember offset of lb in the task structure | |||
3959 | size_t upper_offset = | |||
3960 | (char *)ub - (char *)task; // remember offset of ub in the task structure | |||
3961 | ||||
3962 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3962)); | |||
3963 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3963)); | |||
3964 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 3964)); | |||
3965 | ||||
3966 | // split the loop in two halves | |||
3967 | kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1; | |||
3968 | kmp_uint64 gr_size0 = grainsize; | |||
3969 | kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute | |||
3970 | kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task | |||
3971 | if (n_tsk0 <= extras) { | |||
3972 | gr_size0++; // integrate extras into grainsize | |||
3973 | ext0 = 0; // no extra iters in 1st half | |||
3974 | ext1 = extras - n_tsk0; // remaining extras | |||
3975 | tc0 = gr_size0 * n_tsk0; | |||
3976 | tc1 = tc - tc0; | |||
3977 | } else { // n_tsk0 > extras | |||
3978 | ext1 = 0; // no extra iters in 2nd half | |||
3979 | ext0 = extras; | |||
3980 | tc1 = grainsize * n_tsk1; | |||
3981 | tc0 = tc - tc1; | |||
3982 | } | |||
3983 | ub0 = lower + st * (tc0 - 1); | |||
3984 | lb1 = ub0 + st; | |||
3985 | ||||
3986 | // create pattern task for 2nd half of the loop | |||
3987 | next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task | |||
3988 | // adjust lower bound (upper bound is not changed) for the 2nd half | |||
3989 | *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1; | |||
3990 | if (ptask_dup != NULL__null) // construct fistprivates, etc. | |||
3991 | ptask_dup(next_task, task, 0); | |||
3992 | *ub = ub0; // adjust upper bound for the 1st half | |||
3993 | ||||
3994 | // create auxiliary task for 2nd half of the loop | |||
3995 | kmp_task_t *new_task = | |||
3996 | __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *), | |||
3997 | sizeof(__taskloop_params_t), &__kmp_taskloop_task); | |||
3998 | __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds; | |||
3999 | p->task = next_task; | |||
4000 | p->lb = (kmp_uint64 *)((char *)next_task + lower_offset); | |||
4001 | p->ub = (kmp_uint64 *)((char *)next_task + upper_offset); | |||
4002 | p->task_dup = task_dup; | |||
4003 | p->st = st; | |||
4004 | p->ub_glob = ub_glob; | |||
4005 | p->num_tasks = n_tsk1; | |||
4006 | p->grainsize = grainsize; | |||
4007 | p->extras = ext1; | |||
4008 | p->tc = tc1; | |||
4009 | p->num_t_min = num_t_min; | |||
4010 | #if OMPT_SUPPORT1 | |||
4011 | p->codeptr_ra = codeptr_ra; | |||
4012 | #endif | |||
4013 | ||||
4014 | #if OMPT_SUPPORT1 | |||
4015 | // schedule new task with correct return address for OMPT events | |||
4016 | __kmp_omp_taskloop_task(NULL__null, gtid, new_task, codeptr_ra); | |||
4017 | #else | |||
4018 | __kmp_omp_task(NULL__null, gtid, new_task); // schedule new task | |||
4019 | #endif | |||
4020 | ||||
4021 | // execute the 1st half of current subrange | |||
4022 | if (n_tsk0 > num_t_min) | |||
4023 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0, | |||
4024 | ext0, tc0, num_t_min, | |||
4025 | #if OMPT_SUPPORT1 | |||
4026 | codeptr_ra, | |||
4027 | #endif | |||
4028 | task_dup); | |||
4029 | else | |||
4030 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, | |||
4031 | gr_size0, ext0, tc0, | |||
4032 | #if OMPT_SUPPORT1 | |||
4033 | codeptr_ra, | |||
4034 | #endif | |||
4035 | task_dup); | |||
4036 | ||||
4037 | KA_TRACE(40, ("__kmpc_taskloop_recur(exit): T#%d\n", gtid))if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmpc_taskloop_recur(exit): T#%d\n" , gtid); }; | |||
4038 | } | |||
4039 | ||||
4040 | /*! | |||
4041 | @ingroup TASKING | |||
4042 | @param loc Source location information | |||
4043 | @param gtid Global thread ID | |||
4044 | @param task Task structure | |||
4045 | @param if_val Value of the if clause | |||
4046 | @param lb Pointer to loop lower bound in task structure | |||
4047 | @param ub Pointer to loop upper bound in task structure | |||
4048 | @param st Loop stride | |||
4049 | @param nogroup Flag, 1 if nogroup clause specified, 0 otherwise | |||
4050 | @param sched Schedule specified 0/1/2 for none/grainsize/num_tasks | |||
4051 | @param grainsize Schedule value if specified | |||
4052 | @param task_dup Tasks duplication routine | |||
4053 | ||||
4054 | Execute the taskloop construct. | |||
4055 | */ | |||
4056 | void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, | |||
4057 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, | |||
4058 | int sched, kmp_uint64 grainsize, void *task_dup) { | |||
4059 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
4060 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 4060)); | |||
4061 | ||||
4062 | if (nogroup == 0) { | |||
4063 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
4064 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
4065 | #endif | |||
4066 | __kmpc_taskgroup(loc, gtid); | |||
4067 | } | |||
4068 | ||||
4069 | // ========================================================================= | |||
4070 | // calculate loop parameters | |||
4071 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
4072 | kmp_uint64 tc; | |||
4073 | // compiler provides global bounds here | |||
4074 | kmp_uint64 lower = task_bounds.get_lb(); | |||
4075 | kmp_uint64 upper = task_bounds.get_ub(); | |||
4076 | kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag | |||
4077 | kmp_uint64 num_tasks = 0, extras = 0; | |||
4078 | kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks; | |||
4079 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
4080 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
4081 | ||||
4082 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); } | |||
4083 | "grain %llu(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); } | |||
4084 | gtid, taskdata, lower, upper, st, grainsize, sched, task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); }; | |||
4085 | ||||
4086 | // compute trip count | |||
4087 | if (st == 1) { // most common case | |||
4088 | tc = upper - lower + 1; | |||
4089 | } else if (st < 0) { | |||
4090 | tc = (lower - upper) / (-st) + 1; | |||
4091 | } else { // st > 0 | |||
4092 | tc = (upper - lower) / st + 1; | |||
4093 | } | |||
4094 | if (tc == 0) { | |||
4095 | KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d zero-trip loop\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop(exit): T#%d zero-trip loop\n" , gtid); }; | |||
4096 | // free the pattern task and exit | |||
4097 | __kmp_task_start(gtid, task, current_task); | |||
4098 | // do not execute anything for zero-trip loop | |||
4099 | __kmp_task_finish<false>(gtid, task, current_task); | |||
4100 | return; | |||
4101 | } | |||
4102 | ||||
4103 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
4104 | ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL__null); | |||
4105 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
4106 | if (ompt_enabled.ompt_callback_work) { | |||
4107 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
4108 | ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data), | |||
4109 | &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
4110 | } | |||
4111 | #endif | |||
4112 | ||||
4113 | if (num_tasks_min == 0) | |||
4114 | // TODO: can we choose better default heuristic? | |||
4115 | num_tasks_min = | |||
4116 | KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE)((thread->th.th_team_nproc * 10) < ((1 << 8)) ? ( thread->th.th_team_nproc * 10) : ((1 << 8))); | |||
4117 | ||||
4118 | // compute num_tasks/grainsize based on the input provided | |||
4119 | switch (sched) { | |||
4120 | case 0: // no schedule clause specified, we can choose the default | |||
4121 | // let's try to schedule (team_size*10) tasks | |||
4122 | grainsize = thread->th.th_team_nproc * 10; | |||
4123 | case 2: // num_tasks provided | |||
4124 | if (grainsize > tc) { | |||
4125 | num_tasks = tc; // too big num_tasks requested, adjust values | |||
4126 | grainsize = 1; | |||
4127 | extras = 0; | |||
4128 | } else { | |||
4129 | num_tasks = grainsize; | |||
4130 | grainsize = tc / num_tasks; | |||
4131 | extras = tc % num_tasks; | |||
4132 | } | |||
4133 | break; | |||
4134 | case 1: // grainsize provided | |||
4135 | if (grainsize > tc) { | |||
4136 | num_tasks = 1; // too big grainsize requested, adjust values | |||
4137 | grainsize = tc; | |||
4138 | extras = 0; | |||
4139 | } else { | |||
4140 | num_tasks = tc / grainsize; | |||
4141 | // adjust grainsize for balanced distribution of iterations | |||
4142 | grainsize = tc / num_tasks; | |||
4143 | extras = tc % num_tasks; | |||
4144 | } | |||
4145 | break; | |||
4146 | default: | |||
4147 | KMP_ASSERT2(0, "unknown scheduling of taskloop")((0) ? 0 : __kmp_debug_assert(("unknown scheduling of taskloop" ), "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 4147)); | |||
4148 | } | |||
4149 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 4149)); | |||
4150 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 4150)); | |||
4151 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn338205/projects/openmp/runtime/src/kmp_tasking.cpp" , 4151)); | |||
4152 | // ========================================================================= | |||
4153 | ||||
4154 | // check if clause value first | |||
4155 | // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native) | |||
4156 | if (if_val == 0) { // if(0) specified, mark task as serial | |||
4157 | taskdata->td_flags.task_serial = 1; | |||
4158 | taskdata->td_flags.tiedness = TASK_TIED1; // AC: serial task cannot be untied | |||
4159 | // always start serial tasks linearly | |||
4160 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
4161 | grainsize, extras, tc, | |||
4162 | #if OMPT_SUPPORT1 | |||
4163 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
4164 | #endif | |||
4165 | task_dup); | |||
4166 | // !taskdata->td_flags.native => currently force linear spawning of tasks | |||
4167 | // for GOMP_taskloop | |||
4168 | } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) { | |||
4169 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
4170 | "(%lld), grain %llu, extras %llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
4171 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); }; | |||
4172 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
4173 | grainsize, extras, tc, num_tasks_min, | |||
4174 | #if OMPT_SUPPORT1 | |||
4175 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
4176 | #endif | |||
4177 | task_dup); | |||
4178 | } else { | |||
4179 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
4180 | "(%lld), grain %llu, extras %llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
4181 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); }; | |||
4182 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
4183 | grainsize, extras, tc, | |||
4184 | #if OMPT_SUPPORT1 | |||
4185 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
4186 | #endif | |||
4187 | task_dup); | |||
4188 | } | |||
4189 | ||||
4190 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
4191 | if (ompt_enabled.ompt_callback_work) { | |||
4192 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
4193 | ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data), | |||
4194 | &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
4195 | } | |||
4196 | #endif | |||
4197 | ||||
4198 | if (nogroup == 0) { | |||
4199 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
4200 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
4201 | #endif | |||
4202 | __kmpc_end_taskgroup(loc, gtid); | |||
4203 | } | |||
4204 | KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop(exit): T#%d\n" , gtid); }; | |||
4205 | } | |||
4206 | ||||
4207 | #endif |