| File: | build/source/openmp/runtime/src/kmp_tasking.cpp |
| Warning: | line 3715, column 30 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 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
| 8 | // See https://llvm.org/LICENSE.txt for license information. | |||
| 9 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
| 10 | // | |||
| 11 | //===----------------------------------------------------------------------===// | |||
| 12 | ||||
| 13 | #include "kmp.h" | |||
| 14 | #include "kmp_i18n.h" | |||
| 15 | #include "kmp_itt.h" | |||
| 16 | #include "kmp_stats.h" | |||
| 17 | #include "kmp_wait_release.h" | |||
| 18 | #include "kmp_taskdeps.h" | |||
| 19 | ||||
| 20 | #if OMPT_SUPPORT1 | |||
| 21 | #include "ompt-specific.h" | |||
| 22 | #endif | |||
| 23 | ||||
| 24 | #if ENABLE_LIBOMPTARGET1 | |||
| 25 | static void (*tgt_target_nowait_query)(void **); | |||
| 26 | ||||
| 27 | void __kmp_init_target_task() { | |||
| 28 | *(void **)(&tgt_target_nowait_query) = KMP_DLSYM("__tgt_target_nowait_query")dlsym(((void *) 0), "__tgt_target_nowait_query"); | |||
| 29 | } | |||
| 30 | #endif | |||
| 31 | ||||
| 32 | /* forward declaration */ | |||
| 33 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
| 34 | kmp_info_t *this_thr); | |||
| 35 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
| 36 | kmp_thread_data_t *thread_data); | |||
| 37 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
| 38 | kmp_task_team_t *task_team); | |||
| 39 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask); | |||
| 40 | ||||
| 41 | #ifdef BUILD_TIED_TASK_STACK | |||
| 42 | ||||
| 43 | // __kmp_trace_task_stack: print the tied tasks from the task stack in order | |||
| 44 | // from top do bottom | |||
| 45 | // | |||
| 46 | // gtid: global thread identifier for thread containing stack | |||
| 47 | // thread_data: thread data for task team thread containing stack | |||
| 48 | // threshold: value above which the trace statement triggers | |||
| 49 | // location: string identifying call site of this function (for trace) | |||
| 50 | static void __kmp_trace_task_stack(kmp_int32 gtid, | |||
| 51 | kmp_thread_data_t *thread_data, | |||
| 52 | int threshold, char *location) { | |||
| 53 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 54 | kmp_taskdata_t **stack_top = task_stack->ts_top; | |||
| 55 | kmp_int32 entries = task_stack->ts_entries; | |||
| 56 | kmp_taskdata_t *tied_task; | |||
| 57 | ||||
| 58 | 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); } | |||
| 59 | 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); } | |||
| 60 | ("__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); } | |||
| 61 | "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); } | |||
| 62 | 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); }; | |||
| 63 | ||||
| 64 | KMP_DEBUG_ASSERT(stack_top != NULL)if (!(stack_top != __null)) { __kmp_debug_assert("stack_top != __null" , "openmp/runtime/src/kmp_tasking.cpp", 64); }; | |||
| 65 | KMP_DEBUG_ASSERT(entries > 0)if (!(entries > 0)) { __kmp_debug_assert("entries > 0", "openmp/runtime/src/kmp_tasking.cpp", 65); }; | |||
| 66 | ||||
| 67 | while (entries != 0) { | |||
| 68 | KMP_DEBUG_ASSERT(stack_top != &task_stack->ts_first_block.sb_block[0])if (!(stack_top != &task_stack->ts_first_block.sb_block [0])) { __kmp_debug_assert("stack_top != &task_stack->ts_first_block.sb_block[0]" , "openmp/runtime/src/kmp_tasking.cpp", 68); }; | |||
| 69 | // fix up ts_top if we need to pop from previous block | |||
| 70 | if (entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 71 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(stack_top); | |||
| 72 | ||||
| 73 | stack_block = stack_block->sb_prev; | |||
| 74 | stack_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
| 75 | } | |||
| 76 | ||||
| 77 | // finish bookkeeping | |||
| 78 | stack_top--; | |||
| 79 | entries--; | |||
| 80 | ||||
| 81 | tied_task = *stack_top; | |||
| 82 | ||||
| 83 | KMP_DEBUG_ASSERT(tied_task != NULL)if (!(tied_task != __null)) { __kmp_debug_assert("tied_task != __null" , "openmp/runtime/src/kmp_tasking.cpp", 83); }; | |||
| 84 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)if (!(tied_task->td_flags.tasktype == 1)) { __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 84); }; | |||
| 85 | ||||
| 86 | 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); } | |||
| 87 | ("__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); } | |||
| 88 | "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); } | |||
| 89 | 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); }; | |||
| 90 | } | |||
| 91 | KMP_DEBUG_ASSERT(stack_top == &task_stack->ts_first_block.sb_block[0])if (!(stack_top == &task_stack->ts_first_block.sb_block [0])) { __kmp_debug_assert("stack_top == &task_stack->ts_first_block.sb_block[0]" , "openmp/runtime/src/kmp_tasking.cpp", 91); }; | |||
| 92 | ||||
| 93 | KA_TRACE(threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); } | |||
| 94 | ("__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); } | |||
| 95 | location, gtid))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); }; | |||
| 96 | } | |||
| 97 | ||||
| 98 | // __kmp_init_task_stack: initialize the task stack for the first time | |||
| 99 | // after a thread_data structure is created. | |||
| 100 | // It should not be necessary to do this again (assuming the stack works). | |||
| 101 | // | |||
| 102 | // gtid: global thread identifier of calling thread | |||
| 103 | // thread_data: thread data for task team thread containing stack | |||
| 104 | static void __kmp_init_task_stack(kmp_int32 gtid, | |||
| 105 | kmp_thread_data_t *thread_data) { | |||
| 106 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 107 | kmp_stack_block_t *first_block; | |||
| 108 | ||||
| 109 | // set up the first block of the stack | |||
| 110 | first_block = &task_stack->ts_first_block; | |||
| 111 | task_stack->ts_top = (kmp_taskdata_t **)first_block; | |||
| 112 | memset((void *)first_block, '\0', | |||
| 113 | TASK_STACK_BLOCK_SIZE * sizeof(kmp_taskdata_t *)); | |||
| 114 | ||||
| 115 | // initialize the stack to be empty | |||
| 116 | task_stack->ts_entries = TASK_STACK_EMPTY; | |||
| 117 | first_block->sb_next = NULL__null; | |||
| 118 | first_block->sb_prev = NULL__null; | |||
| 119 | } | |||
| 120 | ||||
| 121 | // __kmp_free_task_stack: free the task stack when thread_data is destroyed. | |||
| 122 | // | |||
| 123 | // gtid: global thread identifier for calling thread | |||
| 124 | // thread_data: thread info for thread containing stack | |||
| 125 | static void __kmp_free_task_stack(kmp_int32 gtid, | |||
| 126 | kmp_thread_data_t *thread_data) { | |||
| 127 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 128 | kmp_stack_block_t *stack_block = &task_stack->ts_first_block; | |||
| 129 | ||||
| 130 | KMP_DEBUG_ASSERT(task_stack->ts_entries == TASK_STACK_EMPTY)if (!(task_stack->ts_entries == TASK_STACK_EMPTY)) { __kmp_debug_assert ("task_stack->ts_entries == TASK_STACK_EMPTY", "openmp/runtime/src/kmp_tasking.cpp" , 130); }; | |||
| 131 | // free from the second block of the stack | |||
| 132 | while (stack_block != NULL__null) { | |||
| 133 | kmp_stack_block_t *next_block = (stack_block) ? stack_block->sb_next : NULL__null; | |||
| 134 | ||||
| 135 | stack_block->sb_next = NULL__null; | |||
| 136 | stack_block->sb_prev = NULL__null; | |||
| 137 | if (stack_block != &task_stack->ts_first_block) { | |||
| 138 | __kmp_thread_free(thread,___kmp_thread_free((thread), (stack_block), "openmp/runtime/src/kmp_tasking.cpp" , 139) | |||
| 139 | stack_block)___kmp_thread_free((thread), (stack_block), "openmp/runtime/src/kmp_tasking.cpp" , 139); // free the block, if not the first | |||
| 140 | } | |||
| 141 | stack_block = next_block; | |||
| 142 | } | |||
| 143 | // initialize the stack to be empty | |||
| 144 | task_stack->ts_entries = 0; | |||
| 145 | task_stack->ts_top = NULL__null; | |||
| 146 | } | |||
| 147 | ||||
| 148 | // __kmp_push_task_stack: Push the tied task onto the task stack. | |||
| 149 | // Grow the stack if necessary by allocating another block. | |||
| 150 | // | |||
| 151 | // gtid: global thread identifier for calling thread | |||
| 152 | // thread: thread info for thread containing stack | |||
| 153 | // tied_task: the task to push on the stack | |||
| 154 | static void __kmp_push_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
| 155 | kmp_taskdata_t *tied_task) { | |||
| 156 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
| 157 | kmp_thread_data_t *thread_data = | |||
| 158 | &thread->th.th_task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)]; | |||
| 159 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 160 | ||||
| 161 | if (tied_task->td_flags.team_serial || tied_task->td_flags.tasking_ser) { | |||
| 162 | return; // Don't push anything on stack if team or team tasks are serialized | |||
| 163 | } | |||
| 164 | ||||
| 165 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)if (!(tied_task->td_flags.tasktype == 1)) { __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 165); }; | |||
| 166 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)if (!(task_stack->ts_top != __null)) { __kmp_debug_assert( "task_stack->ts_top != __null", "openmp/runtime/src/kmp_tasking.cpp" , 166); }; | |||
| 167 | ||||
| 168 | 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); } | |||
| 169 | ("__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); } | |||
| 170 | 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); }; | |||
| 171 | // Store entry | |||
| 172 | *(task_stack->ts_top) = tied_task; | |||
| 173 | ||||
| 174 | // Do bookkeeping for next push | |||
| 175 | task_stack->ts_top++; | |||
| 176 | task_stack->ts_entries++; | |||
| 177 | ||||
| 178 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 179 | // Find beginning of this task block | |||
| 180 | kmp_stack_block_t *stack_block = | |||
| 181 | (kmp_stack_block_t *)(task_stack->ts_top - TASK_STACK_BLOCK_SIZE); | |||
| 182 | ||||
| 183 | // Check if we already have a block | |||
| 184 | if (stack_block->sb_next != | |||
| 185 | NULL__null) { // reset ts_top to beginning of next block | |||
| 186 | task_stack->ts_top = &stack_block->sb_next->sb_block[0]; | |||
| 187 | } else { // Alloc new block and link it up | |||
| 188 | kmp_stack_block_t *new_block = (kmp_stack_block_t *)__kmp_thread_calloc( | |||
| 189 | thread, sizeof(kmp_stack_block_t)); | |||
| 190 | ||||
| 191 | task_stack->ts_top = &new_block->sb_block[0]; | |||
| 192 | stack_block->sb_next = new_block; | |||
| 193 | new_block->sb_prev = stack_block; | |||
| 194 | new_block->sb_next = NULL__null; | |||
| 195 | ||||
| 196 | 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); } | |||
| 197 | 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); } | |||
| 198 | ("__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); } | |||
| 199 | 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); }; | |||
| 200 | } | |||
| 201 | } | |||
| 202 | 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); } | |||
| 203 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
| 204 | } | |||
| 205 | ||||
| 206 | // __kmp_pop_task_stack: Pop the tied task from the task stack. Don't return | |||
| 207 | // the task, just check to make sure it matches the ending task passed in. | |||
| 208 | // | |||
| 209 | // gtid: global thread identifier for the calling thread | |||
| 210 | // thread: thread info structure containing stack | |||
| 211 | // tied_task: the task popped off the stack | |||
| 212 | // ending_task: the task that is ending (should match popped task) | |||
| 213 | static void __kmp_pop_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
| 214 | kmp_taskdata_t *ending_task) { | |||
| 215 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
| 216 | kmp_thread_data_t *thread_data = | |||
| 217 | &thread->th.th_task_team->tt_threads_data[__kmp_tid_from_gtid(gtid)]; | |||
| 218 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 219 | kmp_taskdata_t *tied_task; | |||
| 220 | ||||
| 221 | if (ending_task->td_flags.team_serial || ending_task->td_flags.tasking_ser) { | |||
| 222 | // Don't pop anything from stack if team or team tasks are serialized | |||
| 223 | return; | |||
| 224 | } | |||
| 225 | ||||
| 226 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)if (!(task_stack->ts_top != __null)) { __kmp_debug_assert( "task_stack->ts_top != __null", "openmp/runtime/src/kmp_tasking.cpp" , 226); }; | |||
| 227 | KMP_DEBUG_ASSERT(task_stack->ts_entries > 0)if (!(task_stack->ts_entries > 0)) { __kmp_debug_assert ("task_stack->ts_entries > 0", "openmp/runtime/src/kmp_tasking.cpp" , 227); }; | |||
| 228 | ||||
| 229 | 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); } | |||
| 230 | thread))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n" , gtid, thread); }; | |||
| 231 | ||||
| 232 | // fix up ts_top if we need to pop from previous block | |||
| 233 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 234 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(task_stack->ts_top); | |||
| 235 | ||||
| 236 | stack_block = stack_block->sb_prev; | |||
| 237 | task_stack->ts_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
| 238 | } | |||
| 239 | ||||
| 240 | // finish bookkeeping | |||
| 241 | task_stack->ts_top--; | |||
| 242 | task_stack->ts_entries--; | |||
| 243 | ||||
| 244 | tied_task = *(task_stack->ts_top); | |||
| 245 | ||||
| 246 | KMP_DEBUG_ASSERT(tied_task != NULL)if (!(tied_task != __null)) { __kmp_debug_assert("tied_task != __null" , "openmp/runtime/src/kmp_tasking.cpp", 246); }; | |||
| 247 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)if (!(tied_task->td_flags.tasktype == 1)) { __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 247); }; | |||
| 248 | KMP_DEBUG_ASSERT(tied_task == ending_task)if (!(tied_task == ending_task)) { __kmp_debug_assert("tied_task == ending_task" , "openmp/runtime/src/kmp_tasking.cpp", 248); }; // If we built the stack correctly | |||
| 249 | ||||
| 250 | 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); } | |||
| 251 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
| 252 | return; | |||
| 253 | } | |||
| 254 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 255 | ||||
| 256 | // returns 1 if new task is allowed to execute, 0 otherwise | |||
| 257 | // checks Task Scheduling constraint (if requested) and | |||
| 258 | // mutexinoutset dependencies if any | |||
| 259 | static bool __kmp_task_is_allowed(int gtid, const kmp_int32 is_constrained, | |||
| 260 | const kmp_taskdata_t *tasknew, | |||
| 261 | const kmp_taskdata_t *taskcurr) { | |||
| 262 | if (is_constrained && (tasknew->td_flags.tiedness == TASK_TIED1)) { | |||
| 263 | // Check if the candidate obeys the Task Scheduling Constraints (TSC) | |||
| 264 | // only descendant of all deferred tied tasks can be scheduled, checking | |||
| 265 | // the last one is enough, as it in turn is the descendant of all others | |||
| 266 | kmp_taskdata_t *current = taskcurr->td_last_tied; | |||
| 267 | KMP_DEBUG_ASSERT(current != NULL)if (!(current != __null)) { __kmp_debug_assert("current != __null" , "openmp/runtime/src/kmp_tasking.cpp", 267); }; | |||
| 268 | // check if the task is not suspended on barrier | |||
| 269 | if (current->td_flags.tasktype == TASK_EXPLICIT1 || | |||
| 270 | current->td_taskwait_thread > 0) { // <= 0 on barrier | |||
| 271 | kmp_int32 level = current->td_level; | |||
| 272 | kmp_taskdata_t *parent = tasknew->td_parent; | |||
| 273 | while (parent != current && parent->td_level > level) { | |||
| 274 | // check generation up to the level of the current task | |||
| 275 | parent = parent->td_parent; | |||
| 276 | KMP_DEBUG_ASSERT(parent != NULL)if (!(parent != __null)) { __kmp_debug_assert("parent != __null" , "openmp/runtime/src/kmp_tasking.cpp", 276); }; | |||
| 277 | } | |||
| 278 | if (parent != current) | |||
| 279 | return false; | |||
| 280 | } | |||
| 281 | } | |||
| 282 | // Check mutexinoutset dependencies, acquire locks | |||
| 283 | kmp_depnode_t *node = tasknew->td_depnode; | |||
| 284 | if (UNLIKELY(node && (node->dn.mtx_num_locks > 0))__builtin_expect(!!(node && (node->dn.mtx_num_locks > 0)), 0)) { | |||
| 285 | for (int i = 0; i < node->dn.mtx_num_locks; ++i) { | |||
| 286 | KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL)if (!(node->dn.mtx_locks[i] != __null)) { __kmp_debug_assert ("node->dn.mtx_locks[i] != __null", "openmp/runtime/src/kmp_tasking.cpp" , 286); }; | |||
| 287 | if (__kmp_test_lock(node->dn.mtx_locks[i], gtid)) | |||
| 288 | continue; | |||
| 289 | // could not get the lock, release previous locks | |||
| 290 | for (int j = i - 1; j >= 0; --j) | |||
| 291 | __kmp_release_lock(node->dn.mtx_locks[j], gtid); | |||
| 292 | return false; | |||
| 293 | } | |||
| 294 | // negative num_locks means all locks acquired successfully | |||
| 295 | node->dn.mtx_num_locks = -node->dn.mtx_num_locks; | |||
| 296 | } | |||
| 297 | return true; | |||
| 298 | } | |||
| 299 | ||||
| 300 | // __kmp_realloc_task_deque: | |||
| 301 | // Re-allocates a task deque for a particular thread, copies the content from | |||
| 302 | // the old deque and adjusts the necessary data structures relating to the | |||
| 303 | // deque. This operation must be done with the deque_lock being held | |||
| 304 | static void __kmp_realloc_task_deque(kmp_info_t *thread, | |||
| 305 | kmp_thread_data_t *thread_data) { | |||
| 306 | kmp_int32 size = TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size); | |||
| 307 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == size)if (!((thread_data->td.td_deque_ntasks) == size)) { __kmp_debug_assert ("(thread_data->td.td_deque_ntasks) == size", "openmp/runtime/src/kmp_tasking.cpp" , 307); }; | |||
| 308 | kmp_int32 new_size = 2 * size; | |||
| 309 | ||||
| 310 | 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", __kmp_gtid_from_thread(thread), size , new_size, thread_data); } | |||
| 311 | "%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", __kmp_gtid_from_thread(thread), size , new_size, thread_data); } | |||
| 312 | __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", __kmp_gtid_from_thread(thread), size , new_size, thread_data); }; | |||
| 313 | ||||
| 314 | kmp_taskdata_t **new_deque = | |||
| 315 | (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *))___kmp_allocate((new_size * sizeof(kmp_taskdata_t *)), "openmp/runtime/src/kmp_tasking.cpp" , 315); | |||
| 316 | ||||
| 317 | int i, j; | |||
| 318 | for (i = thread_data->td.td_deque_head, j = 0; j < size; | |||
| 319 | i = (i + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1), j++) | |||
| 320 | new_deque[j] = thread_data->td.td_deque[i]; | |||
| 321 | ||||
| 322 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "openmp/runtime/src/kmp_tasking.cpp" , 322); | |||
| 323 | ||||
| 324 | thread_data->td.td_deque_head = 0; | |||
| 325 | thread_data->td.td_deque_tail = size; | |||
| 326 | thread_data->td.td_deque = new_deque; | |||
| 327 | thread_data->td.td_deque_size = new_size; | |||
| 328 | } | |||
| 329 | ||||
| 330 | static kmp_task_pri_t *__kmp_alloc_task_pri_list() { | |||
| 331 | kmp_task_pri_t *l = (kmp_task_pri_t *)__kmp_allocate(sizeof(kmp_task_pri_t))___kmp_allocate((sizeof(kmp_task_pri_t)), "openmp/runtime/src/kmp_tasking.cpp" , 331); | |||
| 332 | kmp_thread_data_t *thread_data = &l->td; | |||
| 333 | __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 334 | thread_data->td.td_deque_last_stolen = -1; | |||
| 335 | KE_TRACE(20, ("__kmp_alloc_task_pri_list: T#%d allocating deque[%d] "if (kmp_e_debug >= 20) { __kmp_debug_printf ("__kmp_alloc_task_pri_list: T#%d allocating deque[%d] " "for thread_data %p\n", __kmp_get_global_thread_id(), (1 << 8), thread_data); } | |||
| 336 | "for thread_data %p\n",if (kmp_e_debug >= 20) { __kmp_debug_printf ("__kmp_alloc_task_pri_list: T#%d allocating deque[%d] " "for thread_data %p\n", __kmp_get_global_thread_id(), (1 << 8), thread_data); } | |||
| 337 | __kmp_get_gtid(), INITIAL_TASK_DEQUE_SIZE, thread_data))if (kmp_e_debug >= 20) { __kmp_debug_printf ("__kmp_alloc_task_pri_list: T#%d allocating deque[%d] " "for thread_data %p\n", __kmp_get_global_thread_id(), (1 << 8), thread_data); }; | |||
| 338 | thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "openmp/runtime/src/kmp_tasking.cpp" , 339) | |||
| 339 | INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *))___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "openmp/runtime/src/kmp_tasking.cpp" , 339); | |||
| 340 | thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE(1 << 8); | |||
| 341 | return l; | |||
| 342 | } | |||
| 343 | ||||
| 344 | // The function finds the deque of priority tasks with given priority, or | |||
| 345 | // allocates a new deque and put it into sorted (high -> low) list of deques. | |||
| 346 | // Deques of non-default priority tasks are shared between all threads in team, | |||
| 347 | // as opposed to per-thread deques of tasks with default priority. | |||
| 348 | // The function is called under the lock task_team->tt.tt_task_pri_lock. | |||
| 349 | static kmp_thread_data_t * | |||
| 350 | __kmp_get_priority_deque_data(kmp_task_team_t *task_team, kmp_int32 pri) { | |||
| 351 | kmp_thread_data_t *thread_data; | |||
| 352 | kmp_task_pri_t *lst = task_team->tt.tt_task_pri_list; | |||
| 353 | if (lst->priority == pri) { | |||
| 354 | // Found queue of tasks with given priority. | |||
| 355 | thread_data = &lst->td; | |||
| 356 | } else if (lst->priority < pri) { | |||
| 357 | // All current priority queues contain tasks with lower priority. | |||
| 358 | // Allocate new one for given priority tasks. | |||
| 359 | kmp_task_pri_t *list = __kmp_alloc_task_pri_list(); | |||
| 360 | thread_data = &list->td; | |||
| 361 | list->priority = pri; | |||
| 362 | list->next = lst; | |||
| 363 | task_team->tt.tt_task_pri_list = list; | |||
| 364 | } else { // task_team->tt.tt_task_pri_list->priority > pri | |||
| 365 | kmp_task_pri_t *next_queue = lst->next; | |||
| 366 | while (next_queue && next_queue->priority > pri) { | |||
| 367 | lst = next_queue; | |||
| 368 | next_queue = lst->next; | |||
| 369 | } | |||
| 370 | // lst->priority > pri && (next == NULL || pri >= next->priority) | |||
| 371 | if (next_queue == NULL__null) { | |||
| 372 | // No queue with pri priority, need to allocate new one. | |||
| 373 | kmp_task_pri_t *list = __kmp_alloc_task_pri_list(); | |||
| 374 | thread_data = &list->td; | |||
| 375 | list->priority = pri; | |||
| 376 | list->next = NULL__null; | |||
| 377 | lst->next = list; | |||
| 378 | } else if (next_queue->priority == pri) { | |||
| 379 | // Found queue of tasks with given priority. | |||
| 380 | thread_data = &next_queue->td; | |||
| 381 | } else { // lst->priority > pri > next->priority | |||
| 382 | // insert newly allocated between existed queues | |||
| 383 | kmp_task_pri_t *list = __kmp_alloc_task_pri_list(); | |||
| 384 | thread_data = &list->td; | |||
| 385 | list->priority = pri; | |||
| 386 | list->next = next_queue; | |||
| 387 | lst->next = list; | |||
| 388 | } | |||
| 389 | } | |||
| 390 | return thread_data; | |||
| 391 | } | |||
| 392 | ||||
| 393 | // __kmp_push_priority_task: Add a task to the team's priority task deque | |||
| 394 | static kmp_int32 __kmp_push_priority_task(kmp_int32 gtid, kmp_info_t *thread, | |||
| 395 | kmp_taskdata_t *taskdata, | |||
| 396 | kmp_task_team_t *task_team, | |||
| 397 | kmp_int32 pri) { | |||
| 398 | kmp_thread_data_t *thread_data = NULL__null; | |||
| 399 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d trying to push task %p, pri %d.\n" , gtid, taskdata, pri); } | |||
| 400 | ("__kmp_push_priority_task: T#%d trying to push task %p, pri %d.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d trying to push task %p, pri %d.\n" , gtid, taskdata, pri); } | |||
| 401 | gtid, taskdata, pri))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d trying to push task %p, pri %d.\n" , gtid, taskdata, pri); }; | |||
| 402 | ||||
| 403 | // Find task queue specific to priority value | |||
| 404 | kmp_task_pri_t *lst = task_team->tt.tt_task_pri_list; | |||
| 405 | if (UNLIKELY(lst == NULL)__builtin_expect(!!(lst == __null), 0)) { | |||
| 406 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 407 | if (task_team->tt.tt_task_pri_list == NULL__null) { | |||
| 408 | // List of queues is still empty, allocate one. | |||
| 409 | kmp_task_pri_t *list = __kmp_alloc_task_pri_list(); | |||
| 410 | thread_data = &list->td; | |||
| 411 | list->priority = pri; | |||
| 412 | list->next = NULL__null; | |||
| 413 | task_team->tt.tt_task_pri_list = list; | |||
| 414 | } else { | |||
| 415 | // Other thread initialized a queue. Check if it fits and get thread_data. | |||
| 416 | thread_data = __kmp_get_priority_deque_data(task_team, pri); | |||
| 417 | } | |||
| 418 | __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 419 | } else { | |||
| 420 | if (lst->priority == pri) { | |||
| 421 | // Found queue of tasks with given priority. | |||
| 422 | thread_data = &lst->td; | |||
| 423 | } else { | |||
| 424 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 425 | thread_data = __kmp_get_priority_deque_data(task_team, pri); | |||
| 426 | __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 427 | } | |||
| 428 | } | |||
| 429 | KMP_DEBUG_ASSERT(thread_data)if (!(thread_data)) { __kmp_debug_assert("thread_data", "openmp/runtime/src/kmp_tasking.cpp" , 429); }; | |||
| 430 | ||||
| 431 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 432 | // Check if deque is full | |||
| 433 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 434 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 435 | if (__kmp_enable_task_throttling && | |||
| 436 | __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata, | |||
| 437 | thread->th.th_current_task)) { | |||
| 438 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 439 | KA_TRACE(20, ("__kmp_push_priority_task: T#%d deque is full; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 440 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 441 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
| 442 | return TASK_NOT_PUSHED1; | |||
| 443 | } else { | |||
| 444 | // expand deque to push the task which is not allowed to execute | |||
| 445 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 446 | } | |||
| 447 | } | |||
| 448 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <if (!((thread_data->td.td_deque_ntasks) < ((thread_data ->td).td_deque_size))) { __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "openmp/runtime/src/kmp_tasking.cpp", 449); } | |||
| 449 | TASK_DEQUE_SIZE(thread_data->td))if (!((thread_data->td.td_deque_ntasks) < ((thread_data ->td).td_deque_size))) { __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "openmp/runtime/src/kmp_tasking.cpp", 449); }; | |||
| 450 | // Push taskdata. | |||
| 451 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata; | |||
| 452 | // Wrap index. | |||
| 453 | thread_data->td.td_deque_tail = | |||
| 454 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 455 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
| 456 | 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 | |||
| 457 | KMP_FSYNC_RELEASING(thread->th.th_current_task)(!__kmp_itt_fsync_releasing_ptr__3_0) ? (void)0 : __kmp_itt_fsync_releasing_ptr__3_0 ((void *)(thread->th.th_current_task)); // releasing self | |||
| 458 | KMP_FSYNC_RELEASING(taskdata)(!__kmp_itt_fsync_releasing_ptr__3_0) ? (void)0 : __kmp_itt_fsync_releasing_ptr__3_0 ((void *)(taskdata)); // releasing child | |||
| 459 | KA_TRACE(20, ("__kmp_push_priority_task: T#%d returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_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); } | |||
| 460 | "TASK_SUCCESSFULLY_PUSHED: task=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_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); } | |||
| 461 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_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); } | |||
| 462 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_priority_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); }; | |||
| 463 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 464 | task_team->tt.tt_num_task_pri++; // atomic inc | |||
| 465 | return TASK_SUCCESSFULLY_PUSHED0; | |||
| 466 | } | |||
| 467 | ||||
| 468 | // __kmp_push_task: Add a task to the thread's deque | |||
| 469 | static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) { | |||
| 470 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 471 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 472 | ||||
| 473 | // If we encounter a hidden helper task, and the current thread is not a | |||
| 474 | // hidden helper thread, we have to give the task to any hidden helper thread | |||
| 475 | // starting from its shadow one. | |||
| 476 | if (UNLIKELY(taskdata->td_flags.hidden_helper &&__builtin_expect(!!(taskdata->td_flags.hidden_helper && !((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num )), 0) | |||
| 477 | !KMP_HIDDEN_HELPER_THREAD(gtid))__builtin_expect(!!(taskdata->td_flags.hidden_helper && !((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num )), 0)) { | |||
| 478 | kmp_int32 shadow_gtid = KMP_GTID_TO_SHADOW_GTID(gtid)((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2); | |||
| 479 | __kmpc_give_task(task, __kmp_tid_from_gtid(shadow_gtid)); | |||
| 480 | // Signal the hidden helper threads. | |||
| 481 | __kmp_hidden_helper_worker_thread_signal(); | |||
| 482 | return TASK_SUCCESSFULLY_PUSHED0; | |||
| 483 | } | |||
| 484 | ||||
| 485 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 486 | kmp_int32 tid = __kmp_tid_from_gtid(gtid); | |||
| 487 | kmp_thread_data_t *thread_data; | |||
| 488 | ||||
| 489 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d trying to push task %p.\n" , gtid, taskdata); } | |||
| 490 | ("__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); }; | |||
| 491 | ||||
| 492 | if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)__builtin_expect(!!(taskdata->td_flags.tiedness == 0), 0)) { | |||
| 493 | // untied task needs to increment counter so that the task structure is not | |||
| 494 | // freed prematurely | |||
| 495 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 496 | KMP_DEBUG_USE_VAR(counter); | |||
| 497 | 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); } | |||
| 498 | 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); } | |||
| 499 | ("__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); } | |||
| 500 | 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); }; | |||
| 501 | } | |||
| 502 | ||||
| 503 | // The first check avoids building task_team thread data if serialized | |||
| 504 | if (UNLIKELY(taskdata->td_flags.task_serial)__builtin_expect(!!(taskdata->td_flags.task_serial), 0)) { | |||
| 505 | 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); } | |||
| 506 | "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); } | |||
| 507 | 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); }; | |||
| 508 | return TASK_NOT_PUSHED1; | |||
| 509 | } | |||
| 510 | ||||
| 511 | // Now that serialized tasks have returned, we can assume that we are not in | |||
| 512 | // immediate exec mode | |||
| 513 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 513); }; | |||
| 514 | if (UNLIKELY(!KMP_TASKING_ENABLED(task_team))__builtin_expect(!!(!((!0) == ((task_team)->tt.tt_found_tasks ))), 0)) { | |||
| 515 | __kmp_enable_tasking(task_team, thread); | |||
| 516 | } | |||
| 517 | KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE)if (!((task_team->tt.tt_found_tasks) == (!0))) { __kmp_debug_assert ("(task_team->tt.tt_found_tasks) == (!0)", "openmp/runtime/src/kmp_tasking.cpp" , 517); }; | |||
| 518 | KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL)if (!(((void *)(task_team->tt.tt_threads_data)) != __null) ) { __kmp_debug_assert("((void *)(task_team->tt.tt_threads_data)) != __null" , "openmp/runtime/src/kmp_tasking.cpp", 518); }; | |||
| 519 | ||||
| 520 | if (taskdata->td_flags.priority_specified && task->data2.priority > 0 && | |||
| 521 | __kmp_max_task_priority > 0) { | |||
| 522 | int pri = KMP_MIN(task->data2.priority, __kmp_max_task_priority)((task->data2.priority) < (__kmp_max_task_priority) ? ( task->data2.priority) : (__kmp_max_task_priority)); | |||
| 523 | return __kmp_push_priority_task(gtid, thread, taskdata, task_team, pri); | |||
| 524 | } | |||
| 525 | ||||
| 526 | // Find tasking deque specific to encountering thread | |||
| 527 | thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 528 | ||||
| 529 | // No lock needed since only owner can allocate. If the task is hidden_helper, | |||
| 530 | // we don't need it either because we have initialized the dequeue for hidden | |||
| 531 | // helper thread data. | |||
| 532 | if (UNLIKELY(thread_data->td.td_deque == NULL)__builtin_expect(!!(thread_data->td.td_deque == __null), 0 )) { | |||
| 533 | __kmp_alloc_task_deque(thread, thread_data); | |||
| 534 | } | |||
| 535 | ||||
| 536 | int locked = 0; | |||
| 537 | // Check if deque is full | |||
| 538 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 539 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 540 | if (__kmp_enable_task_throttling && | |||
| 541 | __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata, | |||
| 542 | thread->th.th_current_task)) { | |||
| 543 | 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); } | |||
| 544 | "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); } | |||
| 545 | 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); }; | |||
| 546 | return TASK_NOT_PUSHED1; | |||
| 547 | } else { | |||
| 548 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 549 | locked = 1; | |||
| 550 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 551 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 552 | // expand deque to push the task which is not allowed to execute | |||
| 553 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 554 | } | |||
| 555 | } | |||
| 556 | } | |||
| 557 | // Lock the deque for the task push operation | |||
| 558 | if (!locked) { | |||
| 559 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 560 | // Need to recheck as we can get a proxy task from thread outside of OpenMP | |||
| 561 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 562 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 563 | if (__kmp_enable_task_throttling && | |||
| 564 | __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata, | |||
| 565 | thread->th.th_current_task)) { | |||
| 566 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 567 | KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; "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); } | |||
| 568 | "returning 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); } | |||
| 569 | 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); }; | |||
| 570 | return TASK_NOT_PUSHED1; | |||
| 571 | } else { | |||
| 572 | // expand deque to push the task which is not allowed to execute | |||
| 573 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 574 | } | |||
| 575 | } | |||
| 576 | } | |||
| 577 | // Must have room since no thread can add tasks but calling thread | |||
| 578 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <if (!((thread_data->td.td_deque_ntasks) < ((thread_data ->td).td_deque_size))) { __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "openmp/runtime/src/kmp_tasking.cpp", 579); } | |||
| 579 | TASK_DEQUE_SIZE(thread_data->td))if (!((thread_data->td.td_deque_ntasks) < ((thread_data ->td).td_deque_size))) { __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "openmp/runtime/src/kmp_tasking.cpp", 579); }; | |||
| 580 | ||||
| 581 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = | |||
| 582 | taskdata; // Push taskdata | |||
| 583 | // Wrap index. | |||
| 584 | thread_data->td.td_deque_tail = | |||
| 585 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 586 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
| 587 | 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 | |||
| 588 | KMP_FSYNC_RELEASING(thread->th.th_current_task)(!__kmp_itt_fsync_releasing_ptr__3_0) ? (void)0 : __kmp_itt_fsync_releasing_ptr__3_0 ((void *)(thread->th.th_current_task)); // releasing self | |||
| 589 | KMP_FSYNC_RELEASING(taskdata)(!__kmp_itt_fsync_releasing_ptr__3_0) ? (void)0 : __kmp_itt_fsync_releasing_ptr__3_0 ((void *)(taskdata)); // releasing child | |||
| 590 | 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); } | |||
| 591 | "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); } | |||
| 592 | 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); } | |||
| 593 | 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); }; | |||
| 594 | ||||
| 595 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 596 | ||||
| 597 | return TASK_SUCCESSFULLY_PUSHED0; | |||
| 598 | } | |||
| 599 | ||||
| 600 | // __kmp_pop_current_task_from_thread: set up current task from called thread | |||
| 601 | // when team ends | |||
| 602 | // | |||
| 603 | // this_thr: thread structure to set current_task in. | |||
| 604 | void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) { | |||
| 605 | 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); } | |||
| 606 | "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); } | |||
| 607 | "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); } | |||
| 608 | 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); } | |||
| 609 | 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); }; | |||
| 610 | ||||
| 611 | this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent; | |||
| 612 | ||||
| 613 | 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); } | |||
| 614 | "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); } | |||
| 615 | "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); } | |||
| 616 | 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); } | |||
| 617 | 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); }; | |||
| 618 | } | |||
| 619 | ||||
| 620 | // __kmp_push_current_task_to_thread: set up current task in called thread for a | |||
| 621 | // new team | |||
| 622 | // | |||
| 623 | // this_thr: thread structure to set up | |||
| 624 | // team: team for implicit task data | |||
| 625 | // tid: thread within team to set up | |||
| 626 | void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team, | |||
| 627 | int tid) { | |||
| 628 | // current task of the thread is a parent of the new just created implicit | |||
| 629 | // tasks of new team | |||
| 630 | 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); } | |||
| 631 | "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); } | |||
| 632 | "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); } | |||
| 633 | 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); } | |||
| 634 | 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); }; | |||
| 635 | ||||
| 636 | KMP_DEBUG_ASSERT(this_thr != NULL)if (!(this_thr != __null)) { __kmp_debug_assert("this_thr != __null" , "openmp/runtime/src/kmp_tasking.cpp", 636); }; | |||
| 637 | ||||
| 638 | if (tid == 0) { | |||
| 639 | if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) { | |||
| 640 | team->t.t_implicit_task_taskdata[0].td_parent = | |||
| 641 | this_thr->th.th_current_task; | |||
| 642 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0]; | |||
| 643 | } | |||
| 644 | } else { | |||
| 645 | team->t.t_implicit_task_taskdata[tid].td_parent = | |||
| 646 | team->t.t_implicit_task_taskdata[0].td_parent; | |||
| 647 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid]; | |||
| 648 | } | |||
| 649 | ||||
| 650 | 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); } | |||
| 651 | "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); } | |||
| 652 | "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); } | |||
| 653 | 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); } | |||
| 654 | 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); }; | |||
| 655 | } | |||
| 656 | ||||
| 657 | // __kmp_task_start: bookkeeping for a task starting execution | |||
| 658 | // | |||
| 659 | // GTID: global thread id of calling thread | |||
| 660 | // task: task starting execution | |||
| 661 | // current_task: task suspending | |||
| 662 | static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task, | |||
| 663 | kmp_taskdata_t *current_task) { | |||
| 664 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 665 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 666 | ||||
| 667 | 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); } | |||
| 668 | ("__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); } | |||
| 669 | 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); }; | |||
| 670 | ||||
| 671 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 671); }; | |||
| 672 | ||||
| 673 | // mark currently executing task as suspended | |||
| 674 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
| 675 | // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 ); | |||
| 676 | current_task->td_flags.executing = 0; | |||
| 677 | ||||
| 678 | // Add task to stack if tied | |||
| 679 | #ifdef BUILD_TIED_TASK_STACK | |||
| 680 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 681 | __kmp_push_task_stack(gtid, thread, taskdata); | |||
| 682 | } | |||
| 683 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 684 | ||||
| 685 | // mark starting task as executing and as current task | |||
| 686 | thread->th.th_current_task = taskdata; | |||
| 687 | ||||
| 688 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||if (!(taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0)) { __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "openmp/runtime/src/kmp_tasking.cpp", 689); } | |||
| 689 | taskdata->td_flags.tiedness == TASK_UNTIED)if (!(taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0)) { __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "openmp/runtime/src/kmp_tasking.cpp", 689); }; | |||
| 690 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||if (!(taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0)) { __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "openmp/runtime/src/kmp_tasking.cpp", 691); } | |||
| 691 | taskdata->td_flags.tiedness == TASK_UNTIED)if (!(taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0)) { __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "openmp/runtime/src/kmp_tasking.cpp", 691); }; | |||
| 692 | taskdata->td_flags.started = 1; | |||
| 693 | taskdata->td_flags.executing = 1; | |||
| 694 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)if (!(taskdata->td_flags.complete == 0)) { __kmp_debug_assert ("taskdata->td_flags.complete == 0", "openmp/runtime/src/kmp_tasking.cpp" , 694); }; | |||
| 695 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)if (!(taskdata->td_flags.freed == 0)) { __kmp_debug_assert ("taskdata->td_flags.freed == 0", "openmp/runtime/src/kmp_tasking.cpp" , 695); }; | |||
| 696 | ||||
| 697 | // GEH TODO: shouldn't we pass some sort of location identifier here? | |||
| 698 | // APT: yes, we will pass location here. | |||
| 699 | // need to store current thread state (in a thread or taskdata structure) | |||
| 700 | // before setting work_state, otherwise wrong state is set after end of task | |||
| 701 | ||||
| 702 | 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); }; | |||
| 703 | ||||
| 704 | return; | |||
| 705 | } | |||
| 706 | ||||
| 707 | #if OMPT_SUPPORT1 | |||
| 708 | //------------------------------------------------------------------------------ | |||
| 709 | // __ompt_task_init: | |||
| 710 | // Initialize OMPT fields maintained by a task. This will only be called after | |||
| 711 | // ompt_start_tool, so we already know whether ompt is enabled or not. | |||
| 712 | ||||
| 713 | static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) { | |||
| 714 | // The calls to __ompt_task_init already have the ompt_enabled condition. | |||
| 715 | task->ompt_task_info.task_data.value = 0; | |||
| 716 | task->ompt_task_info.frame.exit_frame = ompt_data_none{0}; | |||
| 717 | task->ompt_task_info.frame.enter_frame = ompt_data_none{0}; | |||
| 718 | task->ompt_task_info.frame.exit_frame_flags = | |||
| 719 | ompt_frame_runtime | ompt_frame_framepointer; | |||
| 720 | task->ompt_task_info.frame.enter_frame_flags = | |||
| 721 | ompt_frame_runtime | ompt_frame_framepointer; | |||
| 722 | task->ompt_task_info.dispatch_chunk.start = 0; | |||
| 723 | task->ompt_task_info.dispatch_chunk.iterations = 0; | |||
| 724 | } | |||
| 725 | ||||
| 726 | // __ompt_task_start: | |||
| 727 | // Build and trigger task-begin event | |||
| 728 | static inline void __ompt_task_start(kmp_task_t *task, | |||
| 729 | kmp_taskdata_t *current_task, | |||
| 730 | kmp_int32 gtid) { | |||
| 731 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 732 | ompt_task_status_t status = ompt_task_switch; | |||
| 733 | if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) { | |||
| 734 | status = ompt_task_yield; | |||
| 735 | __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0; | |||
| 736 | } | |||
| 737 | /* let OMPT know that we're about to run this task */ | |||
| 738 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
| 739 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
| 740 | &(current_task->ompt_task_info.task_data), status, | |||
| 741 | &(taskdata->ompt_task_info.task_data)); | |||
| 742 | } | |||
| 743 | taskdata->ompt_task_info.scheduling_parent = current_task; | |||
| 744 | } | |||
| 745 | ||||
| 746 | // __ompt_task_finish: | |||
| 747 | // Build and trigger final task-schedule event | |||
| 748 | static inline void __ompt_task_finish(kmp_task_t *task, | |||
| 749 | kmp_taskdata_t *resumed_task, | |||
| 750 | ompt_task_status_t status) { | |||
| 751 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
| 752 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 753 | if (__kmp_omp_cancellation && taskdata->td_taskgroup && | |||
| 754 | taskdata->td_taskgroup->cancel_request == cancel_taskgroup) { | |||
| 755 | status = ompt_task_cancel; | |||
| 756 | } | |||
| 757 | ||||
| 758 | /* let OMPT know that we're returning to the callee task */ | |||
| 759 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
| 760 | &(taskdata->ompt_task_info.task_data), status, | |||
| 761 | (resumed_task ? &(resumed_task->ompt_task_info.task_data) : NULL__null)); | |||
| 762 | } | |||
| 763 | } | |||
| 764 | #endif | |||
| 765 | ||||
| 766 | template <bool ompt> | |||
| 767 | static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid, | |||
| 768 | kmp_task_t *task, | |||
| 769 | void *frame_address, | |||
| 770 | void *return_address) { | |||
| 771 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 772 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 773 | ||||
| 774 | 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); } | |||
| 775 | "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); } | |||
| 776 | 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); }; | |||
| 777 | ||||
| 778 | if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)__builtin_expect(!!(taskdata->td_flags.tiedness == 0), 0)) { | |||
| 779 | // untied task needs to increment counter so that the task structure is not | |||
| 780 | // freed prematurely | |||
| 781 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 782 | KMP_DEBUG_USE_VAR(counter); | |||
| 783 | 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); } | |||
| 784 | "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); } | |||
| 785 | 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); }; | |||
| 786 | } | |||
| 787 | ||||
| 788 | taskdata->td_flags.task_serial = | |||
| 789 | 1; // Execute this task immediately, not deferred. | |||
| 790 | __kmp_task_start(gtid, task, current_task); | |||
| 791 | ||||
| 792 | #if OMPT_SUPPORT1 | |||
| 793 | if (ompt) { | |||
| 794 | if (current_task->ompt_task_info.frame.enter_frame.ptr == NULL__null) { | |||
| 795 | current_task->ompt_task_info.frame.enter_frame.ptr = | |||
| 796 | taskdata->ompt_task_info.frame.exit_frame.ptr = frame_address; | |||
| 797 | current_task->ompt_task_info.frame.enter_frame_flags = | |||
| 798 | taskdata->ompt_task_info.frame.exit_frame_flags = | |||
| 799 | ompt_frame_application | ompt_frame_framepointer; | |||
| 800 | } | |||
| 801 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 802 | ompt_task_info_t *parent_info = &(current_task->ompt_task_info); | |||
| 803 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 804 | &(parent_info->task_data), &(parent_info->frame), | |||
| 805 | &(taskdata->ompt_task_info.task_data), | |||
| 806 | 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, | |||
| 807 | return_address); | |||
| 808 | } | |||
| 809 | __ompt_task_start(task, current_task, gtid); | |||
| 810 | } | |||
| 811 | #endif // OMPT_SUPPORT | |||
| 812 | ||||
| 813 | 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); } | |||
| 814 | 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); }; | |||
| 815 | } | |||
| 816 | ||||
| 817 | #if OMPT_SUPPORT1 | |||
| 818 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 819 | static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 820 | kmp_task_t *task, | |||
| 821 | void *frame_address, | |||
| 822 | void *return_address) { | |||
| 823 | __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address, | |||
| 824 | return_address); | |||
| 825 | } | |||
| 826 | #endif // OMPT_SUPPORT | |||
| 827 | ||||
| 828 | // __kmpc_omp_task_begin_if0: report that a given serialized task has started | |||
| 829 | // execution | |||
| 830 | // | |||
| 831 | // loc_ref: source location information; points to beginning of task block. | |||
| 832 | // gtid: global thread number. | |||
| 833 | // task: task thunk for the started task. | |||
| 834 | void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
| 835 | kmp_task_t *task) { | |||
| 836 | #if OMPT_SUPPORT1 | |||
| 837 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 838 | OMPT_STORE_RETURN_ADDRESS(gtid)OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address (0)};; | |||
| 839 | __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task, | |||
| 840 | OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1), | |||
| 841 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 842 | return; | |||
| 843 | } | |||
| 844 | #endif | |||
| 845 | __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL__null, NULL__null); | |||
| 846 | } | |||
| 847 | ||||
| 848 | #ifdef TASK_UNUSED | |||
| 849 | // __kmpc_omp_task_begin: report that a given task has started execution | |||
| 850 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
| 851 | void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) { | |||
| 852 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 853 | ||||
| 854 | 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 ); } | |||
| 855 | 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 ); } | |||
| 856 | ("__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 ); } | |||
| 857 | 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 ); }; | |||
| 858 | ||||
| 859 | __kmp_task_start(gtid, task, current_task); | |||
| 860 | ||||
| 861 | 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)); } | |||
| 862 | 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)); }; | |||
| 863 | return; | |||
| 864 | } | |||
| 865 | #endif // TASK_UNUSED | |||
| 866 | ||||
| 867 | // __kmp_free_task: free the current task space and the space for shareds | |||
| 868 | // | |||
| 869 | // gtid: Global thread ID of calling thread | |||
| 870 | // taskdata: task to free | |||
| 871 | // thread: thread data structure of caller | |||
| 872 | static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata, | |||
| 873 | kmp_info_t *thread) { | |||
| 874 | 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); } | |||
| 875 | taskdata))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_free_task: T#%d freeing data from task %p\n" , gtid, taskdata); }; | |||
| 876 | ||||
| 877 | // Check to make sure all flags and counters have the correct values | |||
| 878 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 878); }; | |||
| 879 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0)if (!(taskdata->td_flags.executing == 0)) { __kmp_debug_assert ("taskdata->td_flags.executing == 0", "openmp/runtime/src/kmp_tasking.cpp" , 879); }; | |||
| 880 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1)if (!(taskdata->td_flags.complete == 1)) { __kmp_debug_assert ("taskdata->td_flags.complete == 1", "openmp/runtime/src/kmp_tasking.cpp" , 880); }; | |||
| 881 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)if (!(taskdata->td_flags.freed == 0)) { __kmp_debug_assert ("taskdata->td_flags.freed == 0", "openmp/runtime/src/kmp_tasking.cpp" , 881); }; | |||
| 882 | KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||if (!(taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1)) { __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "openmp/runtime/src/kmp_tasking.cpp", 883); } | |||
| 883 | taskdata->td_flags.task_serial == 1)if (!(taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1)) { __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "openmp/runtime/src/kmp_tasking.cpp", 883); }; | |||
| 884 | KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0)if (!(taskdata->td_incomplete_child_tasks == 0)) { __kmp_debug_assert ("taskdata->td_incomplete_child_tasks == 0", "openmp/runtime/src/kmp_tasking.cpp" , 884); }; | |||
| 885 | kmp_task_t *task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 886 | // Clear data to not be re-used later by mistake. | |||
| 887 | task->data1.destructors = NULL__null; | |||
| 888 | task->data2.priority = 0; | |||
| 889 | ||||
| 890 | taskdata->td_flags.freed = 1; | |||
| 891 | // deallocate the taskdata and shared variable blocks associated with this task | |||
| 892 | #if USE_FAST_MEMORY3 | |||
| 893 | __kmp_fast_free(thread, taskdata)___kmp_fast_free((thread), (taskdata), "openmp/runtime/src/kmp_tasking.cpp" , 893); | |||
| 894 | #else /* ! USE_FAST_MEMORY */ | |||
| 895 | __kmp_thread_free(thread, taskdata)___kmp_thread_free((thread), (taskdata), "openmp/runtime/src/kmp_tasking.cpp" , 895); | |||
| 896 | #endif | |||
| 897 | 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); }; | |||
| 898 | } | |||
| 899 | ||||
| 900 | // __kmp_free_task_and_ancestors: free the current task and ancestors without | |||
| 901 | // children | |||
| 902 | // | |||
| 903 | // gtid: Global thread ID of calling thread | |||
| 904 | // taskdata: task to free | |||
| 905 | // thread: thread data structure of caller | |||
| 906 | static void __kmp_free_task_and_ancestors(kmp_int32 gtid, | |||
| 907 | kmp_taskdata_t *taskdata, | |||
| 908 | kmp_info_t *thread) { | |||
| 909 | // Proxy tasks must always be allowed to free their parents | |||
| 910 | // because they can be run in background even in serial mode. | |||
| 911 | kmp_int32 team_serial = | |||
| 912 | (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) && | |||
| 913 | !taskdata->td_flags.proxy; | |||
| 914 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 914); }; | |||
| 915 | ||||
| 916 | 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; | |||
| 917 | KMP_DEBUG_ASSERT(children >= 0)if (!(children >= 0)) { __kmp_debug_assert("children >= 0" , "openmp/runtime/src/kmp_tasking.cpp", 917); }; | |||
| 918 | ||||
| 919 | // Now, go up the ancestor tree to see if any ancestors can now be freed. | |||
| 920 | while (children == 0) { | |||
| 921 | kmp_taskdata_t *parent_taskdata = taskdata->td_parent; | |||
| 922 | ||||
| 923 | 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); } | |||
| 924 | "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); } | |||
| 925 | 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); }; | |||
| 926 | ||||
| 927 | // --- Deallocate my ancestor task --- | |||
| 928 | __kmp_free_task(gtid, taskdata, thread); | |||
| 929 | ||||
| 930 | taskdata = parent_taskdata; | |||
| 931 | ||||
| 932 | if (team_serial) | |||
| 933 | return; | |||
| 934 | // Stop checking ancestors at implicit task instead of walking up ancestor | |||
| 935 | // tree to avoid premature deallocation of ancestors. | |||
| 936 | if (taskdata->td_flags.tasktype == TASK_IMPLICIT0) { | |||
| 937 | if (taskdata->td_dephash) { // do we need to cleanup dephash? | |||
| 938 | int children = KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ); | |||
| 939 | kmp_tasking_flags_t flags_old = taskdata->td_flags; | |||
| 940 | if (children == 0 && flags_old.complete == 1) { | |||
| 941 | kmp_tasking_flags_t flags_new = flags_old; | |||
| 942 | flags_new.complete = 0; | |||
| 943 | if (KMP_COMPARE_AND_STORE_ACQ32(__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&taskdata->td_flags)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_old)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_new))) | |||
| 944 | RCAST(kmp_int32 *, &taskdata->td_flags),__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&taskdata->td_flags)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_old)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_new))) | |||
| 945 | *RCAST(kmp_int32 *, &flags_old),__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&taskdata->td_flags)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_old)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_new))) | |||
| 946 | *RCAST(kmp_int32 *, &flags_new))__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&taskdata->td_flags)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_old)), (kmp_uint32 )(*reinterpret_cast<kmp_int32 *>(&flags_new)))) { | |||
| 947 | KA_TRACE(100, ("__kmp_free_task_and_ancestors: T#%d cleans "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_task_and_ancestors: T#%d cleans " "dephash of implicit task %p\n", gtid, taskdata); } | |||
| 948 | "dephash of implicit task %p\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_task_and_ancestors: T#%d cleans " "dephash of implicit task %p\n", gtid, taskdata); } | |||
| 949 | gtid, taskdata))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_free_task_and_ancestors: T#%d cleans " "dephash of implicit task %p\n", gtid, taskdata); }; | |||
| 950 | // cleanup dephash of finished implicit task | |||
| 951 | __kmp_dephash_free_entries(thread, taskdata->td_dephash); | |||
| 952 | } | |||
| 953 | } | |||
| 954 | } | |||
| 955 | return; | |||
| 956 | } | |||
| 957 | // Predecrement simulated by "- 1" calculation | |||
| 958 | children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks)(&taskdata->td_allocated_child_tasks)->fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
| 959 | KMP_DEBUG_ASSERT(children >= 0)if (!(children >= 0)) { __kmp_debug_assert("children >= 0" , "openmp/runtime/src/kmp_tasking.cpp", 959); }; | |||
| 960 | } | |||
| 961 | ||||
| 962 | 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); } | |||
| 963 | 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); } | |||
| 964 | "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); } | |||
| 965 | 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); }; | |||
| 966 | } | |||
| 967 | ||||
| 968 | // Only need to keep track of child task counts if any of the following: | |||
| 969 | // 1. team parallel and tasking not serialized; | |||
| 970 | // 2. it is a proxy or detachable or hidden helper task | |||
| 971 | // 3. the children counter of its parent task is greater than 0. | |||
| 972 | // The reason for the 3rd one is for serialized team that found detached task, | |||
| 973 | // hidden helper task, T. In this case, the execution of T is still deferred, | |||
| 974 | // and it is also possible that a regular task depends on T. In this case, if we | |||
| 975 | // don't track the children, task synchronization will be broken. | |||
| 976 | static bool __kmp_track_children_task(kmp_taskdata_t *taskdata) { | |||
| 977 | kmp_tasking_flags_t flags = taskdata->td_flags; | |||
| 978 | bool ret = !(flags.team_serial || flags.tasking_ser); | |||
| 979 | ret = ret || flags.proxy == TASK_PROXY1 || | |||
| 980 | flags.detachable == TASK_DETACHABLE1 || flags.hidden_helper; | |||
| 981 | ret = ret || | |||
| 982 | KMP_ATOMIC_LD_ACQ(&taskdata->td_parent->td_incomplete_child_tasks)(&taskdata->td_parent->td_incomplete_child_tasks)-> load(std::memory_order_acquire) > 0; | |||
| 983 | return ret; | |||
| 984 | } | |||
| 985 | ||||
| 986 | // __kmp_task_finish: bookkeeping to do when a task finishes execution | |||
| 987 | // | |||
| 988 | // gtid: global thread ID for calling thread | |||
| 989 | // task: task to be finished | |||
| 990 | // resumed_task: task to be resumed. (may be NULL if task is serialized) | |||
| 991 | // | |||
| 992 | // template<ompt>: effectively ompt_enabled.enabled!=0 | |||
| 993 | // the version with ompt=false is inlined, allowing to optimize away all ompt | |||
| 994 | // code in this case | |||
| 995 | template <bool ompt> | |||
| 996 | static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task, | |||
| 997 | kmp_taskdata_t *resumed_task) { | |||
| 998 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 999 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1000 | kmp_task_team_t *task_team = | |||
| 1001 | thread->th.th_task_team; // might be NULL for serial teams... | |||
| 1002 | #if KMP_DEBUG1 | |||
| 1003 | kmp_int32 children = 0; | |||
| 1004 | #endif | |||
| 1005 | 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); } | |||
| 1006 | "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); } | |||
| 1007 | 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); }; | |||
| 1008 | ||||
| 1009 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 1009); }; | |||
| 1010 | ||||
| 1011 | // Pop task from stack if tied | |||
| 1012 | #ifdef BUILD_TIED_TASK_STACK | |||
| 1013 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 1014 | __kmp_pop_task_stack(gtid, thread, taskdata); | |||
| 1015 | } | |||
| 1016 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 1017 | ||||
| 1018 | if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)__builtin_expect(!!(taskdata->td_flags.tiedness == 0), 0)) { | |||
| 1019 | // untied task needs to check the counter so that the task structure is not | |||
| 1020 | // freed prematurely | |||
| 1021 | kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_sub(1, std::memory_order_acq_rel ) - 1; | |||
| 1022 | 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); } | |||
| 1023 | 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); } | |||
| 1024 | ("__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); } | |||
| 1025 | 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); }; | |||
| 1026 | if (counter > 0) { | |||
| 1027 | // untied task is not done, to be continued possibly by other thread, do | |||
| 1028 | // not free it now | |||
| 1029 | if (resumed_task == NULL__null) { | |||
| 1030 | KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial)if (!(taskdata->td_flags.task_serial)) { __kmp_debug_assert ("taskdata->td_flags.task_serial", "openmp/runtime/src/kmp_tasking.cpp" , 1030); }; | |||
| 1031 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
| 1032 | // task is the parent | |||
| 1033 | } | |||
| 1034 | thread->th.th_current_task = resumed_task; // restore current_task | |||
| 1035 | resumed_task->td_flags.executing = 1; // resume previous task | |||
| 1036 | 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); } | |||
| 1037 | "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); } | |||
| 1038 | 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); }; | |||
| 1039 | return; | |||
| 1040 | } | |||
| 1041 | } | |||
| 1042 | ||||
| 1043 | // bookkeeping for resuming task: | |||
| 1044 | // GEH - note tasking_ser => task_serial | |||
| 1045 | KMP_DEBUG_ASSERT(if (!((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial)) { __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "openmp/runtime/src/kmp_tasking.cpp", 1047); } | |||
| 1046 | (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==if (!((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial)) { __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "openmp/runtime/src/kmp_tasking.cpp", 1047); } | |||
| 1047 | taskdata->td_flags.task_serial)if (!((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial)) { __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "openmp/runtime/src/kmp_tasking.cpp", 1047); }; | |||
| 1048 | if (taskdata->td_flags.task_serial) { | |||
| 1049 | if (resumed_task == NULL__null) { | |||
| 1050 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
| 1051 | // task is the parent | |||
| 1052 | } | |||
| 1053 | } else { | |||
| 1054 | KMP_DEBUG_ASSERT(resumed_task !=if (!(resumed_task != __null)) { __kmp_debug_assert("resumed_task != __null" , "openmp/runtime/src/kmp_tasking.cpp", 1055); } | |||
| 1055 | NULL)if (!(resumed_task != __null)) { __kmp_debug_assert("resumed_task != __null" , "openmp/runtime/src/kmp_tasking.cpp", 1055); }; // verify that resumed task is passed as argument | |||
| 1056 | } | |||
| 1057 | ||||
| 1058 | /* If the tasks' destructor thunk flag has been set, we need to invoke the | |||
| 1059 | destructor thunk that has been generated by the compiler. The code is | |||
| 1060 | placed here, since at this point other tasks might have been released | |||
| 1061 | hence overlapping the destructor invocations with some other work in the | |||
| 1062 | released tasks. The OpenMP spec is not specific on when the destructors | |||
| 1063 | are invoked, so we should be free to choose. */ | |||
| 1064 | if (UNLIKELY(taskdata->td_flags.destructors_thunk)__builtin_expect(!!(taskdata->td_flags.destructors_thunk), 0)) { | |||
| 1065 | kmp_routine_entry_t destr_thunk = task->data1.destructors; | |||
| 1066 | KMP_ASSERT(destr_thunk)if (!(destr_thunk)) { __kmp_debug_assert("destr_thunk", "openmp/runtime/src/kmp_tasking.cpp" , 1066); }; | |||
| 1067 | destr_thunk(gtid, task); | |||
| 1068 | } | |||
| 1069 | ||||
| 1070 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)if (!(taskdata->td_flags.complete == 0)) { __kmp_debug_assert ("taskdata->td_flags.complete == 0", "openmp/runtime/src/kmp_tasking.cpp" , 1070); }; | |||
| 1071 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1)if (!(taskdata->td_flags.started == 1)) { __kmp_debug_assert ("taskdata->td_flags.started == 1", "openmp/runtime/src/kmp_tasking.cpp" , 1071); }; | |||
| 1072 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)if (!(taskdata->td_flags.freed == 0)) { __kmp_debug_assert ("taskdata->td_flags.freed == 0", "openmp/runtime/src/kmp_tasking.cpp" , 1072); }; | |||
| 1073 | ||||
| 1074 | bool completed = true; | |||
| 1075 | if (UNLIKELY(taskdata->td_flags.detachable == TASK_DETACHABLE)__builtin_expect(!!(taskdata->td_flags.detachable == 1), 0 )) { | |||
| 1076 | if (taskdata->td_allow_completion_event.type == | |||
| 1077 | KMP_EVENT_ALLOW_COMPLETION) { | |||
| 1078 | // event hasn't been fulfilled yet. Try to detach task. | |||
| 1079 | __kmp_acquire_tas_lock(&taskdata->td_allow_completion_event.lock, gtid); | |||
| 1080 | if (taskdata->td_allow_completion_event.type == | |||
| 1081 | KMP_EVENT_ALLOW_COMPLETION) { | |||
| 1082 | // task finished execution | |||
| 1083 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1)if (!(taskdata->td_flags.executing == 1)) { __kmp_debug_assert ("taskdata->td_flags.executing == 1", "openmp/runtime/src/kmp_tasking.cpp" , 1083); }; | |||
| 1084 | taskdata->td_flags.executing = 0; // suspend the finishing task | |||
| 1085 | ||||
| 1086 | #if OMPT_SUPPORT1 | |||
| 1087 | // For a detached task, which is not completed, we switch back | |||
| 1088 | // the omp_fulfill_event signals completion | |||
| 1089 | // locking is necessary to avoid a race with ompt_task_late_fulfill | |||
| 1090 | if (ompt) | |||
| 1091 | __ompt_task_finish(task, resumed_task, ompt_task_detach); | |||
| 1092 | #endif | |||
| 1093 | ||||
| 1094 | // no access to taskdata after this point! | |||
| 1095 | // __kmp_fulfill_event might free taskdata at any time from now | |||
| 1096 | ||||
| 1097 | taskdata->td_flags.proxy = TASK_PROXY1; // proxify! | |||
| 1098 | completed = false; | |||
| 1099 | } | |||
| 1100 | __kmp_release_tas_lock(&taskdata->td_allow_completion_event.lock, gtid); | |||
| 1101 | } | |||
| 1102 | } | |||
| 1103 | ||||
| 1104 | // Tasks with valid target async handles must be re-enqueued. | |||
| 1105 | if (taskdata->td_target_data.async_handle != NULL__null) { | |||
| 1106 | // Note: no need to translate gtid to its shadow. If the current thread is a | |||
| 1107 | // hidden helper one, then the gtid is already correct. Otherwise, hidden | |||
| 1108 | // helper threads are disabled, and gtid refers to a OpenMP thread. | |||
| 1109 | __kmpc_give_task(task, __kmp_tid_from_gtid(gtid)); | |||
| 1110 | if (KMP_HIDDEN_HELPER_THREAD(gtid)((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num )) | |||
| 1111 | __kmp_hidden_helper_worker_thread_signal(); | |||
| 1112 | completed = false; | |||
| 1113 | } | |||
| 1114 | ||||
| 1115 | if (completed) { | |||
| 1116 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
| 1117 | ||||
| 1118 | #if OMPT_SUPPORT1 | |||
| 1119 | // This is not a detached task, we are done here | |||
| 1120 | if (ompt) | |||
| 1121 | __ompt_task_finish(task, resumed_task, ompt_task_complete); | |||
| 1122 | #endif | |||
| 1123 | // TODO: What would be the balance between the conditions in the function | |||
| 1124 | // and an atomic operation? | |||
| 1125 | if (__kmp_track_children_task(taskdata)) { | |||
| 1126 | __kmp_release_deps(gtid, taskdata); | |||
| 1127 | // Predecrement simulated by "- 1" calculation | |||
| 1128 | #if KMP_DEBUG1 | |||
| 1129 | children = -1 + | |||
| 1130 | #endif | |||
| 1131 | 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); | |||
| 1132 | KMP_DEBUG_ASSERT(children >= 0)if (!(children >= 0)) { __kmp_debug_assert("children >= 0" , "openmp/runtime/src/kmp_tasking.cpp", 1132); }; | |||
| 1133 | if (taskdata->td_taskgroup) | |||
| 1134 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
| 1135 | } else if (task_team && (task_team->tt.tt_found_proxy_tasks || | |||
| 1136 | task_team->tt.tt_hidden_helper_task_encountered)) { | |||
| 1137 | // if we found proxy or hidden helper tasks there could exist a dependency | |||
| 1138 | // chain with the proxy task as origin | |||
| 1139 | __kmp_release_deps(gtid, taskdata); | |||
| 1140 | } | |||
| 1141 | // td_flags.executing must be marked as 0 after __kmp_release_deps has been | |||
| 1142 | // called. Othertwise, if a task is executed immediately from the | |||
| 1143 | // release_deps code, the flag will be reset to 1 again by this same | |||
| 1144 | // function | |||
| 1145 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1)if (!(taskdata->td_flags.executing == 1)) { __kmp_debug_assert ("taskdata->td_flags.executing == 1", "openmp/runtime/src/kmp_tasking.cpp" , 1145); }; | |||
| 1146 | taskdata->td_flags.executing = 0; // suspend the finishing task | |||
| 1147 | ||||
| 1148 | // Decrement the counter of hidden helper tasks to be executed. | |||
| 1149 | if (taskdata->td_flags.hidden_helper) { | |||
| 1150 | // Hidden helper tasks can only be executed by hidden helper threads. | |||
| 1151 | KMP_ASSERT(KMP_HIDDEN_HELPER_THREAD(gtid))if (!(((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num ))) { __kmp_debug_assert("KMP_HIDDEN_HELPER_THREAD(gtid)", "openmp/runtime/src/kmp_tasking.cpp" , 1151); }; | |||
| 1152 | KMP_ATOMIC_DEC(&__kmp_unexecuted_hidden_helper_tasks)(&__kmp_unexecuted_hidden_helper_tasks)->fetch_sub(1, std ::memory_order_acq_rel); | |||
| 1153 | } | |||
| 1154 | } | |||
| 1155 | ||||
| 1156 | 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); } | |||
| 1157 | 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); } | |||
| 1158 | 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); }; | |||
| 1159 | ||||
| 1160 | // Free this task and then ancestor tasks if they have no children. | |||
| 1161 | // Restore th_current_task first as suggested by John: | |||
| 1162 | // johnmc: if an asynchronous inquiry peers into the runtime system | |||
| 1163 | // it doesn't see the freed task as the current task. | |||
| 1164 | thread->th.th_current_task = resumed_task; | |||
| 1165 | if (completed) | |||
| 1166 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
| 1167 | ||||
| 1168 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
| 1169 | // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 ); | |||
| 1170 | resumed_task->td_flags.executing = 1; // resume previous task | |||
| 1171 | ||||
| 1172 | 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); } | |||
| 1173 | 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); } | |||
| 1174 | 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); }; | |||
| 1175 | ||||
| 1176 | return; | |||
| 1177 | } | |||
| 1178 | ||||
| 1179 | template <bool ompt> | |||
| 1180 | static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref, | |||
| 1181 | kmp_int32 gtid, | |||
| 1182 | kmp_task_t *task) { | |||
| 1183 | 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)); } | |||
| 1184 | 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)); }; | |||
| 1185 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_tasking.cpp" , 1185); }; | |||
| 1186 | // this routine will provide task to resume | |||
| 1187 | __kmp_task_finish<ompt>(gtid, task, NULL__null); | |||
| 1188 | ||||
| 1189 | 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)); } | |||
| 1190 | 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)); }; | |||
| 1191 | ||||
| 1192 | #if OMPT_SUPPORT1 | |||
| 1193 | if (ompt) { | |||
| 1194 | ompt_frame_t *ompt_frame; | |||
| 1195 | __ompt_get_task_info_internal(0, NULL__null, NULL__null, &ompt_frame, NULL__null, NULL__null); | |||
| 1196 | ompt_frame->enter_frame = ompt_data_none{0}; | |||
| 1197 | ompt_frame->enter_frame_flags = | |||
| 1198 | ompt_frame_runtime | ompt_frame_framepointer; | |||
| 1199 | } | |||
| 1200 | #endif | |||
| 1201 | ||||
| 1202 | return; | |||
| 1203 | } | |||
| 1204 | ||||
| 1205 | #if OMPT_SUPPORT1 | |||
| 1206 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 1207 | void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1208 | kmp_task_t *task) { | |||
| 1209 | __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task); | |||
| 1210 | } | |||
| 1211 | #endif // OMPT_SUPPORT | |||
| 1212 | ||||
| 1213 | // __kmpc_omp_task_complete_if0: report that a task has completed execution | |||
| 1214 | // | |||
| 1215 | // loc_ref: source location information; points to end of task block. | |||
| 1216 | // gtid: global thread number. | |||
| 1217 | // task: task thunk for the completed task. | |||
| 1218 | void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1219 | kmp_task_t *task) { | |||
| 1220 | #if OMPT_SUPPORT1 | |||
| 1221 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1222 | __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task); | |||
| 1223 | return; | |||
| 1224 | } | |||
| 1225 | #endif | |||
| 1226 | __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task); | |||
| 1227 | } | |||
| 1228 | ||||
| 1229 | #ifdef TASK_UNUSED | |||
| 1230 | // __kmpc_omp_task_complete: report that a task has completed execution | |||
| 1231 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
| 1232 | void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1233 | kmp_task_t *task) { | |||
| 1234 | 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)); } | |||
| 1235 | 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)); }; | |||
| 1236 | ||||
| 1237 | __kmp_task_finish<false>(gtid, task, | |||
| 1238 | NULL__null); // Not sure how to find task to resume | |||
| 1239 | ||||
| 1240 | 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)); } | |||
| 1241 | 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)); }; | |||
| 1242 | return; | |||
| 1243 | } | |||
| 1244 | #endif // TASK_UNUSED | |||
| 1245 | ||||
| 1246 | // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit | |||
| 1247 | // task for a given thread | |||
| 1248 | // | |||
| 1249 | // loc_ref: reference to source location of parallel region | |||
| 1250 | // this_thr: thread data structure corresponding to implicit task | |||
| 1251 | // team: team for this_thr | |||
| 1252 | // tid: thread id of given thread within team | |||
| 1253 | // set_curr_task: TRUE if need to push current task to thread | |||
| 1254 | // NOTE: Routine does not set up the implicit task ICVS. This is assumed to | |||
| 1255 | // have already been done elsewhere. | |||
| 1256 | // TODO: Get better loc_ref. Value passed in may be NULL | |||
| 1257 | void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, | |||
| 1258 | kmp_team_t *team, int tid, int set_curr_task) { | |||
| 1259 | kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid]; | |||
| 1260 | ||||
| 1261 | 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"); } | |||
| 1262 | 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"); } | |||
| 1263 | ("__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"); } | |||
| 1264 | 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"); }; | |||
| 1265 | ||||
| 1266 | task->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 1267 | task->td_team = team; | |||
| 1268 | // task->td_parent = NULL; // fix for CQ230101 (broken parent task info | |||
| 1269 | // in debugger) | |||
| 1270 | task->td_ident = loc_ref; | |||
| 1271 | task->td_taskwait_ident = NULL__null; | |||
| 1272 | task->td_taskwait_counter = 0; | |||
| 1273 | task->td_taskwait_thread = 0; | |||
| 1274 | ||||
| 1275 | task->td_flags.tiedness = TASK_TIED1; | |||
| 1276 | task->td_flags.tasktype = TASK_IMPLICIT0; | |||
| 1277 | task->td_flags.proxy = TASK_FULL0; | |||
| 1278 | ||||
| 1279 | // All implicit tasks are executed immediately, not deferred | |||
| 1280 | task->td_flags.task_serial = 1; | |||
| 1281 | task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
| 1282 | task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
| 1283 | ||||
| 1284 | task->td_flags.started = 1; | |||
| 1285 | task->td_flags.executing = 1; | |||
| 1286 | task->td_flags.complete = 0; | |||
| 1287 | task->td_flags.freed = 0; | |||
| 1288 | ||||
| 1289 | task->td_depnode = NULL__null; | |||
| 1290 | task->td_last_tied = task; | |||
| 1291 | task->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED; | |||
| 1292 | ||||
| 1293 | if (set_curr_task) { // only do this init first time thread is created | |||
| 1294 | KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0)(&task->td_incomplete_child_tasks)->store(0, std::memory_order_release ); | |||
| 1295 | // Not used: don't need to deallocate implicit task | |||
| 1296 | KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0)(&task->td_allocated_child_tasks)->store(0, std::memory_order_release ); | |||
| 1297 | task->td_taskgroup = NULL__null; // An implicit task does not have taskgroup | |||
| 1298 | task->td_dephash = NULL__null; | |||
| 1299 | __kmp_push_current_task_to_thread(this_thr, team, tid); | |||
| 1300 | } else { | |||
| 1301 | KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0)if (!(task->td_incomplete_child_tasks == 0)) { __kmp_debug_assert ("task->td_incomplete_child_tasks == 0", "openmp/runtime/src/kmp_tasking.cpp" , 1301); }; | |||
| 1302 | KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0)if (!(task->td_allocated_child_tasks == 0)) { __kmp_debug_assert ("task->td_allocated_child_tasks == 0", "openmp/runtime/src/kmp_tasking.cpp" , 1302); }; | |||
| 1303 | } | |||
| 1304 | ||||
| 1305 | #if OMPT_SUPPORT1 | |||
| 1306 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1307 | __ompt_task_init(task, tid); | |||
| 1308 | #endif | |||
| 1309 | ||||
| 1310 | 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); } | |||
| 1311 | 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); }; | |||
| 1312 | } | |||
| 1313 | ||||
| 1314 | // __kmp_finish_implicit_task: Release resources associated to implicit tasks | |||
| 1315 | // at the end of parallel regions. Some resources are kept for reuse in the next | |||
| 1316 | // parallel region. | |||
| 1317 | // | |||
| 1318 | // thread: thread data structure corresponding to implicit task | |||
| 1319 | void __kmp_finish_implicit_task(kmp_info_t *thread) { | |||
| 1320 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
| 1321 | if (task->td_dephash) { | |||
| 1322 | int children; | |||
| 1323 | task->td_flags.complete = 1; | |||
| 1324 | children = KMP_ATOMIC_LD_ACQ(&task->td_incomplete_child_tasks)(&task->td_incomplete_child_tasks)->load(std::memory_order_acquire ); | |||
| 1325 | kmp_tasking_flags_t flags_old = task->td_flags; | |||
| 1326 | if (children == 0 && flags_old.complete == 1) { | |||
| 1327 | kmp_tasking_flags_t flags_new = flags_old; | |||
| 1328 | flags_new.complete = 0; | |||
| 1329 | if (KMP_COMPARE_AND_STORE_ACQ32(RCAST(kmp_int32 *, &task->td_flags),__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&task->td_flags)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_old)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_new))) | |||
| 1330 | *RCAST(kmp_int32 *, &flags_old),__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&task->td_flags)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_old)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_new))) | |||
| 1331 | *RCAST(kmp_int32 *, &flags_new))__sync_bool_compare_and_swap((volatile kmp_uint32 *)(reinterpret_cast <kmp_int32 *>(&task->td_flags)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_old)), (kmp_uint32)(*reinterpret_cast <kmp_int32 *>(&flags_new)))) { | |||
| 1332 | KA_TRACE(100, ("__kmp_finish_implicit_task: T#%d cleans "if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_finish_implicit_task: T#%d cleans " "dephash of implicit task %p\n", thread->th.th_info.ds.ds_gtid , task); } | |||
| 1333 | "dephash of implicit task %p\n",if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_finish_implicit_task: T#%d cleans " "dephash of implicit task %p\n", thread->th.th_info.ds.ds_gtid , task); } | |||
| 1334 | thread->th.th_info.ds.ds_gtid, task))if (kmp_a_debug >= 100) { __kmp_debug_printf ("__kmp_finish_implicit_task: T#%d cleans " "dephash of implicit task %p\n", thread->th.th_info.ds.ds_gtid , task); }; | |||
| 1335 | __kmp_dephash_free_entries(thread, task->td_dephash); | |||
| 1336 | } | |||
| 1337 | } | |||
| 1338 | } | |||
| 1339 | } | |||
| 1340 | ||||
| 1341 | // __kmp_free_implicit_task: Release resources associated to implicit tasks | |||
| 1342 | // when these are destroyed regions | |||
| 1343 | // | |||
| 1344 | // thread: thread data structure corresponding to implicit task | |||
| 1345 | void __kmp_free_implicit_task(kmp_info_t *thread) { | |||
| 1346 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
| 1347 | if (task && task->td_dephash) { | |||
| 1348 | __kmp_dephash_free(thread, task->td_dephash); | |||
| 1349 | task->td_dephash = NULL__null; | |||
| 1350 | } | |||
| 1351 | } | |||
| 1352 | ||||
| 1353 | // Round up a size to a power of two specified by val: Used to insert padding | |||
| 1354 | // between structures co-allocated using a single malloc() call | |||
| 1355 | static size_t __kmp_round_up_to_val(size_t size, size_t val) { | |||
| 1356 | if (size & (val - 1)) { | |||
| 1357 | size &= ~(val - 1); | |||
| 1358 | if (size <= KMP_SIZE_T_MAX(0xFFFFFFFFFFFFFFFF) - val) { | |||
| 1359 | size += val; // Round up if there is no overflow. | |||
| 1360 | } | |||
| 1361 | } | |||
| 1362 | return size; | |||
| 1363 | } // __kmp_round_up_to_va | |||
| 1364 | ||||
| 1365 | // __kmp_task_alloc: Allocate the taskdata and task data structures for a task | |||
| 1366 | // | |||
| 1367 | // loc_ref: source location information | |||
| 1368 | // gtid: global thread number. | |||
| 1369 | // flags: include tiedness & task type (explicit vs. implicit) of the ''new'' | |||
| 1370 | // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine. | |||
| 1371 | // sizeof_kmp_task_t: Size in bytes of kmp_task_t data structure including | |||
| 1372 | // private vars accessed in task. | |||
| 1373 | // sizeof_shareds: Size in bytes of array of pointers to shared vars accessed | |||
| 1374 | // in task. | |||
| 1375 | // task_entry: Pointer to task code entry point generated by compiler. | |||
| 1376 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
| 1377 | kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1378 | kmp_tasking_flags_t *flags, | |||
| 1379 | size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
| 1380 | kmp_routine_entry_t task_entry) { | |||
| 1381 | kmp_task_t *task; | |||
| 1382 | kmp_taskdata_t *taskdata; | |||
| 1383 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1384 | kmp_team_t *team = thread->th.th_team; | |||
| 1385 | kmp_taskdata_t *parent_task = thread->th.th_current_task; | |||
| 1386 | size_t shareds_offset; | |||
| 1387 | ||||
| 1388 | if (UNLIKELY(!TCR_4(__kmp_init_middle))__builtin_expect(!!(!(__kmp_init_middle)), 0)) | |||
| 1389 | __kmp_middle_initialize(); | |||
| 1390 | ||||
| 1391 | if (flags->hidden_helper) { | |||
| 1392 | if (__kmp_enable_hidden_helper) { | |||
| 1393 | if (!TCR_4(__kmp_init_hidden_helper)(__kmp_init_hidden_helper)) | |||
| 1394 | __kmp_hidden_helper_initialize(); | |||
| 1395 | } else { | |||
| 1396 | // If the hidden helper task is not enabled, reset the flag to FALSE. | |||
| 1397 | flags->hidden_helper = FALSE0; | |||
| 1398 | } | |||
| 1399 | } | |||
| 1400 | ||||
| 1401 | 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 ); } | |||
| 1402 | "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 ); } | |||
| 1403 | 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 ); } | |||
| 1404 | 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 ); }; | |||
| 1405 | ||||
| 1406 | KMP_DEBUG_ASSERT(parent_task)if (!(parent_task)) { __kmp_debug_assert("parent_task", "openmp/runtime/src/kmp_tasking.cpp" , 1406); }; | |||
| 1407 | if (parent_task->td_flags.final) { | |||
| 1408 | if (flags->merged_if0) { | |||
| 1409 | } | |||
| 1410 | flags->final = 1; | |||
| 1411 | } | |||
| 1412 | ||||
| 1413 | if (flags->tiedness == TASK_UNTIED0 && !team->t.t_serialized) { | |||
| 1414 | // Untied task encountered causes the TSC algorithm to check entire deque of | |||
| 1415 | // the victim thread. If no untied task encountered, then checking the head | |||
| 1416 | // of the deque should be enough. | |||
| 1417 | 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); | |||
| 1418 | } | |||
| 1419 | ||||
| 1420 | // Detachable tasks are not proxy tasks yet but could be in the future. Doing | |||
| 1421 | // the tasking setup | |||
| 1422 | // when that happens is too late. | |||
| 1423 | if (UNLIKELY(flags->proxy == TASK_PROXY ||__builtin_expect(!!(flags->proxy == 1 || flags->detachable == 1 || flags->hidden_helper), 0) | |||
| 1424 | flags->detachable == TASK_DETACHABLE || flags->hidden_helper)__builtin_expect(!!(flags->proxy == 1 || flags->detachable == 1 || flags->hidden_helper), 0)) { | |||
| 1425 | if (flags->proxy == TASK_PROXY1) { | |||
| 1426 | flags->tiedness = TASK_UNTIED0; | |||
| 1427 | flags->merged_if0 = 1; | |||
| 1428 | } | |||
| 1429 | /* are we running in a sequential parallel or tskm_immediate_exec... we need | |||
| 1430 | tasking support enabled */ | |||
| 1431 | if ((thread->th.th_task_team) == NULL__null) { | |||
| 1432 | /* This should only happen if the team is serialized | |||
| 1433 | setup a task team and propagate it to the thread */ | |||
| 1434 | KMP_DEBUG_ASSERT(team->t.t_serialized)if (!(team->t.t_serialized)) { __kmp_debug_assert("team->t.t_serialized" , "openmp/runtime/src/kmp_tasking.cpp", 1434); }; | |||
| 1435 | 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); } | |||
| 1436 | ("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); } | |||
| 1437 | gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d creating task team in __kmp_task_alloc for proxy task\n" , gtid); }; | |||
| 1438 | // 1 indicates setup the current team regardless of nthreads | |||
| 1439 | __kmp_task_team_setup(thread, team, 1); | |||
| 1440 | thread->th.th_task_team = team->t.t_task_team[thread->th.th_task_state]; | |||
| 1441 | } | |||
| 1442 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 1443 | ||||
| 1444 | /* tasking must be enabled now as the task might not be pushed */ | |||
| 1445 | if (!KMP_TASKING_ENABLED(task_team)((!0) == ((task_team)->tt.tt_found_tasks))) { | |||
| 1446 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
| 1447 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
| 1448 | ("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); }; | |||
| 1449 | __kmp_enable_tasking(task_team, thread); | |||
| 1450 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
| 1451 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 1452 | // No lock needed since only owner can allocate | |||
| 1453 | if (thread_data->td.td_deque == NULL__null) { | |||
| 1454 | __kmp_alloc_task_deque(thread, thread_data); | |||
| 1455 | } | |||
| 1456 | } | |||
| 1457 | ||||
| 1458 | if ((flags->proxy == TASK_PROXY1 || flags->detachable == TASK_DETACHABLE1) && | |||
| 1459 | task_team->tt.tt_found_proxy_tasks == FALSE0) | |||
| 1460 | TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE)(task_team->tt.tt_found_proxy_tasks) = ((!0)); | |||
| 1461 | if (flags->hidden_helper && | |||
| 1462 | task_team->tt.tt_hidden_helper_task_encountered == FALSE0) | |||
| 1463 | TCW_4(task_team->tt.tt_hidden_helper_task_encountered, TRUE)(task_team->tt.tt_hidden_helper_task_encountered) = ((!0)); | |||
| 1464 | } | |||
| 1465 | ||||
| 1466 | // Calculate shared structure offset including padding after kmp_task_t struct | |||
| 1467 | // to align pointers in shared struct | |||
| 1468 | shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t; | |||
| 1469 | shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *)); | |||
| 1470 | ||||
| 1471 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
| 1472 | 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); } | |||
| 1473 | shareds_offset))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d First malloc size: %ld\n" , gtid, shareds_offset); }; | |||
| 1474 | 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); } | |||
| 1475 | sizeof_shareds))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d Second malloc size: %ld\n" , gtid, sizeof_shareds); }; | |||
| 1476 | ||||
| 1477 | // Avoid double allocation here by combining shareds with taskdata | |||
| 1478 | #if USE_FAST_MEMORY3 | |||
| 1479 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, shareds_offset +___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "openmp/runtime/src/kmp_tasking.cpp", 1480) | |||
| 1480 | sizeof_shareds)___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "openmp/runtime/src/kmp_tasking.cpp", 1480); | |||
| 1481 | #else /* ! USE_FAST_MEMORY */ | |||
| 1482 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, shareds_offset +___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "openmp/runtime/src/kmp_tasking.cpp", 1483) | |||
| 1483 | sizeof_shareds)___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "openmp/runtime/src/kmp_tasking.cpp", 1483); | |||
| 1484 | #endif /* USE_FAST_MEMORY */ | |||
| 1485 | ||||
| 1486 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 1487 | ||||
| 1488 | // Make sure task & taskdata are aligned appropriately | |||
| 1489 | #if KMP_ARCH_X860 || KMP_ARCH_PPC64(0 || 0) || !KMP_HAVE_QUAD0 | |||
| 1490 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0)if (!((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1490); }; | |||
| 1491 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0)if (!((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0 )) { __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1491); }; | |||
| 1492 | #else | |||
| 1493 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0)if (!((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1493); }; | |||
| 1494 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0)if (!((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0) ) { __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1494); }; | |||
| 1495 | #endif | |||
| 1496 | if (sizeof_shareds > 0) { | |||
| 1497 | // Avoid double allocation here by combining shareds with taskdata | |||
| 1498 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
| 1499 | // Make sure shareds struct is aligned to pointer size | |||
| 1500 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==if (!((((kmp_uintptr_t)task->shareds) & (sizeof(void * ) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1501); } | |||
| 1501 | 0)if (!((((kmp_uintptr_t)task->shareds) & (sizeof(void * ) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 1501); }; | |||
| 1502 | } else { | |||
| 1503 | task->shareds = NULL__null; | |||
| 1504 | } | |||
| 1505 | task->routine = task_entry; | |||
| 1506 | task->part_id = 0; // AC: Always start with 0 part id | |||
| 1507 | ||||
| 1508 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 1509 | taskdata->td_team = thread->th.th_team; | |||
| 1510 | taskdata->td_alloc_thread = thread; | |||
| 1511 | taskdata->td_parent = parent_task; | |||
| 1512 | taskdata->td_level = parent_task->td_level + 1; // increment nesting level | |||
| 1513 | KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0)(&taskdata->td_untied_count)->store(0, std::memory_order_relaxed ); | |||
| 1514 | taskdata->td_ident = loc_ref; | |||
| 1515 | taskdata->td_taskwait_ident = NULL__null; | |||
| 1516 | taskdata->td_taskwait_counter = 0; | |||
| 1517 | taskdata->td_taskwait_thread = 0; | |||
| 1518 | KMP_DEBUG_ASSERT(taskdata->td_parent != NULL)if (!(taskdata->td_parent != __null)) { __kmp_debug_assert ("taskdata->td_parent != __null", "openmp/runtime/src/kmp_tasking.cpp" , 1518); }; | |||
| 1519 | // avoid copying icvs for proxy tasks | |||
| 1520 | if (flags->proxy == TASK_FULL0) | |||
| 1521 | copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs); | |||
| 1522 | ||||
| 1523 | taskdata->td_flags = *flags; | |||
| 1524 | taskdata->td_task_team = thread->th.th_task_team; | |||
| 1525 | taskdata->td_size_alloc = shareds_offset + sizeof_shareds; | |||
| 1526 | taskdata->td_flags.tasktype = TASK_EXPLICIT1; | |||
| 1527 | // If it is hidden helper task, we need to set the team and task team | |||
| 1528 | // correspondingly. | |||
| 1529 | if (flags->hidden_helper) { | |||
| 1530 | kmp_info_t *shadow_thread = __kmp_threads[KMP_GTID_TO_SHADOW_GTID(gtid)((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)]; | |||
| 1531 | taskdata->td_team = shadow_thread->th.th_team; | |||
| 1532 | taskdata->td_task_team = shadow_thread->th.th_task_team; | |||
| 1533 | } | |||
| 1534 | ||||
| 1535 | // GEH - TODO: fix this to copy parent task's value of tasking_ser flag | |||
| 1536 | taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
| 1537 | ||||
| 1538 | // GEH - TODO: fix this to copy parent task's value of team_serial flag | |||
| 1539 | taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
| 1540 | ||||
| 1541 | // GEH - Note we serialize the task if the team is serialized to make sure | |||
| 1542 | // implicit parallel region tasks are not left until program termination to | |||
| 1543 | // execute. Also, it helps locality to execute immediately. | |||
| 1544 | ||||
| 1545 | taskdata->td_flags.task_serial = | |||
| 1546 | (parent_task->td_flags.final || taskdata->td_flags.team_serial || | |||
| 1547 | taskdata->td_flags.tasking_ser || flags->merged_if0); | |||
| 1548 | ||||
| 1549 | taskdata->td_flags.started = 0; | |||
| 1550 | taskdata->td_flags.executing = 0; | |||
| 1551 | taskdata->td_flags.complete = 0; | |||
| 1552 | taskdata->td_flags.freed = 0; | |||
| 1553 | ||||
| 1554 | KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0)(&taskdata->td_incomplete_child_tasks)->store(0, std ::memory_order_relaxed); | |||
| 1555 | // start at one because counts current task and children | |||
| 1556 | KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1)(&taskdata->td_allocated_child_tasks)->store(1, std ::memory_order_relaxed); | |||
| 1557 | taskdata->td_taskgroup = | |||
| 1558 | parent_task->td_taskgroup; // task inherits taskgroup from the parent task | |||
| 1559 | taskdata->td_dephash = NULL__null; | |||
| 1560 | taskdata->td_depnode = NULL__null; | |||
| 1561 | taskdata->td_target_data.async_handle = NULL__null; | |||
| 1562 | if (flags->tiedness == TASK_UNTIED0) | |||
| 1563 | taskdata->td_last_tied = NULL__null; // will be set when the task is scheduled | |||
| 1564 | else | |||
| 1565 | taskdata->td_last_tied = taskdata; | |||
| 1566 | taskdata->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED; | |||
| 1567 | #if OMPT_SUPPORT1 | |||
| 1568 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1569 | __ompt_task_init(taskdata, gtid); | |||
| 1570 | #endif | |||
| 1571 | // TODO: What would be the balance between the conditions in the function and | |||
| 1572 | // an atomic operation? | |||
| 1573 | if (__kmp_track_children_task(taskdata)) { | |||
| 1574 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
| 1575 | if (parent_task->td_taskgroup) | |||
| 1576 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
| 1577 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
| 1578 | // implicit not deallocated | |||
| 1579 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) { | |||
| 1580 | 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); | |||
| 1581 | } | |||
| 1582 | if (flags->hidden_helper) { | |||
| 1583 | taskdata->td_flags.task_serial = FALSE0; | |||
| 1584 | // Increment the number of hidden helper tasks to be executed | |||
| 1585 | KMP_ATOMIC_INC(&__kmp_unexecuted_hidden_helper_tasks)(&__kmp_unexecuted_hidden_helper_tasks)->fetch_add(1, std ::memory_order_acq_rel); | |||
| 1586 | } | |||
| 1587 | } | |||
| 1588 | ||||
| 1589 | 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); } | |||
| 1590 | 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); }; | |||
| 1591 | ||||
| 1592 | return task; | |||
| 1593 | } | |||
| 1594 | ||||
| 1595 | kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1596 | kmp_int32 flags, size_t sizeof_kmp_task_t, | |||
| 1597 | size_t sizeof_shareds, | |||
| 1598 | kmp_routine_entry_t task_entry) { | |||
| 1599 | kmp_task_t *retval; | |||
| 1600 | kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; | |||
| 1601 | __kmp_assert_valid_gtid(gtid); | |||
| 1602 | input_flags->native = FALSE0; | |||
| 1603 | // __kmp_task_alloc() sets up all other runtime flags | |||
| 1604 | KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1605 | "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 %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1606 | 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 %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1607 | input_flags->proxy ? "proxy" : "",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1608 | input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1609 | sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); }; | |||
| 1610 | ||||
| 1611 | retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t, | |||
| 1612 | sizeof_shareds, task_entry); | |||
| 1613 | ||||
| 1614 | 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); }; | |||
| 1615 | ||||
| 1616 | return retval; | |||
| 1617 | } | |||
| 1618 | ||||
| 1619 | kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1620 | kmp_int32 flags, | |||
| 1621 | size_t sizeof_kmp_task_t, | |||
| 1622 | size_t sizeof_shareds, | |||
| 1623 | kmp_routine_entry_t task_entry, | |||
| 1624 | kmp_int64 device_id) { | |||
| 1625 | auto &input_flags = reinterpret_cast<kmp_tasking_flags_t &>(flags); | |||
| 1626 | // target task is untied defined in the specification | |||
| 1627 | input_flags.tiedness = TASK_UNTIED0; | |||
| 1628 | ||||
| 1629 | if (__kmp_enable_hidden_helper) | |||
| 1630 | input_flags.hidden_helper = TRUE(!0); | |||
| 1631 | ||||
| 1632 | return __kmpc_omp_task_alloc(loc_ref, gtid, flags, sizeof_kmp_task_t, | |||
| 1633 | sizeof_shareds, task_entry); | |||
| 1634 | } | |||
| 1635 | ||||
| 1636 | /*! | |||
| 1637 | @ingroup TASKING | |||
| 1638 | @param loc_ref location of the original task directive | |||
| 1639 | @param gtid Global Thread ID of encountering thread | |||
| 1640 | @param new_task task thunk allocated by __kmpc_omp_task_alloc() for the ''new | |||
| 1641 | task'' | |||
| 1642 | @param naffins Number of affinity items | |||
| 1643 | @param affin_list List of affinity items | |||
| 1644 | @return Returns non-zero if registering affinity information was not successful. | |||
| 1645 | Returns 0 if registration was successful | |||
| 1646 | This entry registers the affinity information attached to a task with the task | |||
| 1647 | thunk structure kmp_taskdata_t. | |||
| 1648 | */ | |||
| 1649 | kmp_int32 | |||
| 1650 | __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1651 | kmp_task_t *new_task, kmp_int32 naffins, | |||
| 1652 | kmp_task_affinity_info_t *affin_list) { | |||
| 1653 | return 0; | |||
| 1654 | } | |||
| 1655 | ||||
| 1656 | // __kmp_invoke_task: invoke the specified task | |||
| 1657 | // | |||
| 1658 | // gtid: global thread ID of caller | |||
| 1659 | // task: the task to invoke | |||
| 1660 | // current_task: the task to resume after task invocation | |||
| 1661 | static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task, | |||
| 1662 | kmp_taskdata_t *current_task) { | |||
| 1663 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 1664 | kmp_info_t *thread; | |||
| 1665 | int discard = 0 /* false */; | |||
| 1666 | 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); } | |||
| 1667 | 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); } | |||
| 1668 | 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); }; | |||
| 1669 | KMP_DEBUG_ASSERT(task)if (!(task)) { __kmp_debug_assert("task", "openmp/runtime/src/kmp_tasking.cpp" , 1669); }; | |||
| 1670 | if (UNLIKELY(taskdata->td_flags.proxy == TASK_PROXY &&__builtin_expect(!!(taskdata->td_flags.proxy == 1 && taskdata->td_flags.complete == 1), 0) | |||
| 1671 | taskdata->td_flags.complete == 1)__builtin_expect(!!(taskdata->td_flags.proxy == 1 && taskdata->td_flags.complete == 1), 0)) { | |||
| 1672 | // This is a proxy task that was already completed but it needs to run | |||
| 1673 | // its bottom-half finish | |||
| 1674 | 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); } | |||
| 1675 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); } | |||
| 1676 | ("__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); } | |||
| 1677 | 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); }; | |||
| 1678 | ||||
| 1679 | __kmp_bottom_half_finish_proxy(gtid, task); | |||
| 1680 | ||||
| 1681 | 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 ); } | |||
| 1682 | "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 ); } | |||
| 1683 | 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 ); }; | |||
| 1684 | ||||
| 1685 | return; | |||
| 1686 | } | |||
| 1687 | ||||
| 1688 | #if OMPT_SUPPORT1 | |||
| 1689 | // For untied tasks, the first task executed only calls __kmpc_omp_task and | |||
| 1690 | // does not execute code. | |||
| 1691 | ompt_thread_info_t oldInfo; | |||
| 1692 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1693 | // Store the threads states and restore them after the task | |||
| 1694 | thread = __kmp_threads[gtid]; | |||
| 1695 | oldInfo = thread->th.ompt_thread_info; | |||
| 1696 | thread->th.ompt_thread_info.wait_id = 0; | |||
| 1697 | thread->th.ompt_thread_info.state = (thread->th.th_team_serialized) | |||
| 1698 | ? ompt_state_work_serial | |||
| 1699 | : ompt_state_work_parallel; | |||
| 1700 | taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
| 1701 | } | |||
| 1702 | #endif | |||
| 1703 | ||||
| 1704 | // Proxy tasks are not handled by the runtime | |||
| 1705 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
| 1706 | __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded | |||
| 1707 | } | |||
| 1708 | ||||
| 1709 | // TODO: cancel tasks if the parallel region has also been cancelled | |||
| 1710 | // TODO: check if this sequence can be hoisted above __kmp_task_start | |||
| 1711 | // if cancellation has been enabled for this run ... | |||
| 1712 | if (UNLIKELY(__kmp_omp_cancellation)__builtin_expect(!!(__kmp_omp_cancellation), 0)) { | |||
| 1713 | thread = __kmp_threads[gtid]; | |||
| 1714 | kmp_team_t *this_team = thread->th.th_team; | |||
| 1715 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
| 1716 | if ((taskgroup && taskgroup->cancel_request) || | |||
| 1717 | (this_team->t.t_cancel_request == cancel_parallel)) { | |||
| 1718 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1719 | ompt_data_t *task_data; | |||
| 1720 | if (UNLIKELY(ompt_enabled.ompt_callback_cancel)__builtin_expect(!!(ompt_enabled.ompt_callback_cancel), 0)) { | |||
| 1721 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, NULL__null, NULL__null); | |||
| 1722 | ompt_callbacks.ompt_callback(ompt_callback_cancel)ompt_callback_cancel_callback( | |||
| 1723 | task_data, | |||
| 1724 | ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup | |||
| 1725 | : ompt_cancel_parallel) | | |||
| 1726 | ompt_cancel_discarded_task, | |||
| 1727 | NULL__null); | |||
| 1728 | } | |||
| 1729 | #endif | |||
| 1730 | KMP_COUNT_BLOCK(TASK_cancelled)((void)0); | |||
| 1731 | // this task belongs to a task group and we need to cancel it | |||
| 1732 | discard = 1 /* true */; | |||
| 1733 | } | |||
| 1734 | } | |||
| 1735 | ||||
| 1736 | // Invoke the task routine and pass in relevant data. | |||
| 1737 | // Thunks generated by gcc take a different argument list. | |||
| 1738 | if (!discard) { | |||
| 1739 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
| 1740 | taskdata->td_last_tied = current_task->td_last_tied; | |||
| 1741 | KMP_DEBUG_ASSERT(taskdata->td_last_tied)if (!(taskdata->td_last_tied)) { __kmp_debug_assert("taskdata->td_last_tied" , "openmp/runtime/src/kmp_tasking.cpp", 1741); }; | |||
| 1742 | } | |||
| 1743 | #if KMP_STATS_ENABLED0 | |||
| 1744 | KMP_COUNT_BLOCK(TASK_executed)((void)0); | |||
| 1745 | switch (KMP_GET_THREAD_STATE()((void)0)) { | |||
| 1746 | case FORK_JOIN_BARRIER: | |||
| 1747 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar)((void)0); | |||
| 1748 | break; | |||
| 1749 | case PLAIN_BARRIER: | |||
| 1750 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar)((void)0); | |||
| 1751 | break; | |||
| 1752 | case TASKYIELD: | |||
| 1753 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield)((void)0); | |||
| 1754 | break; | |||
| 1755 | case TASKWAIT: | |||
| 1756 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait)((void)0); | |||
| 1757 | break; | |||
| 1758 | case TASKGROUP: | |||
| 1759 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup)((void)0); | |||
| 1760 | break; | |||
| 1761 | default: | |||
| 1762 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate)((void)0); | |||
| 1763 | break; | |||
| 1764 | } | |||
| 1765 | #endif // KMP_STATS_ENABLED | |||
| 1766 | ||||
| 1767 | // OMPT task begin | |||
| 1768 | #if OMPT_SUPPORT1 | |||
| 1769 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1770 | __ompt_task_start(task, current_task, gtid); | |||
| 1771 | #endif | |||
| 1772 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1773 | if (UNLIKELY(ompt_enabled.ompt_callback_dispatch &&__builtin_expect(!!(ompt_enabled.ompt_callback_dispatch && taskdata->ompt_task_info.dispatch_chunk.iterations > 0 ), 0) | |||
| 1774 | taskdata->ompt_task_info.dispatch_chunk.iterations > 0)__builtin_expect(!!(ompt_enabled.ompt_callback_dispatch && taskdata->ompt_task_info.dispatch_chunk.iterations > 0 ), 0)) { | |||
| 1775 | ompt_data_t instance = ompt_data_none{0}; | |||
| 1776 | instance.ptr = &(taskdata->ompt_task_info.dispatch_chunk); | |||
| 1777 | ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL__null); | |||
| 1778 | ompt_callbacks.ompt_callback(ompt_callback_dispatch)ompt_callback_dispatch_callback( | |||
| 1779 | &(team_info->parallel_data), &(taskdata->ompt_task_info.task_data), | |||
| 1780 | ompt_dispatch_taskloop_chunk, instance); | |||
| 1781 | taskdata->ompt_task_info.dispatch_chunk = {0, 0}; | |||
| 1782 | } | |||
| 1783 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 1784 | ||||
| 1785 | #if OMPD_SUPPORT1 | |||
| 1786 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
| 1787 | ompd_bp_task_begin(); | |||
| 1788 | #endif | |||
| 1789 | ||||
| 1790 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 1791 | kmp_uint64 cur_time; | |||
| 1792 | kmp_int32 kmp_itt_count_task = | |||
| 1793 | __kmp_forkjoin_frames_mode == 3 && !taskdata->td_flags.task_serial && | |||
| 1794 | current_task->td_flags.tasktype == TASK_IMPLICIT0; | |||
| 1795 | if (kmp_itt_count_task) { | |||
| 1796 | thread = __kmp_threads[gtid]; | |||
| 1797 | // Time outer level explicit task on barrier for adjusting imbalance time | |||
| 1798 | if (thread->th.th_bar_arrive_time) | |||
| 1799 | cur_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | |||
| 1800 | else | |||
| 1801 | kmp_itt_count_task = 0; // thread is not on a barrier - skip timing | |||
| 1802 | } | |||
| 1803 | KMP_FSYNC_ACQUIRED(taskdata)(!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)(taskdata)); // acquired self (new task) | |||
| 1804 | #endif | |||
| 1805 | ||||
| 1806 | #if ENABLE_LIBOMPTARGET1 | |||
| 1807 | if (taskdata->td_target_data.async_handle != NULL__null) { | |||
| 1808 | // If we have a valid target async handle, that means that we have already | |||
| 1809 | // executed the task routine once. We must query for the handle completion | |||
| 1810 | // instead of re-executing the routine. | |||
| 1811 | KMP_ASSERT(tgt_target_nowait_query)if (!(tgt_target_nowait_query)) { __kmp_debug_assert("tgt_target_nowait_query" , "openmp/runtime/src/kmp_tasking.cpp", 1811); }; | |||
| 1812 | tgt_target_nowait_query(&taskdata->td_target_data.async_handle); | |||
| 1813 | } else | |||
| 1814 | #endif | |||
| 1815 | if (task->routine != NULL__null) { | |||
| 1816 | #ifdef KMP_GOMP_COMPAT | |||
| 1817 | if (taskdata->td_flags.native) { | |||
| 1818 | ((void (*)(void *))(*(task->routine)))(task->shareds); | |||
| 1819 | } else | |||
| 1820 | #endif /* KMP_GOMP_COMPAT */ | |||
| 1821 | { | |||
| 1822 | (*(task->routine))(gtid, task); | |||
| 1823 | } | |||
| 1824 | } | |||
| 1825 | KMP_POP_PARTITIONED_TIMER()((void)0); | |||
| 1826 | ||||
| 1827 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 1828 | if (kmp_itt_count_task) { | |||
| 1829 | // Barrier imbalance - adjust arrive time with the task duration | |||
| 1830 | thread->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); | |||
| 1831 | } | |||
| 1832 | KMP_FSYNC_CANCEL(taskdata)(!__kmp_itt_fsync_cancel_ptr__3_0) ? (void)0 : __kmp_itt_fsync_cancel_ptr__3_0 ((void *)(taskdata)); // destroy self (just executed) | |||
| 1833 | KMP_FSYNC_RELEASING(taskdata->td_parent)(!__kmp_itt_fsync_releasing_ptr__3_0) ? (void)0 : __kmp_itt_fsync_releasing_ptr__3_0 ((void *)(taskdata->td_parent)); // releasing parent | |||
| 1834 | #endif | |||
| 1835 | } | |||
| 1836 | ||||
| 1837 | #if OMPD_SUPPORT1 | |||
| 1838 | if (ompd_state & OMPD_ENABLE_BP0x1) | |||
| 1839 | ompd_bp_task_end(); | |||
| 1840 | #endif | |||
| 1841 | ||||
| 1842 | // Proxy tasks are not handled by the runtime | |||
| 1843 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
| 1844 | #if OMPT_SUPPORT1 | |||
| 1845 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1846 | thread->th.ompt_thread_info = oldInfo; | |||
| 1847 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 1848 | taskdata->ompt_task_info.frame.exit_frame = ompt_data_none{0}; | |||
| 1849 | } | |||
| 1850 | __kmp_task_finish<true>(gtid, task, current_task); | |||
| 1851 | } else | |||
| 1852 | #endif | |||
| 1853 | __kmp_task_finish<false>(gtid, task, current_task); | |||
| 1854 | } | |||
| 1855 | ||||
| 1856 | 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); } | |||
| 1857 | 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); } | |||
| 1858 | ("__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); } | |||
| 1859 | 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); }; | |||
| 1860 | return; | |||
| 1861 | } | |||
| 1862 | ||||
| 1863 | // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution | |||
| 1864 | // | |||
| 1865 | // loc_ref: location of original task pragma (ignored) | |||
| 1866 | // gtid: Global Thread ID of encountering thread | |||
| 1867 | // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task'' | |||
| 1868 | // Returns: | |||
| 1869 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1870 | // be resumed later. | |||
| 1871 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1872 | // resumed later. | |||
| 1873 | kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1874 | kmp_task_t *new_task) { | |||
| 1875 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1876 | ||||
| 1877 | 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); } | |||
| 1878 | 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); }; | |||
| 1879 | ||||
| 1880 | #if OMPT_SUPPORT1 | |||
| 1881 | kmp_taskdata_t *parent; | |||
| 1882 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1883 | parent = new_taskdata->td_parent; | |||
| 1884 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 1885 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 1886 | &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame), | |||
| 1887 | &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0, | |||
| 1888 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 1889 | } | |||
| 1890 | } | |||
| 1891 | #endif | |||
| 1892 | ||||
| 1893 | /* Should we execute the new task or queue it? For now, let's just always try | |||
| 1894 | to queue it. If the queue fills up, then we'll execute it. */ | |||
| 1895 | ||||
| 1896 | if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
| 1897 | { // Execute this task immediately | |||
| 1898 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 1899 | new_taskdata->td_flags.task_serial = 1; | |||
| 1900 | __kmp_invoke_task(gtid, new_task, current_task); | |||
| 1901 | } | |||
| 1902 | ||||
| 1903 | 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); } | |||
| 1904 | 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); } | |||
| 1905 | ("__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); } | |||
| 1906 | "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); } | |||
| 1907 | 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); }; | |||
| 1908 | ||||
| 1909 | #if OMPT_SUPPORT1 | |||
| 1910 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1911 | parent->ompt_task_info.frame.enter_frame = ompt_data_none{0}; | |||
| 1912 | } | |||
| 1913 | #endif | |||
| 1914 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1915 | } | |||
| 1916 | ||||
| 1917 | // __kmp_omp_task: Schedule a non-thread-switchable task for execution | |||
| 1918 | // | |||
| 1919 | // gtid: Global Thread ID of encountering thread | |||
| 1920 | // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc() | |||
| 1921 | // serialize_immediate: if TRUE then if the task is executed immediately its | |||
| 1922 | // execution will be serialized | |||
| 1923 | // Returns: | |||
| 1924 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1925 | // be resumed later. | |||
| 1926 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1927 | // resumed later. | |||
| 1928 | kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, | |||
| 1929 | bool serialize_immediate) { | |||
| 1930 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1931 | ||||
| 1932 | /* Should we execute the new task or queue it? For now, let's just always try | |||
| 1933 | to queue it. If the queue fills up, then we'll execute it. */ | |||
| 1934 | if (new_taskdata->td_flags.proxy == TASK_PROXY1 || | |||
| 1935 | __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
| 1936 | { // Execute this task immediately | |||
| 1937 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 1938 | if (serialize_immediate) | |||
| 1939 | new_taskdata->td_flags.task_serial = 1; | |||
| 1940 | __kmp_invoke_task(gtid, new_task, current_task); | |||
| 1941 | } else if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647) && | |||
| 1942 | __kmp_wpolicy_passive) { | |||
| 1943 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
| 1944 | kmp_team_t *team = this_thr->th.th_team; | |||
| 1945 | kmp_int32 nthreads = this_thr->th.th_team_nproc; | |||
| 1946 | for (int i = 0; i < nthreads; ++i) { | |||
| 1947 | kmp_info_t *thread = team->t.t_threads[i]; | |||
| 1948 | if (thread == this_thr) | |||
| 1949 | continue; | |||
| 1950 | if (thread->th.th_sleep_loc != NULL__null) { | |||
| 1951 | __kmp_null_resume_wrapper(thread); | |||
| 1952 | break; // awake one thread at a time | |||
| 1953 | } | |||
| 1954 | } | |||
| 1955 | } | |||
| 1956 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1957 | } | |||
| 1958 | ||||
| 1959 | // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a | |||
| 1960 | // non-thread-switchable task from the parent thread only! | |||
| 1961 | // | |||
| 1962 | // loc_ref: location of original task pragma (ignored) | |||
| 1963 | // gtid: Global Thread ID of encountering thread | |||
| 1964 | // new_task: non-thread-switchable task thunk allocated by | |||
| 1965 | // __kmp_omp_task_alloc() | |||
| 1966 | // Returns: | |||
| 1967 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1968 | // be resumed later. | |||
| 1969 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1970 | // resumed later. | |||
| 1971 | kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1972 | kmp_task_t *new_task) { | |||
| 1973 | kmp_int32 res; | |||
| 1974 | KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK)((void)0); | |||
| 1975 | ||||
| 1976 | #if KMP_DEBUG1 || OMPT_SUPPORT1 | |||
| 1977 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1978 | #endif | |||
| 1979 | 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); } | |||
| 1980 | 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); }; | |||
| 1981 | __kmp_assert_valid_gtid(gtid); | |||
| 1982 | ||||
| 1983 | #if OMPT_SUPPORT1 | |||
| 1984 | kmp_taskdata_t *parent = NULL__null; | |||
| 1985 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1986 | if (!new_taskdata->td_flags.started) { | |||
| 1987 | OMPT_STORE_RETURN_ADDRESS(gtid)OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address (0)};; | |||
| 1988 | parent = new_taskdata->td_parent; | |||
| 1989 | if (!parent->ompt_task_info.frame.enter_frame.ptr) { | |||
| 1990 | parent->ompt_task_info.frame.enter_frame.ptr = | |||
| 1991 | OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
| 1992 | } | |||
| 1993 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 1994 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 1995 | &(parent->ompt_task_info.task_data), | |||
| 1996 | &(parent->ompt_task_info.frame), | |||
| 1997 | &(new_taskdata->ompt_task_info.task_data), | |||
| 1998 | 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, | |||
| 1999 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 2000 | } | |||
| 2001 | } else { | |||
| 2002 | // We are scheduling the continuation of an UNTIED task. | |||
| 2003 | // Scheduling back to the parent task. | |||
| 2004 | __ompt_task_finish(new_task, | |||
| 2005 | new_taskdata->ompt_task_info.scheduling_parent, | |||
| 2006 | ompt_task_switch); | |||
| 2007 | new_taskdata->ompt_task_info.frame.exit_frame = ompt_data_none{0}; | |||
| 2008 | } | |||
| 2009 | } | |||
| 2010 | #endif | |||
| 2011 | ||||
| 2012 | res = __kmp_omp_task(gtid, new_task, true); | |||
| 2013 | ||||
| 2014 | 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 ); } | |||
| 2015 | "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 ); } | |||
| 2016 | 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 ); }; | |||
| 2017 | #if OMPT_SUPPORT1 | |||
| 2018 | if (UNLIKELY(ompt_enabled.enabled && parent != NULL)__builtin_expect(!!(ompt_enabled.enabled && parent != __null), 0)) { | |||
| 2019 | parent->ompt_task_info.frame.enter_frame = ompt_data_none{0}; | |||
| 2020 | } | |||
| 2021 | #endif | |||
| 2022 | return res; | |||
| 2023 | } | |||
| 2024 | ||||
| 2025 | // __kmp_omp_taskloop_task: Wrapper around __kmp_omp_task to schedule | |||
| 2026 | // a taskloop task with the correct OMPT return address | |||
| 2027 | // | |||
| 2028 | // loc_ref: location of original task pragma (ignored) | |||
| 2029 | // gtid: Global Thread ID of encountering thread | |||
| 2030 | // new_task: non-thread-switchable task thunk allocated by | |||
| 2031 | // __kmp_omp_task_alloc() | |||
| 2032 | // codeptr_ra: return address for OMPT callback | |||
| 2033 | // Returns: | |||
| 2034 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 2035 | // be resumed later. | |||
| 2036 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 2037 | // resumed later. | |||
| 2038 | kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid, | |||
| 2039 | kmp_task_t *new_task, void *codeptr_ra) { | |||
| 2040 | kmp_int32 res; | |||
| 2041 | KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK)((void)0); | |||
| 2042 | ||||
| 2043 | #if KMP_DEBUG1 || OMPT_SUPPORT1 | |||
| 2044 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 2045 | #endif | |||
| 2046 | 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); } | |||
| 2047 | 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); }; | |||
| 2048 | ||||
| 2049 | #if OMPT_SUPPORT1 | |||
| 2050 | kmp_taskdata_t *parent = NULL__null; | |||
| 2051 | if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)__builtin_expect(!!(ompt_enabled.enabled && !new_taskdata ->td_flags.started), 0)) { | |||
| 2052 | parent = new_taskdata->td_parent; | |||
| 2053 | if (!parent->ompt_task_info.frame.enter_frame.ptr) | |||
| 2054 | parent->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
| 2055 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 2056 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 2057 | &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame), | |||
| 2058 | &(new_taskdata->ompt_task_info.task_data), | |||
| 2059 | 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, | |||
| 2060 | codeptr_ra); | |||
| 2061 | } | |||
| 2062 | } | |||
| 2063 | #endif | |||
| 2064 | ||||
| 2065 | res = __kmp_omp_task(gtid, new_task, true); | |||
| 2066 | ||||
| 2067 | 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 ); } | |||
| 2068 | "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 ); } | |||
| 2069 | 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 ); }; | |||
| 2070 | #if OMPT_SUPPORT1 | |||
| 2071 | if (UNLIKELY(ompt_enabled.enabled && parent != NULL)__builtin_expect(!!(ompt_enabled.enabled && parent != __null), 0)) { | |||
| 2072 | parent->ompt_task_info.frame.enter_frame = ompt_data_none{0}; | |||
| 2073 | } | |||
| 2074 | #endif | |||
| 2075 | return res; | |||
| 2076 | } | |||
| 2077 | ||||
| 2078 | template <bool ompt> | |||
| 2079 | static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid, | |||
| 2080 | void *frame_address, | |||
| 2081 | void *return_address) { | |||
| 2082 | kmp_taskdata_t *taskdata = nullptr; | |||
| 2083 | kmp_info_t *thread; | |||
| 2084 | int thread_finished = FALSE0; | |||
| 2085 | KMP_SET_THREAD_STATE_BLOCK(TASKWAIT)((void)0); | |||
| 2086 | ||||
| 2087 | 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); }; | |||
| 2088 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_tasking.cpp" , 2088); }; | |||
| 2089 | ||||
| 2090 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
| 2091 | thread = __kmp_threads[gtid]; | |||
| 2092 | taskdata = thread->th.th_current_task; | |||
| 2093 | ||||
| 2094 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2095 | ompt_data_t *my_task_data; | |||
| 2096 | ompt_data_t *my_parallel_data; | |||
| 2097 | ||||
| 2098 | if (ompt) { | |||
| 2099 | my_task_data = &(taskdata->ompt_task_info.task_data); | |||
| 2100 | my_parallel_data = OMPT_CUR_TEAM_DATA(thread)(&(thread->th.th_team->t.ompt_team_info.parallel_data )); | |||
| 2101 | ||||
| 2102 | taskdata->ompt_task_info.frame.enter_frame.ptr = frame_address; | |||
| 2103 | ||||
| 2104 | if (ompt_enabled.ompt_callback_sync_region) { | |||
| 2105 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2106 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
| 2107 | my_task_data, return_address); | |||
| 2108 | } | |||
| 2109 | ||||
| 2110 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
| 2111 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2112 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
| 2113 | my_task_data, return_address); | |||
| 2114 | } | |||
| 2115 | } | |||
| 2116 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 2117 | ||||
| 2118 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
| 2119 | // taskwait. | |||
| 2120 | #if USE_ITT_BUILD1 | |||
| 2121 | // Note: These values are used by ITT events as well. | |||
| 2122 | #endif /* USE_ITT_BUILD */ | |||
| 2123 | taskdata->td_taskwait_counter += 1; | |||
| 2124 | taskdata->td_taskwait_ident = loc_ref; | |||
| 2125 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 2126 | ||||
| 2127 | #if USE_ITT_BUILD1 | |||
| 2128 | void *itt_sync_obj = NULL__null; | |||
| 2129 | #if USE_ITT_NOTIFY1 | |||
| 2130 | KMP_ITT_TASKWAIT_STARTING(itt_sync_obj)if (__builtin_expect(!!(__kmp_itt_sync_create_ptr__3_0), 0)) { itt_sync_obj = __kmp_itt_taskwait_object(gtid); if (itt_sync_obj != __null) { __kmp_itt_taskwait_starting(gtid, itt_sync_obj) ; } }; | |||
| 2131 | #endif /* USE_ITT_NOTIFY */ | |||
| 2132 | #endif /* USE_ITT_BUILD */ | |||
| 2133 | ||||
| 2134 | bool must_wait = | |||
| 2135 | !taskdata->td_flags.team_serial && !taskdata->td_flags.final; | |||
| 2136 | ||||
| 2137 | must_wait = must_wait || (thread->th.th_task_team != NULL__null && | |||
| 2138 | thread->th.th_task_team->tt.tt_found_proxy_tasks); | |||
| 2139 | // If hidden helper thread is encountered, we must enable wait here. | |||
| 2140 | must_wait = | |||
| 2141 | must_wait || | |||
| 2142 | (__kmp_enable_hidden_helper && thread->th.th_task_team != NULL__null && | |||
| 2143 | thread->th.th_task_team->tt.tt_hidden_helper_task_encountered); | |||
| 2144 | ||||
| 2145 | if (must_wait) { | |||
| 2146 | kmp_flag_32<false, false> flag( | |||
| 2147 | RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)) | |||
| 2148 | &(taskdata->td_incomplete_child_tasks))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)), | |||
| 2149 | 0U); | |||
| 2150 | while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) != 0) { | |||
| 2151 | flag.execute_tasks(thread, gtid, FALSE0, | |||
| 2152 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 2153 | __kmp_task_stealing_constraint); | |||
| 2154 | } | |||
| 2155 | } | |||
| 2156 | #if USE_ITT_BUILD1 | |||
| 2157 | KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj)if (__builtin_expect(!!(itt_sync_obj != __null), 0)) __kmp_itt_taskwait_finished (gtid, itt_sync_obj);; | |||
| 2158 | KMP_FSYNC_ACQUIRED(taskdata)(!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)(taskdata)); // acquire self - sync with children | |||
| 2159 | #endif /* USE_ITT_BUILD */ | |||
| 2160 | ||||
| 2161 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
| 2162 | // negated. | |||
| 2163 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
| 2164 | ||||
| 2165 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2166 | if (ompt) { | |||
| 2167 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
| 2168 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2169 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
| 2170 | my_task_data, return_address); | |||
| 2171 | } | |||
| 2172 | if (ompt_enabled.ompt_callback_sync_region) { | |||
| 2173 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2174 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
| 2175 | my_task_data, return_address); | |||
| 2176 | } | |||
| 2177 | taskdata->ompt_task_info.frame.enter_frame = ompt_data_none{0}; | |||
| 2178 | } | |||
| 2179 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 2180 | ||||
| 2181 | } | |||
| 2182 | ||||
| 2183 | 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); } | |||
| 2184 | "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); } | |||
| 2185 | 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); }; | |||
| 2186 | ||||
| 2187 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 2188 | } | |||
| 2189 | ||||
| 2190 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2191 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 2192 | static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 2193 | void *frame_address, | |||
| 2194 | void *return_address) { | |||
| 2195 | return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address, | |||
| 2196 | return_address); | |||
| 2197 | } | |||
| 2198 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 2199 | ||||
| 2200 | // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are | |||
| 2201 | // complete | |||
| 2202 | kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) { | |||
| 2203 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2204 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 2205 | OMPT_STORE_RETURN_ADDRESS(gtid)OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address (0)};; | |||
| 2206 | return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0), | |||
| 2207 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 2208 | } | |||
| 2209 | #endif | |||
| 2210 | return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL__null, NULL__null); | |||
| 2211 | } | |||
| 2212 | ||||
| 2213 | // __kmpc_omp_taskyield: switch to a different task | |||
| 2214 | kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) { | |||
| 2215 | kmp_taskdata_t *taskdata = NULL__null; | |||
| 2216 | kmp_info_t *thread; | |||
| 2217 | int thread_finished = FALSE0; | |||
| 2218 | ||||
| 2219 | KMP_COUNT_BLOCK(OMP_TASKYIELD)((void)0); | |||
| 2220 | KMP_SET_THREAD_STATE_BLOCK(TASKYIELD)((void)0); | |||
| 2221 | ||||
| 2222 | 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); } | |||
| 2223 | 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); }; | |||
| 2224 | __kmp_assert_valid_gtid(gtid); | |||
| 2225 | ||||
| 2226 | if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) { | |||
| 2227 | thread = __kmp_threads[gtid]; | |||
| 2228 | taskdata = thread->th.th_current_task; | |||
| 2229 | // Should we model this as a task wait or not? | |||
| 2230 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
| 2231 | // taskwait. | |||
| 2232 | #if USE_ITT_BUILD1 | |||
| 2233 | // Note: These values are used by ITT events as well. | |||
| 2234 | #endif /* USE_ITT_BUILD */ | |||
| 2235 | taskdata->td_taskwait_counter += 1; | |||
| 2236 | taskdata->td_taskwait_ident = loc_ref; | |||
| 2237 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 2238 | ||||
| 2239 | #if USE_ITT_BUILD1 | |||
| 2240 | void *itt_sync_obj = NULL__null; | |||
| 2241 | #if USE_ITT_NOTIFY1 | |||
| 2242 | KMP_ITT_TASKWAIT_STARTING(itt_sync_obj)if (__builtin_expect(!!(__kmp_itt_sync_create_ptr__3_0), 0)) { itt_sync_obj = __kmp_itt_taskwait_object(gtid); if (itt_sync_obj != __null) { __kmp_itt_taskwait_starting(gtid, itt_sync_obj) ; } }; | |||
| 2243 | #endif /* USE_ITT_NOTIFY */ | |||
| 2244 | #endif /* USE_ITT_BUILD */ | |||
| 2245 | if (!taskdata->td_flags.team_serial) { | |||
| 2246 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 2247 | if (task_team != NULL__null) { | |||
| 2248 | if (KMP_TASKING_ENABLED(task_team)((!0) == ((task_team)->tt.tt_found_tasks))) { | |||
| 2249 | #if OMPT_SUPPORT1 | |||
| 2250 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 2251 | thread->th.ompt_thread_info.ompt_task_yielded = 1; | |||
| 2252 | #endif | |||
| 2253 | __kmp_execute_tasks_32( | |||
| 2254 | thread, gtid, (kmp_flag_32<> *)NULL__null, FALSE0, | |||
| 2255 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 2256 | __kmp_task_stealing_constraint); | |||
| 2257 | #if OMPT_SUPPORT1 | |||
| 2258 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 2259 | thread->th.ompt_thread_info.ompt_task_yielded = 0; | |||
| 2260 | #endif | |||
| 2261 | } | |||
| 2262 | } | |||
| 2263 | } | |||
| 2264 | #if USE_ITT_BUILD1 | |||
| 2265 | KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj)if (__builtin_expect(!!(itt_sync_obj != __null), 0)) __kmp_itt_taskwait_finished (gtid, itt_sync_obj);; | |||
| 2266 | #endif /* USE_ITT_BUILD */ | |||
| 2267 | ||||
| 2268 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
| 2269 | // negated. | |||
| 2270 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
| 2271 | } | |||
| 2272 | ||||
| 2273 | 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); } | |||
| 2274 | "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); } | |||
| 2275 | 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); }; | |||
| 2276 | ||||
| 2277 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 2278 | } | |||
| 2279 | ||||
| 2280 | // Task Reduction implementation | |||
| 2281 | // | |||
| 2282 | // Note: initial implementation didn't take into account the possibility | |||
| 2283 | // to specify omp_orig for initializer of the UDR (user defined reduction). | |||
| 2284 | // Corrected implementation takes into account the omp_orig object. | |||
| 2285 | // Compiler is free to use old implementation if omp_orig is not specified. | |||
| 2286 | ||||
| 2287 | /*! | |||
| 2288 | @ingroup BASIC_TYPES | |||
| 2289 | @{ | |||
| 2290 | */ | |||
| 2291 | ||||
| 2292 | /*! | |||
| 2293 | Flags for special info per task reduction item. | |||
| 2294 | */ | |||
| 2295 | typedef struct kmp_taskred_flags { | |||
| 2296 | /*! 1 - use lazy alloc/init (e.g. big objects, num tasks < num threads) */ | |||
| 2297 | unsigned lazy_priv : 1; | |||
| 2298 | unsigned reserved31 : 31; | |||
| 2299 | } kmp_taskred_flags_t; | |||
| 2300 | ||||
| 2301 | /*! | |||
| 2302 | Internal struct for reduction data item related info set up by compiler. | |||
| 2303 | */ | |||
| 2304 | typedef struct kmp_task_red_input { | |||
| 2305 | void *reduce_shar; /**< shared between tasks item to reduce into */ | |||
| 2306 | size_t reduce_size; /**< size of data item in bytes */ | |||
| 2307 | // three compiler-generated routines (init, fini are optional): | |||
| 2308 | void *reduce_init; /**< data initialization routine (single parameter) */ | |||
| 2309 | void *reduce_fini; /**< data finalization routine */ | |||
| 2310 | void *reduce_comb; /**< data combiner routine */ | |||
| 2311 | kmp_taskred_flags_t flags; /**< flags for additional info from compiler */ | |||
| 2312 | } kmp_task_red_input_t; | |||
| 2313 | ||||
| 2314 | /*! | |||
| 2315 | Internal struct for reduction data item related info saved by the library. | |||
| 2316 | */ | |||
| 2317 | typedef struct kmp_taskred_data { | |||
| 2318 | void *reduce_shar; /**< shared between tasks item to reduce into */ | |||
| 2319 | size_t reduce_size; /**< size of data item */ | |||
| 2320 | kmp_taskred_flags_t flags; /**< flags for additional info from compiler */ | |||
| 2321 | void *reduce_priv; /**< array of thread specific items */ | |||
| 2322 | void *reduce_pend; /**< end of private data for faster comparison op */ | |||
| 2323 | // three compiler-generated routines (init, fini are optional): | |||
| 2324 | void *reduce_comb; /**< data combiner routine */ | |||
| 2325 | void *reduce_init; /**< data initialization routine (two parameters) */ | |||
| 2326 | void *reduce_fini; /**< data finalization routine */ | |||
| 2327 | void *reduce_orig; /**< original item (can be used in UDR initializer) */ | |||
| 2328 | } kmp_taskred_data_t; | |||
| 2329 | ||||
| 2330 | /*! | |||
| 2331 | Internal struct for reduction data item related info set up by compiler. | |||
| 2332 | ||||
| 2333 | New interface: added reduce_orig field to provide omp_orig for UDR initializer. | |||
| 2334 | */ | |||
| 2335 | typedef struct kmp_taskred_input { | |||
| 2336 | void *reduce_shar; /**< shared between tasks item to reduce into */ | |||
| 2337 | void *reduce_orig; /**< original reduction item used for initialization */ | |||
| 2338 | size_t reduce_size; /**< size of data item */ | |||
| 2339 | // three compiler-generated routines (init, fini are optional): | |||
| 2340 | void *reduce_init; /**< data initialization routine (two parameters) */ | |||
| 2341 | void *reduce_fini; /**< data finalization routine */ | |||
| 2342 | void *reduce_comb; /**< data combiner routine */ | |||
| 2343 | kmp_taskred_flags_t flags; /**< flags for additional info from compiler */ | |||
| 2344 | } kmp_taskred_input_t; | |||
| 2345 | /*! | |||
| 2346 | @} | |||
| 2347 | */ | |||
| 2348 | ||||
| 2349 | template <typename T> void __kmp_assign_orig(kmp_taskred_data_t &item, T &src); | |||
| 2350 | template <> | |||
| 2351 | void __kmp_assign_orig<kmp_task_red_input_t>(kmp_taskred_data_t &item, | |||
| 2352 | kmp_task_red_input_t &src) { | |||
| 2353 | item.reduce_orig = NULL__null; | |||
| 2354 | } | |||
| 2355 | template <> | |||
| 2356 | void __kmp_assign_orig<kmp_taskred_input_t>(kmp_taskred_data_t &item, | |||
| 2357 | kmp_taskred_input_t &src) { | |||
| 2358 | if (src.reduce_orig != NULL__null) { | |||
| 2359 | item.reduce_orig = src.reduce_orig; | |||
| 2360 | } else { | |||
| 2361 | item.reduce_orig = src.reduce_shar; | |||
| 2362 | } // non-NULL reduce_orig means new interface used | |||
| 2363 | } | |||
| 2364 | ||||
| 2365 | template <typename T> void __kmp_call_init(kmp_taskred_data_t &item, size_t j); | |||
| 2366 | template <> | |||
| 2367 | void __kmp_call_init<kmp_task_red_input_t>(kmp_taskred_data_t &item, | |||
| 2368 | size_t offset) { | |||
| 2369 | ((void (*)(void *))item.reduce_init)((char *)(item.reduce_priv) + offset); | |||
| 2370 | } | |||
| 2371 | template <> | |||
| 2372 | void __kmp_call_init<kmp_taskred_input_t>(kmp_taskred_data_t &item, | |||
| 2373 | size_t offset) { | |||
| 2374 | ((void (*)(void *, void *))item.reduce_init)( | |||
| 2375 | (char *)(item.reduce_priv) + offset, item.reduce_orig); | |||
| 2376 | } | |||
| 2377 | ||||
| 2378 | template <typename T> | |||
| 2379 | void *__kmp_task_reduction_init(int gtid, int num, T *data) { | |||
| 2380 | __kmp_assert_valid_gtid(gtid); | |||
| 2381 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 2382 | kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup; | |||
| 2383 | kmp_uint32 nth = thread->th.th_team_nproc; | |||
| 2384 | kmp_taskred_data_t *arr; | |||
| 2385 | ||||
| 2386 | // check input data just in case | |||
| 2387 | KMP_ASSERT(tg != NULL)if (!(tg != __null)) { __kmp_debug_assert("tg != NULL", "openmp/runtime/src/kmp_tasking.cpp" , 2387); }; | |||
| 2388 | KMP_ASSERT(data != NULL)if (!(data != __null)) { __kmp_debug_assert("data != NULL", "openmp/runtime/src/kmp_tasking.cpp" , 2388); }; | |||
| 2389 | KMP_ASSERT(num > 0)if (!(num > 0)) { __kmp_debug_assert("num > 0", "openmp/runtime/src/kmp_tasking.cpp" , 2389); }; | |||
| 2390 | if (nth == 1) { | |||
| 2391 | 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); } | |||
| 2392 | gtid, tg))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n" , gtid, tg); }; | |||
| 2393 | return (void *)tg; | |||
| 2394 | } | |||
| 2395 | 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); } | |||
| 2396 | 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); }; | |||
| 2397 | arr = (kmp_taskred_data_t *)__kmp_thread_malloc(___kmp_thread_malloc((thread), (num * sizeof(kmp_taskred_data_t )), "openmp/runtime/src/kmp_tasking.cpp", 2398) | |||
| 2398 | thread, num * sizeof(kmp_taskred_data_t))___kmp_thread_malloc((thread), (num * sizeof(kmp_taskred_data_t )), "openmp/runtime/src/kmp_tasking.cpp", 2398); | |||
| 2399 | for (int i = 0; i < num; ++i) { | |||
| 2400 | size_t size = data[i].reduce_size - 1; | |||
| 2401 | // round the size up to cache line per thread-specific item | |||
| 2402 | size += CACHE_LINE64 - size % CACHE_LINE64; | |||
| 2403 | KMP_ASSERT(data[i].reduce_comb != NULL)if (!(data[i].reduce_comb != __null)) { __kmp_debug_assert("data[i].reduce_comb != NULL" , "openmp/runtime/src/kmp_tasking.cpp", 2403); }; // combiner is mandatory | |||
| 2404 | arr[i].reduce_shar = data[i].reduce_shar; | |||
| 2405 | arr[i].reduce_size = size; | |||
| 2406 | arr[i].flags = data[i].flags; | |||
| 2407 | arr[i].reduce_comb = data[i].reduce_comb; | |||
| 2408 | arr[i].reduce_init = data[i].reduce_init; | |||
| 2409 | arr[i].reduce_fini = data[i].reduce_fini; | |||
| 2410 | __kmp_assign_orig<T>(arr[i], data[i]); | |||
| 2411 | if (!arr[i].flags.lazy_priv) { | |||
| 2412 | // allocate cache-line aligned block and fill it with zeros | |||
| 2413 | arr[i].reduce_priv = __kmp_allocate(nth * size)___kmp_allocate((nth * size), "openmp/runtime/src/kmp_tasking.cpp" , 2413); | |||
| 2414 | arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size; | |||
| 2415 | if (arr[i].reduce_init != NULL__null) { | |||
| 2416 | // initialize all thread-specific items | |||
| 2417 | for (size_t j = 0; j < nth; ++j) { | |||
| 2418 | __kmp_call_init<T>(arr[i], j * size); | |||
| 2419 | } | |||
| 2420 | } | |||
| 2421 | } else { | |||
| 2422 | // only allocate space for pointers now, | |||
| 2423 | // objects will be lazily allocated/initialized if/when requested | |||
| 2424 | // note that __kmp_allocate zeroes the allocated memory | |||
| 2425 | arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *))___kmp_allocate((nth * sizeof(void *)), "openmp/runtime/src/kmp_tasking.cpp" , 2425); | |||
| 2426 | } | |||
| 2427 | } | |||
| 2428 | tg->reduce_data = (void *)arr; | |||
| 2429 | tg->reduce_num_data = num; | |||
| 2430 | return (void *)tg; | |||
| 2431 | } | |||
| 2432 | ||||
| 2433 | /*! | |||
| 2434 | @ingroup TASKING | |||
| 2435 | @param gtid Global thread ID | |||
| 2436 | @param num Number of data items to reduce | |||
| 2437 | @param data Array of data for reduction | |||
| 2438 | @return The taskgroup identifier | |||
| 2439 | ||||
| 2440 | Initialize task reduction for the taskgroup. | |||
| 2441 | ||||
| 2442 | Note: this entry supposes the optional compiler-generated initializer routine | |||
| 2443 | has single parameter - pointer to object to be initialized. That means | |||
| 2444 | the reduction either does not use omp_orig object, or the omp_orig is accessible | |||
| 2445 | without help of the runtime library. | |||
| 2446 | */ | |||
| 2447 | void *__kmpc_task_reduction_init(int gtid, int num, void *data) { | |||
| 2448 | return __kmp_task_reduction_init(gtid, num, (kmp_task_red_input_t *)data); | |||
| 2449 | } | |||
| 2450 | ||||
| 2451 | /*! | |||
| 2452 | @ingroup TASKING | |||
| 2453 | @param gtid Global thread ID | |||
| 2454 | @param num Number of data items to reduce | |||
| 2455 | @param data Array of data for reduction | |||
| 2456 | @return The taskgroup identifier | |||
| 2457 | ||||
| 2458 | Initialize task reduction for the taskgroup. | |||
| 2459 | ||||
| 2460 | Note: this entry supposes the optional compiler-generated initializer routine | |||
| 2461 | has two parameters, pointer to object to be initialized and pointer to omp_orig | |||
| 2462 | */ | |||
| 2463 | void *__kmpc_taskred_init(int gtid, int num, void *data) { | |||
| 2464 | return __kmp_task_reduction_init(gtid, num, (kmp_taskred_input_t *)data); | |||
| 2465 | } | |||
| 2466 | ||||
| 2467 | // Copy task reduction data (except for shared pointers). | |||
| 2468 | template <typename T> | |||
| 2469 | void __kmp_task_reduction_init_copy(kmp_info_t *thr, int num, T *data, | |||
| 2470 | kmp_taskgroup_t *tg, void *reduce_data) { | |||
| 2471 | kmp_taskred_data_t *arr; | |||
| 2472 | KA_TRACE(20, ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p,"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p," " from data %p\n", thr, tg, reduce_data); } | |||
| 2473 | " from data %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p," " from data %p\n", thr, tg, reduce_data); } | |||
| 2474 | thr, tg, reduce_data))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p," " from data %p\n", thr, tg, reduce_data); }; | |||
| 2475 | arr = (kmp_taskred_data_t *)__kmp_thread_malloc(___kmp_thread_malloc((thr), (num * sizeof(kmp_taskred_data_t) ), "openmp/runtime/src/kmp_tasking.cpp", 2476) | |||
| 2476 | thr, num * sizeof(kmp_taskred_data_t))___kmp_thread_malloc((thr), (num * sizeof(kmp_taskred_data_t) ), "openmp/runtime/src/kmp_tasking.cpp", 2476); | |||
| 2477 | // threads will share private copies, thunk routines, sizes, flags, etc.: | |||
| 2478 | KMP_MEMCPYmemcpy(arr, reduce_data, num * sizeof(kmp_taskred_data_t)); | |||
| 2479 | for (int i = 0; i < num; ++i) { | |||
| 2480 | arr[i].reduce_shar = data[i].reduce_shar; // init unique shared pointers | |||
| 2481 | } | |||
| 2482 | tg->reduce_data = (void *)arr; | |||
| 2483 | tg->reduce_num_data = num; | |||
| 2484 | } | |||
| 2485 | ||||
| 2486 | /*! | |||
| 2487 | @ingroup TASKING | |||
| 2488 | @param gtid Global thread ID | |||
| 2489 | @param tskgrp The taskgroup ID (optional) | |||
| 2490 | @param data Shared location of the item | |||
| 2491 | @return The pointer to per-thread data | |||
| 2492 | ||||
| 2493 | Get thread-specific location of data item | |||
| 2494 | */ | |||
| 2495 | void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) { | |||
| 2496 | __kmp_assert_valid_gtid(gtid); | |||
| 2497 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 2498 | kmp_int32 nth = thread->th.th_team_nproc; | |||
| 2499 | if (nth == 1) | |||
| 2500 | return data; // nothing to do | |||
| 2501 | ||||
| 2502 | kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp; | |||
| 2503 | if (tg == NULL__null) | |||
| 2504 | tg = thread->th.th_current_task->td_taskgroup; | |||
| 2505 | KMP_ASSERT(tg != NULL)if (!(tg != __null)) { __kmp_debug_assert("tg != NULL", "openmp/runtime/src/kmp_tasking.cpp" , 2505); }; | |||
| 2506 | kmp_taskred_data_t *arr = (kmp_taskred_data_t *)(tg->reduce_data); | |||
| 2507 | kmp_int32 num = tg->reduce_num_data; | |||
| 2508 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
| 2509 | ||||
| 2510 | KMP_ASSERT(data != NULL)if (!(data != __null)) { __kmp_debug_assert("data != NULL", "openmp/runtime/src/kmp_tasking.cpp" , 2510); }; | |||
| 2511 | while (tg != NULL__null) { | |||
| 2512 | for (int i = 0; i < num; ++i) { | |||
| 2513 | if (!arr[i].flags.lazy_priv) { | |||
| 2514 | if (data == arr[i].reduce_shar || | |||
| 2515 | (data >= arr[i].reduce_priv && data < arr[i].reduce_pend)) | |||
| 2516 | return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size; | |||
| 2517 | } else { | |||
| 2518 | // check shared location first | |||
| 2519 | void **p_priv = (void **)(arr[i].reduce_priv); | |||
| 2520 | if (data == arr[i].reduce_shar) | |||
| 2521 | goto found; | |||
| 2522 | // check if we get some thread specific location as parameter | |||
| 2523 | for (int j = 0; j < nth; ++j) | |||
| 2524 | if (data == p_priv[j]) | |||
| 2525 | goto found; | |||
| 2526 | continue; // not found, continue search | |||
| 2527 | found: | |||
| 2528 | if (p_priv[tid] == NULL__null) { | |||
| 2529 | // allocate thread specific object lazily | |||
| 2530 | p_priv[tid] = __kmp_allocate(arr[i].reduce_size)___kmp_allocate((arr[i].reduce_size), "openmp/runtime/src/kmp_tasking.cpp" , 2530); | |||
| 2531 | if (arr[i].reduce_init != NULL__null) { | |||
| 2532 | if (arr[i].reduce_orig != NULL__null) { // new interface | |||
| 2533 | ((void (*)(void *, void *))arr[i].reduce_init)( | |||
| 2534 | p_priv[tid], arr[i].reduce_orig); | |||
| 2535 | } else { // old interface (single parameter) | |||
| 2536 | ((void (*)(void *))arr[i].reduce_init)(p_priv[tid]); | |||
| 2537 | } | |||
| 2538 | } | |||
| 2539 | } | |||
| 2540 | return p_priv[tid]; | |||
| 2541 | } | |||
| 2542 | } | |||
| 2543 | tg = tg->parent; | |||
| 2544 | arr = (kmp_taskred_data_t *)(tg->reduce_data); | |||
| 2545 | num = tg->reduce_num_data; | |||
| 2546 | } | |||
| 2547 | KMP_ASSERT2(0, "Unknown task reduction item")if (!(0)) { __kmp_debug_assert(("Unknown task reduction item" ), "openmp/runtime/src/kmp_tasking.cpp", 2547); }; | |||
| 2548 | return NULL__null; // ERROR, this line never executed | |||
| 2549 | } | |||
| 2550 | ||||
| 2551 | // Finalize task reduction. | |||
| 2552 | // Called from __kmpc_end_taskgroup() | |||
| 2553 | static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) { | |||
| 2554 | kmp_int32 nth = th->th.th_team_nproc; | |||
| 2555 | KMP_DEBUG_ASSERT(nth > 1)if (!(nth > 1)) { __kmp_debug_assert("nth > 1", "openmp/runtime/src/kmp_tasking.cpp" , 2555); }; // should not be called if nth == 1 | |||
| 2556 | kmp_taskred_data_t *arr = (kmp_taskred_data_t *)tg->reduce_data; | |||
| 2557 | kmp_int32 num = tg->reduce_num_data; | |||
| 2558 | for (int i = 0; i < num; ++i) { | |||
| 2559 | void *sh_data = arr[i].reduce_shar; | |||
| 2560 | void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini); | |||
| 2561 | void (*f_comb)(void *, void *) = | |||
| 2562 | (void (*)(void *, void *))(arr[i].reduce_comb); | |||
| 2563 | if (!arr[i].flags.lazy_priv) { | |||
| 2564 | void *pr_data = arr[i].reduce_priv; | |||
| 2565 | size_t size = arr[i].reduce_size; | |||
| 2566 | for (int j = 0; j < nth; ++j) { | |||
| 2567 | void *priv_data = (char *)pr_data + j * size; | |||
| 2568 | f_comb(sh_data, priv_data); // combine results | |||
| 2569 | if (f_fini) | |||
| 2570 | f_fini(priv_data); // finalize if needed | |||
| 2571 | } | |||
| 2572 | } else { | |||
| 2573 | void **pr_data = (void **)(arr[i].reduce_priv); | |||
| 2574 | for (int j = 0; j < nth; ++j) { | |||
| 2575 | if (pr_data[j] != NULL__null) { | |||
| 2576 | f_comb(sh_data, pr_data[j]); // combine results | |||
| 2577 | if (f_fini) | |||
| 2578 | f_fini(pr_data[j]); // finalize if needed | |||
| 2579 | __kmp_free(pr_data[j])___kmp_free((pr_data[j]), "openmp/runtime/src/kmp_tasking.cpp" , 2579); | |||
| 2580 | } | |||
| 2581 | } | |||
| 2582 | } | |||
| 2583 | __kmp_free(arr[i].reduce_priv)___kmp_free((arr[i].reduce_priv), "openmp/runtime/src/kmp_tasking.cpp" , 2583); | |||
| 2584 | } | |||
| 2585 | __kmp_thread_free(th, arr)___kmp_thread_free((th), (arr), "openmp/runtime/src/kmp_tasking.cpp" , 2585); | |||
| 2586 | tg->reduce_data = NULL__null; | |||
| 2587 | tg->reduce_num_data = 0; | |||
| 2588 | } | |||
| 2589 | ||||
| 2590 | // Cleanup task reduction data for parallel or worksharing, | |||
| 2591 | // do not touch task private data other threads still working with. | |||
| 2592 | // Called from __kmpc_end_taskgroup() | |||
| 2593 | static void __kmp_task_reduction_clean(kmp_info_t *th, kmp_taskgroup_t *tg) { | |||
| 2594 | __kmp_thread_free(th, tg->reduce_data)___kmp_thread_free((th), (tg->reduce_data), "openmp/runtime/src/kmp_tasking.cpp" , 2594); | |||
| 2595 | tg->reduce_data = NULL__null; | |||
| 2596 | tg->reduce_num_data = 0; | |||
| 2597 | } | |||
| 2598 | ||||
| 2599 | template <typename T> | |||
| 2600 | void *__kmp_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, | |||
| 2601 | int num, T *data) { | |||
| 2602 | __kmp_assert_valid_gtid(gtid); | |||
| 2603 | kmp_info_t *thr = __kmp_threads[gtid]; | |||
| 2604 | kmp_int32 nth = thr->th.th_team_nproc; | |||
| 2605 | __kmpc_taskgroup(loc, gtid); // form new taskgroup first | |||
| 2606 | if (nth == 1) { | |||
| 2607 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n" , gtid, thr->th.th_current_task->td_taskgroup); } | |||
| 2608 | ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n" , gtid, thr->th.th_current_task->td_taskgroup); } | |||
| 2609 | gtid, thr->th.th_current_task->td_taskgroup))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n" , gtid, thr->th.th_current_task->td_taskgroup); }; | |||
| 2610 | return (void *)thr->th.th_current_task->td_taskgroup; | |||
| 2611 | } | |||
| 2612 | kmp_team_t *team = thr->th.th_team; | |||
| 2613 | void *reduce_data; | |||
| 2614 | kmp_taskgroup_t *tg; | |||
| 2615 | reduce_data = KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[is_ws])(&team->t.t_tg_reduce_data[is_ws])->load(std::memory_order_relaxed ); | |||
| 2616 | if (reduce_data == NULL__null && | |||
| 2617 | __kmp_atomic_compare_store(&team->t.t_tg_reduce_data[is_ws], reduce_data, | |||
| 2618 | (void *)1)) { | |||
| 2619 | // single thread enters this block to initialize common reduction data | |||
| 2620 | KMP_DEBUG_ASSERT(reduce_data == NULL)if (!(reduce_data == __null)) { __kmp_debug_assert("reduce_data == __null" , "openmp/runtime/src/kmp_tasking.cpp", 2620); }; | |||
| 2621 | // first initialize own data, then make a copy other threads can use | |||
| 2622 | tg = (kmp_taskgroup_t *)__kmp_task_reduction_init<T>(gtid, num, data); | |||
| 2623 | reduce_data = __kmp_thread_malloc(thr, num * sizeof(kmp_taskred_data_t))___kmp_thread_malloc((thr), (num * sizeof(kmp_taskred_data_t) ), "openmp/runtime/src/kmp_tasking.cpp", 2623); | |||
| 2624 | KMP_MEMCPYmemcpy(reduce_data, tg->reduce_data, num * sizeof(kmp_taskred_data_t)); | |||
| 2625 | // fini counters should be 0 at this point | |||
| 2626 | KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[0]) == 0)if (!((&team->t.t_tg_fini_counter[0])->load(std::memory_order_relaxed ) == 0)) { __kmp_debug_assert("(&team->t.t_tg_fini_counter[0])->load(std::memory_order_relaxed) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 2626); }; | |||
| 2627 | KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[1]) == 0)if (!((&team->t.t_tg_fini_counter[1])->load(std::memory_order_relaxed ) == 0)) { __kmp_debug_assert("(&team->t.t_tg_fini_counter[1])->load(std::memory_order_relaxed) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 2627); }; | |||
| 2628 | KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[is_ws], reduce_data)(&team->t.t_tg_reduce_data[is_ws])->store(reduce_data , std::memory_order_release); | |||
| 2629 | } else { | |||
| 2630 | while ( | |||
| 2631 | (reduce_data = KMP_ATOMIC_LD_ACQ(&team->t.t_tg_reduce_data[is_ws])(&team->t.t_tg_reduce_data[is_ws])->load(std::memory_order_acquire )) == | |||
| 2632 | (void *)1) { // wait for task reduction initialization | |||
| 2633 | KMP_CPU_PAUSE()__kmp_x86_pause(); | |||
| 2634 | } | |||
| 2635 | KMP_DEBUG_ASSERT(reduce_data > (void *)1)if (!(reduce_data > (void *)1)) { __kmp_debug_assert("reduce_data > (void *)1" , "openmp/runtime/src/kmp_tasking.cpp", 2635); }; // should be valid pointer here | |||
| 2636 | tg = thr->th.th_current_task->td_taskgroup; | |||
| 2637 | __kmp_task_reduction_init_copy<T>(thr, num, data, tg, reduce_data); | |||
| 2638 | } | |||
| 2639 | return tg; | |||
| 2640 | } | |||
| 2641 | ||||
| 2642 | /*! | |||
| 2643 | @ingroup TASKING | |||
| 2644 | @param loc Source location info | |||
| 2645 | @param gtid Global thread ID | |||
| 2646 | @param is_ws Is 1 if the reduction is for worksharing, 0 otherwise | |||
| 2647 | @param num Number of data items to reduce | |||
| 2648 | @param data Array of data for reduction | |||
| 2649 | @return The taskgroup identifier | |||
| 2650 | ||||
| 2651 | Initialize task reduction for a parallel or worksharing. | |||
| 2652 | ||||
| 2653 | Note: this entry supposes the optional compiler-generated initializer routine | |||
| 2654 | has single parameter - pointer to object to be initialized. That means | |||
| 2655 | the reduction either does not use omp_orig object, or the omp_orig is accessible | |||
| 2656 | without help of the runtime library. | |||
| 2657 | */ | |||
| 2658 | void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, | |||
| 2659 | int num, void *data) { | |||
| 2660 | return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num, | |||
| 2661 | (kmp_task_red_input_t *)data); | |||
| 2662 | } | |||
| 2663 | ||||
| 2664 | /*! | |||
| 2665 | @ingroup TASKING | |||
| 2666 | @param loc Source location info | |||
| 2667 | @param gtid Global thread ID | |||
| 2668 | @param is_ws Is 1 if the reduction is for worksharing, 0 otherwise | |||
| 2669 | @param num Number of data items to reduce | |||
| 2670 | @param data Array of data for reduction | |||
| 2671 | @return The taskgroup identifier | |||
| 2672 | ||||
| 2673 | Initialize task reduction for a parallel or worksharing. | |||
| 2674 | ||||
| 2675 | Note: this entry supposes the optional compiler-generated initializer routine | |||
| 2676 | has two parameters, pointer to object to be initialized and pointer to omp_orig | |||
| 2677 | */ | |||
| 2678 | void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, | |||
| 2679 | void *data) { | |||
| 2680 | return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num, | |||
| 2681 | (kmp_taskred_input_t *)data); | |||
| 2682 | } | |||
| 2683 | ||||
| 2684 | /*! | |||
| 2685 | @ingroup TASKING | |||
| 2686 | @param loc Source location info | |||
| 2687 | @param gtid Global thread ID | |||
| 2688 | @param is_ws Is 1 if the reduction is for worksharing, 0 otherwise | |||
| 2689 | ||||
| 2690 | Finalize task reduction for a parallel or worksharing. | |||
| 2691 | */ | |||
| 2692 | void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws) { | |||
| 2693 | __kmpc_end_taskgroup(loc, gtid); | |||
| 2694 | } | |||
| 2695 | ||||
| 2696 | // __kmpc_taskgroup: Start a new taskgroup | |||
| 2697 | void __kmpc_taskgroup(ident_t *loc, int gtid) { | |||
| 2698 | __kmp_assert_valid_gtid(gtid); | |||
| 2699 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 2700 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 2701 | kmp_taskgroup_t *tg_new = | |||
| 2702 | (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t))___kmp_thread_malloc((thread), (sizeof(kmp_taskgroup_t)), "openmp/runtime/src/kmp_tasking.cpp" , 2702); | |||
| 2703 | 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); }; | |||
| 2704 | KMP_ATOMIC_ST_RLX(&tg_new->count, 0)(&tg_new->count)->store(0, std::memory_order_relaxed ); | |||
| 2705 | KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq)(&tg_new->cancel_request)->store(cancel_noreq, std:: memory_order_relaxed); | |||
| 2706 | tg_new->parent = taskdata->td_taskgroup; | |||
| 2707 | tg_new->reduce_data = NULL__null; | |||
| 2708 | tg_new->reduce_num_data = 0; | |||
| 2709 | tg_new->gomp_data = NULL__null; | |||
| 2710 | taskdata->td_taskgroup = tg_new; | |||
| 2711 | ||||
| 2712 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2713 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
| 2714 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
| 2715 | if (!codeptr) | |||
| 2716 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
| 2717 | kmp_team_t *team = thread->th.th_team; | |||
| 2718 | ompt_data_t my_task_data = taskdata->ompt_task_info.task_data; | |||
| 2719 | // FIXME: I think this is wrong for lwt! | |||
| 2720 | ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
| 2721 | ||||
| 2722 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2723 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
| 2724 | &(my_task_data), codeptr); | |||
| 2725 | } | |||
| 2726 | #endif | |||
| 2727 | } | |||
| 2728 | ||||
| 2729 | // __kmpc_end_taskgroup: Wait until all tasks generated by the current task | |||
| 2730 | // and its descendants are complete | |||
| 2731 | void __kmpc_end_taskgroup(ident_t *loc, int gtid) { | |||
| 2732 | __kmp_assert_valid_gtid(gtid); | |||
| 2733 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 2734 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 2735 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
| 2736 | int thread_finished = FALSE0; | |||
| 2737 | ||||
| 2738 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2739 | kmp_team_t *team; | |||
| 2740 | ompt_data_t my_task_data; | |||
| 2741 | ompt_data_t my_parallel_data; | |||
| 2742 | void *codeptr = nullptr; | |||
| 2743 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 2744 | team = thread->th.th_team; | |||
| 2745 | my_task_data = taskdata->ompt_task_info.task_data; | |||
| 2746 | // FIXME: I think this is wrong for lwt! | |||
| 2747 | my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
| 2748 | codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
| 2749 | if (!codeptr) | |||
| 2750 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
| 2751 | } | |||
| 2752 | #endif | |||
| 2753 | ||||
| 2754 | 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); }; | |||
| 2755 | KMP_DEBUG_ASSERT(taskgroup != NULL)if (!(taskgroup != __null)) { __kmp_debug_assert("taskgroup != __null" , "openmp/runtime/src/kmp_tasking.cpp", 2755); }; | |||
| 2756 | KMP_SET_THREAD_STATE_BLOCK(TASKGROUP)((void)0); | |||
| 2757 | ||||
| 2758 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
| 2759 | // mark task as waiting not on a barrier | |||
| 2760 | taskdata->td_taskwait_counter += 1; | |||
| 2761 | taskdata->td_taskwait_ident = loc; | |||
| 2762 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 2763 | #if USE_ITT_BUILD1 | |||
| 2764 | // For ITT the taskgroup wait is similar to taskwait until we need to | |||
| 2765 | // distinguish them | |||
| 2766 | void *itt_sync_obj = NULL__null; | |||
| 2767 | #if USE_ITT_NOTIFY1 | |||
| 2768 | KMP_ITT_TASKWAIT_STARTING(itt_sync_obj)if (__builtin_expect(!!(__kmp_itt_sync_create_ptr__3_0), 0)) { itt_sync_obj = __kmp_itt_taskwait_object(gtid); if (itt_sync_obj != __null) { __kmp_itt_taskwait_starting(gtid, itt_sync_obj) ; } }; | |||
| 2769 | #endif /* USE_ITT_NOTIFY */ | |||
| 2770 | #endif /* USE_ITT_BUILD */ | |||
| 2771 | ||||
| 2772 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2773 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
| 2774 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2775 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
| 2776 | &(my_task_data), codeptr); | |||
| 2777 | } | |||
| 2778 | #endif | |||
| 2779 | ||||
| 2780 | if (!taskdata->td_flags.team_serial || | |||
| 2781 | (thread->th.th_task_team != NULL__null && | |||
| 2782 | (thread->th.th_task_team->tt.tt_found_proxy_tasks || | |||
| 2783 | thread->th.th_task_team->tt.tt_hidden_helper_task_encountered))) { | |||
| 2784 | kmp_flag_32<false, false> flag( | |||
| 2785 | RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskgroup->count)), 0U); | |||
| 2786 | while (KMP_ATOMIC_LD_ACQ(&taskgroup->count)(&taskgroup->count)->load(std::memory_order_acquire ) != 0) { | |||
| 2787 | flag.execute_tasks(thread, gtid, FALSE0, | |||
| 2788 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 2789 | __kmp_task_stealing_constraint); | |||
| 2790 | } | |||
| 2791 | } | |||
| 2792 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting | |||
| 2793 | ||||
| 2794 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2795 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
| 2796 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2797 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
| 2798 | &(my_task_data), codeptr); | |||
| 2799 | } | |||
| 2800 | #endif | |||
| 2801 | ||||
| 2802 | #if USE_ITT_BUILD1 | |||
| 2803 | KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj)if (__builtin_expect(!!(itt_sync_obj != __null), 0)) __kmp_itt_taskwait_finished (gtid, itt_sync_obj);; | |||
| 2804 | KMP_FSYNC_ACQUIRED(taskdata)(!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)(taskdata)); // acquire self - sync with descendants | |||
| 2805 | #endif /* USE_ITT_BUILD */ | |||
| 2806 | } | |||
| 2807 | KMP_DEBUG_ASSERT(taskgroup->count == 0)if (!(taskgroup->count == 0)) { __kmp_debug_assert("taskgroup->count == 0" , "openmp/runtime/src/kmp_tasking.cpp", 2807); }; | |||
| 2808 | ||||
| 2809 | if (taskgroup->reduce_data != NULL__null && | |||
| 2810 | !taskgroup->gomp_data) { // need to reduce? | |||
| 2811 | int cnt; | |||
| 2812 | void *reduce_data; | |||
| 2813 | kmp_team_t *t = thread->th.th_team; | |||
| 2814 | kmp_taskred_data_t *arr = (kmp_taskred_data_t *)taskgroup->reduce_data; | |||
| 2815 | // check if <priv> data of the first reduction variable shared for the team | |||
| 2816 | void *priv0 = arr[0].reduce_priv; | |||
| 2817 | if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[0])(&t->t.t_tg_reduce_data[0])->load(std::memory_order_acquire )) != NULL__null && | |||
| 2818 | ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) { | |||
| 2819 | // finishing task reduction on parallel | |||
| 2820 | cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[0])(&t->t.t_tg_fini_counter[0])->fetch_add(1, std::memory_order_acq_rel ); | |||
| 2821 | if (cnt == thread->th.th_team_nproc - 1) { | |||
| 2822 | // we are the last thread passing __kmpc_reduction_modifier_fini() | |||
| 2823 | // finalize task reduction: | |||
| 2824 | __kmp_task_reduction_fini(thread, taskgroup); | |||
| 2825 | // cleanup fields in the team structure: | |||
| 2826 | // TODO: is relaxed store enough here (whole barrier should follow)? | |||
| 2827 | __kmp_thread_free(thread, reduce_data)___kmp_thread_free((thread), (reduce_data), "openmp/runtime/src/kmp_tasking.cpp" , 2827); | |||
| 2828 | KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[0], NULL)(&t->t.t_tg_reduce_data[0])->store(__null, std::memory_order_release ); | |||
| 2829 | KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[0], 0)(&t->t.t_tg_fini_counter[0])->store(0, std::memory_order_release ); | |||
| 2830 | } else { | |||
| 2831 | // we are not the last thread passing __kmpc_reduction_modifier_fini(), | |||
| 2832 | // so do not finalize reduction, just clean own copy of the data | |||
| 2833 | __kmp_task_reduction_clean(thread, taskgroup); | |||
| 2834 | } | |||
| 2835 | } else if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[1])(&t->t.t_tg_reduce_data[1])->load(std::memory_order_acquire )) != | |||
| 2836 | NULL__null && | |||
| 2837 | ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) { | |||
| 2838 | // finishing task reduction on worksharing | |||
| 2839 | cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[1])(&t->t.t_tg_fini_counter[1])->fetch_add(1, std::memory_order_acq_rel ); | |||
| 2840 | if (cnt == thread->th.th_team_nproc - 1) { | |||
| 2841 | // we are the last thread passing __kmpc_reduction_modifier_fini() | |||
| 2842 | __kmp_task_reduction_fini(thread, taskgroup); | |||
| 2843 | // cleanup fields in team structure: | |||
| 2844 | // TODO: is relaxed store enough here (whole barrier should follow)? | |||
| 2845 | __kmp_thread_free(thread, reduce_data)___kmp_thread_free((thread), (reduce_data), "openmp/runtime/src/kmp_tasking.cpp" , 2845); | |||
| 2846 | KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[1], NULL)(&t->t.t_tg_reduce_data[1])->store(__null, std::memory_order_release ); | |||
| 2847 | KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[1], 0)(&t->t.t_tg_fini_counter[1])->store(0, std::memory_order_release ); | |||
| 2848 | } else { | |||
| 2849 | // we are not the last thread passing __kmpc_reduction_modifier_fini(), | |||
| 2850 | // so do not finalize reduction, just clean own copy of the data | |||
| 2851 | __kmp_task_reduction_clean(thread, taskgroup); | |||
| 2852 | } | |||
| 2853 | } else { | |||
| 2854 | // finishing task reduction on taskgroup | |||
| 2855 | __kmp_task_reduction_fini(thread, taskgroup); | |||
| 2856 | } | |||
| 2857 | } | |||
| 2858 | // Restore parent taskgroup for the current task | |||
| 2859 | taskdata->td_taskgroup = taskgroup->parent; | |||
| 2860 | __kmp_thread_free(thread, taskgroup)___kmp_thread_free((thread), (taskgroup), "openmp/runtime/src/kmp_tasking.cpp" , 2860); | |||
| 2861 | ||||
| 2862 | 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); } | |||
| 2863 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n" , gtid, taskdata); }; | |||
| 2864 | ||||
| 2865 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2866 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
| 2867 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2868 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
| 2869 | &(my_task_data), codeptr); | |||
| 2870 | } | |||
| 2871 | #endif | |||
| 2872 | } | |||
| 2873 | ||||
| 2874 | static kmp_task_t *__kmp_get_priority_task(kmp_int32 gtid, | |||
| 2875 | kmp_task_team_t *task_team, | |||
| 2876 | kmp_int32 is_constrained) { | |||
| 2877 | kmp_task_t *task = NULL__null; | |||
| 2878 | kmp_taskdata_t *taskdata; | |||
| 2879 | kmp_taskdata_t *current; | |||
| 2880 | kmp_thread_data_t *thread_data; | |||
| 2881 | int ntasks = task_team->tt.tt_num_task_pri; | |||
| 2882 | if (ntasks == 0) { | |||
| 2883 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #1): T#%d No tasks to get\n" , gtid); } | |||
| 2884 | 20, ("__kmp_get_priority_task(exit #1): T#%d No tasks to get\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #1): T#%d No tasks to get\n" , gtid); }; | |||
| 2885 | return NULL__null; | |||
| 2886 | } | |||
| 2887 | do { | |||
| 2888 | // decrement num_tasks to "reserve" one task to get for execution | |||
| 2889 | if (__kmp_atomic_compare_store(&task_team->tt.tt_num_task_pri, ntasks, | |||
| 2890 | ntasks - 1)) | |||
| 2891 | break; | |||
| 2892 | } while (ntasks > 0); | |||
| 2893 | if (ntasks == 0) { | |||
| 2894 | KA_TRACE(20, ("__kmp_get_priority_task(exit #2): T#%d No tasks to get\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #2): T#%d No tasks to get\n" , __kmp_get_global_thread_id()); } | |||
| 2895 | __kmp_get_gtid()))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #2): T#%d No tasks to get\n" , __kmp_get_global_thread_id()); }; | |||
| 2896 | return NULL__null; | |||
| 2897 | } | |||
| 2898 | // We got a "ticket" to get a "reserved" priority task | |||
| 2899 | int deque_ntasks; | |||
| 2900 | kmp_task_pri_t *list = task_team->tt.tt_task_pri_list; | |||
| 2901 | do { | |||
| 2902 | KMP_ASSERT(list != NULL)if (!(list != __null)) { __kmp_debug_assert("list != NULL", "openmp/runtime/src/kmp_tasking.cpp" , 2902); }; | |||
| 2903 | thread_data = &list->td; | |||
| 2904 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2905 | deque_ntasks = thread_data->td.td_deque_ntasks; | |||
| 2906 | if (deque_ntasks == 0) { | |||
| 2907 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2908 | KA_TRACE(20, ("__kmp_get_priority_task: T#%d No tasks to get from %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task: T#%d No tasks to get from %p\n" , __kmp_get_global_thread_id(), thread_data); } | |||
| 2909 | __kmp_get_gtid(), thread_data))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task: T#%d No tasks to get from %p\n" , __kmp_get_global_thread_id(), thread_data); }; | |||
| 2910 | list = list->next; | |||
| 2911 | } | |||
| 2912 | } while (deque_ntasks == 0); | |||
| 2913 | KMP_DEBUG_ASSERT(deque_ntasks)if (!(deque_ntasks)) { __kmp_debug_assert("deque_ntasks", "openmp/runtime/src/kmp_tasking.cpp" , 2913); }; | |||
| 2914 | int target = thread_data->td.td_deque_head; | |||
| 2915 | current = __kmp_threads[gtid]->th.th_current_task; | |||
| 2916 | taskdata = thread_data->td.td_deque[target]; | |||
| 2917 | if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) { | |||
| 2918 | // Bump head pointer and Wrap. | |||
| 2919 | thread_data->td.td_deque_head = | |||
| 2920 | (target + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 2921 | } else { | |||
| 2922 | if (!task_team->tt.tt_untied_task_encountered) { | |||
| 2923 | // The TSC does not allow to steal victim task | |||
| 2924 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2925 | KA_TRACE(20, ("__kmp_get_priority_task(exit #3): T#%d could not get task "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #3): T#%d could not get task " "from %p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, target, thread_data->td.td_deque_tail ); } | |||
| 2926 | "from %p: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #3): T#%d could not get task " "from %p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, target, thread_data->td.td_deque_tail ); } | |||
| 2927 | gtid, thread_data, task_team, deque_ntasks, target,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #3): T#%d could not get task " "from %p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, target, thread_data->td.td_deque_tail ); } | |||
| 2928 | thread_data->td.td_deque_tail))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_get_priority_task(exit #3): T#%d could not get task " "from %p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, target, thread_data->td.td_deque_tail ); }; | |||
| 2929 | task_team->tt.tt_num_task_pri++; // atomic inc, restore value | |||
| 2930 | return NULL__null; | |||
| 2931 | } | |||
| 2932 | int i; | |||
| 2933 | // walk through the deque trying to steal any task | |||
| 2934 | taskdata = NULL__null; | |||
| 2935 | for (i = 1; i < deque_ntasks; ++i) { | |||
| 2936 | target = (target + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 2937 | taskdata = thread_data->td.td_deque[target]; | |||
| 2938 | if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) { | |||
| 2939 | break; // found task to execute | |||
| 2940 | } else { | |||
| 2941 | taskdata = NULL__null; | |||
| 2942 | } | |||
| 2943 | } | |||
| 2944 | if (taskdata == NULL__null) { | |||
| 2945 | // No appropriate candidate found to execute | |||
| 2946 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2947 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_priority_task(exit #4): T#%d could not get task from " "%p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2948 | 10, ("__kmp_get_priority_task(exit #4): T#%d could not get task from "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_priority_task(exit #4): T#%d could not get task from " "%p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2949 | "%p: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_priority_task(exit #4): T#%d could not get task from " "%p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2950 | gtid, thread_data, task_team, deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_priority_task(exit #4): T#%d could not get task from " "%p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2951 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_get_priority_task(exit #4): T#%d could not get task from " "%p: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, thread_data , task_team, deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); }; | |||
| 2952 | task_team->tt.tt_num_task_pri++; // atomic inc, restore value | |||
| 2953 | return NULL__null; | |||
| 2954 | } | |||
| 2955 | int prev = target; | |||
| 2956 | for (i = i + 1; i < deque_ntasks; ++i) { | |||
| 2957 | // shift remaining tasks in the deque left by 1 | |||
| 2958 | target = (target + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 2959 | thread_data->td.td_deque[prev] = thread_data->td.td_deque[target]; | |||
| 2960 | prev = target; | |||
| 2961 | } | |||
| 2962 | KMP_DEBUG_ASSERT(if (!(thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1)))) { __kmp_debug_assert ("thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 2964); } | |||
| 2963 | thread_data->td.td_deque_tail ==if (!(thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1)))) { __kmp_debug_assert ("thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 2964); } | |||
| 2964 | (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(thread_data->td)))if (!(thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1)))) { __kmp_debug_assert ("thread_data->td.td_deque_tail == (kmp_uint32)((target + 1) & ((thread_data->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 2964); }; | |||
| 2965 | thread_data->td.td_deque_tail = target; // tail -= 1 (wrapped)) | |||
| 2966 | } | |||
| 2967 | thread_data->td.td_deque_ntasks = deque_ntasks - 1; | |||
| 2968 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2969 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 2970 | return task; | |||
| 2971 | } | |||
| 2972 | ||||
| 2973 | // __kmp_remove_my_task: remove a task from my own deque | |||
| 2974 | static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid, | |||
| 2975 | kmp_task_team_t *task_team, | |||
| 2976 | kmp_int32 is_constrained) { | |||
| 2977 | kmp_task_t *task; | |||
| 2978 | kmp_taskdata_t *taskdata; | |||
| 2979 | kmp_thread_data_t *thread_data; | |||
| 2980 | kmp_uint32 tail; | |||
| 2981 | ||||
| 2982 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 2982); }; | |||
| 2983 | KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=if (!(task_team->tt.tt_threads_data != __null)) { __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "openmp/runtime/src/kmp_tasking.cpp" , 2984); } | |||
| 2984 | NULL)if (!(task_team->tt.tt_threads_data != __null)) { __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "openmp/runtime/src/kmp_tasking.cpp" , 2984); }; // Caller should check this condition | |||
| 2985 | ||||
| 2986 | thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)]; | |||
| 2987 | ||||
| 2988 | 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); } | |||
| 2989 | 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); } | |||
| 2990 | 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); }; | |||
| 2991 | ||||
| 2992 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
| 2993 | 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 ); } | |||
| 2994 | ("__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 ); } | |||
| 2995 | "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 ); } | |||
| 2996 | 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 ); } | |||
| 2997 | 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 ); }; | |||
| 2998 | return NULL__null; | |||
| 2999 | } | |||
| 3000 | ||||
| 3001 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3002 | ||||
| 3003 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
| 3004 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3005 | 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 ); } | |||
| 3006 | ("__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 ); } | |||
| 3007 | "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 ); } | |||
| 3008 | 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 ); } | |||
| 3009 | 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 ); }; | |||
| 3010 | return NULL__null; | |||
| 3011 | } | |||
| 3012 | ||||
| 3013 | tail = (thread_data->td.td_deque_tail - 1) & | |||
| 3014 | TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); // Wrap index. | |||
| 3015 | taskdata = thread_data->td.td_deque[tail]; | |||
| 3016 | ||||
| 3017 | if (!__kmp_task_is_allowed(gtid, is_constrained, taskdata, | |||
| 3018 | thread->th.th_current_task)) { | |||
| 3019 | // The TSC does not allow to steal victim task | |||
| 3020 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3021 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: " "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 ); } | |||
| 3022 | ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: " "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 ); } | |||
| 3023 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: " "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 ); } | |||
| 3024 | gtid, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: " "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 ); } | |||
| 3025 | 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 #3): T#%d TSC blocks tail task: " "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 ); }; | |||
| 3026 | return NULL__null; | |||
| 3027 | } | |||
| 3028 | ||||
| 3029 | thread_data->td.td_deque_tail = tail; | |||
| 3030 | 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); | |||
| 3031 | ||||
| 3032 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3033 | ||||
| 3034 | KA_TRACE(10, ("__kmp_remove_my_task(exit #4): T#%d task %p removed: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #4): 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); } | |||
| 3035 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #4): 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); } | |||
| 3036 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #4): 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); } | |||
| 3037 | 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 #4): 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); }; | |||
| 3038 | ||||
| 3039 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 3040 | return task; | |||
| 3041 | } | |||
| 3042 | ||||
| 3043 | // __kmp_steal_task: remove a task from another thread's deque | |||
| 3044 | // Assume that calling thread has already checked existence of | |||
| 3045 | // task_team thread_data before calling this routine. | |||
| 3046 | static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid, | |||
| 3047 | kmp_task_team_t *task_team, | |||
| 3048 | std::atomic<kmp_int32> *unfinished_threads, | |||
| 3049 | int *thread_finished, | |||
| 3050 | kmp_int32 is_constrained) { | |||
| 3051 | kmp_task_t *task; | |||
| 3052 | kmp_taskdata_t *taskdata; | |||
| 3053 | kmp_taskdata_t *current; | |||
| 3054 | kmp_thread_data_t *victim_td, *threads_data; | |||
| 3055 | kmp_int32 target; | |||
| 3056 | kmp_int32 victim_tid; | |||
| 3057 | ||||
| 3058 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 3058); }; | |||
| 3059 | ||||
| 3060 | threads_data = task_team->tt.tt_threads_data; | |||
| 3061 | KMP_DEBUG_ASSERT(threads_data != NULL)if (!(threads_data != __null)) { __kmp_debug_assert("threads_data != __null" , "openmp/runtime/src/kmp_tasking.cpp", 3061); }; // Caller should check this condition | |||
| 3062 | ||||
| 3063 | victim_tid = victim_thr->th.th_info.ds.ds_tid; | |||
| 3064 | victim_td = &threads_data[victim_tid]; | |||
| 3065 | ||||
| 3066 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3067 | "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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3068 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3069 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3070 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); }; | |||
| 3071 | ||||
| 3072 | if (TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks) == 0) { | |||
| 3073 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3074 | "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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3075 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3076 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); } | |||
| 3077 | 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, __kmp_gtid_from_thread (victim_thr), task_team, victim_td->td.td_deque_ntasks, victim_td ->td.td_deque_head, victim_td->td.td_deque_tail); }; | |||
| 3078 | return NULL__null; | |||
| 3079 | } | |||
| 3080 | ||||
| 3081 | __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 3082 | ||||
| 3083 | int ntasks = TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks); | |||
| 3084 | // Check again after we acquire the lock | |||
| 3085 | if (ntasks == 0) { | |||
| 3086 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 3087 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3088 | "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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3089 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3090 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 3091 | return NULL__null; | |||
| 3092 | } | |||
| 3093 | ||||
| 3094 | KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL)if (!(victim_td->td.td_deque != __null)) { __kmp_debug_assert ("victim_td->td.td_deque != __null", "openmp/runtime/src/kmp_tasking.cpp" , 3094); }; | |||
| 3095 | current = __kmp_threads[gtid]->th.th_current_task; | |||
| 3096 | taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head]; | |||
| 3097 | if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) { | |||
| 3098 | // Bump head pointer and Wrap. | |||
| 3099 | victim_td->td.td_deque_head = | |||
| 3100 | (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 3101 | } else { | |||
| 3102 | if (!task_team->tt.tt_untied_task_encountered) { | |||
| 3103 | // The TSC does not allow to steal victim task | |||
| 3104 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 3105 | KA_TRACE(10, ("__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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3106 | "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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3107 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3108 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 3109 | return NULL__null; | |||
| 3110 | } | |||
| 3111 | int i; | |||
| 3112 | // walk through victim's deque trying to steal any task | |||
| 3113 | target = victim_td->td.td_deque_head; | |||
| 3114 | taskdata = NULL__null; | |||
| 3115 | for (i = 1; i < ntasks; ++i) { | |||
| 3116 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 3117 | taskdata = victim_td->td.td_deque[target]; | |||
| 3118 | if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) { | |||
| 3119 | break; // found victim task | |||
| 3120 | } else { | |||
| 3121 | taskdata = NULL__null; | |||
| 3122 | } | |||
| 3123 | } | |||
| 3124 | if (taskdata == NULL__null) { | |||
| 3125 | // No appropriate candidate to steal found | |||
| 3126 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 3127 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3128 | "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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3129 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3130 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 3131 | return NULL__null; | |||
| 3132 | } | |||
| 3133 | int prev = target; | |||
| 3134 | for (i = i + 1; i < ntasks; ++i) { | |||
| 3135 | // shift remaining tasks in the deque left by 1 | |||
| 3136 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 3137 | victim_td->td.td_deque[prev] = victim_td->td.td_deque[target]; | |||
| 3138 | prev = target; | |||
| 3139 | } | |||
| 3140 | KMP_DEBUG_ASSERT(if (!(victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1)))) { __kmp_debug_assert ("victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 3142); } | |||
| 3141 | victim_td->td.td_deque_tail ==if (!(victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1)))) { __kmp_debug_assert ("victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 3142); } | |||
| 3142 | (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(victim_td->td)))if (!(victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1)))) { __kmp_debug_assert ("victim_td->td.td_deque_tail == (kmp_uint32)((target + 1) & ((victim_td->td).td_deque_size - 1))" , "openmp/runtime/src/kmp_tasking.cpp", 3142); }; | |||
| 3143 | victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped)) | |||
| 3144 | } | |||
| 3145 | if (*thread_finished) { | |||
| 3146 | // We need to un-mark this victim as a finished victim. This must be done | |||
| 3147 | // before releasing the lock, or else other threads (starting with the | |||
| 3148 | // primary thread victim) might be prematurely released from the barrier!!! | |||
| 3149 | #if KMP_DEBUG1 | |||
| 3150 | kmp_int32 count = | |||
| 3151 | #endif | |||
| 3152 | KMP_ATOMIC_INC(unfinished_threads)(unfinished_threads)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 3153 | 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); } | |||
| 3154 | 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); } | |||
| 3155 | ("__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); } | |||
| 3156 | 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); }; | |||
| 3157 | *thread_finished = FALSE0; | |||
| 3158 | } | |||
| 3159 | TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1)(victim_td->td.td_deque_ntasks) = (ntasks - 1); | |||
| 3160 | ||||
| 3161 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 3162 | ||||
| 3163 | KMP_COUNT_BLOCK(TASK_stolen)((void)0); | |||
| 3164 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3165 | ("__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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3166 | "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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3167 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 3168 | 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, __kmp_gtid_from_thread (victim_thr), task_team, ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 3169 | ||||
| 3170 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 3171 | return task; | |||
| 3172 | } | |||
| 3173 | ||||
| 3174 | // __kmp_execute_tasks_template: Choose and execute tasks until either the | |||
| 3175 | // condition is statisfied (return true) or there are none left (return false). | |||
| 3176 | // | |||
| 3177 | // final_spin is TRUE if this is the spin at the release barrier. | |||
| 3178 | // thread_finished indicates whether the thread is finished executing all | |||
| 3179 | // the tasks it has on its deque, and is at the release barrier. | |||
| 3180 | // spinner is the location on which to spin. | |||
| 3181 | // spinner == NULL means only execute a single task and return. | |||
| 3182 | // checker is the value to check to terminate the spin. | |||
| 3183 | template <class C> | |||
| 3184 | static inline int __kmp_execute_tasks_template( | |||
| 3185 | kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin, | |||
| 3186 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 3187 | kmp_int32 is_constrained) { | |||
| 3188 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 3189 | kmp_thread_data_t *threads_data; | |||
| 3190 | kmp_task_t *task; | |||
| 3191 | kmp_info_t *other_thread; | |||
| 3192 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 3193 | std::atomic<kmp_int32> *unfinished_threads; | |||
| 3194 | kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0, | |||
| 3195 | tid = thread->th.th_info.ds.ds_tid; | |||
| 3196 | ||||
| 3197 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 3197); }; | |||
| 3198 | KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid])if (!(thread == __kmp_threads[gtid])) { __kmp_debug_assert("thread == __kmp_threads[gtid]" , "openmp/runtime/src/kmp_tasking.cpp", 3198); }; | |||
| 3199 | ||||
| 3200 | if (task_team == NULL__null || current_task == NULL__null) | |||
| 3201 | return FALSE0; | |||
| 3202 | ||||
| 3203 | 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) ; } | |||
| 3204 | "*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) ; } | |||
| 3205 | 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) ; }; | |||
| 3206 | ||||
| 3207 | thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; | |||
| 3208 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
| 3209 | ||||
| 3210 | KMP_DEBUG_ASSERT(threads_data != NULL)if (!(threads_data != __null)) { __kmp_debug_assert("threads_data != __null" , "openmp/runtime/src/kmp_tasking.cpp", 3210); }; | |||
| 3211 | ||||
| 3212 | nthreads = task_team->tt.tt_nproc; | |||
| 3213 | unfinished_threads = &(task_team->tt.tt_unfinished_threads); | |||
| 3214 | KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks ||if (!(nthreads > 1 || task_team->tt.tt_found_proxy_tasks || task_team->tt.tt_hidden_helper_task_encountered)) { __kmp_debug_assert ("nthreads > 1 || task_team->tt.tt_found_proxy_tasks || task_team->tt.tt_hidden_helper_task_encountered" , "openmp/runtime/src/kmp_tasking.cpp", 3215); } | |||
| 3215 | task_team->tt.tt_hidden_helper_task_encountered)if (!(nthreads > 1 || task_team->tt.tt_found_proxy_tasks || task_team->tt.tt_hidden_helper_task_encountered)) { __kmp_debug_assert ("nthreads > 1 || task_team->tt.tt_found_proxy_tasks || task_team->tt.tt_hidden_helper_task_encountered" , "openmp/runtime/src/kmp_tasking.cpp", 3215); }; | |||
| 3216 | KMP_DEBUG_ASSERT(*unfinished_threads >= 0)if (!(*unfinished_threads >= 0)) { __kmp_debug_assert("*unfinished_threads >= 0" , "openmp/runtime/src/kmp_tasking.cpp", 3216); }; | |||
| 3217 | ||||
| 3218 | while (1) { // Outer loop keeps trying to find tasks in case of single thread | |||
| 3219 | // getting tasks from target constructs | |||
| 3220 | while (1) { // Inner loop to find a task and execute it | |||
| 3221 | task = NULL__null; | |||
| 3222 | if (task_team->tt.tt_num_task_pri) { // get priority task first | |||
| 3223 | task = __kmp_get_priority_task(gtid, task_team, is_constrained); | |||
| 3224 | } | |||
| 3225 | if (task == NULL__null && use_own_tasks) { // check own queue next | |||
| 3226 | task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained); | |||
| 3227 | } | |||
| 3228 | if ((task == NULL__null) && (nthreads > 1)) { // Steal a task finally | |||
| 3229 | int asleep = 1; | |||
| 3230 | use_own_tasks = 0; | |||
| 3231 | // Try to steal from the last place I stole from successfully. | |||
| 3232 | if (victim_tid == -2) { // haven't stolen anything yet | |||
| 3233 | victim_tid = threads_data[tid].td.td_deque_last_stolen; | |||
| 3234 | if (victim_tid != | |||
| 3235 | -1) // if we have a last stolen from victim, get the thread | |||
| 3236 | other_thread = threads_data[victim_tid].td.td_thr; | |||
| 3237 | } | |||
| 3238 | if (victim_tid != -1) { // found last victim | |||
| 3239 | asleep = 0; | |||
| 3240 | } else if (!new_victim) { // no recent steals and we haven't already | |||
| 3241 | // used a new victim; select a random thread | |||
| 3242 | do { // Find a different thread to steal work from. | |||
| 3243 | // Pick a random thread. Initial plan was to cycle through all the | |||
| 3244 | // threads, and only return if we tried to steal from every thread, | |||
| 3245 | // and failed. Arch says that's not such a great idea. | |||
| 3246 | victim_tid = __kmp_get_random(thread) % (nthreads - 1); | |||
| 3247 | if (victim_tid >= tid) { | |||
| 3248 | ++victim_tid; // Adjusts random distribution to exclude self | |||
| 3249 | } | |||
| 3250 | // Found a potential victim | |||
| 3251 | other_thread = threads_data[victim_tid].td.td_thr; | |||
| 3252 | // There is a slight chance that __kmp_enable_tasking() did not wake | |||
| 3253 | // up all threads waiting at the barrier. If victim is sleeping, | |||
| 3254 | // then wake it up. Since we were going to pay the cache miss | |||
| 3255 | // penalty for referencing another thread's kmp_info_t struct | |||
| 3256 | // anyway, | |||
| 3257 | // the check shouldn't cost too much performance at this point. In | |||
| 3258 | // extra barrier mode, tasks do not sleep at the separate tasking | |||
| 3259 | // barrier, so this isn't a problem. | |||
| 3260 | asleep = 0; | |||
| 3261 | if ((__kmp_tasking_mode == tskm_task_teams) && | |||
| 3262 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) && | |||
| 3263 | (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc))((void *)(const_cast<void *>(other_thread->th.th_sleep_loc ))) != | |||
| 3264 | NULL__null)) { | |||
| 3265 | asleep = 1; | |||
| 3266 | __kmp_null_resume_wrapper(other_thread); | |||
| 3267 | // A sleeping thread should not have any tasks on it's queue. | |||
| 3268 | // There is a slight possibility that it resumes, steals a task | |||
| 3269 | // from another thread, which spawns more tasks, all in the time | |||
| 3270 | // that it takes this thread to check => don't write an assertion | |||
| 3271 | // that the victim's queue is empty. Try stealing from a | |||
| 3272 | // different thread. | |||
| 3273 | } | |||
| 3274 | } while (asleep); | |||
| 3275 | } | |||
| 3276 | ||||
| 3277 | if (!asleep) { | |||
| 3278 | // We have a victim to try to steal from | |||
| 3279 | task = __kmp_steal_task(other_thread, gtid, task_team, | |||
| 3280 | unfinished_threads, thread_finished, | |||
| 3281 | is_constrained); | |||
| 3282 | } | |||
| 3283 | if (task != NULL__null) { // set last stolen to victim | |||
| 3284 | if (threads_data[tid].td.td_deque_last_stolen != victim_tid) { | |||
| 3285 | threads_data[tid].td.td_deque_last_stolen = victim_tid; | |||
| 3286 | // The pre-refactored code did not try more than 1 successful new | |||
| 3287 | // vicitm, unless the last one generated more local tasks; | |||
| 3288 | // new_victim keeps track of this | |||
| 3289 | new_victim = 1; | |||
| 3290 | } | |||
| 3291 | } else { // No tasks found; unset last_stolen | |||
| 3292 | 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); | |||
| 3293 | victim_tid = -2; // no successful victim found | |||
| 3294 | } | |||
| 3295 | } | |||
| 3296 | ||||
| 3297 | if (task == NULL__null) | |||
| 3298 | break; // break out of tasking loop | |||
| 3299 | ||||
| 3300 | // Found a task; execute it | |||
| 3301 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 3302 | if (__itt_sync_create_ptr__kmp_itt_sync_create_ptr__3_0 || KMP_ITT_DEBUG0) { | |||
| 3303 | if (itt_sync_obj == NULL__null) { // we are at fork barrier where we could not | |||
| 3304 | // get the object reliably | |||
| 3305 | itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier); | |||
| 3306 | } | |||
| 3307 | __kmp_itt_task_starting(itt_sync_obj); | |||
| 3308 | } | |||
| 3309 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ | |||
| 3310 | __kmp_invoke_task(gtid, task, current_task); | |||
| 3311 | #if USE_ITT_BUILD1 | |||
| 3312 | if (itt_sync_obj != NULL__null) | |||
| 3313 | __kmp_itt_task_finished(itt_sync_obj); | |||
| 3314 | #endif /* USE_ITT_BUILD */ | |||
| 3315 | // If this thread is only partway through the barrier and the condition is | |||
| 3316 | // met, then return now, so that the barrier gather/release pattern can | |||
| 3317 | // proceed. If this thread is in the last spin loop in the barrier, | |||
| 3318 | // waiting to be released, we know that the termination condition will not | |||
| 3319 | // be satisfied, so don't waste any cycles checking it. | |||
| 3320 | if (flag == NULL__null || (!final_spin && flag->done_check())) { | |||
| 3321 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 3322 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 3323 | ("__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); } | |||
| 3324 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
| 3325 | return TRUE(!0); | |||
| 3326 | } | |||
| 3327 | if (thread->th.th_task_team == NULL__null) { | |||
| 3328 | break; | |||
| 3329 | } | |||
| 3330 | KMP_YIELD(__kmp_library == library_throughput){ __kmp_x86_pause(); if ((__kmp_library == library_throughput ) && (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))))))) __kmp_yield(); }; // Yield before next task | |||
| 3331 | // If execution of a stolen task results in more tasks being placed on our | |||
| 3332 | // run queue, reset use_own_tasks | |||
| 3333 | if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks)(threads_data[tid].td.td_deque_ntasks) != 0) { | |||
| 3334 | 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); } | |||
| 3335 | "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); } | |||
| 3336 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d stolen task spawned " "other tasks, restart\n", gtid); }; | |||
| 3337 | use_own_tasks = 1; | |||
| 3338 | new_victim = 0; | |||
| 3339 | } | |||
| 3340 | } | |||
| 3341 | ||||
| 3342 | // The task source has been exhausted. If in final spin loop of barrier, | |||
| 3343 | // check if termination condition is satisfied. The work queue may be empty | |||
| 3344 | // but there might be proxy tasks still executing. | |||
| 3345 | if (final_spin && | |||
| 3346 | KMP_ATOMIC_LD_ACQ(¤t_task->td_incomplete_child_tasks)(¤t_task->td_incomplete_child_tasks)->load(std ::memory_order_acquire) == 0) { | |||
| 3347 | // First, decrement the #unfinished threads, if that has not already been | |||
| 3348 | // done. This decrement might be to the spin location, and result in the | |||
| 3349 | // termination condition being satisfied. | |||
| 3350 | if (!*thread_finished) { | |||
| 3351 | #if KMP_DEBUG1 | |||
| 3352 | kmp_int32 count = -1 + | |||
| 3353 | #endif | |||
| 3354 | KMP_ATOMIC_DEC(unfinished_threads)(unfinished_threads)->fetch_sub(1, std::memory_order_acq_rel ); | |||
| 3355 | 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 ); } | |||
| 3356 | "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 ); } | |||
| 3357 | 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 ); }; | |||
| 3358 | *thread_finished = TRUE(!0); | |||
| 3359 | } | |||
| 3360 | ||||
| 3361 | // It is now unsafe to reference thread->th.th_team !!! | |||
| 3362 | // Decrementing task_team->tt.tt_unfinished_threads can allow the primary | |||
| 3363 | // thread to pass through the barrier, where it might reset each thread's | |||
| 3364 | // th.th_team field for the next parallel region. If we can steal more | |||
| 3365 | // work, we know that this has not happened yet. | |||
| 3366 | if (flag != NULL__null && flag->done_check()) { | |||
| 3367 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 3368 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 3369 | ("__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); } | |||
| 3370 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
| 3371 | return TRUE(!0); | |||
| 3372 | } | |||
| 3373 | } | |||
| 3374 | ||||
| 3375 | // If this thread's task team is NULL, primary thread has recognized that | |||
| 3376 | // there are no more tasks; bail out | |||
| 3377 | if (thread->th.th_task_team == NULL__null) { | |||
| 3378 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d no more tasks\n" , gtid); } | |||
| 3379 | ("__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); }; | |||
| 3380 | return FALSE0; | |||
| 3381 | } | |||
| 3382 | ||||
| 3383 | // Check the flag again to see if it has already done in case to be trapped | |||
| 3384 | // into infinite loop when a if0 task depends on a hidden helper task | |||
| 3385 | // outside any parallel region. Detached tasks are not impacted in this case | |||
| 3386 | // because the only thread executing this function has to execute the proxy | |||
| 3387 | // task so it is in another code path that has the same check. | |||
| 3388 | if (flag == NULL__null || (!final_spin && flag->done_check())) { | |||
| 3389 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 3390 | ("__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); } | |||
| 3391 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
| 3392 | return TRUE(!0); | |||
| 3393 | } | |||
| 3394 | ||||
| 3395 | // We could be getting tasks from target constructs; if this is the only | |||
| 3396 | // thread, keep trying to execute tasks from own queue | |||
| 3397 | if (nthreads == 1 && | |||
| 3398 | KMP_ATOMIC_LD_ACQ(¤t_task->td_incomplete_child_tasks)(¤t_task->td_incomplete_child_tasks)->load(std ::memory_order_acquire)) | |||
| 3399 | use_own_tasks = 1; | |||
| 3400 | else { | |||
| 3401 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d can't find work\n" , gtid); } | |||
| 3402 | ("__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); }; | |||
| 3403 | return FALSE0; | |||
| 3404 | } | |||
| 3405 | } | |||
| 3406 | } | |||
| 3407 | ||||
| 3408 | template <bool C, bool S> | |||
| 3409 | int __kmp_execute_tasks_32( | |||
| 3410 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32<C, S> *flag, int final_spin, | |||
| 3411 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 3412 | kmp_int32 is_constrained) { | |||
| 3413 | return __kmp_execute_tasks_template( | |||
| 3414 | thread, gtid, flag, final_spin, | |||
| 3415 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 3416 | } | |||
| 3417 | ||||
| 3418 | template <bool C, bool S> | |||
| 3419 | int __kmp_execute_tasks_64( | |||
| 3420 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64<C, S> *flag, int final_spin, | |||
| 3421 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 3422 | kmp_int32 is_constrained) { | |||
| 3423 | return __kmp_execute_tasks_template( | |||
| 3424 | thread, gtid, flag, final_spin, | |||
| 3425 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 3426 | } | |||
| 3427 | ||||
| 3428 | template <bool C, bool S> | |||
| 3429 | int __kmp_atomic_execute_tasks_64( | |||
| 3430 | kmp_info_t *thread, kmp_int32 gtid, kmp_atomic_flag_64<C, S> *flag, | |||
| 3431 | int final_spin, int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 3432 | kmp_int32 is_constrained) { | |||
| 3433 | return __kmp_execute_tasks_template( | |||
| 3434 | thread, gtid, flag, final_spin, | |||
| 3435 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 3436 | } | |||
| 3437 | ||||
| 3438 | int __kmp_execute_tasks_oncore( | |||
| 3439 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin, | |||
| 3440 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 3441 | kmp_int32 is_constrained) { | |||
| 3442 | return __kmp_execute_tasks_template( | |||
| 3443 | thread, gtid, flag, final_spin, | |||
| 3444 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 3445 | } | |||
| 3446 | ||||
| 3447 | template int | |||
| 3448 | __kmp_execute_tasks_32<false, false>(kmp_info_t *, kmp_int32, | |||
| 3449 | kmp_flag_32<false, false> *, int, | |||
| 3450 | int *USE_ITT_BUILD_ARG(void *), void *, kmp_int32); | |||
| 3451 | ||||
| 3452 | template int __kmp_execute_tasks_64<false, true>(kmp_info_t *, kmp_int32, | |||
| 3453 | kmp_flag_64<false, true> *, | |||
| 3454 | int, | |||
| 3455 | int *USE_ITT_BUILD_ARG(void *), void *, | |||
| 3456 | kmp_int32); | |||
| 3457 | ||||
| 3458 | template int __kmp_execute_tasks_64<true, false>(kmp_info_t *, kmp_int32, | |||
| 3459 | kmp_flag_64<true, false> *, | |||
| 3460 | int, | |||
| 3461 | int *USE_ITT_BUILD_ARG(void *), void *, | |||
| 3462 | kmp_int32); | |||
| 3463 | ||||
| 3464 | template int __kmp_atomic_execute_tasks_64<false, true>( | |||
| 3465 | kmp_info_t *, kmp_int32, kmp_atomic_flag_64<false, true> *, int, | |||
| 3466 | int *USE_ITT_BUILD_ARG(void *), void *, kmp_int32); | |||
| 3467 | ||||
| 3468 | template int __kmp_atomic_execute_tasks_64<true, false>( | |||
| 3469 | kmp_info_t *, kmp_int32, kmp_atomic_flag_64<true, false> *, int, | |||
| 3470 | int *USE_ITT_BUILD_ARG(void *), void *, kmp_int32); | |||
| 3471 | ||||
| 3472 | // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the | |||
| 3473 | // next barrier so they can assist in executing enqueued tasks. | |||
| 3474 | // First thread in allocates the task team atomically. | |||
| 3475 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
| 3476 | kmp_info_t *this_thr) { | |||
| 3477 | kmp_thread_data_t *threads_data; | |||
| 3478 | int nthreads, i, is_init_thread; | |||
| 3479 | ||||
| 3480 | 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" , __kmp_gtid_from_thread(this_thr)); } | |||
| 3481 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(enter): T#%d\n" , __kmp_gtid_from_thread(this_thr)); }; | |||
| 3482 | ||||
| 3483 | KMP_DEBUG_ASSERT(task_team != NULL)if (!(task_team != __null)) { __kmp_debug_assert("task_team != __null" , "openmp/runtime/src/kmp_tasking.cpp", 3483); }; | |||
| 3484 | KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL)if (!(this_thr->th.th_team != __null)) { __kmp_debug_assert ("this_thr->th.th_team != __null", "openmp/runtime/src/kmp_tasking.cpp" , 3484); }; | |||
| 3485 | ||||
| 3486 | nthreads = task_team->tt.tt_nproc; | |||
| 3487 | KMP_DEBUG_ASSERT(nthreads > 0)if (!(nthreads > 0)) { __kmp_debug_assert("nthreads > 0" , "openmp/runtime/src/kmp_tasking.cpp", 3487); }; | |||
| 3488 | KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc)if (!(nthreads == this_thr->th.th_team->t.t_nproc)) { __kmp_debug_assert ("nthreads == this_thr->th.th_team->t.t_nproc", "openmp/runtime/src/kmp_tasking.cpp" , 3488); }; | |||
| 3489 | ||||
| 3490 | // Allocate or increase the size of threads_data if necessary | |||
| 3491 | is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team); | |||
| 3492 | ||||
| 3493 | if (!is_init_thread) { | |||
| 3494 | // Some other thread already set up the array. | |||
| 3495 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , __kmp_gtid_from_thread(this_thr)); } | |||
| 3496 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , __kmp_gtid_from_thread(this_thr)); } | |||
| 3497 | ("__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" , __kmp_gtid_from_thread(this_thr)); } | |||
| 3498 | __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" , __kmp_gtid_from_thread(this_thr)); }; | |||
| 3499 | return; | |||
| 3500 | } | |||
| 3501 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
| 3502 | KMP_DEBUG_ASSERT(threads_data != NULL)if (!(threads_data != __null)) { __kmp_debug_assert("threads_data != __null" , "openmp/runtime/src/kmp_tasking.cpp", 3502); }; | |||
| 3503 | ||||
| 3504 | if (__kmp_tasking_mode == tskm_task_teams && | |||
| 3505 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647))) { | |||
| 3506 | // Release any threads sleeping at the barrier, so that they can steal | |||
| 3507 | // tasks and execute them. In extra barrier mode, tasks do not sleep | |||
| 3508 | // at the separate tasking barrier, so this isn't a problem. | |||
| 3509 | for (i = 0; i < nthreads; i++) { | |||
| 3510 | void *sleep_loc; | |||
| 3511 | kmp_info_t *thread = threads_data[i].td.td_thr; | |||
| 3512 | ||||
| 3513 | if (i == this_thr->th.th_info.ds.ds_tid) { | |||
| 3514 | continue; | |||
| 3515 | } | |||
| 3516 | // Since we haven't locked the thread's suspend mutex lock at this | |||
| 3517 | // point, there is a small window where a thread might be putting | |||
| 3518 | // itself to sleep, but hasn't set the th_sleep_loc field yet. | |||
| 3519 | // To work around this, __kmp_execute_tasks_template() periodically checks | |||
| 3520 | // see if other threads are sleeping (using the same random mechanism that | |||
| 3521 | // is used for task stealing) and awakens them if they are. | |||
| 3522 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
| 3523 | NULL__null) { | |||
| 3524 | 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" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); } | |||
| 3525 | __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" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); } | |||
| 3526 | __kmp_gtid_from_thread(thread)))if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); }; | |||
| 3527 | __kmp_null_resume_wrapper(thread); | |||
| 3528 | } else { | |||
| 3529 | 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" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); } | |||
| 3530 | __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" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); } | |||
| 3531 | __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" , __kmp_gtid_from_thread(this_thr), __kmp_gtid_from_thread(thread )); }; | |||
| 3532 | } | |||
| 3533 | } | |||
| 3534 | } | |||
| 3535 | ||||
| 3536 | 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" , __kmp_gtid_from_thread(this_thr)); } | |||
| 3537 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d\n" , __kmp_gtid_from_thread(this_thr)); }; | |||
| 3538 | } | |||
| 3539 | ||||
| 3540 | /* // TODO: Check the comment consistency | |||
| 3541 | * Utility routines for "task teams". A task team (kmp_task_t) is kind of | |||
| 3542 | * like a shadow of the kmp_team_t data struct, with a different lifetime. | |||
| 3543 | * After a child * thread checks into a barrier and calls __kmp_release() from | |||
| 3544 | * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no | |||
| 3545 | * longer assume that the kmp_team_t structure is intact (at any moment, the | |||
| 3546 | * primary thread may exit the barrier code and free the team data structure, | |||
| 3547 | * and return the threads to the thread pool). | |||
| 3548 | * | |||
| 3549 | * This does not work with the tasking code, as the thread is still | |||
| 3550 | * expected to participate in the execution of any tasks that may have been | |||
| 3551 | * spawned my a member of the team, and the thread still needs access to all | |||
| 3552 | * to each thread in the team, so that it can steal work from it. | |||
| 3553 | * | |||
| 3554 | * Enter the existence of the kmp_task_team_t struct. It employs a reference | |||
| 3555 | * counting mechanism, and is allocated by the primary thread before calling | |||
| 3556 | * __kmp_<barrier_kind>_release, and then is release by the last thread to | |||
| 3557 | * exit __kmp_<barrier_kind>_release at the next barrier. I.e. the lifetimes | |||
| 3558 | * of the kmp_task_team_t structs for consecutive barriers can overlap | |||
| 3559 | * (and will, unless the primary thread is the last thread to exit the barrier | |||
| 3560 | * release phase, which is not typical). The existence of such a struct is | |||
| 3561 | * useful outside the context of tasking. | |||
| 3562 | * | |||
| 3563 | * We currently use the existence of the threads array as an indicator that | |||
| 3564 | * tasks were spawned since the last barrier. If the structure is to be | |||
| 3565 | * useful outside the context of tasking, then this will have to change, but | |||
| 3566 | * not setting the field minimizes the performance impact of tasking on | |||
| 3567 | * barriers, when no explicit tasks were spawned (pushed, actually). | |||
| 3568 | */ | |||
| 3569 | ||||
| 3570 | static kmp_task_team_t *__kmp_free_task_teams = | |||
| 3571 | NULL__null; // Free list for task_team data structures | |||
| 3572 | // Lock for task team data structures | |||
| 3573 | kmp_bootstrap_lock_t __kmp_task_team_lock = | |||
| 3574 | KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock){ { true, &((__kmp_task_team_lock)), __null, 0U, 0U, 0, - 1 } }; | |||
| 3575 | ||||
| 3576 | // __kmp_alloc_task_deque: | |||
| 3577 | // Allocates a task deque for a particular thread, and initialize the necessary | |||
| 3578 | // data structures relating to the deque. This only happens once per thread | |||
| 3579 | // per task team since task teams are recycled. No lock is needed during | |||
| 3580 | // allocation since each thread allocates its own deque. | |||
| 3581 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
| 3582 | kmp_thread_data_t *thread_data) { | |||
| 3583 | __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3584 | KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL)if (!(thread_data->td.td_deque == __null)) { __kmp_debug_assert ("thread_data->td.td_deque == __null", "openmp/runtime/src/kmp_tasking.cpp" , 3584); }; | |||
| 3585 | ||||
| 3586 | // Initialize last stolen task field to "none" | |||
| 3587 | thread_data->td.td_deque_last_stolen = -1; | |||
| 3588 | ||||
| 3589 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0)if (!((thread_data->td.td_deque_ntasks) == 0)) { __kmp_debug_assert ("(thread_data->td.td_deque_ntasks) == 0", "openmp/runtime/src/kmp_tasking.cpp" , 3589); }; | |||
| 3590 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0)if (!(thread_data->td.td_deque_head == 0)) { __kmp_debug_assert ("thread_data->td.td_deque_head == 0", "openmp/runtime/src/kmp_tasking.cpp" , 3590); }; | |||
| 3591 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0)if (!(thread_data->td.td_deque_tail == 0)) { __kmp_debug_assert ("thread_data->td.td_deque_tail == 0", "openmp/runtime/src/kmp_tasking.cpp" , 3591); }; | |||
| 3592 | ||||
| 3593 | KE_TRACE(if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , __kmp_gtid_from_thread(thread), (1 << 8), thread_data ); } | |||
| 3594 | 10,if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , __kmp_gtid_from_thread(thread), (1 << 8), thread_data ); } | |||
| 3595 | ("__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" , __kmp_gtid_from_thread(thread), (1 << 8), thread_data ); } | |||
| 3596 | __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" , __kmp_gtid_from_thread(thread), (1 << 8), thread_data ); }; | |||
| 3597 | // Allocate space for task deque, and zero the deque | |||
| 3598 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
| 3599 | // kmp_reap_task_team( ). | |||
| 3600 | thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "openmp/runtime/src/kmp_tasking.cpp" , 3601) | |||
| 3601 | INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *))___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "openmp/runtime/src/kmp_tasking.cpp" , 3601); | |||
| 3602 | thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE(1 << 8); | |||
| 3603 | } | |||
| 3604 | ||||
| 3605 | // __kmp_free_task_deque: | |||
| 3606 | // Deallocates a task deque for a particular thread. Happens at library | |||
| 3607 | // deallocation so don't need to reset all thread data fields. | |||
| 3608 | static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) { | |||
| 3609 | if (thread_data->td.td_deque != NULL__null) { | |||
| 3610 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3611 | TCW_4(thread_data->td.td_deque_ntasks, 0)(thread_data->td.td_deque_ntasks) = (0); | |||
| 3612 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "openmp/runtime/src/kmp_tasking.cpp" , 3612); | |||
| 3613 | thread_data->td.td_deque = NULL__null; | |||
| 3614 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3615 | } | |||
| 3616 | ||||
| 3617 | #ifdef BUILD_TIED_TASK_STACK | |||
| 3618 | // GEH: Figure out what to do here for td_susp_tied_tasks | |||
| 3619 | if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) { | |||
| 3620 | __kmp_free_task_stack(__kmp_thread_from_gtid(gtid), thread_data); | |||
| 3621 | } | |||
| 3622 | #endif // BUILD_TIED_TASK_STACK | |||
| 3623 | } | |||
| 3624 | ||||
| 3625 | // __kmp_realloc_task_threads_data: | |||
| 3626 | // Allocates a threads_data array for a task team, either by allocating an | |||
| 3627 | // initial array or enlarging an existing array. Only the first thread to get | |||
| 3628 | // the lock allocs or enlarges the array and re-initializes the array elements. | |||
| 3629 | // That thread returns "TRUE", the rest return "FALSE". | |||
| 3630 | // Assumes that the new array size is given by task_team -> tt.tt_nproc. | |||
| 3631 | // The current size is given by task_team -> tt.tt_max_threads. | |||
| 3632 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
| 3633 | kmp_task_team_t *task_team) { | |||
| 3634 | kmp_thread_data_t **threads_data_p; | |||
| 3635 | kmp_int32 nthreads, maxthreads; | |||
| 3636 | int is_init_thread = FALSE0; | |||
| 3637 | ||||
| 3638 | if (TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
| ||||
| 3639 | // Already reallocated and initialized. | |||
| 3640 | return FALSE0; | |||
| 3641 | } | |||
| 3642 | ||||
| 3643 | threads_data_p = &task_team->tt.tt_threads_data; | |||
| 3644 | nthreads = task_team->tt.tt_nproc; | |||
| 3645 | maxthreads = task_team->tt.tt_max_threads; | |||
| 3646 | ||||
| 3647 | // All threads must lock when they encounter the first task of the implicit | |||
| 3648 | // task region to make sure threads_data fields are (re)initialized before | |||
| 3649 | // used. | |||
| 3650 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 3651 | ||||
| 3652 | if (!TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
| 3653 | // first thread to enable tasking | |||
| 3654 | kmp_team_t *team = thread->th.th_team; | |||
| 3655 | int i; | |||
| 3656 | ||||
| 3657 | is_init_thread = TRUE(!0); | |||
| 3658 | if (maxthreads < nthreads) { | |||
| 3659 | ||||
| 3660 | if (*threads_data_p != NULL__null) { | |||
| 3661 | kmp_thread_data_t *old_data = *threads_data_p; | |||
| 3662 | kmp_thread_data_t *new_data = NULL__null; | |||
| 3663 | ||||
| 3664 | 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" , __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads ); } | |||
| 3665 | 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" , __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads ); } | |||
| 3666 | ("__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" , __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads ); } | |||
| 3667 | "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" , __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads ); } | |||
| 3668 | __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" , __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads ); }; | |||
| 3669 | // Reallocate threads_data to have more elements than current array | |||
| 3670 | // Cannot use __kmp_thread_realloc() because threads not around for | |||
| 3671 | // kmp_reap_task_team( ). Note all new array entries are initialized | |||
| 3672 | // to zero by __kmp_allocate(). | |||
| 3673 | new_data = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "openmp/runtime/src/kmp_tasking.cpp" , 3674) | |||
| 3674 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "openmp/runtime/src/kmp_tasking.cpp" , 3674); | |||
| 3675 | // copy old data to new data | |||
| 3676 | 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)) | |||
| 3677 | (void *)old_data, maxthreads * sizeof(kmp_thread_data_t))memcpy((void *)new_data, (void *)old_data, maxthreads * sizeof (kmp_thread_data_t)); | |||
| 3678 | ||||
| 3679 | #ifdef BUILD_TIED_TASK_STACK | |||
| 3680 | // GEH: Figure out if this is the right thing to do | |||
| 3681 | for (i = maxthreads; i < nthreads; i++) { | |||
| 3682 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 3683 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data); | |||
| 3684 | } | |||
| 3685 | #endif // BUILD_TIED_TASK_STACK | |||
| 3686 | // Install the new data and free the old data | |||
| 3687 | (*threads_data_p) = new_data; | |||
| 3688 | __kmp_free(old_data)___kmp_free((old_data), "openmp/runtime/src/kmp_tasking.cpp", 3688); | |||
| 3689 | } else { | |||
| 3690 | 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", __kmp_gtid_from_thread (thread), task_team, nthreads); } | |||
| 3691 | "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", __kmp_gtid_from_thread (thread), task_team, nthreads); } | |||
| 3692 | __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", __kmp_gtid_from_thread (thread), task_team, nthreads); }; | |||
| 3693 | // Make the initial allocate for threads_data array, and zero entries | |||
| 3694 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
| 3695 | // kmp_reap_task_team( ). | |||
| 3696 | *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "openmp/runtime/src/kmp_tasking.cpp" , 3697) | |||
| 3697 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "openmp/runtime/src/kmp_tasking.cpp" , 3697); | |||
| 3698 | #ifdef BUILD_TIED_TASK_STACK | |||
| 3699 | // GEH: Figure out if this is the right thing to do | |||
| 3700 | for (i = 0; i < nthreads; i++) { | |||
| 3701 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 3702 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data); | |||
| 3703 | } | |||
| 3704 | #endif // BUILD_TIED_TASK_STACK | |||
| 3705 | } | |||
| 3706 | task_team->tt.tt_max_threads = nthreads; | |||
| 3707 | } else { | |||
| 3708 | // If array has (more than) enough elements, go ahead and use it | |||
| 3709 | KMP_DEBUG_ASSERT(*threads_data_p != NULL)if (!(*threads_data_p != __null)) { __kmp_debug_assert("*threads_data_p != __null" , "openmp/runtime/src/kmp_tasking.cpp", 3709); }; | |||
| 3710 | } | |||
| 3711 | ||||
| 3712 | // initialize threads_data pointers back to thread_info structures | |||
| 3713 | for (i = 0; i < nthreads; i++) { | |||
| 3714 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 3715 | thread_data->td.td_thr = team->t.t_threads[i]; | |||
| ||||
| 3716 | ||||
| 3717 | if (thread_data->td.td_deque_last_stolen >= nthreads) { | |||
| 3718 | // The last stolen field survives across teams / barrier, and the number | |||
| 3719 | // of threads may have changed. It's possible (likely?) that a new | |||
| 3720 | // parallel region will exhibit the same behavior as previous region. | |||
| 3721 | thread_data->td.td_deque_last_stolen = -1; | |||
| 3722 | } | |||
| 3723 | } | |||
| 3724 | ||||
| 3725 | KMP_MB(); | |||
| 3726 | TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE)(task_team->tt.tt_found_tasks) = ((!0)); | |||
| 3727 | } | |||
| 3728 | ||||
| 3729 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 3730 | return is_init_thread; | |||
| 3731 | } | |||
| 3732 | ||||
| 3733 | // __kmp_free_task_threads_data: | |||
| 3734 | // Deallocates a threads_data array for a task team, including any attached | |||
| 3735 | // tasking deques. Only occurs at library shutdown. | |||
| 3736 | static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) { | |||
| 3737 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 3738 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
| 3739 | int i; | |||
| 3740 | for (i = 0; i < task_team->tt.tt_max_threads; i++) { | |||
| 3741 | __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]); | |||
| 3742 | } | |||
| 3743 | __kmp_free(task_team->tt.tt_threads_data)___kmp_free((task_team->tt.tt_threads_data), "openmp/runtime/src/kmp_tasking.cpp" , 3743); | |||
| 3744 | task_team->tt.tt_threads_data = NULL__null; | |||
| 3745 | } | |||
| 3746 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 3747 | } | |||
| 3748 | ||||
| 3749 | // __kmp_free_task_pri_list: | |||
| 3750 | // Deallocates tasking deques used for priority tasks. | |||
| 3751 | // Only occurs at library shutdown. | |||
| 3752 | static void __kmp_free_task_pri_list(kmp_task_team_t *task_team) { | |||
| 3753 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 3754 | if (task_team->tt.tt_task_pri_list != NULL__null) { | |||
| 3755 | kmp_task_pri_t *list = task_team->tt.tt_task_pri_list; | |||
| 3756 | while (list != NULL__null) { | |||
| 3757 | kmp_task_pri_t *next = list->next; | |||
| 3758 | __kmp_free_task_deque(&list->td); | |||
| 3759 | __kmp_free(list)___kmp_free((list), "openmp/runtime/src/kmp_tasking.cpp", 3759 ); | |||
| 3760 | list = next; | |||
| 3761 | } | |||
| 3762 | task_team->tt.tt_task_pri_list = NULL__null; | |||
| 3763 | } | |||
| 3764 | __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 3765 | } | |||
| 3766 | ||||
| 3767 | // __kmp_allocate_task_team: | |||
| 3768 | // Allocates a task team associated with a specific team, taking it from | |||
| 3769 | // the global task team free list if possible. Also initializes data | |||
| 3770 | // structures. | |||
| 3771 | static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread, | |||
| 3772 | kmp_team_t *team) { | |||
| 3773 | kmp_task_team_t *task_team = NULL__null; | |||
| 3774 | int nthreads; | |||
| 3775 | ||||
| 3776 | 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 ? __kmp_gtid_from_thread(thread) : -1), team); } | |||
| 3777 | (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 ? __kmp_gtid_from_thread(thread) : -1), team); }; | |||
| 3778 | ||||
| 3779 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
| 3780 | // Take a task team from the task team pool | |||
| 3781 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3782 | if (__kmp_free_task_teams != NULL__null) { | |||
| 3783 | task_team = __kmp_free_task_teams; | |||
| 3784 | TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next)((__kmp_free_task_teams)) = ((task_team->tt.tt_next)); | |||
| 3785 | task_team->tt.tt_next = NULL__null; | |||
| 3786 | } | |||
| 3787 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3788 | } | |||
| 3789 | ||||
| 3790 | if (task_team == NULL__null) { | |||
| 3791 | 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", __kmp_gtid_from_thread(thread), team ); } | |||
| 3792 | "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", __kmp_gtid_from_thread(thread), team ); } | |||
| 3793 | __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", __kmp_gtid_from_thread(thread), team ); }; | |||
| 3794 | // Allocate a new task team if one is not available. Cannot use | |||
| 3795 | // __kmp_thread_malloc because threads not around for kmp_reap_task_team. | |||
| 3796 | task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t))___kmp_allocate((sizeof(kmp_task_team_t)), "openmp/runtime/src/kmp_tasking.cpp" , 3796); | |||
| 3797 | __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 3798 | __kmp_init_bootstrap_lock(&task_team->tt.tt_task_pri_lock); | |||
| 3799 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 && KMP_DEBUG1 | |||
| 3800 | // suppress race conditions detection on synchronization flags in debug mode | |||
| 3801 | // this helps to analyze library internals eliminating false positives | |||
| 3802 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0( | |||
| 3803 | __itt_suppress_range, __itt_suppress_threading_errors0x000000ff, | |||
| 3804 | &task_team->tt.tt_found_tasks, sizeof(task_team->tt.tt_found_tasks)); | |||
| 3805 | __itt_suppress_mark_range(!__kmp_itt_suppress_mark_range_ptr__3_0) ? (void)0 : __kmp_itt_suppress_mark_range_ptr__3_0(__itt_suppress_range, | |||
| 3806 | __itt_suppress_threading_errors0x000000ff, | |||
| 3807 | CCAST(kmp_uint32 *, &task_team->tt.tt_active)const_cast<kmp_uint32 *>(&task_team->tt.tt_active ), | |||
| 3808 | sizeof(task_team->tt.tt_active)); | |||
| 3809 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */ | |||
| 3810 | // Note: __kmp_allocate zeroes returned memory, othewise we would need: | |||
| 3811 | // task_team->tt.tt_threads_data = NULL; | |||
| 3812 | // task_team->tt.tt_max_threads = 0; | |||
| 3813 | // task_team->tt.tt_next = NULL; | |||
| 3814 | } | |||
| 3815 | ||||
| 3816 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
| 3817 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 3818 | TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE)(task_team->tt.tt_hidden_helper_task_encountered) = (0); | |||
| 3819 | task_team->tt.tt_nproc = nthreads = team->t.t_nproc; | |||
| 3820 | ||||
| 3821 | KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads)(&task_team->tt.tt_unfinished_threads)->store(nthreads , std::memory_order_release); | |||
| 3822 | TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE)(task_team->tt.tt_hidden_helper_task_encountered) = (0); | |||
| 3823 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
| 3824 | ||||
| 3825 | 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 ? __kmp_gtid_from_thread (thread) : -1), task_team, (&task_team->tt.tt_unfinished_threads )->load(std::memory_order_relaxed)); } | |||
| 3826 | "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 ? __kmp_gtid_from_thread (thread) : -1), task_team, (&task_team->tt.tt_unfinished_threads )->load(std::memory_order_relaxed)); } | |||
| 3827 | (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 ? __kmp_gtid_from_thread (thread) : -1), task_team, (&task_team->tt.tt_unfinished_threads )->load(std::memory_order_relaxed)); } | |||
| 3828 | 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 ? __kmp_gtid_from_thread (thread) : -1), task_team, (&task_team->tt.tt_unfinished_threads )->load(std::memory_order_relaxed)); }; | |||
| 3829 | return task_team; | |||
| 3830 | } | |||
| 3831 | ||||
| 3832 | // __kmp_free_task_team: | |||
| 3833 | // Frees the task team associated with a specific thread, and adds it | |||
| 3834 | // to the global task team free list. | |||
| 3835 | void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) { | |||
| 3836 | 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 ? __kmp_gtid_from_thread(thread) : -1, task_team); } | |||
| 3837 | 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 ? __kmp_gtid_from_thread(thread) : -1, task_team); }; | |||
| 3838 | ||||
| 3839 | // Put task team back on free list | |||
| 3840 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3841 | ||||
| 3842 | KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL)if (!(task_team->tt.tt_next == __null)) { __kmp_debug_assert ("task_team->tt.tt_next == __null", "openmp/runtime/src/kmp_tasking.cpp" , 3842); }; | |||
| 3843 | task_team->tt.tt_next = __kmp_free_task_teams; | |||
| 3844 | TCW_PTR(__kmp_free_task_teams, task_team)((__kmp_free_task_teams)) = ((task_team)); | |||
| 3845 | ||||
| 3846 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3847 | } | |||
| 3848 | ||||
| 3849 | // __kmp_reap_task_teams: | |||
| 3850 | // Free all the task teams on the task team free list. | |||
| 3851 | // Should only be done during library shutdown. | |||
| 3852 | // Cannot do anything that needs a thread structure or gtid since they are | |||
| 3853 | // already gone. | |||
| 3854 | void __kmp_reap_task_teams(void) { | |||
| 3855 | kmp_task_team_t *task_team; | |||
| 3856 | ||||
| 3857 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
| 3858 | // Free all task_teams on the free list | |||
| 3859 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3860 | while ((task_team = __kmp_free_task_teams) != NULL__null) { | |||
| 3861 | __kmp_free_task_teams = task_team->tt.tt_next; | |||
| 3862 | task_team->tt.tt_next = NULL__null; | |||
| 3863 | ||||
| 3864 | // Free threads_data if necessary | |||
| 3865 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
| 3866 | __kmp_free_task_threads_data(task_team); | |||
| 3867 | } | |||
| 3868 | if (task_team->tt.tt_task_pri_list != NULL__null) { | |||
| 3869 | __kmp_free_task_pri_list(task_team); | |||
| 3870 | } | |||
| 3871 | __kmp_free(task_team)___kmp_free((task_team), "openmp/runtime/src/kmp_tasking.cpp" , 3871); | |||
| 3872 | } | |||
| 3873 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3874 | } | |||
| 3875 | } | |||
| 3876 | ||||
| 3877 | // __kmp_wait_to_unref_task_teams: | |||
| 3878 | // Some threads could still be in the fork barrier release code, possibly | |||
| 3879 | // trying to steal tasks. Wait for each thread to unreference its task team. | |||
| 3880 | void __kmp_wait_to_unref_task_teams(void) { | |||
| 3881 | kmp_info_t *thread; | |||
| 3882 | kmp_uint32 spins; | |||
| 3883 | kmp_uint64 time; | |||
| 3884 | int done; | |||
| 3885 | ||||
| 3886 | KMP_INIT_YIELD(spins){ (spins) = __kmp_yield_init; }; | |||
| 3887 | KMP_INIT_BACKOFF(time){ (time) = __kmp_pause_init; }; | |||
| 3888 | ||||
| 3889 | for (;;) { | |||
| 3890 | done = TRUE(!0); | |||
| 3891 | ||||
| 3892 | // TODO: GEH - this may be is wrong because some sync would be necessary | |||
| 3893 | // in case threads are added to the pool during the traversal. Need to | |||
| 3894 | // verify that lock for thread pool is held when calling this routine. | |||
| 3895 | for (thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); thread != NULL__null; | |||
| 3896 | thread = thread->th.th_next_pool) { | |||
| 3897 | #if KMP_OS_WINDOWS0 | |||
| 3898 | DWORD exit_val; | |||
| 3899 | #endif | |||
| 3900 | if (TCR_PTR(thread->th.th_task_team)((void *)(thread->th.th_task_team)) == NULL__null) { | |||
| 3901 | 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" , __kmp_gtid_from_thread(thread)); } | |||
| 3902 | __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" , __kmp_gtid_from_thread(thread)); }; | |||
| 3903 | continue; | |||
| 3904 | } | |||
| 3905 | #if KMP_OS_WINDOWS0 | |||
| 3906 | // TODO: GEH - add this check for Linux* OS / OS X* as well? | |||
| 3907 | if (!__kmp_is_thread_alive(thread, &exit_val)) { | |||
| 3908 | thread->th.th_task_team = NULL__null; | |||
| 3909 | continue; | |||
| 3910 | } | |||
| 3911 | #endif | |||
| 3912 | ||||
| 3913 | done = FALSE0; // Because th_task_team pointer is not NULL for this thread | |||
| 3914 | ||||
| 3915 | 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", __kmp_gtid_from_thread(thread)); } | |||
| 3916 | "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", __kmp_gtid_from_thread(thread)); } | |||
| 3917 | __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", __kmp_gtid_from_thread(thread)); }; | |||
| 3918 | ||||
| 3919 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | |||
| 3920 | void *sleep_loc; | |||
| 3921 | // If the thread is sleeping, awaken it. | |||
| 3922 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
| 3923 | NULL__null) { | |||
| 3924 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread )); } | |||
| 3925 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread )); } | |||
| 3926 | ("__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" , __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread )); } | |||
| 3927 | __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" , __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread )); }; | |||
| 3928 | __kmp_null_resume_wrapper(thread); | |||
| 3929 | } | |||
| 3930 | } | |||
| 3931 | } | |||
| 3932 | if (done) { | |||
| 3933 | break; | |||
| 3934 | } | |||
| 3935 | ||||
| 3936 | // If oversubscribed or have waited a bit, yield. | |||
| 3937 | KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time){ if (__kmp_tpause_enabled) { if (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))) { __kmp_tpause(0, (time) ); } else { __kmp_tpause(__kmp_tpause_hint, (time)); } (time) = (time << 1 | 1) & ((kmp_uint64)0xFFFF); } else { __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { __kmp_yield(); } else if (__kmp_use_yield == 1) { (spins) -= 2; if (!(spins)) { __kmp_yield(); (spins) = __kmp_yield_next; } } } }; | |||
| 3938 | } | |||
| 3939 | } | |||
| 3940 | ||||
| 3941 | void __kmp_shift_task_state_stack(kmp_info_t *this_thr, kmp_uint8 value) { | |||
| 3942 | // Shift values from th_task_state_top+1 to task_state_stack_sz | |||
| 3943 | if (this_thr->th.th_task_state_top + 1 >= | |||
| 3944 | this_thr->th.th_task_state_stack_sz) { // increase size | |||
| 3945 | kmp_uint32 new_size = 2 * this_thr->th.th_task_state_stack_sz; | |||
| 3946 | kmp_uint8 *old_stack, *new_stack; | |||
| 3947 | kmp_uint32 i; | |||
| 3948 | new_stack = (kmp_uint8 *)__kmp_allocate(new_size)___kmp_allocate((new_size), "openmp/runtime/src/kmp_tasking.cpp" , 3948); | |||
| 3949 | for (i = 0; i <= this_thr->th.th_task_state_top; ++i) { | |||
| 3950 | new_stack[i] = this_thr->th.th_task_state_memo_stack[i]; | |||
| 3951 | } | |||
| 3952 | // If we need to reallocate do the shift at the same time. | |||
| 3953 | for (; i < this_thr->th.th_task_state_stack_sz; ++i) { | |||
| 3954 | new_stack[i + 1] = this_thr->th.th_task_state_memo_stack[i]; | |||
| 3955 | } | |||
| 3956 | for (i = this_thr->th.th_task_state_stack_sz; i < new_size; | |||
| 3957 | ++i) { // zero-init rest of stack | |||
| 3958 | new_stack[i] = 0; | |||
| 3959 | } | |||
| 3960 | old_stack = this_thr->th.th_task_state_memo_stack; | |||
| 3961 | this_thr->th.th_task_state_memo_stack = new_stack; | |||
| 3962 | this_thr->th.th_task_state_stack_sz = new_size; | |||
| 3963 | __kmp_free(old_stack)___kmp_free((old_stack), "openmp/runtime/src/kmp_tasking.cpp" , 3963); | |||
| 3964 | } else { | |||
| 3965 | kmp_uint8 *end; | |||
| 3966 | kmp_uint32 i; | |||
| 3967 | ||||
| 3968 | end = &this_thr->th | |||
| 3969 | .th_task_state_memo_stack[this_thr->th.th_task_state_stack_sz]; | |||
| 3970 | ||||
| 3971 | for (i = this_thr->th.th_task_state_stack_sz - 1; | |||
| 3972 | i > this_thr->th.th_task_state_top; i--, end--) | |||
| 3973 | end[0] = end[-1]; | |||
| 3974 | } | |||
| 3975 | this_thr->th.th_task_state_memo_stack[this_thr->th.th_task_state_top + 1] = | |||
| 3976 | value; | |||
| 3977 | } | |||
| 3978 | ||||
| 3979 | // __kmp_task_team_setup: Create a task_team for the current team, but use | |||
| 3980 | // an already created, unused one if it already exists. | |||
| 3981 | void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) { | |||
| 3982 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 3982); }; | |||
| 3983 | ||||
| 3984 | // If this task_team hasn't been created yet, allocate it. It will be used in | |||
| 3985 | // the region after the next. | |||
| 3986 | // If it exists, it is the current task team and shouldn't be touched yet as | |||
| 3987 | // it may still be in use. | |||
| 3988 | if (team->t.t_task_team[this_thr->th.th_task_state] == NULL__null && | |||
| 3989 | (always || team->t.t_nproc > 1)) { | |||
| 3990 | team->t.t_task_team[this_thr->th.th_task_state] = | |||
| 3991 | __kmp_allocate_task_team(this_thr, team); | |||
| 3992 | KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created new task_team %p"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created new task_team %p" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[this_thr->th.th_task_state], team ->t.t_id, this_thr->th.th_task_state); } | |||
| 3993 | " for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created new task_team %p" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[this_thr->th.th_task_state], team ->t.t_id, this_thr->th.th_task_state); } | |||
| 3994 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created new task_team %p" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[this_thr->th.th_task_state], team ->t.t_id, this_thr->th.th_task_state); } | |||
| 3995 | team->t.t_task_team[this_thr->th.th_task_state], team->t.t_id,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created new task_team %p" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[this_thr->th.th_task_state], team ->t.t_id, this_thr->th.th_task_state); } | |||
| 3996 | this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created new task_team %p" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[this_thr->th.th_task_state], team ->t.t_id, this_thr->th.th_task_state); }; | |||
| 3997 | } | |||
| 3998 | if (this_thr->th.th_task_state == 1 && always && team->t.t_nproc == 1) { | |||
| 3999 | // fix task state stack to adjust for proxy and helper tasks | |||
| 4000 | KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d needs to shift stack"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d needs to shift stack" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_id, this_thr->th.th_task_state); } | |||
| 4001 | " for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d needs to shift stack" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_id, this_thr->th.th_task_state); } | |||
| 4002 | __kmp_gtid_from_thread(this_thr), team->t.t_id,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d needs to shift stack" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_id, this_thr->th.th_task_state); } | |||
| 4003 | this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d needs to shift stack" " for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_id, this_thr->th.th_task_state); }; | |||
| 4004 | __kmp_shift_task_state_stack(this_thr, this_thr->th.th_task_state); | |||
| 4005 | } | |||
| 4006 | ||||
| 4007 | // After threads exit the release, they will call sync, and then point to this | |||
| 4008 | // other task_team; make sure it is allocated and properly initialized. As | |||
| 4009 | // threads spin in the barrier release phase, they will continue to use the | |||
| 4010 | // previous task_team struct(above), until they receive the signal to stop | |||
| 4011 | // checking for tasks (they can't safely reference the kmp_team_t struct, | |||
| 4012 | // which could be reallocated by the primary thread). No task teams are formed | |||
| 4013 | // for serialized teams. | |||
| 4014 | if (team->t.t_nproc > 1) { | |||
| 4015 | int other_team = 1 - this_thr->th.th_task_state; | |||
| 4016 | KMP_DEBUG_ASSERT(other_team >= 0 && other_team < 2)if (!(other_team >= 0 && other_team < 2)) { __kmp_debug_assert ("other_team >= 0 && other_team < 2", "openmp/runtime/src/kmp_tasking.cpp" , 4016); }; | |||
| 4017 | if (team->t.t_task_team[other_team] == NULL__null) { // setup other team as well | |||
| 4018 | team->t.t_task_team[other_team] = | |||
| 4019 | __kmp_allocate_task_team(this_thr, team); | |||
| 4020 | KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created second new "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created second new " "task_team %p for team %d at parity=%d\n", __kmp_gtid_from_thread (this_thr), team->t.t_task_team[other_team], team->t.t_id , other_team); } | |||
| 4021 | "task_team %p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created second new " "task_team %p for team %d at parity=%d\n", __kmp_gtid_from_thread (this_thr), team->t.t_task_team[other_team], team->t.t_id , other_team); } | |||
| 4022 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created second new " "task_team %p for team %d at parity=%d\n", __kmp_gtid_from_thread (this_thr), team->t.t_task_team[other_team], team->t.t_id , other_team); } | |||
| 4023 | team->t.t_task_team[other_team], team->t.t_id, other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d created second new " "task_team %p for team %d at parity=%d\n", __kmp_gtid_from_thread (this_thr), team->t.t_task_team[other_team], team->t.t_id , other_team); }; | |||
| 4024 | } else { // Leave the old task team struct in place for the upcoming region; | |||
| 4025 | // adjust as needed | |||
| 4026 | kmp_task_team_t *task_team = team->t.t_task_team[other_team]; | |||
| 4027 | if (!task_team->tt.tt_active || | |||
| 4028 | team->t.t_nproc != task_team->tt.tt_nproc) { | |||
| 4029 | TCW_4(task_team->tt.tt_nproc, team->t.t_nproc)(task_team->tt.tt_nproc) = (team->t.t_nproc); | |||
| 4030 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
| 4031 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 4032 | TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE)(task_team->tt.tt_hidden_helper_task_encountered) = (0); | |||
| 4033 | 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) | |||
| 4034 | team->t.t_nproc)(&task_team->tt.tt_unfinished_threads)->store(team-> t.t_nproc, std::memory_order_release); | |||
| 4035 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
| 4036 | } | |||
| 4037 | // if team size has changed, the first thread to enable tasking will | |||
| 4038 | // realloc threads_data if necessary | |||
| 4039 | KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d reset next task_team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d reset next task_team " "%p for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[other_team], team->t.t_id, other_team ); } | |||
| 4040 | "%p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d reset next task_team " "%p for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[other_team], team->t.t_id, other_team ); } | |||
| 4041 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d reset next task_team " "%p for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[other_team], team->t.t_id, other_team ); } | |||
| 4042 | team->t.t_task_team[other_team], team->t.t_id, other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Primary T#%d reset next task_team " "%p for team %d at parity=%d\n", __kmp_gtid_from_thread(this_thr ), team->t.t_task_team[other_team], team->t.t_id, other_team ); }; | |||
| 4043 | } | |||
| 4044 | } | |||
| 4045 | ||||
| 4046 | // For regular thread, task enabling should be called when the task is going | |||
| 4047 | // to be pushed to a dequeue. However, for the hidden helper thread, we need | |||
| 4048 | // it ahead of time so that some operations can be performed without race | |||
| 4049 | // condition. | |||
| 4050 | if (this_thr == __kmp_hidden_helper_main_thread) { | |||
| 4051 | for (int i = 0; i < 2; ++i) { | |||
| 4052 | kmp_task_team_t *task_team = team->t.t_task_team[i]; | |||
| 4053 | if (KMP_TASKING_ENABLED(task_team)((!0) == ((task_team)->tt.tt_found_tasks))) { | |||
| 4054 | continue; | |||
| 4055 | } | |||
| 4056 | __kmp_enable_tasking(task_team, this_thr); | |||
| 4057 | for (int j = 0; j < task_team->tt.tt_nproc; ++j) { | |||
| 4058 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[j]; | |||
| 4059 | if (thread_data->td.td_deque == NULL__null) { | |||
| 4060 | __kmp_alloc_task_deque(__kmp_hidden_helper_threads[j], thread_data); | |||
| 4061 | } | |||
| 4062 | } | |||
| 4063 | } | |||
| 4064 | } | |||
| 4065 | } | |||
| 4066 | ||||
| 4067 | // __kmp_task_team_sync: Propagation of task team data from team to threads | |||
| 4068 | // which happens just after the release phase of a team barrier. This may be | |||
| 4069 | // called by any thread, but only for teams with # threads > 1. | |||
| 4070 | void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) { | |||
| 4071 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 4071); }; | |||
| 4072 | ||||
| 4073 | // Toggle the th_task_state field, to switch which task_team this thread | |||
| 4074 | // refers to | |||
| 4075 | this_thr->th.th_task_state = (kmp_uint8)(1 - this_thr->th.th_task_state); | |||
| 4076 | ||||
| 4077 | // It is now safe to propagate the task team pointer from the team struct to | |||
| 4078 | // the current thread. | |||
| 4079 | 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])) | |||
| 4080 | 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])); | |||
| 4081 | 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", __kmp_gtid_from_thread(this_thr ), this_thr->th.th_task_team, team->t.t_id, this_thr-> th.th_task_state); } | |||
| 4082 | ("__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", __kmp_gtid_from_thread(this_thr ), this_thr->th.th_task_team, team->t.t_id, this_thr-> th.th_task_state); } | |||
| 4083 | "%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", __kmp_gtid_from_thread(this_thr ), this_thr->th.th_task_team, team->t.t_id, this_thr-> th.th_task_state); } | |||
| 4084 | __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", __kmp_gtid_from_thread(this_thr ), this_thr->th.th_task_team, team->t.t_id, this_thr-> th.th_task_state); } | |||
| 4085 | team->t.t_id, 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", __kmp_gtid_from_thread(this_thr ), this_thr->th.th_task_team, team->t.t_id, this_thr-> th.th_task_state); }; | |||
| 4086 | } | |||
| 4087 | ||||
| 4088 | // __kmp_task_team_wait: Primary thread waits for outstanding tasks after the | |||
| 4089 | // barrier gather phase. Only called by primary thread if #threads in team > 1 | |||
| 4090 | // or if proxy tasks were created. | |||
| 4091 | // | |||
| 4092 | // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off | |||
| 4093 | // by passing in 0 optionally as the last argument. When wait is zero, primary | |||
| 4094 | // thread does not wait for unfinished_threads to reach 0. | |||
| 4095 | void __kmp_task_team_wait( | |||
| 4096 | kmp_info_t *this_thr, | |||
| 4097 | kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, int wait) { | |||
| 4098 | kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state]; | |||
| 4099 | ||||
| 4100 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)if (!(__kmp_tasking_mode != tskm_immediate_exec)) { __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "openmp/runtime/src/kmp_tasking.cpp" , 4100); }; | |||
| 4101 | KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team)if (!(task_team == this_thr->th.th_task_team)) { __kmp_debug_assert ("task_team == this_thr->th.th_task_team", "openmp/runtime/src/kmp_tasking.cpp" , 4101); }; | |||
| 4102 | ||||
| 4103 | if ((task_team != NULL__null) && KMP_TASKING_ENABLED(task_team)((!0) == ((task_team)->tt.tt_found_tasks))) { | |||
| 4104 | if (wait) { | |||
| 4105 | KA_TRACE(20, ("__kmp_task_team_wait: Primary T#%d waiting for all tasks "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", __kmp_gtid_from_thread (this_thr), task_team); } | |||
| 4106 | "(for unfinished_threads to reach 0) on task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", __kmp_gtid_from_thread (this_thr), task_team); } | |||
| 4107 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", __kmp_gtid_from_thread (this_thr), task_team); }; | |||
| 4108 | // Worker threads may have dropped through to release phase, but could | |||
| 4109 | // still be executing tasks. Wait here for tasks to complete. To avoid | |||
| 4110 | // memory contention, only primary thread checks termination condition. | |||
| 4111 | kmp_flag_32<false, false> flag( | |||
| 4112 | RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads) | |||
| 4113 | &task_team->tt.tt_unfinished_threads)reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads), | |||
| 4114 | 0U); | |||
| 4115 | flag.wait(this_thr, TRUE(!0) USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj); | |||
| 4116 | } | |||
| 4117 | // Deactivate the old task team, so that the worker threads will stop | |||
| 4118 | // referencing it while spinning. | |||
| 4119 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , __kmp_gtid_from_thread(this_thr), task_team); } | |||
| 4120 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , __kmp_gtid_from_thread(this_thr), task_team); } | |||
| 4121 | ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , __kmp_gtid_from_thread(this_thr), task_team); } | |||
| 4122 | "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: Primary T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , __kmp_gtid_from_thread(this_thr), task_team); } | |||
| 4123 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , __kmp_gtid_from_thread(this_thr), task_team); }; | |||
| 4124 | KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||if (!(task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0))) { __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0)" , "openmp/runtime/src/kmp_tasking.cpp", 4126); } | |||
| 4125 | task_team->tt.tt_found_proxy_tasks == TRUE ||if (!(task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0))) { __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0)" , "openmp/runtime/src/kmp_tasking.cpp", 4126); } | |||
| 4126 | task_team->tt.tt_hidden_helper_task_encountered == TRUE)if (!(task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0))) { __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0) || task_team->tt.tt_hidden_helper_task_encountered == (!0)" , "openmp/runtime/src/kmp_tasking.cpp", 4126); }; | |||
| 4127 | TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 4128 | TCW_SYNC_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE)(task_team->tt.tt_hidden_helper_task_encountered) = (0); | |||
| 4129 | 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); | |||
| 4130 | TCW_SYNC_4(task_team->tt.tt_active, FALSE)(task_team->tt.tt_active) = (0); | |||
| 4131 | KMP_MB(); | |||
| 4132 | ||||
| 4133 | TCW_PTR(this_thr->th.th_task_team, NULL)((this_thr->th.th_task_team)) = ((__null)); | |||
| 4134 | } | |||
| 4135 | } | |||
| 4136 | ||||
| 4137 | // __kmp_tasking_barrier: | |||
| 4138 | // This routine is called only when __kmp_tasking_mode == tskm_extra_barrier. | |||
| 4139 | // Internal function to execute all tasks prior to a regular barrier or a join | |||
| 4140 | // barrier. It is a full barrier itself, which unfortunately turns regular | |||
| 4141 | // barriers into double barriers and join barriers into 1 1/2 barriers. | |||
| 4142 | void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) { | |||
| 4143 | 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 ) | |||
| 4144 | std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads ) | |||
| 4145 | &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 ); | |||
| 4146 | int flag = FALSE0; | |||
| 4147 | KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier)if (!(__kmp_tasking_mode == tskm_extra_barrier)) { __kmp_debug_assert ("__kmp_tasking_mode == tskm_extra_barrier", "openmp/runtime/src/kmp_tasking.cpp" , 4147); }; | |||
| 4148 | ||||
| 4149 | #if USE_ITT_BUILD1 | |||
| 4150 | 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"); | |||
| 4151 | #endif /* USE_ITT_BUILD */ | |||
| 4152 | kmp_flag_32<false, false> spin_flag(spin, 0U); | |||
| 4153 | while (!spin_flag.execute_tasks(thread, gtid, TRUE(!0), | |||
| 4154 | &flag USE_ITT_BUILD_ARG(NULL), __null, 0)) { | |||
| 4155 | #if USE_ITT_BUILD1 | |||
| 4156 | // TODO: What about itt_sync_obj?? | |||
| 4157 | 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); | |||
| 4158 | #endif /* USE_ITT_BUILD */ | |||
| 4159 | ||||
| 4160 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
| 4161 | if (__kmp_global.g.g_abort) | |||
| 4162 | __kmp_abort_thread(); | |||
| 4163 | break; | |||
| 4164 | } | |||
| 4165 | KMP_YIELD(TRUE){ __kmp_x86_pause(); if (((!0)) && (((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))))))) __kmp_yield (); }; | |||
| 4166 | } | |||
| 4167 | #if USE_ITT_BUILD1 | |||
| 4168 | 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); | |||
| 4169 | #endif /* USE_ITT_BUILD */ | |||
| 4170 | } | |||
| 4171 | ||||
| 4172 | // __kmp_give_task puts a task into a given thread queue if: | |||
| 4173 | // - the queue for that thread was created | |||
| 4174 | // - there's space in that queue | |||
| 4175 | // Because of this, __kmp_push_task needs to check if there's space after | |||
| 4176 | // getting the lock | |||
| 4177 | static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task, | |||
| 4178 | kmp_int32 pass) { | |||
| 4179 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 4180 | kmp_task_team_t *task_team = taskdata->td_task_team; | |||
| 4181 | ||||
| 4182 | 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); } | |||
| 4183 | taskdata, tid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_give_task: trying to give task %p to thread %d.\n" , taskdata, tid); }; | |||
| 4184 | ||||
| 4185 | // If task_team is NULL something went really bad... | |||
| 4186 | KMP_DEBUG_ASSERT(task_team != NULL)if (!(task_team != __null)) { __kmp_debug_assert("task_team != __null" , "openmp/runtime/src/kmp_tasking.cpp", 4186); }; | |||
| 4187 | ||||
| 4188 | bool result = false; | |||
| 4189 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 4190 | ||||
| 4191 | if (thread_data->td.td_deque == NULL__null) { | |||
| 4192 | // There's no queue in this thread, go find another one | |||
| 4193 | // We're guaranteed that at least one thread has a queue | |||
| 4194 | 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); } | |||
| 4195 | ("__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); } | |||
| 4196 | 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); }; | |||
| 4197 | return result; | |||
| 4198 | } | |||
| 4199 | ||||
| 4200 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 4201 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 4202 | 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); } | |||
| 4203 | 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); } | |||
| 4204 | ("__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); } | |||
| 4205 | 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); }; | |||
| 4206 | ||||
| 4207 | // if this deque is bigger than the pass ratio give a chance to another | |||
| 4208 | // thread | |||
| 4209 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
| 4210 | return result; | |||
| 4211 | ||||
| 4212 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 4213 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 4214 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 4215 | // expand deque to push the task which is not allowed to execute | |||
| 4216 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 4217 | } | |||
| 4218 | ||||
| 4219 | } else { | |||
| 4220 | ||||
| 4221 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 4222 | ||||
| 4223 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 4224 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 4225 | 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); } | |||
| 4226 | "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); } | |||
| 4227 | 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); }; | |||
| 4228 | ||||
| 4229 | // if this deque is bigger than the pass ratio give a chance to another | |||
| 4230 | // thread | |||
| 4231 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
| 4232 | goto release_and_exit; | |||
| 4233 | ||||
| 4234 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 4235 | } | |||
| 4236 | } | |||
| 4237 | ||||
| 4238 | // lock is held here, and there is space in the deque | |||
| 4239 | ||||
| 4240 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata; | |||
| 4241 | // Wrap index. | |||
| 4242 | thread_data->td.td_deque_tail = | |||
| 4243 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 4244 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
| 4245 | TCR_4(thread_data->td.td_deque_ntasks) + 1)(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1); | |||
| 4246 | ||||
| 4247 | result = true; | |||
| 4248 | 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); } | |||
| 4249 | taskdata, tid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: successfully gave task %p to thread %d.\n" , taskdata, tid); }; | |||
| 4250 | ||||
| 4251 | release_and_exit: | |||
| 4252 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 4253 | ||||
| 4254 | return result; | |||
| 4255 | } | |||
| 4256 | ||||
| 4257 | #define PROXY_TASK_FLAG0x40000000 0x40000000 | |||
| 4258 | /* The finish of the proxy tasks is divided in two pieces: | |||
| 4259 | - the top half is the one that can be done from a thread outside the team | |||
| 4260 | - the bottom half must be run from a thread within the team | |||
| 4261 | ||||
| 4262 | In order to run the bottom half the task gets queued back into one of the | |||
| 4263 | threads of the team. Once the td_incomplete_child_task counter of the parent | |||
| 4264 | is decremented the threads can leave the barriers. So, the bottom half needs | |||
| 4265 | to be queued before the counter is decremented. The top half is therefore | |||
| 4266 | divided in two parts: | |||
| 4267 | - things that can be run before queuing the bottom half | |||
| 4268 | - things that must be run after queuing the bottom half | |||
| 4269 | ||||
| 4270 | This creates a second race as the bottom half can free the task before the | |||
| 4271 | second top half is executed. To avoid this we use the | |||
| 4272 | td_incomplete_child_task of the proxy task to synchronize the top and bottom | |||
| 4273 | half. */ | |||
| 4274 | static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
| 4275 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4275); }; | |||
| 4276 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)if (!(taskdata->td_flags.proxy == 1)) { __kmp_debug_assert ("taskdata->td_flags.proxy == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4276); }; | |||
| 4277 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)if (!(taskdata->td_flags.complete == 0)) { __kmp_debug_assert ("taskdata->td_flags.complete == 0", "openmp/runtime/src/kmp_tasking.cpp" , 4277); }; | |||
| 4278 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)if (!(taskdata->td_flags.freed == 0)) { __kmp_debug_assert ("taskdata->td_flags.freed == 0", "openmp/runtime/src/kmp_tasking.cpp" , 4278); }; | |||
| 4279 | ||||
| 4280 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
| 4281 | ||||
| 4282 | if (taskdata->td_taskgroup) | |||
| 4283 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
| 4284 | ||||
| 4285 | // Create an imaginary children for this task so the bottom half cannot | |||
| 4286 | // release the task before we have completed the second top half | |||
| 4287 | KMP_ATOMIC_OR(&taskdata->td_incomplete_child_tasks, PROXY_TASK_FLAG)(&taskdata->td_incomplete_child_tasks)->fetch_or(0x40000000 , std::memory_order_acq_rel); | |||
| 4288 | } | |||
| 4289 | ||||
| 4290 | static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
| 4291 | #if KMP_DEBUG1 | |||
| 4292 | kmp_int32 children = 0; | |||
| 4293 | // Predecrement simulated by "- 1" calculation | |||
| 4294 | children = -1 + | |||
| 4295 | #endif | |||
| 4296 | 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); | |||
| 4297 | KMP_DEBUG_ASSERT(children >= 0)if (!(children >= 0)) { __kmp_debug_assert("children >= 0" , "openmp/runtime/src/kmp_tasking.cpp", 4297); }; | |||
| 4298 | ||||
| 4299 | // Remove the imaginary children | |||
| 4300 | KMP_ATOMIC_AND(&taskdata->td_incomplete_child_tasks, ~PROXY_TASK_FLAG)(&taskdata->td_incomplete_child_tasks)->fetch_and(~ 0x40000000, std::memory_order_acq_rel); | |||
| 4301 | } | |||
| 4302 | ||||
| 4303 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) { | |||
| 4304 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 4305 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 4306 | ||||
| 4307 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)if (!(taskdata->td_flags.proxy == 1)) { __kmp_debug_assert ("taskdata->td_flags.proxy == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4307); }; | |||
| 4308 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==if (!(taskdata->td_flags.complete == 1)) { __kmp_debug_assert ("taskdata->td_flags.complete == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4309); } | |||
| 4309 | 1)if (!(taskdata->td_flags.complete == 1)) { __kmp_debug_assert ("taskdata->td_flags.complete == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4309); }; // top half must run before bottom half | |||
| 4310 | ||||
| 4311 | // We need to wait to make sure the top half is finished | |||
| 4312 | // Spinning here should be ok as this should happen quickly | |||
| 4313 | while ((KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) & | |||
| 4314 | PROXY_TASK_FLAG0x40000000) > 0) | |||
| 4315 | ; | |||
| 4316 | ||||
| 4317 | __kmp_release_deps(gtid, taskdata); | |||
| 4318 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
| 4319 | } | |||
| 4320 | ||||
| 4321 | /*! | |||
| 4322 | @ingroup TASKING | |||
| 4323 | @param gtid Global Thread ID of encountering thread | |||
| 4324 | @param ptask Task which execution is completed | |||
| 4325 | ||||
| 4326 | Execute the completion of a proxy task from a thread of that is part of the | |||
| 4327 | team. Run first and bottom halves directly. | |||
| 4328 | */ | |||
| 4329 | void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) { | |||
| 4330 | KMP_DEBUG_ASSERT(ptask != NULL)if (!(ptask != __null)) { __kmp_debug_assert("ptask != __null" , "openmp/runtime/src/kmp_tasking.cpp", 4330); }; | |||
| 4331 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 4332 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
| 4333 | 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); } | |||
| 4334 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
| 4335 | __kmp_assert_valid_gtid(gtid); | |||
| 4336 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)if (!(taskdata->td_flags.proxy == 1)) { __kmp_debug_assert ("taskdata->td_flags.proxy == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4336); }; | |||
| 4337 | ||||
| 4338 | __kmp_first_top_half_finish_proxy(taskdata); | |||
| 4339 | __kmp_second_top_half_finish_proxy(taskdata); | |||
| 4340 | __kmp_bottom_half_finish_proxy(gtid, ptask); | |||
| 4341 | ||||
| 4342 | 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); } | |||
| 4343 | ("__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); } | |||
| 4344 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
| 4345 | } | |||
| 4346 | ||||
| 4347 | void __kmpc_give_task(kmp_task_t *ptask, kmp_int32 start = 0) { | |||
| 4348 | KMP_DEBUG_ASSERT(ptask != NULL)if (!(ptask != __null)) { __kmp_debug_assert("ptask != __null" , "openmp/runtime/src/kmp_tasking.cpp", 4348); }; | |||
| 4349 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 4350 | ||||
| 4351 | // Enqueue task to complete bottom half completion from a thread within the | |||
| 4352 | // corresponding team | |||
| 4353 | kmp_team_t *team = taskdata->td_team; | |||
| 4354 | kmp_int32 nthreads = team->t.t_nproc; | |||
| 4355 | kmp_info_t *thread; | |||
| 4356 | ||||
| 4357 | // This should be similar to start_k = __kmp_get_random( thread ) % nthreads | |||
| 4358 | // but we cannot use __kmp_get_random here | |||
| 4359 | kmp_int32 start_k = start % nthreads; | |||
| 4360 | kmp_int32 pass = 1; | |||
| 4361 | kmp_int32 k = start_k; | |||
| 4362 | ||||
| 4363 | do { | |||
| 4364 | // For now we're just linearly trying to find a thread | |||
| 4365 | thread = team->t.t_threads[k]; | |||
| 4366 | k = (k + 1) % nthreads; | |||
| 4367 | ||||
| 4368 | // we did a full pass through all the threads | |||
| 4369 | if (k == start_k) | |||
| 4370 | pass = pass << 1; | |||
| 4371 | ||||
| 4372 | } while (!__kmp_give_task(thread, k, ptask, pass)); | |||
| 4373 | ||||
| 4374 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647) && __kmp_wpolicy_passive) { | |||
| 4375 | // awake at least one thread to execute given task | |||
| 4376 | for (int i = 0; i < nthreads; ++i) { | |||
| 4377 | thread = team->t.t_threads[i]; | |||
| 4378 | if (thread->th.th_sleep_loc != NULL__null) { | |||
| 4379 | __kmp_null_resume_wrapper(thread); | |||
| 4380 | break; | |||
| 4381 | } | |||
| 4382 | } | |||
| 4383 | } | |||
| 4384 | } | |||
| 4385 | ||||
| 4386 | /*! | |||
| 4387 | @ingroup TASKING | |||
| 4388 | @param ptask Task which execution is completed | |||
| 4389 | ||||
| 4390 | Execute the completion of a proxy task from a thread that could not belong to | |||
| 4391 | the team. | |||
| 4392 | */ | |||
| 4393 | void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) { | |||
| 4394 | KMP_DEBUG_ASSERT(ptask != NULL)if (!(ptask != __null)) { __kmp_debug_assert("ptask != __null" , "openmp/runtime/src/kmp_tasking.cpp", 4394); }; | |||
| 4395 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 4396 | ||||
| 4397 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
| 4398 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
| 4399 | ("__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); } | |||
| 4400 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); }; | |||
| 4401 | ||||
| 4402 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)if (!(taskdata->td_flags.proxy == 1)) { __kmp_debug_assert ("taskdata->td_flags.proxy == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4402); }; | |||
| 4403 | ||||
| 4404 | __kmp_first_top_half_finish_proxy(taskdata); | |||
| 4405 | ||||
| 4406 | __kmpc_give_task(ptask); | |||
| 4407 | ||||
| 4408 | __kmp_second_top_half_finish_proxy(taskdata); | |||
| 4409 | ||||
| 4410 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
| 4411 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
| 4412 | ("__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); } | |||
| 4413 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); }; | |||
| 4414 | } | |||
| 4415 | ||||
| 4416 | kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, int gtid, | |||
| 4417 | kmp_task_t *task) { | |||
| 4418 | kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 4419 | if (td->td_allow_completion_event.type == KMP_EVENT_UNINITIALIZED) { | |||
| 4420 | td->td_allow_completion_event.type = KMP_EVENT_ALLOW_COMPLETION; | |||
| 4421 | td->td_allow_completion_event.ed.task = task; | |||
| 4422 | __kmp_init_tas_lock(&td->td_allow_completion_event.lock); | |||
| 4423 | } | |||
| 4424 | return &td->td_allow_completion_event; | |||
| 4425 | } | |||
| 4426 | ||||
| 4427 | void __kmp_fulfill_event(kmp_event_t *event) { | |||
| 4428 | if (event->type == KMP_EVENT_ALLOW_COMPLETION) { | |||
| 4429 | kmp_task_t *ptask = event->ed.task; | |||
| 4430 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 4431 | bool detached = false; | |||
| 4432 | int gtid = __kmp_get_gtid()__kmp_get_global_thread_id(); | |||
| 4433 | ||||
| 4434 | // The associated task might have completed or could be completing at this | |||
| 4435 | // point. | |||
| 4436 | // We need to take the lock to avoid races | |||
| 4437 | __kmp_acquire_tas_lock(&event->lock, gtid); | |||
| 4438 | if (taskdata->td_flags.proxy == TASK_PROXY1) { | |||
| 4439 | detached = true; | |||
| 4440 | } else { | |||
| 4441 | #if OMPT_SUPPORT1 | |||
| 4442 | // The OMPT event must occur under mutual exclusion, | |||
| 4443 | // otherwise the tool might access ptask after free | |||
| 4444 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 4445 | __ompt_task_finish(ptask, NULL__null, ompt_task_early_fulfill); | |||
| 4446 | #endif | |||
| 4447 | } | |||
| 4448 | event->type = KMP_EVENT_UNINITIALIZED; | |||
| 4449 | __kmp_release_tas_lock(&event->lock, gtid); | |||
| 4450 | ||||
| 4451 | if (detached) { | |||
| 4452 | #if OMPT_SUPPORT1 | |||
| 4453 | // We free ptask afterwards and know the task is finished, | |||
| 4454 | // so locking is not necessary | |||
| 4455 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 4456 | __ompt_task_finish(ptask, NULL__null, ompt_task_late_fulfill); | |||
| 4457 | #endif | |||
| 4458 | // If the task detached complete the proxy task | |||
| 4459 | if (gtid >= 0) { | |||
| 4460 | kmp_team_t *team = taskdata->td_team; | |||
| 4461 | kmp_info_t *thread = __kmp_get_thread()(__kmp_thread_from_gtid(__kmp_get_global_thread_id())); | |||
| 4462 | if (thread->th.th_team == team) { | |||
| 4463 | __kmpc_proxy_task_completed(gtid, ptask); | |||
| 4464 | return; | |||
| 4465 | } | |||
| 4466 | } | |||
| 4467 | ||||
| 4468 | // fallback | |||
| 4469 | __kmpc_proxy_task_completed_ooo(ptask); | |||
| 4470 | } | |||
| 4471 | } | |||
| 4472 | } | |||
| 4473 | ||||
| 4474 | // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task | |||
| 4475 | // for taskloop | |||
| 4476 | // | |||
| 4477 | // thread: allocating thread | |||
| 4478 | // task_src: pointer to source task to be duplicated | |||
| 4479 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
| 4480 | kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) { | |||
| 4481 | kmp_task_t *task; | |||
| 4482 | kmp_taskdata_t *taskdata; | |||
| 4483 | kmp_taskdata_t *taskdata_src = KMP_TASK_TO_TASKDATA(task_src)(((kmp_taskdata_t *)task_src) - 1); | |||
| 4484 | kmp_taskdata_t *parent_task = taskdata_src->td_parent; // same parent task | |||
| 4485 | size_t shareds_offset; | |||
| 4486 | size_t task_size; | |||
| 4487 | ||||
| 4488 | 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); } | |||
| 4489 | task_src))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n" , thread, task_src); }; | |||
| 4490 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==if (!(taskdata_src->td_flags.proxy == 0)) { __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "openmp/runtime/src/kmp_tasking.cpp" , 4491); } | |||
| 4491 | TASK_FULL)if (!(taskdata_src->td_flags.proxy == 0)) { __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "openmp/runtime/src/kmp_tasking.cpp" , 4491); }; // it should not be proxy task | |||
| 4492 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT)if (!(taskdata_src->td_flags.tasktype == 1)) { __kmp_debug_assert ("taskdata_src->td_flags.tasktype == 1", "openmp/runtime/src/kmp_tasking.cpp" , 4492); }; | |||
| 4493 | task_size = taskdata_src->td_size_alloc; | |||
| 4494 | ||||
| 4495 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
| 4496 | 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); } | |||
| 4497 | task_size))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n" , thread, task_size); }; | |||
| 4498 | #if USE_FAST_MEMORY3 | |||
| 4499 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size)___kmp_fast_allocate((thread), (task_size), "openmp/runtime/src/kmp_tasking.cpp" , 4499); | |||
| 4500 | #else | |||
| 4501 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size)___kmp_thread_malloc((thread), (task_size), "openmp/runtime/src/kmp_tasking.cpp" , 4501); | |||
| 4502 | #endif /* USE_FAST_MEMORY */ | |||
| 4503 | KMP_MEMCPYmemcpy(taskdata, taskdata_src, task_size); | |||
| 4504 | ||||
| 4505 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 4506 | ||||
| 4507 | // Initialize new task (only specific fields not affected by memcpy) | |||
| 4508 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 4509 | if (task->shareds != NULL__null) { // need setup shareds pointer | |||
| 4510 | shareds_offset = (char *)task_src->shareds - (char *)taskdata_src; | |||
| 4511 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
| 4512 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==if (!((((kmp_uintptr_t)task->shareds) & (sizeof(void * ) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 4513); } | |||
| 4513 | 0)if (!((((kmp_uintptr_t)task->shareds) & (sizeof(void * ) - 1)) == 0)) { __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "openmp/runtime/src/kmp_tasking.cpp", 4513); }; | |||
| 4514 | } | |||
| 4515 | taskdata->td_alloc_thread = thread; | |||
| 4516 | taskdata->td_parent = parent_task; | |||
| 4517 | // task inherits the taskgroup from the parent task | |||
| 4518 | taskdata->td_taskgroup = parent_task->td_taskgroup; | |||
| 4519 | // tied task needs to initialize the td_last_tied at creation, | |||
| 4520 | // untied one does this when it is scheduled for execution | |||
| 4521 | if (taskdata->td_flags.tiedness == TASK_TIED1) | |||
| 4522 | taskdata->td_last_tied = taskdata; | |||
| 4523 | ||||
| 4524 | // Only need to keep track of child task counts if team parallel and tasking | |||
| 4525 | // not serialized | |||
| 4526 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) { | |||
| 4527 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
| 4528 | if (parent_task->td_taskgroup) | |||
| 4529 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
| 4530 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
| 4531 | // implicit not deallocated | |||
| 4532 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) | |||
| 4533 | 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); | |||
| 4534 | } | |||
| 4535 | ||||
| 4536 | 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); } | |||
| 4537 | ("__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); } | |||
| 4538 | 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); }; | |||
| 4539 | #if OMPT_SUPPORT1 | |||
| 4540 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 4541 | __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid); | |||
| 4542 | #endif | |||
| 4543 | return task; | |||
| 4544 | } | |||
| 4545 | ||||
| 4546 | // Routine optionally generated by the compiler for setting the lastprivate flag | |||
| 4547 | // and calling needed constructors for private/firstprivate objects | |||
| 4548 | // (used to form taskloop tasks from pattern task) | |||
| 4549 | // Parameters: dest task, src task, lastprivate flag. | |||
| 4550 | typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); | |||
| 4551 | ||||
| 4552 | KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8)static_assert(sizeof(long) == 4 || sizeof(long) == 8, "Build condition error" ); | |||
| 4553 | ||||
| 4554 | // class to encapsulate manipulating loop bounds in a taskloop task. | |||
| 4555 | // this abstracts away the Intel vs GOMP taskloop interface for setting/getting | |||
| 4556 | // the loop bound variables. | |||
| 4557 | class kmp_taskloop_bounds_t { | |||
| 4558 | kmp_task_t *task; | |||
| 4559 | const kmp_taskdata_t *taskdata; | |||
| 4560 | size_t lower_offset; | |||
| 4561 | size_t upper_offset; | |||
| 4562 | ||||
| 4563 | public: | |||
| 4564 | kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub) | |||
| 4565 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1)), | |||
| 4566 | lower_offset((char *)lb - (char *)task), | |||
| 4567 | upper_offset((char *)ub - (char *)task) { | |||
| 4568 | KMP_DEBUG_ASSERT((char *)lb > (char *)_task)if (!((char *)lb > (char *)_task)) { __kmp_debug_assert("(char *)lb > (char *)_task" , "openmp/runtime/src/kmp_tasking.cpp", 4568); }; | |||
| 4569 | KMP_DEBUG_ASSERT((char *)ub > (char *)_task)if (!((char *)ub > (char *)_task)) { __kmp_debug_assert("(char *)ub > (char *)_task" , "openmp/runtime/src/kmp_tasking.cpp", 4569); }; | |||
| 4570 | } | |||
| 4571 | kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds) | |||
| 4572 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)(((kmp_taskdata_t *)_task) - 1)), | |||
| 4573 | lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {} | |||
| 4574 | size_t get_lower_offset() const { return lower_offset; } | |||
| 4575 | size_t get_upper_offset() const { return upper_offset; } | |||
| 4576 | kmp_uint64 get_lb() const { | |||
| 4577 | kmp_int64 retval; | |||
| 4578 | #if defined(KMP_GOMP_COMPAT) | |||
| 4579 | // Intel task just returns the lower bound normally | |||
| 4580 | if (!taskdata->td_flags.native) { | |||
| 4581 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
| 4582 | } else { | |||
| 4583 | // GOMP task has to take into account the sizeof(long) | |||
| 4584 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 4585 | kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds); | |||
| 4586 | retval = (kmp_int64)*lb; | |||
| 4587 | } else { | |||
| 4588 | kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds); | |||
| 4589 | retval = (kmp_int64)*lb; | |||
| 4590 | } | |||
| 4591 | } | |||
| 4592 | #else | |||
| 4593 | (void)taskdata; | |||
| 4594 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
| 4595 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 4596 | return retval; | |||
| 4597 | } | |||
| 4598 | kmp_uint64 get_ub() const { | |||
| 4599 | kmp_int64 retval; | |||
| 4600 | #if defined(KMP_GOMP_COMPAT) | |||
| 4601 | // Intel task just returns the upper bound normally | |||
| 4602 | if (!taskdata->td_flags.native) { | |||
| 4603 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
| 4604 | } else { | |||
| 4605 | // GOMP task has to take into account the sizeof(long) | |||
| 4606 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 4607 | kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds) + 1; | |||
| 4608 | retval = (kmp_int64)*ub; | |||
| 4609 | } else { | |||
| 4610 | kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds) + 1; | |||
| 4611 | retval = (kmp_int64)*ub; | |||
| 4612 | } | |||
| 4613 | } | |||
| 4614 | #else | |||
| 4615 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
| 4616 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 4617 | return retval; | |||
| 4618 | } | |||
| 4619 | void set_lb(kmp_uint64 lb) { | |||
| 4620 | #if defined(KMP_GOMP_COMPAT) | |||
| 4621 | // Intel task just sets the lower bound normally | |||
| 4622 | if (!taskdata->td_flags.native) { | |||
| 4623 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
| 4624 | } else { | |||
| 4625 | // GOMP task has to take into account the sizeof(long) | |||
| 4626 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 4627 | kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds); | |||
| 4628 | *lower = (kmp_uint32)lb; | |||
| 4629 | } else { | |||
| 4630 | kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds); | |||
| 4631 | *lower = (kmp_uint64)lb; | |||
| 4632 | } | |||
| 4633 | } | |||
| 4634 | #else | |||
| 4635 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
| 4636 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 4637 | } | |||
| 4638 | void set_ub(kmp_uint64 ub) { | |||
| 4639 | #if defined(KMP_GOMP_COMPAT) | |||
| 4640 | // Intel task just sets the upper bound normally | |||
| 4641 | if (!taskdata->td_flags.native) { | |||
| 4642 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
| 4643 | } else { | |||
| 4644 | // GOMP task has to take into account the sizeof(long) | |||
| 4645 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 4646 | kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds) + 1; | |||
| 4647 | *upper = (kmp_uint32)ub; | |||
| 4648 | } else { | |||
| 4649 | kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds) + 1; | |||
| 4650 | *upper = (kmp_uint64)ub; | |||
| 4651 | } | |||
| 4652 | } | |||
| 4653 | #else | |||
| 4654 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
| 4655 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 4656 | } | |||
| 4657 | }; | |||
| 4658 | ||||
| 4659 | // __kmp_taskloop_linear: Start tasks of the taskloop linearly | |||
| 4660 | // | |||
| 4661 | // loc Source location information | |||
| 4662 | // gtid Global thread ID | |||
| 4663 | // task Pattern task, exposes the loop iteration range | |||
| 4664 | // lb Pointer to loop lower bound in task structure | |||
| 4665 | // ub Pointer to loop upper bound in task structure | |||
| 4666 | // st Loop stride | |||
| 4667 | // ub_glob Global upper bound (used for lastprivate check) | |||
| 4668 | // num_tasks Number of tasks to execute | |||
| 4669 | // grainsize Number of loop iterations per task | |||
| 4670 | // extras Number of chunks with grainsize+1 iterations | |||
| 4671 | // last_chunk Reduction of grainsize for last task | |||
| 4672 | // tc Iterations count | |||
| 4673 | // task_dup Tasks duplication routine | |||
| 4674 | // codeptr_ra Return address for OMPT events | |||
| 4675 | void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task, | |||
| 4676 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 4677 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
| 4678 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
| 4679 | kmp_int64 last_chunk, kmp_uint64 tc, | |||
| 4680 | #if OMPT_SUPPORT1 | |||
| 4681 | void *codeptr_ra, | |||
| 4682 | #endif | |||
| 4683 | void *task_dup) { | |||
| 4684 | KMP_COUNT_BLOCK(OMP_TASKLOOP)((void)0); | |||
| 4685 | KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling)((void)0); | |||
| 4686 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 4687 | // compiler provides global bounds here | |||
| 4688 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
| 4689 | kmp_uint64 lower = task_bounds.get_lb(); | |||
| 4690 | kmp_uint64 upper = task_bounds.get_ub(); | |||
| 4691 | kmp_uint64 i; | |||
| 4692 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 4693 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 4694 | kmp_task_t *next_task; | |||
| 4695 | kmp_int32 lastpriv = 0; | |||
| 4696 | ||||
| 4697 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 4698); } | |||
| 4698 | (last_chunk < 0 ? last_chunk : extras))if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 4698); }; | |||
| 4699 | KMP_DEBUG_ASSERT(num_tasks > extras)if (!(num_tasks > extras)) { __kmp_debug_assert("num_tasks > extras" , "openmp/runtime/src/kmp_tasking.cpp", 4699); }; | |||
| 4700 | KMP_DEBUG_ASSERT(num_tasks > 0)if (!(num_tasks > 0)) { __kmp_debug_assert("num_tasks > 0" , "openmp/runtime/src/kmp_tasking.cpp", 4700); }; | |||
| 4701 | 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, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n" , gtid, num_tasks, grainsize, extras, last_chunk, lower, upper , ub_glob, st, task_dup); } | |||
| 4702 | "extras %lld, last_chunk %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, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n" , gtid, num_tasks, grainsize, extras, last_chunk, lower, upper , ub_glob, st, task_dup); } | |||
| 4703 | gtid, num_tasks, grainsize, extras, last_chunk, lower, upper,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n" , gtid, num_tasks, grainsize, extras, last_chunk, lower, upper , ub_glob, st, task_dup); } | |||
| 4704 | ub_glob, st, task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n" , gtid, num_tasks, grainsize, extras, last_chunk, lower, upper , ub_glob, st, task_dup); }; | |||
| 4705 | ||||
| 4706 | // Launch num_tasks tasks, assign grainsize iterations each task | |||
| 4707 | for (i = 0; i < num_tasks; ++i) { | |||
| 4708 | kmp_uint64 chunk_minus_1; | |||
| 4709 | if (extras == 0) { | |||
| 4710 | chunk_minus_1 = grainsize - 1; | |||
| 4711 | } else { | |||
| 4712 | chunk_minus_1 = grainsize; | |||
| 4713 | --extras; // first extras iterations get bigger chunk (grainsize+1) | |||
| 4714 | } | |||
| 4715 | upper = lower + st * chunk_minus_1; | |||
| 4716 | if (upper > *ub) { | |||
| 4717 | upper = *ub; | |||
| 4718 | } | |||
| 4719 | if (i == num_tasks - 1) { | |||
| 4720 | // schedule the last task, set lastprivate flag if needed | |||
| 4721 | if (st == 1) { // most common case | |||
| 4722 | KMP_DEBUG_ASSERT(upper == *ub)if (!(upper == *ub)) { __kmp_debug_assert("upper == *ub", "openmp/runtime/src/kmp_tasking.cpp" , 4722); }; | |||
| 4723 | if (upper == ub_glob) | |||
| 4724 | lastpriv = 1; | |||
| 4725 | } else if (st > 0) { // positive loop stride | |||
| 4726 | KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper)if (!((kmp_uint64)st > *ub - upper)) { __kmp_debug_assert( "(kmp_uint64)st > *ub - upper", "openmp/runtime/src/kmp_tasking.cpp" , 4726); }; | |||
| 4727 | if ((kmp_uint64)st > ub_glob - upper) | |||
| 4728 | lastpriv = 1; | |||
| 4729 | } else { // negative loop stride | |||
| 4730 | KMP_DEBUG_ASSERT(upper + st < *ub)if (!(upper + st < *ub)) { __kmp_debug_assert("upper + st < *ub" , "openmp/runtime/src/kmp_tasking.cpp", 4730); }; | |||
| 4731 | if (upper - ub_glob < (kmp_uint64)(-st)) | |||
| 4732 | lastpriv = 1; | |||
| 4733 | } | |||
| 4734 | } | |||
| 4735 | next_task = __kmp_task_dup_alloc(thread, task); // allocate new task | |||
| 4736 | kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task)(((kmp_taskdata_t *)next_task) - 1); | |||
| 4737 | kmp_taskloop_bounds_t next_task_bounds = | |||
| 4738 | kmp_taskloop_bounds_t(next_task, task_bounds); | |||
| 4739 | ||||
| 4740 | // adjust task-specific bounds | |||
| 4741 | next_task_bounds.set_lb(lower); | |||
| 4742 | if (next_taskdata->td_flags.native) { | |||
| 4743 | next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1)); | |||
| 4744 | } else { | |||
| 4745 | next_task_bounds.set_ub(upper); | |||
| 4746 | } | |||
| 4747 | if (ptask_dup != NULL__null) // set lastprivate flag, construct firstprivates, | |||
| 4748 | // etc. | |||
| 4749 | ptask_dup(next_task, task, lastpriv); | |||
| 4750 | 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()); } | |||
| 4751 | ("__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()); } | |||
| 4752 | "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()); } | |||
| 4753 | 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()); } | |||
| 4754 | 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()); } | |||
| 4755 | 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()); }; | |||
| 4756 | #if OMPT_SUPPORT1 | |||
| 4757 | __kmp_omp_taskloop_task(NULL__null, gtid, next_task, | |||
| 4758 | codeptr_ra); // schedule new task | |||
| 4759 | #if OMPT_OPTIONAL1 | |||
| 4760 | if (ompt_enabled.ompt_callback_dispatch) { | |||
| 4761 | OMPT_GET_DISPATCH_CHUNK(next_taskdata->ompt_task_info.dispatch_chunk,do { if (st > 0) { next_taskdata->ompt_task_info.dispatch_chunk .start = static_cast<uint64_t>(lower); next_taskdata-> ompt_task_info.dispatch_chunk.iterations = static_cast<uint64_t >(((upper) - (lower)) / (st) + 1); } else { next_taskdata-> ompt_task_info.dispatch_chunk.start = static_cast<uint64_t >(upper); next_taskdata->ompt_task_info.dispatch_chunk. iterations = static_cast<uint64_t>(((lower) - (upper)) / -(st) + 1); } } while (0) | |||
| 4762 | lower, upper, st)do { if (st > 0) { next_taskdata->ompt_task_info.dispatch_chunk .start = static_cast<uint64_t>(lower); next_taskdata-> ompt_task_info.dispatch_chunk.iterations = static_cast<uint64_t >(((upper) - (lower)) / (st) + 1); } else { next_taskdata-> ompt_task_info.dispatch_chunk.start = static_cast<uint64_t >(upper); next_taskdata->ompt_task_info.dispatch_chunk. iterations = static_cast<uint64_t>(((lower) - (upper)) / -(st) + 1); } } while (0); | |||
| 4763 | } | |||
| 4764 | #endif // OMPT_OPTIONAL | |||
| 4765 | #else | |||
| 4766 | __kmp_omp_task(gtid, next_task, true); // schedule new task | |||
| 4767 | #endif | |||
| 4768 | lower = upper + st; // adjust lower bound for the next iteration | |||
| 4769 | } | |||
| 4770 | // free the pattern task and exit | |||
| 4771 | __kmp_task_start(gtid, task, current_task); // make internal bookkeeping | |||
| 4772 | // do not execute the pattern task, just do internal bookkeeping | |||
| 4773 | __kmp_task_finish<false>(gtid, task, current_task); | |||
| 4774 | } | |||
| 4775 | ||||
| 4776 | // Structure to keep taskloop parameters for auxiliary task | |||
| 4777 | // kept in the shareds of the task structure. | |||
| 4778 | typedef struct __taskloop_params { | |||
| 4779 | kmp_task_t *task; | |||
| 4780 | kmp_uint64 *lb; | |||
| 4781 | kmp_uint64 *ub; | |||
| 4782 | void *task_dup; | |||
| 4783 | kmp_int64 st; | |||
| 4784 | kmp_uint64 ub_glob; | |||
| 4785 | kmp_uint64 num_tasks; | |||
| 4786 | kmp_uint64 grainsize; | |||
| 4787 | kmp_uint64 extras; | |||
| 4788 | kmp_int64 last_chunk; | |||
| 4789 | kmp_uint64 tc; | |||
| 4790 | kmp_uint64 num_t_min; | |||
| 4791 | #if OMPT_SUPPORT1 | |||
| 4792 | void *codeptr_ra; | |||
| 4793 | #endif | |||
| 4794 | } __taskloop_params_t; | |||
| 4795 | ||||
| 4796 | void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *, | |||
| 4797 | kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64, | |||
| 4798 | kmp_uint64, kmp_uint64, kmp_int64, kmp_uint64, | |||
| 4799 | kmp_uint64, | |||
| 4800 | #if OMPT_SUPPORT1 | |||
| 4801 | void *, | |||
| 4802 | #endif | |||
| 4803 | void *); | |||
| 4804 | ||||
| 4805 | // Execute part of the taskloop submitted as a task. | |||
| 4806 | int __kmp_taskloop_task(int gtid, void *ptask) { | |||
| 4807 | __taskloop_params_t *p = | |||
| 4808 | (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds; | |||
| 4809 | kmp_task_t *task = p->task; | |||
| 4810 | kmp_uint64 *lb = p->lb; | |||
| 4811 | kmp_uint64 *ub = p->ub; | |||
| 4812 | void *task_dup = p->task_dup; | |||
| 4813 | // p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 4814 | kmp_int64 st = p->st; | |||
| 4815 | kmp_uint64 ub_glob = p->ub_glob; | |||
| 4816 | kmp_uint64 num_tasks = p->num_tasks; | |||
| 4817 | kmp_uint64 grainsize = p->grainsize; | |||
| 4818 | kmp_uint64 extras = p->extras; | |||
| 4819 | kmp_int64 last_chunk = p->last_chunk; | |||
| 4820 | kmp_uint64 tc = p->tc; | |||
| 4821 | kmp_uint64 num_t_min = p->num_t_min; | |||
| 4822 | #if OMPT_SUPPORT1 | |||
| 4823 | void *codeptr_ra = p->codeptr_ra; | |||
| 4824 | #endif | |||
| 4825 | #if KMP_DEBUG1 | |||
| 4826 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 4827 | KMP_DEBUG_ASSERT(task != NULL)if (!(task != __null)) { __kmp_debug_assert("task != __null", "openmp/runtime/src/kmp_tasking.cpp", 4827); }; | |||
| 4828 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4829 | ("__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, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4830 | " %lld, extras %lld, last_chunk %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, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4831 | gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4832 | st, task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); }; | |||
| 4833 | #endif | |||
| 4834 | KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min)if (!(num_tasks * 2 + 1 > num_t_min)) { __kmp_debug_assert ("num_tasks * 2 + 1 > num_t_min", "openmp/runtime/src/kmp_tasking.cpp" , 4834); }; | |||
| 4835 | if (num_tasks > num_t_min) | |||
| 4836 | __kmp_taskloop_recur(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 4837 | grainsize, extras, last_chunk, tc, num_t_min, | |||
| 4838 | #if OMPT_SUPPORT1 | |||
| 4839 | codeptr_ra, | |||
| 4840 | #endif | |||
| 4841 | task_dup); | |||
| 4842 | else | |||
| 4843 | __kmp_taskloop_linear(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 4844 | grainsize, extras, last_chunk, tc, | |||
| 4845 | #if OMPT_SUPPORT1 | |||
| 4846 | codeptr_ra, | |||
| 4847 | #endif | |||
| 4848 | task_dup); | |||
| 4849 | ||||
| 4850 | 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); }; | |||
| 4851 | return 0; | |||
| 4852 | } | |||
| 4853 | ||||
| 4854 | // Schedule part of the taskloop as a task, | |||
| 4855 | // execute the rest of the taskloop. | |||
| 4856 | // | |||
| 4857 | // loc Source location information | |||
| 4858 | // gtid Global thread ID | |||
| 4859 | // task Pattern task, exposes the loop iteration range | |||
| 4860 | // lb Pointer to loop lower bound in task structure | |||
| 4861 | // ub Pointer to loop upper bound in task structure | |||
| 4862 | // st Loop stride | |||
| 4863 | // ub_glob Global upper bound (used for lastprivate check) | |||
| 4864 | // num_tasks Number of tasks to execute | |||
| 4865 | // grainsize Number of loop iterations per task | |||
| 4866 | // extras Number of chunks with grainsize+1 iterations | |||
| 4867 | // last_chunk Reduction of grainsize for last task | |||
| 4868 | // tc Iterations count | |||
| 4869 | // num_t_min Threshold to launch tasks recursively | |||
| 4870 | // task_dup Tasks duplication routine | |||
| 4871 | // codeptr_ra Return address for OMPT events | |||
| 4872 | void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task, | |||
| 4873 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 4874 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
| 4875 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
| 4876 | kmp_int64 last_chunk, kmp_uint64 tc, | |||
| 4877 | kmp_uint64 num_t_min, | |||
| 4878 | #if OMPT_SUPPORT1 | |||
| 4879 | void *codeptr_ra, | |||
| 4880 | #endif | |||
| 4881 | void *task_dup) { | |||
| 4882 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 4883 | KMP_DEBUG_ASSERT(task != NULL)if (!(task != __null)) { __kmp_debug_assert("task != __null", "openmp/runtime/src/kmp_tasking.cpp", 4883); }; | |||
| 4884 | KMP_DEBUG_ASSERT(num_tasks > num_t_min)if (!(num_tasks > num_t_min)) { __kmp_debug_assert("num_tasks > num_t_min" , "openmp/runtime/src/kmp_tasking.cpp", 4884); }; | |||
| 4885 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4886 | ("__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, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4887 | " %lld, extras %lld, last_chunk %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, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4888 | gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); } | |||
| 4889 | st, task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n" , gtid, taskdata, num_tasks, grainsize, extras, last_chunk, * lb, *ub, st, task_dup); }; | |||
| 4890 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 4891 | kmp_uint64 lower = *lb; | |||
| 4892 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 4893 | // kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 4894 | kmp_task_t *next_task; | |||
| 4895 | size_t lower_offset = | |||
| 4896 | (char *)lb - (char *)task; // remember offset of lb in the task structure | |||
| 4897 | size_t upper_offset = | |||
| 4898 | (char *)ub - (char *)task; // remember offset of ub in the task structure | |||
| 4899 | ||||
| 4900 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 4901); } | |||
| 4901 | (last_chunk < 0 ? last_chunk : extras))if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 4901); }; | |||
| 4902 | KMP_DEBUG_ASSERT(num_tasks > extras)if (!(num_tasks > extras)) { __kmp_debug_assert("num_tasks > extras" , "openmp/runtime/src/kmp_tasking.cpp", 4902); }; | |||
| 4903 | KMP_DEBUG_ASSERT(num_tasks > 0)if (!(num_tasks > 0)) { __kmp_debug_assert("num_tasks > 0" , "openmp/runtime/src/kmp_tasking.cpp", 4903); }; | |||
| 4904 | ||||
| 4905 | // split the loop in two halves | |||
| 4906 | kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1; | |||
| 4907 | kmp_int64 last_chunk0 = 0, last_chunk1 = 0; | |||
| 4908 | kmp_uint64 gr_size0 = grainsize; | |||
| 4909 | kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute | |||
| 4910 | kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task | |||
| 4911 | if (last_chunk < 0) { | |||
| 4912 | ext0 = ext1 = 0; | |||
| 4913 | last_chunk1 = last_chunk; | |||
| 4914 | tc0 = grainsize * n_tsk0; | |||
| 4915 | tc1 = tc - tc0; | |||
| 4916 | } else if (n_tsk0 <= extras) { | |||
| 4917 | gr_size0++; // integrate extras into grainsize | |||
| 4918 | ext0 = 0; // no extra iters in 1st half | |||
| 4919 | ext1 = extras - n_tsk0; // remaining extras | |||
| 4920 | tc0 = gr_size0 * n_tsk0; | |||
| 4921 | tc1 = tc - tc0; | |||
| 4922 | } else { // n_tsk0 > extras | |||
| 4923 | ext1 = 0; // no extra iters in 2nd half | |||
| 4924 | ext0 = extras; | |||
| 4925 | tc1 = grainsize * n_tsk1; | |||
| 4926 | tc0 = tc - tc1; | |||
| 4927 | } | |||
| 4928 | ub0 = lower + st * (tc0 - 1); | |||
| 4929 | lb1 = ub0 + st; | |||
| 4930 | ||||
| 4931 | // create pattern task for 2nd half of the loop | |||
| 4932 | next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task | |||
| 4933 | // adjust lower bound (upper bound is not changed) for the 2nd half | |||
| 4934 | *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1; | |||
| 4935 | if (ptask_dup != NULL__null) // construct firstprivates, etc. | |||
| 4936 | ptask_dup(next_task, task, 0); | |||
| 4937 | *ub = ub0; // adjust upper bound for the 1st half | |||
| 4938 | ||||
| 4939 | // create auxiliary task for 2nd half of the loop | |||
| 4940 | // make sure new task has same parent task as the pattern task | |||
| 4941 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 4942 | thread->th.th_current_task = taskdata->td_parent; | |||
| 4943 | kmp_task_t *new_task = | |||
| 4944 | __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *), | |||
| 4945 | sizeof(__taskloop_params_t), &__kmp_taskloop_task); | |||
| 4946 | // restore current task | |||
| 4947 | thread->th.th_current_task = current_task; | |||
| 4948 | __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds; | |||
| 4949 | p->task = next_task; | |||
| 4950 | p->lb = (kmp_uint64 *)((char *)next_task + lower_offset); | |||
| 4951 | p->ub = (kmp_uint64 *)((char *)next_task + upper_offset); | |||
| 4952 | p->task_dup = task_dup; | |||
| 4953 | p->st = st; | |||
| 4954 | p->ub_glob = ub_glob; | |||
| 4955 | p->num_tasks = n_tsk1; | |||
| 4956 | p->grainsize = grainsize; | |||
| 4957 | p->extras = ext1; | |||
| 4958 | p->last_chunk = last_chunk1; | |||
| 4959 | p->tc = tc1; | |||
| 4960 | p->num_t_min = num_t_min; | |||
| 4961 | #if OMPT_SUPPORT1 | |||
| 4962 | p->codeptr_ra = codeptr_ra; | |||
| 4963 | #endif | |||
| 4964 | ||||
| 4965 | #if OMPT_SUPPORT1 | |||
| 4966 | // schedule new task with correct return address for OMPT events | |||
| 4967 | __kmp_omp_taskloop_task(NULL__null, gtid, new_task, codeptr_ra); | |||
| 4968 | #else | |||
| 4969 | __kmp_omp_task(gtid, new_task, true); // schedule new task | |||
| 4970 | #endif | |||
| 4971 | ||||
| 4972 | // execute the 1st half of current subrange | |||
| 4973 | if (n_tsk0 > num_t_min) | |||
| 4974 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0, | |||
| 4975 | ext0, last_chunk0, tc0, num_t_min, | |||
| 4976 | #if OMPT_SUPPORT1 | |||
| 4977 | codeptr_ra, | |||
| 4978 | #endif | |||
| 4979 | task_dup); | |||
| 4980 | else | |||
| 4981 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, | |||
| 4982 | gr_size0, ext0, last_chunk0, tc0, | |||
| 4983 | #if OMPT_SUPPORT1 | |||
| 4984 | codeptr_ra, | |||
| 4985 | #endif | |||
| 4986 | task_dup); | |||
| 4987 | ||||
| 4988 | KA_TRACE(40, ("__kmp_taskloop_recur(exit): T#%d\n", gtid))if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmp_taskloop_recur(exit): T#%d\n" , gtid); }; | |||
| 4989 | } | |||
| 4990 | ||||
| 4991 | static void __kmp_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, | |||
| 4992 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 4993 | int nogroup, int sched, kmp_uint64 grainsize, | |||
| 4994 | int modifier, void *task_dup) { | |||
| 4995 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 4996 | KMP_DEBUG_ASSERT(task != NULL)if (!(task != __null)) { __kmp_debug_assert("task != __null", "openmp/runtime/src/kmp_tasking.cpp", 4996); }; | |||
| 4997 | if (nogroup == 0) { | |||
| 4998 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4999 | OMPT_STORE_RETURN_ADDRESS(gtid)OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address (0)};; | |||
| 5000 | #endif | |||
| 5001 | __kmpc_taskgroup(loc, gtid); | |||
| 5002 | } | |||
| 5003 | ||||
| 5004 | // ========================================================================= | |||
| 5005 | // calculate loop parameters | |||
| 5006 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
| 5007 | kmp_uint64 tc; | |||
| 5008 | // compiler provides global bounds here | |||
| 5009 | kmp_uint64 lower = task_bounds.get_lb(); | |||
| 5010 | kmp_uint64 upper = task_bounds.get_ub(); | |||
| 5011 | kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag | |||
| 5012 | kmp_uint64 num_tasks = 0, extras = 0; | |||
| 5013 | kmp_int64 last_chunk = | |||
| 5014 | 0; // reduce grainsize of last task by last_chunk in strict mode | |||
| 5015 | kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks; | |||
| 5016 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 5017 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 5018 | ||||
| 5019 | KA_TRACE(20, ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d, %d), dup %p\n", gtid, taskdata, lower, upper , st, grainsize, sched, modifier, task_dup); } | |||
| 5020 | "grain %llu(%d, %d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d, %d), dup %p\n", gtid, taskdata, lower, upper , st, grainsize, sched, modifier, task_dup); } | |||
| 5021 | gtid, taskdata, lower, upper, st, grainsize, sched, modifier,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d, %d), dup %p\n", gtid, taskdata, lower, upper , st, grainsize, sched, modifier, task_dup); } | |||
| 5022 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d, %d), dup %p\n", gtid, taskdata, lower, upper , st, grainsize, sched, modifier, task_dup); }; | |||
| 5023 | ||||
| 5024 | // compute trip count | |||
| 5025 | if (st == 1) { // most common case | |||
| 5026 | tc = upper - lower + 1; | |||
| 5027 | } else if (st < 0) { | |||
| 5028 | tc = (lower - upper) / (-st) + 1; | |||
| 5029 | } else { // st > 0 | |||
| 5030 | tc = (upper - lower) / st + 1; | |||
| 5031 | } | |||
| 5032 | if (tc == 0) { | |||
| 5033 | KA_TRACE(20, ("__kmp_taskloop(exit): T#%d zero-trip loop\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop(exit): T#%d zero-trip loop\n" , gtid); }; | |||
| 5034 | // free the pattern task and exit | |||
| 5035 | __kmp_task_start(gtid, task, current_task); | |||
| 5036 | // do not execute anything for zero-trip loop | |||
| 5037 | __kmp_task_finish<false>(gtid, task, current_task); | |||
| 5038 | return; | |||
| 5039 | } | |||
| 5040 | ||||
| 5041 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 5042 | ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL__null); | |||
| 5043 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
| 5044 | if (ompt_enabled.ompt_callback_work) { | |||
| 5045 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
| 5046 | ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data), | |||
| 5047 | &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 5048 | } | |||
| 5049 | #endif | |||
| 5050 | ||||
| 5051 | if (num_tasks_min == 0) | |||
| 5052 | // TODO: can we choose better default heuristic? | |||
| 5053 | num_tasks_min = | |||
| 5054 | 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))); | |||
| 5055 | ||||
| 5056 | // compute num_tasks/grainsize based on the input provided | |||
| 5057 | switch (sched) { | |||
| 5058 | case 0: // no schedule clause specified, we can choose the default | |||
| 5059 | // let's try to schedule (team_size*10) tasks | |||
| 5060 | grainsize = thread->th.th_team_nproc * 10; | |||
| 5061 | KMP_FALLTHROUGH()[[fallthrough]]; | |||
| 5062 | case 2: // num_tasks provided | |||
| 5063 | if (grainsize > tc) { | |||
| 5064 | num_tasks = tc; // too big num_tasks requested, adjust values | |||
| 5065 | grainsize = 1; | |||
| 5066 | extras = 0; | |||
| 5067 | } else { | |||
| 5068 | num_tasks = grainsize; | |||
| 5069 | grainsize = tc / num_tasks; | |||
| 5070 | extras = tc % num_tasks; | |||
| 5071 | } | |||
| 5072 | break; | |||
| 5073 | case 1: // grainsize provided | |||
| 5074 | if (grainsize > tc) { | |||
| 5075 | num_tasks = 1; | |||
| 5076 | grainsize = tc; // too big grainsize requested, adjust values | |||
| 5077 | extras = 0; | |||
| 5078 | } else { | |||
| 5079 | if (modifier) { | |||
| 5080 | num_tasks = (tc + grainsize - 1) / grainsize; | |||
| 5081 | last_chunk = tc - (num_tasks * grainsize); | |||
| 5082 | extras = 0; | |||
| 5083 | } else { | |||
| 5084 | num_tasks = tc / grainsize; | |||
| 5085 | // adjust grainsize for balanced distribution of iterations | |||
| 5086 | grainsize = tc / num_tasks; | |||
| 5087 | extras = tc % num_tasks; | |||
| 5088 | } | |||
| 5089 | } | |||
| 5090 | break; | |||
| 5091 | default: | |||
| 5092 | KMP_ASSERT2(0, "unknown scheduling of taskloop")if (!(0)) { __kmp_debug_assert(("unknown scheduling of taskloop" ), "openmp/runtime/src/kmp_tasking.cpp", 5092); }; | |||
| 5093 | } | |||
| 5094 | ||||
| 5095 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 5096); } | |||
| 5096 | (last_chunk < 0 ? last_chunk : extras))if (!(tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras))) { __kmp_debug_assert("tc == num_tasks * grainsize + (last_chunk < 0 ? last_chunk : extras)" , "openmp/runtime/src/kmp_tasking.cpp", 5096); }; | |||
| 5097 | KMP_DEBUG_ASSERT(num_tasks > extras)if (!(num_tasks > extras)) { __kmp_debug_assert("num_tasks > extras" , "openmp/runtime/src/kmp_tasking.cpp", 5097); }; | |||
| 5098 | KMP_DEBUG_ASSERT(num_tasks > 0)if (!(num_tasks > 0)) { __kmp_debug_assert("num_tasks > 0" , "openmp/runtime/src/kmp_tasking.cpp", 5098); }; | |||
| 5099 | // ========================================================================= | |||
| 5100 | ||||
| 5101 | // check if clause value first | |||
| 5102 | // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native) | |||
| 5103 | if (if_val == 0) { // if(0) specified, mark task as serial | |||
| 5104 | taskdata->td_flags.task_serial = 1; | |||
| 5105 | taskdata->td_flags.tiedness = TASK_TIED1; // AC: serial task cannot be untied | |||
| 5106 | // always start serial tasks linearly | |||
| 5107 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 5108 | grainsize, extras, last_chunk, tc, | |||
| 5109 | #if OMPT_SUPPORT1 | |||
| 5110 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
| 5111 | #endif | |||
| 5112 | task_dup); | |||
| 5113 | // !taskdata->td_flags.native => currently force linear spawning of tasks | |||
| 5114 | // for GOMP_taskloop | |||
| 5115 | } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) { | |||
| 5116 | KA_TRACE(20, ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5117 | "(%lld), grain %llu, extras %llu, last_chunk %lld\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5118 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5119 | last_chunk))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); }; | |||
| 5120 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 5121 | grainsize, extras, last_chunk, tc, num_tasks_min, | |||
| 5122 | #if OMPT_SUPPORT1 | |||
| 5123 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
| 5124 | #endif | |||
| 5125 | task_dup); | |||
| 5126 | } else { | |||
| 5127 | KA_TRACE(20, ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5128 | "(%lld), grain %llu, extras %llu, last_chunk %lld\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5129 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); } | |||
| 5130 | last_chunk))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu, last_chunk %lld\n", gtid, tc , num_tasks, num_tasks_min, grainsize, extras, last_chunk); }; | |||
| 5131 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 5132 | grainsize, extras, last_chunk, tc, | |||
| 5133 | #if OMPT_SUPPORT1 | |||
| 5134 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0), | |||
| 5135 | #endif | |||
| 5136 | task_dup); | |||
| 5137 | } | |||
| 5138 | ||||
| 5139 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 5140 | if (ompt_enabled.ompt_callback_work) { | |||
| 5141 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
| 5142 | ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data), | |||
| 5143 | &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 5144 | } | |||
| 5145 | #endif | |||
| 5146 | ||||
| 5147 | if (nogroup == 0) { | |||
| 5148 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 5149 | OMPT_STORE_RETURN_ADDRESS(gtid)OmptReturnAddressGuard ReturnAddressGuard{gtid, __builtin_return_address (0)};; | |||
| 5150 | #endif | |||
| 5151 | __kmpc_end_taskgroup(loc, gtid); | |||
| 5152 | } | |||
| 5153 | KA_TRACE(20, ("__kmp_taskloop(exit): T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop(exit): T#%d\n" , gtid); }; | |||
| 5154 | } | |||
| 5155 | ||||
| 5156 | /*! | |||
| 5157 | @ingroup TASKING | |||
| 5158 | @param loc Source location information | |||
| 5159 | @param gtid Global thread ID | |||
| 5160 | @param task Task structure | |||
| 5161 | @param if_val Value of the if clause | |||
| 5162 | @param lb Pointer to loop lower bound in task structure | |||
| 5163 | @param ub Pointer to loop upper bound in task structure | |||
| 5164 | @param st Loop stride | |||
| 5165 | @param nogroup Flag, 1 if nogroup clause specified, 0 otherwise | |||
| 5166 | @param sched Schedule specified 0/1/2 for none/grainsize/num_tasks | |||
| 5167 | @param grainsize Schedule value if specified | |||
| 5168 | @param task_dup Tasks duplication routine | |||
| 5169 | ||||
| 5170 | Execute the taskloop construct. | |||
| 5171 | */ | |||
| 5172 | void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, | |||
| 5173 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, | |||
| 5174 | int sched, kmp_uint64 grainsize, void *task_dup) { | |||
| 5175 | __kmp_assert_valid_gtid(gtid); | |||
| 5176 | KA_TRACE(20, ("__kmpc_taskloop(enter): T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop(enter): T#%d\n" , gtid); }; | |||
| 5177 | __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize, | |||
| 5178 | 0, task_dup); | |||
| 5179 | 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); }; | |||
| 5180 | } | |||
| 5181 | ||||
| 5182 | /*! | |||
| 5183 | @ingroup TASKING | |||
| 5184 | @param loc Source location information | |||
| 5185 | @param gtid Global thread ID | |||
| 5186 | @param task Task structure | |||
| 5187 | @param if_val Value of the if clause | |||
| 5188 | @param lb Pointer to loop lower bound in task structure | |||
| 5189 | @param ub Pointer to loop upper bound in task structure | |||
| 5190 | @param st Loop stride | |||
| 5191 | @param nogroup Flag, 1 if nogroup clause specified, 0 otherwise | |||
| 5192 | @param sched Schedule specified 0/1/2 for none/grainsize/num_tasks | |||
| 5193 | @param grainsize Schedule value if specified | |||
| 5194 | @param modifier Modifier 'strict' for sched, 1 if present, 0 otherwise | |||
| 5195 | @param task_dup Tasks duplication routine | |||
| 5196 | ||||
| 5197 | Execute the taskloop construct. | |||
| 5198 | */ | |||
| 5199 | void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val, | |||
| 5200 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 5201 | int nogroup, int sched, kmp_uint64 grainsize, | |||
| 5202 | int modifier, void *task_dup) { | |||
| 5203 | __kmp_assert_valid_gtid(gtid); | |||
| 5204 | KA_TRACE(20, ("__kmpc_taskloop_5(enter): T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop_5(enter): T#%d\n" , gtid); }; | |||
| 5205 | __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize, | |||
| 5206 | modifier, task_dup); | |||
| 5207 | KA_TRACE(20, ("__kmpc_taskloop_5(exit): T#%d\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop_5(exit): T#%d\n" , gtid); }; | |||
| 5208 | } | |||
| 5209 | ||||
| 5210 | /*! | |||
| 5211 | @ingroup TASKING | |||
| 5212 | @param gtid Global Thread ID of current thread | |||
| 5213 | @return Returns a pointer to the thread's current task async handle. If no task | |||
| 5214 | is present or gtid is invalid, returns NULL. | |||
| 5215 | ||||
| 5216 | Acqurires a pointer to the target async handle from the current task. | |||
| 5217 | */ | |||
| 5218 | void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid) { | |||
| 5219 | if (gtid == KMP_GTID_DNE(-2)) | |||
| 5220 | return NULL__null; | |||
| 5221 | ||||
| 5222 | kmp_info_t *thread = __kmp_thread_from_gtid(gtid); | |||
| 5223 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 5224 | ||||
| 5225 | if (!taskdata) | |||
| 5226 | return NULL__null; | |||
| 5227 | ||||
| 5228 | return &taskdata->td_target_data.async_handle; | |||
| 5229 | } | |||
| 5230 | ||||
| 5231 | /*! | |||
| 5232 | @ingroup TASKING | |||
| 5233 | @param gtid Global Thread ID of current thread | |||
| 5234 | @return Returns TRUE if the current task being executed of the given thread has | |||
| 5235 | a task team allocated to it. Otherwise, returns FALSE. | |||
| 5236 | ||||
| 5237 | Checks if the current thread has a task team. | |||
| 5238 | */ | |||
| 5239 | bool __kmpc_omp_has_task_team(kmp_int32 gtid) { | |||
| 5240 | if (gtid == KMP_GTID_DNE(-2)) | |||
| 5241 | return FALSE0; | |||
| 5242 | ||||
| 5243 | kmp_info_t *thread = __kmp_thread_from_gtid(gtid); | |||
| 5244 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 5245 | ||||
| 5246 | if (!taskdata) | |||
| 5247 | return FALSE0; | |||
| 5248 | ||||
| 5249 | return taskdata->td_task_team != NULL__null; | |||
| 5250 | } |
| 1 | /* |
| 2 | * kmp_lock.h -- lock header file |
| 3 | */ |
| 4 | |
| 5 | //===----------------------------------------------------------------------===// |
| 6 | // |
| 7 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 8 | // See https://llvm.org/LICENSE.txt for license information. |
| 9 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef KMP_LOCK_H |
| 14 | #define KMP_LOCK_H |
| 15 | |
| 16 | #include <limits.h> // CHAR_BIT |
| 17 | #include <stddef.h> // offsetof |
| 18 | |
| 19 | #include "kmp_debug.h" |
| 20 | #include "kmp_os.h" |
| 21 | |
| 22 | #ifdef __cplusplus201703L |
| 23 | #include <atomic> |
| 24 | |
| 25 | extern "C" { |
| 26 | #endif // __cplusplus |
| 27 | |
| 28 | // ---------------------------------------------------------------------------- |
| 29 | // Have to copy these definitions from kmp.h because kmp.h cannot be included |
| 30 | // due to circular dependencies. Will undef these at end of file. |
| 31 | |
| 32 | #define KMP_PAD(type, sz)(sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) \ |
| 33 | (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) |
| 34 | #define KMP_GTID_DNE(-2) (-2) |
| 35 | |
| 36 | // Forward declaration of ident and ident_t |
| 37 | |
| 38 | struct ident; |
| 39 | typedef struct ident ident_t; |
| 40 | |
| 41 | // End of copied code. |
| 42 | // ---------------------------------------------------------------------------- |
| 43 | |
| 44 | // We need to know the size of the area we can assume that the compiler(s) |
| 45 | // allocated for objects of type omp_lock_t and omp_nest_lock_t. The Intel |
| 46 | // compiler always allocates a pointer-sized area, as does visual studio. |
| 47 | // |
| 48 | // gcc however, only allocates 4 bytes for regular locks, even on 64-bit |
| 49 | // intel archs. It allocates at least 8 bytes for nested lock (more on |
| 50 | // recent versions), but we are bounded by the pointer-sized chunks that |
| 51 | // the Intel compiler allocates. |
| 52 | |
| 53 | #if KMP_OS_LINUX1 && defined(KMP_GOMP_COMPAT) |
| 54 | #define OMP_LOCK_T_SIZEsizeof(int) sizeof(int) |
| 55 | #define OMP_NEST_LOCK_T_SIZEsizeof(void *) sizeof(void *) |
| 56 | #else |
| 57 | #define OMP_LOCK_T_SIZEsizeof(int) sizeof(void *) |
| 58 | #define OMP_NEST_LOCK_T_SIZEsizeof(void *) sizeof(void *) |
| 59 | #endif |
| 60 | |
| 61 | // The Intel compiler allocates a 32-byte chunk for a critical section. |
| 62 | // Both gcc and visual studio only allocate enough space for a pointer. |
| 63 | // Sometimes we know that the space was allocated by the Intel compiler. |
| 64 | #define OMP_CRITICAL_SIZEsizeof(void *) sizeof(void *) |
| 65 | #define INTEL_CRITICAL_SIZE32 32 |
| 66 | |
| 67 | // lock flags |
| 68 | typedef kmp_uint32 kmp_lock_flags_t; |
| 69 | |
| 70 | #define kmp_lf_critical_section1 1 |
| 71 | |
| 72 | // When a lock table is used, the indices are of kmp_lock_index_t |
| 73 | typedef kmp_uint32 kmp_lock_index_t; |
| 74 | |
| 75 | // When memory allocated for locks are on the lock pool (free list), |
| 76 | // it is treated as structs of this type. |
| 77 | struct kmp_lock_pool { |
| 78 | union kmp_user_lock *next; |
| 79 | kmp_lock_index_t index; |
| 80 | }; |
| 81 | |
| 82 | typedef struct kmp_lock_pool kmp_lock_pool_t; |
| 83 | |
| 84 | extern void __kmp_validate_locks(void); |
| 85 | |
| 86 | // ---------------------------------------------------------------------------- |
| 87 | // There are 5 lock implementations: |
| 88 | // 1. Test and set locks. |
| 89 | // 2. futex locks (Linux* OS on x86 and |
| 90 | // Intel(R) Many Integrated Core Architecture) |
| 91 | // 3. Ticket (Lamport bakery) locks. |
| 92 | // 4. Queuing locks (with separate spin fields). |
| 93 | // 5. DRPA (Dynamically Reconfigurable Distributed Polling Area) locks |
| 94 | // |
| 95 | // and 3 lock purposes: |
| 96 | // 1. Bootstrap locks -- Used for a few locks available at library |
| 97 | // startup-shutdown time. |
| 98 | // These do not require non-negative global thread ID's. |
| 99 | // 2. Internal RTL locks -- Used everywhere else in the RTL |
| 100 | // 3. User locks (includes critical sections) |
| 101 | // ---------------------------------------------------------------------------- |
| 102 | |
| 103 | // ============================================================================ |
| 104 | // Lock implementations. |
| 105 | // |
| 106 | // Test and set locks. |
| 107 | // |
| 108 | // Non-nested test and set locks differ from the other lock kinds (except |
| 109 | // futex) in that we use the memory allocated by the compiler for the lock, |
| 110 | // rather than a pointer to it. |
| 111 | // |
| 112 | // On lin32, lin_32e, and win_32, the space allocated may be as small as 4 |
| 113 | // bytes, so we have to use a lock table for nested locks, and avoid accessing |
| 114 | // the depth_locked field for non-nested locks. |
| 115 | // |
| 116 | // Information normally available to the tools, such as lock location, lock |
| 117 | // usage (normal lock vs. critical section), etc. is not available with test and |
| 118 | // set locks. |
| 119 | // ---------------------------------------------------------------------------- |
| 120 | |
| 121 | struct kmp_base_tas_lock { |
| 122 | // KMP_LOCK_FREE(tas) => unlocked; locked: (gtid+1) of owning thread |
| 123 | std::atomic<kmp_int32> poll; |
| 124 | kmp_int32 depth_locked; // depth locked, for nested locks only |
| 125 | }; |
| 126 | |
| 127 | typedef struct kmp_base_tas_lock kmp_base_tas_lock_t; |
| 128 | |
| 129 | union kmp_tas_lock { |
| 130 | kmp_base_tas_lock_t lk; |
| 131 | kmp_lock_pool_t pool; // make certain struct is large enough |
| 132 | double lk_align; // use worst case alignment; no cache line padding |
| 133 | }; |
| 134 | |
| 135 | typedef union kmp_tas_lock kmp_tas_lock_t; |
| 136 | |
| 137 | // Static initializer for test and set lock variables. Usage: |
| 138 | // kmp_tas_lock_t xlock = KMP_TAS_LOCK_INITIALIZER( xlock ); |
| 139 | #define KMP_TAS_LOCK_INITIALIZER(lock){ { (locktag_tas), 0 } } \ |
| 140 | { \ |
| 141 | { KMP_LOCK_FREE(tas)(locktag_tas), 0 } \ |
| 142 | } |
| 143 | |
| 144 | extern int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 145 | extern int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 146 | extern int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 147 | extern void __kmp_init_tas_lock(kmp_tas_lock_t *lck); |
| 148 | extern void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck); |
| 149 | |
| 150 | extern int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 151 | extern int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 152 | extern int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid); |
| 153 | extern void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck); |
| 154 | extern void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck); |
| 155 | |
| 156 | #define KMP_LOCK_RELEASED1 1 |
| 157 | #define KMP_LOCK_STILL_HELD0 0 |
| 158 | #define KMP_LOCK_ACQUIRED_FIRST1 1 |
| 159 | #define KMP_LOCK_ACQUIRED_NEXT0 0 |
| 160 | #ifndef KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 161 | #define KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) \ |
| 162 | (KMP_OS_LINUX1 && \ |
| 163 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640)) |
| 164 | #endif |
| 165 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 166 | |
| 167 | // ---------------------------------------------------------------------------- |
| 168 | // futex locks. futex locks are only available on Linux* OS. |
| 169 | // |
| 170 | // Like non-nested test and set lock, non-nested futex locks use the memory |
| 171 | // allocated by the compiler for the lock, rather than a pointer to it. |
| 172 | // |
| 173 | // Information normally available to the tools, such as lock location, lock |
| 174 | // usage (normal lock vs. critical section), etc. is not available with test and |
| 175 | // set locks. With non-nested futex locks, the lock owner is not even available. |
| 176 | // ---------------------------------------------------------------------------- |
| 177 | |
| 178 | struct kmp_base_futex_lock { |
| 179 | volatile kmp_int32 poll; // KMP_LOCK_FREE(futex) => unlocked |
| 180 | // 2*(gtid+1) of owning thread, 0 if unlocked |
| 181 | // locked: (gtid+1) of owning thread |
| 182 | kmp_int32 depth_locked; // depth locked, for nested locks only |
| 183 | }; |
| 184 | |
| 185 | typedef struct kmp_base_futex_lock kmp_base_futex_lock_t; |
| 186 | |
| 187 | union kmp_futex_lock { |
| 188 | kmp_base_futex_lock_t lk; |
| 189 | kmp_lock_pool_t pool; // make certain struct is large enough |
| 190 | double lk_align; // use worst case alignment |
| 191 | // no cache line padding |
| 192 | }; |
| 193 | |
| 194 | typedef union kmp_futex_lock kmp_futex_lock_t; |
| 195 | |
| 196 | // Static initializer for futex lock variables. Usage: |
| 197 | // kmp_futex_lock_t xlock = KMP_FUTEX_LOCK_INITIALIZER( xlock ); |
| 198 | #define KMP_FUTEX_LOCK_INITIALIZER(lock){ { (locktag_futex), 0 } } \ |
| 199 | { \ |
| 200 | { KMP_LOCK_FREE(futex)(locktag_futex), 0 } \ |
| 201 | } |
| 202 | |
| 203 | extern int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid); |
| 204 | extern int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid); |
| 205 | extern int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid); |
| 206 | extern void __kmp_init_futex_lock(kmp_futex_lock_t *lck); |
| 207 | extern void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck); |
| 208 | |
| 209 | extern int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, |
| 210 | kmp_int32 gtid); |
| 211 | extern int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid); |
| 212 | extern int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, |
| 213 | kmp_int32 gtid); |
| 214 | extern void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck); |
| 215 | extern void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck); |
| 216 | |
| 217 | #endif // KMP_USE_FUTEX |
| 218 | |
| 219 | // ---------------------------------------------------------------------------- |
| 220 | // Ticket locks. |
| 221 | |
| 222 | #ifdef __cplusplus201703L |
| 223 | |
| 224 | #ifdef _MSC_VER |
| 225 | // MSVC won't allow use of std::atomic<> in a union since it has non-trivial |
| 226 | // copy constructor. |
| 227 | |
| 228 | struct kmp_base_ticket_lock { |
| 229 | // `initialized' must be the first entry in the lock data structure! |
| 230 | std::atomic_bool initialized; |
| 231 | volatile union kmp_ticket_lock *self; // points to the lock union |
| 232 | ident_t const *location; // Source code location of omp_init_lock(). |
| 233 | std::atomic_uint |
| 234 | next_ticket; // ticket number to give to next thread which acquires |
| 235 | std::atomic_uint now_serving; // ticket number for thread which holds the lock |
| 236 | std::atomic_int owner_id; // (gtid+1) of owning thread, 0 if unlocked |
| 237 | std::atomic_int depth_locked; // depth locked, for nested locks only |
| 238 | kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock |
| 239 | }; |
| 240 | #else |
| 241 | struct kmp_base_ticket_lock { |
| 242 | // `initialized' must be the first entry in the lock data structure! |
| 243 | std::atomic<bool> initialized; |
| 244 | volatile union kmp_ticket_lock *self; // points to the lock union |
| 245 | ident_t const *location; // Source code location of omp_init_lock(). |
| 246 | std::atomic<unsigned> |
| 247 | next_ticket; // ticket number to give to next thread which acquires |
| 248 | std::atomic<unsigned> |
| 249 | now_serving; // ticket number for thread which holds the lock |
| 250 | std::atomic<int> owner_id; // (gtid+1) of owning thread, 0 if unlocked |
| 251 | std::atomic<int> depth_locked; // depth locked, for nested locks only |
| 252 | kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock |
| 253 | }; |
| 254 | #endif |
| 255 | |
| 256 | #else // __cplusplus |
| 257 | |
| 258 | struct kmp_base_ticket_lock; |
| 259 | |
| 260 | #endif // !__cplusplus |
| 261 | |
| 262 | typedef struct kmp_base_ticket_lock kmp_base_ticket_lock_t; |
| 263 | |
| 264 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_ticket_lock { |
| 265 | kmp_base_ticket_lock_t |
| 266 | lk; // This field must be first to allow static initializing. |
| 267 | kmp_lock_pool_t pool; |
| 268 | double lk_align; // use worst case alignment |
| 269 | char lk_pad[KMP_PAD(kmp_base_ticket_lock_t, CACHE_LINE)(sizeof(kmp_base_ticket_lock_t) + (64 - ((sizeof(kmp_base_ticket_lock_t ) - 1) % (64)) - 1))]; |
| 270 | }; |
| 271 | |
| 272 | typedef union kmp_ticket_lock kmp_ticket_lock_t; |
| 273 | |
| 274 | // Static initializer for simple ticket lock variables. Usage: |
| 275 | // kmp_ticket_lock_t xlock = KMP_TICKET_LOCK_INITIALIZER( xlock ); |
| 276 | // Note the macro argument. It is important to make var properly initialized. |
| 277 | #define KMP_TICKET_LOCK_INITIALIZER(lock){ { true, &(lock), __null, 0U, 0U, 0, -1 } } \ |
| 278 | { \ |
| 279 | { true, &(lock), NULL__null, 0U, 0U, 0, -1 } \ |
| 280 | } |
| 281 | |
| 282 | extern int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid); |
| 283 | extern int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid); |
| 284 | extern int __kmp_test_ticket_lock_with_cheks(kmp_ticket_lock_t *lck, |
| 285 | kmp_int32 gtid); |
| 286 | extern int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid); |
| 287 | extern void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck); |
| 288 | extern void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck); |
| 289 | |
| 290 | extern int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, |
| 291 | kmp_int32 gtid); |
| 292 | extern int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, |
| 293 | kmp_int32 gtid); |
| 294 | extern int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, |
| 295 | kmp_int32 gtid); |
| 296 | extern void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck); |
| 297 | extern void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck); |
| 298 | |
| 299 | // ---------------------------------------------------------------------------- |
| 300 | // Queuing locks. |
| 301 | |
| 302 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
| 303 | |
| 304 | struct kmp_adaptive_lock_info; |
| 305 | |
| 306 | typedef struct kmp_adaptive_lock_info kmp_adaptive_lock_info_t; |
| 307 | |
| 308 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 |
| 309 | |
| 310 | struct kmp_adaptive_lock_statistics { |
| 311 | /* So we can get stats from locks that haven't been destroyed. */ |
| 312 | kmp_adaptive_lock_info_t *next; |
| 313 | kmp_adaptive_lock_info_t *prev; |
| 314 | |
| 315 | /* Other statistics */ |
| 316 | kmp_uint32 successfulSpeculations; |
| 317 | kmp_uint32 hardFailedSpeculations; |
| 318 | kmp_uint32 softFailedSpeculations; |
| 319 | kmp_uint32 nonSpeculativeAcquires; |
| 320 | kmp_uint32 nonSpeculativeAcquireAttempts; |
| 321 | kmp_uint32 lemmingYields; |
| 322 | }; |
| 323 | |
| 324 | typedef struct kmp_adaptive_lock_statistics kmp_adaptive_lock_statistics_t; |
| 325 | |
| 326 | extern void __kmp_print_speculative_stats(); |
| 327 | extern void __kmp_init_speculative_stats(); |
| 328 | |
| 329 | #endif // KMP_DEBUG_ADAPTIVE_LOCKS |
| 330 | |
| 331 | struct kmp_adaptive_lock_info { |
| 332 | /* Values used for adaptivity. |
| 333 | Although these are accessed from multiple threads we don't access them |
| 334 | atomically, because if we miss updates it probably doesn't matter much. (It |
| 335 | just affects our decision about whether to try speculation on the lock). */ |
| 336 | kmp_uint32 volatile badness; |
| 337 | kmp_uint32 volatile acquire_attempts; |
| 338 | /* Parameters of the lock. */ |
| 339 | kmp_uint32 max_badness; |
| 340 | kmp_uint32 max_soft_retries; |
| 341 | |
| 342 | #if KMP_DEBUG_ADAPTIVE_LOCKS0 |
| 343 | kmp_adaptive_lock_statistics_t volatile stats; |
| 344 | #endif |
| 345 | }; |
| 346 | |
| 347 | #endif // KMP_USE_ADAPTIVE_LOCKS |
| 348 | |
| 349 | struct kmp_base_queuing_lock { |
| 350 | |
| 351 | // `initialized' must be the first entry in the lock data structure! |
| 352 | volatile union kmp_queuing_lock |
| 353 | *initialized; // Points to the lock union if in initialized state. |
| 354 | |
| 355 | ident_t const *location; // Source code location of omp_init_lock(). |
| 356 | |
| 357 | KMP_ALIGN(8)__attribute__((aligned(8))) // tail_id must be 8-byte aligned! |
| 358 | |
| 359 | volatile kmp_int32 |
| 360 | tail_id; // (gtid+1) of thread at tail of wait queue, 0 if empty |
| 361 | // Must be no padding here since head/tail used in 8-byte CAS |
| 362 | volatile kmp_int32 |
| 363 | head_id; // (gtid+1) of thread at head of wait queue, 0 if empty |
| 364 | // Decl order assumes little endian |
| 365 | // bakery-style lock |
| 366 | volatile kmp_uint32 |
| 367 | next_ticket; // ticket number to give to next thread which acquires |
| 368 | volatile kmp_uint32 |
| 369 | now_serving; // ticket number for thread which holds the lock |
| 370 | volatile kmp_int32 owner_id; // (gtid+1) of owning thread, 0 if unlocked |
| 371 | kmp_int32 depth_locked; // depth locked, for nested locks only |
| 372 | |
| 373 | kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock |
| 374 | }; |
| 375 | |
| 376 | typedef struct kmp_base_queuing_lock kmp_base_queuing_lock_t; |
| 377 | |
| 378 | KMP_BUILD_ASSERT(offsetof(kmp_base_queuing_lock_t, tail_id) % 8 == 0)static_assert(__builtin_offsetof(kmp_base_queuing_lock_t, tail_id ) % 8 == 0, "Build condition error"); |
| 379 | |
| 380 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_queuing_lock { |
| 381 | kmp_base_queuing_lock_t |
| 382 | lk; // This field must be first to allow static initializing. |
| 383 | kmp_lock_pool_t pool; |
| 384 | double lk_align; // use worst case alignment |
| 385 | char lk_pad[KMP_PAD(kmp_base_queuing_lock_t, CACHE_LINE)(sizeof(kmp_base_queuing_lock_t) + (64 - ((sizeof(kmp_base_queuing_lock_t ) - 1) % (64)) - 1))]; |
| 386 | }; |
| 387 | |
| 388 | typedef union kmp_queuing_lock kmp_queuing_lock_t; |
| 389 | |
| 390 | extern int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid); |
| 391 | extern int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid); |
| 392 | extern int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid); |
| 393 | extern void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck); |
| 394 | extern void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck); |
| 395 | |
| 396 | extern int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, |
| 397 | kmp_int32 gtid); |
| 398 | extern int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, |
| 399 | kmp_int32 gtid); |
| 400 | extern int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, |
| 401 | kmp_int32 gtid); |
| 402 | extern void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck); |
| 403 | extern void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck); |
| 404 | |
| 405 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
| 406 | |
| 407 | // ---------------------------------------------------------------------------- |
| 408 | // Adaptive locks. |
| 409 | struct kmp_base_adaptive_lock { |
| 410 | kmp_base_queuing_lock qlk; |
| 411 | KMP_ALIGN(CACHE_LINE)__attribute__((aligned(64))) |
| 412 | kmp_adaptive_lock_info_t |
| 413 | adaptive; // Information for the speculative adaptive lock |
| 414 | }; |
| 415 | |
| 416 | typedef struct kmp_base_adaptive_lock kmp_base_adaptive_lock_t; |
| 417 | |
| 418 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_adaptive_lock { |
| 419 | kmp_base_adaptive_lock_t lk; |
| 420 | kmp_lock_pool_t pool; |
| 421 | double lk_align; |
| 422 | char lk_pad[KMP_PAD(kmp_base_adaptive_lock_t, CACHE_LINE)(sizeof(kmp_base_adaptive_lock_t) + (64 - ((sizeof(kmp_base_adaptive_lock_t ) - 1) % (64)) - 1))]; |
| 423 | }; |
| 424 | typedef union kmp_adaptive_lock kmp_adaptive_lock_t; |
| 425 | |
| 426 | #define GET_QLK_PTR(l)((kmp_queuing_lock_t *)&(l)->lk.qlk) ((kmp_queuing_lock_t *)&(l)->lk.qlk) |
| 427 | |
| 428 | #endif // KMP_USE_ADAPTIVE_LOCKS |
| 429 | |
| 430 | // ---------------------------------------------------------------------------- |
| 431 | // DRDPA ticket locks. |
| 432 | struct kmp_base_drdpa_lock { |
| 433 | // All of the fields on the first cache line are only written when |
| 434 | // initializing or reconfiguring the lock. These are relatively rare |
| 435 | // operations, so data from the first cache line will usually stay resident in |
| 436 | // the cache of each thread trying to acquire the lock. |
| 437 | // |
| 438 | // initialized must be the first entry in the lock data structure! |
| 439 | KMP_ALIGN_CACHE__attribute__((aligned(64))) |
| 440 | |
| 441 | volatile union kmp_drdpa_lock |
| 442 | *initialized; // points to the lock union if in initialized state |
| 443 | ident_t const *location; // Source code location of omp_init_lock(). |
| 444 | std::atomic<std::atomic<kmp_uint64> *> polls; |
| 445 | std::atomic<kmp_uint64> mask; // is 2**num_polls-1 for mod op |
| 446 | kmp_uint64 cleanup_ticket; // thread with cleanup ticket |
| 447 | std::atomic<kmp_uint64> *old_polls; // will deallocate old_polls |
| 448 | kmp_uint32 num_polls; // must be power of 2 |
| 449 | |
| 450 | // next_ticket it needs to exist in a separate cache line, as it is |
| 451 | // invalidated every time a thread takes a new ticket. |
| 452 | KMP_ALIGN_CACHE__attribute__((aligned(64))) |
| 453 | |
| 454 | std::atomic<kmp_uint64> next_ticket; |
| 455 | |
| 456 | // now_serving is used to store our ticket value while we hold the lock. It |
| 457 | // has a slightly different meaning in the DRDPA ticket locks (where it is |
| 458 | // written by the acquiring thread) than it does in the simple ticket locks |
| 459 | // (where it is written by the releasing thread). |
| 460 | // |
| 461 | // Since now_serving is only read and written in the critical section, |
| 462 | // it is non-volatile, but it needs to exist on a separate cache line, |
| 463 | // as it is invalidated at every lock acquire. |
| 464 | // |
| 465 | // Likewise, the vars used for nested locks (owner_id and depth_locked) are |
| 466 | // only written by the thread owning the lock, so they are put in this cache |
| 467 | // line. owner_id is read by other threads, so it must be declared volatile. |
| 468 | KMP_ALIGN_CACHE__attribute__((aligned(64))) |
| 469 | kmp_uint64 now_serving; // doesn't have to be volatile |
| 470 | volatile kmp_uint32 owner_id; // (gtid+1) of owning thread, 0 if unlocked |
| 471 | kmp_int32 depth_locked; // depth locked |
| 472 | kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock |
| 473 | }; |
| 474 | |
| 475 | typedef struct kmp_base_drdpa_lock kmp_base_drdpa_lock_t; |
| 476 | |
| 477 | union KMP_ALIGN_CACHE__attribute__((aligned(64))) kmp_drdpa_lock { |
| 478 | kmp_base_drdpa_lock_t |
| 479 | lk; // This field must be first to allow static initializing. */ |
| 480 | kmp_lock_pool_t pool; |
| 481 | double lk_align; // use worst case alignment |
| 482 | char lk_pad[KMP_PAD(kmp_base_drdpa_lock_t, CACHE_LINE)(sizeof(kmp_base_drdpa_lock_t) + (64 - ((sizeof(kmp_base_drdpa_lock_t ) - 1) % (64)) - 1))]; |
| 483 | }; |
| 484 | |
| 485 | typedef union kmp_drdpa_lock kmp_drdpa_lock_t; |
| 486 | |
| 487 | extern int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid); |
| 488 | extern int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid); |
| 489 | extern int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid); |
| 490 | extern void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck); |
| 491 | extern void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck); |
| 492 | |
| 493 | extern int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, |
| 494 | kmp_int32 gtid); |
| 495 | extern int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid); |
| 496 | extern int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, |
| 497 | kmp_int32 gtid); |
| 498 | extern void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck); |
| 499 | extern void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck); |
| 500 | |
| 501 | // ============================================================================ |
| 502 | // Lock purposes. |
| 503 | // ============================================================================ |
| 504 | |
| 505 | // Bootstrap locks. |
| 506 | // |
| 507 | // Bootstrap locks -- very few locks used at library initialization time. |
| 508 | // Bootstrap locks are currently implemented as ticket locks. |
| 509 | // They could also be implemented as test and set lock, but cannot be |
| 510 | // implemented with other lock kinds as they require gtids which are not |
| 511 | // available at initialization time. |
| 512 | |
| 513 | typedef kmp_ticket_lock_t kmp_bootstrap_lock_t; |
| 514 | |
| 515 | #define KMP_BOOTSTRAP_LOCK_INITIALIZER(lock){ { true, &((lock)), __null, 0U, 0U, 0, -1 } } KMP_TICKET_LOCK_INITIALIZER((lock)){ { true, &((lock)), __null, 0U, 0U, 0, -1 } } |
| 516 | #define KMP_BOOTSTRAP_LOCK_INIT(lock)kmp_bootstrap_lock_t lock = { { true, &(lock), __null, 0U , 0U, 0, -1 } } \ |
| 517 | kmp_bootstrap_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock){ { true, &(lock), __null, 0U, 0U, 0, -1 } } |
| 518 | |
| 519 | static inline int __kmp_acquire_bootstrap_lock(kmp_bootstrap_lock_t *lck) { |
| 520 | return __kmp_acquire_ticket_lock(lck, KMP_GTID_DNE(-2)); |
| 521 | } |
| 522 | |
| 523 | static inline int __kmp_test_bootstrap_lock(kmp_bootstrap_lock_t *lck) { |
| 524 | return __kmp_test_ticket_lock(lck, KMP_GTID_DNE(-2)); |
| 525 | } |
| 526 | |
| 527 | static inline void __kmp_release_bootstrap_lock(kmp_bootstrap_lock_t *lck) { |
| 528 | __kmp_release_ticket_lock(lck, KMP_GTID_DNE(-2)); |
| 529 | } |
| 530 | |
| 531 | static inline void __kmp_init_bootstrap_lock(kmp_bootstrap_lock_t *lck) { |
| 532 | __kmp_init_ticket_lock(lck); |
| 533 | } |
| 534 | |
| 535 | static inline void __kmp_destroy_bootstrap_lock(kmp_bootstrap_lock_t *lck) { |
| 536 | __kmp_destroy_ticket_lock(lck); |
| 537 | } |
| 538 | |
| 539 | // Internal RTL locks. |
| 540 | // |
| 541 | // Internal RTL locks are also implemented as ticket locks, for now. |
| 542 | // |
| 543 | // FIXME - We should go through and figure out which lock kind works best for |
| 544 | // each internal lock, and use the type declaration and function calls for |
| 545 | // that explicit lock kind (and get rid of this section). |
| 546 | |
| 547 | typedef kmp_ticket_lock_t kmp_lock_t; |
| 548 | |
| 549 | #define KMP_LOCK_INIT(lock)kmp_lock_t lock = { { true, &(lock), __null, 0U, 0U, 0, - 1 } } kmp_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock){ { true, &(lock), __null, 0U, 0U, 0, -1 } } |
| 550 | |
| 551 | static inline int __kmp_acquire_lock(kmp_lock_t *lck, kmp_int32 gtid) { |
| 552 | return __kmp_acquire_ticket_lock(lck, gtid); |
| 553 | } |
| 554 | |
| 555 | static inline int __kmp_test_lock(kmp_lock_t *lck, kmp_int32 gtid) { |
| 556 | return __kmp_test_ticket_lock(lck, gtid); |
| 557 | } |
| 558 | |
| 559 | static inline void __kmp_release_lock(kmp_lock_t *lck, kmp_int32 gtid) { |
| 560 | __kmp_release_ticket_lock(lck, gtid); |
| 561 | } |
| 562 | |
| 563 | static inline void __kmp_init_lock(kmp_lock_t *lck) { |
| 564 | __kmp_init_ticket_lock(lck); |
| 565 | } |
| 566 | |
| 567 | static inline void __kmp_destroy_lock(kmp_lock_t *lck) { |
| 568 | __kmp_destroy_ticket_lock(lck); |
| 569 | } |
| 570 | |
| 571 | // User locks. |
| 572 | // |
| 573 | // Do not allocate objects of type union kmp_user_lock!!! This will waste space |
| 574 | // unless __kmp_user_lock_kind == lk_drdpa. Instead, check the value of |
| 575 | // __kmp_user_lock_kind and allocate objects of the type of the appropriate |
| 576 | // union member, and cast their addresses to kmp_user_lock_p. |
| 577 | |
| 578 | enum kmp_lock_kind { |
| 579 | lk_default = 0, |
| 580 | lk_tas, |
| 581 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 582 | lk_futex, |
| 583 | #endif |
| 584 | #if KMP_USE_DYNAMIC_LOCK1 && KMP_USE_TSX(0 || 1) && !0 |
| 585 | lk_hle, |
| 586 | lk_rtm_queuing, |
| 587 | lk_rtm_spin, |
| 588 | #endif |
| 589 | lk_ticket, |
| 590 | lk_queuing, |
| 591 | lk_drdpa, |
| 592 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
| 593 | lk_adaptive |
| 594 | #endif // KMP_USE_ADAPTIVE_LOCKS |
| 595 | }; |
| 596 | |
| 597 | typedef enum kmp_lock_kind kmp_lock_kind_t; |
| 598 | |
| 599 | extern kmp_lock_kind_t __kmp_user_lock_kind; |
| 600 | |
| 601 | union kmp_user_lock { |
| 602 | kmp_tas_lock_t tas; |
| 603 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 604 | kmp_futex_lock_t futex; |
| 605 | #endif |
| 606 | kmp_ticket_lock_t ticket; |
| 607 | kmp_queuing_lock_t queuing; |
| 608 | kmp_drdpa_lock_t drdpa; |
| 609 | #if KMP_USE_ADAPTIVE_LOCKS(0 || 1) && !0 |
| 610 | kmp_adaptive_lock_t adaptive; |
| 611 | #endif // KMP_USE_ADAPTIVE_LOCKS |
| 612 | kmp_lock_pool_t pool; |
| 613 | }; |
| 614 | |
| 615 | typedef union kmp_user_lock *kmp_user_lock_p; |
| 616 | |
| 617 | #if !KMP_USE_DYNAMIC_LOCK1 |
| 618 | |
| 619 | extern size_t __kmp_base_user_lock_size; |
| 620 | extern size_t __kmp_user_lock_size; |
| 621 | |
| 622 | extern kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck); |
| 623 | |
| 624 | static inline kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck) { |
| 625 | KMP_DEBUG_ASSERT(__kmp_get_user_lock_owner_ != NULL)if (!(__kmp_get_user_lock_owner_ != __null)) { __kmp_debug_assert ("__kmp_get_user_lock_owner_ != __null", "openmp/runtime/src/kmp_lock.h" , 625); }; |
| 626 | return (*__kmp_get_user_lock_owner_)(lck); |
| 627 | } |
| 628 | |
| 629 | extern int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 630 | kmp_int32 gtid); |
| 631 | |
| 632 | #if KMP_OS_LINUX1 && \ |
| 633 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) |
| 634 | |
| 635 | #define __kmp_acquire_user_lock_with_checks(lck, gtid) \ |
| 636 | if (__kmp_user_lock_kind == lk_tas) { \ |
| 637 | if (__kmp_env_consistency_check) { \ |
| 638 | char const *const func = "omp_set_lock"; \ |
| 639 | if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZEsizeof(int)) && \ |
| 640 | lck->tas.lk.depth_locked != -1) { \ |
| 641 | KMP_FATAL(LockNestableUsedAsSimple, func)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_LockNestableUsedAsSimple , func), __kmp_msg_null); \ |
| 642 | } \ |
| 643 | if ((gtid >= 0) && (lck->tas.lk.poll - 1 == gtid)) { \ |
| 644 | KMP_FATAL(LockIsAlreadyOwned, func)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_LockIsAlreadyOwned, func), __kmp_msg_null); \ |
| 645 | } \ |
| 646 | } \ |
| 647 | if (lck->tas.lk.poll != 0 || \ |
| 648 | !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) { \ |
| 649 | kmp_uint32 spins; \ |
| 650 | kmp_uint64 time; \ |
| 651 | KMP_FSYNC_PREPARE(lck)(!__kmp_itt_fsync_prepare_ptr__3_0) ? (void)0 : __kmp_itt_fsync_prepare_ptr__3_0 ((void *)(lck)); \ |
| 652 | KMP_INIT_YIELD(spins){ (spins) = __kmp_yield_init; }; \ |
| 653 | KMP_INIT_BACKOFF(time){ (time) = __kmp_pause_init; }; \ |
| 654 | do { \ |
| 655 | KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time){ if (__kmp_tpause_enabled) { if (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))) { __kmp_tpause(0, (time) ); } else { __kmp_tpause(__kmp_tpause_hint, (time)); } (time) = (time << 1 | 1) & ((kmp_uint64)0xFFFF); } else { __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { __kmp_yield(); } else if (__kmp_use_yield == 1) { (spins) -= 2; if (!(spins)) { __kmp_yield(); (spins) = __kmp_yield_next; } } } }; \ |
| 656 | } while ( \ |
| 657 | lck->tas.lk.poll != 0 || \ |
| 658 | !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)); \ |
| 659 | } \ |
| 660 | KMP_FSYNC_ACQUIRED(lck)(!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)(lck)); \ |
| 661 | } else { \ |
| 662 | KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL)if (!(__kmp_acquire_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_acquire_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 662); }; \ |
| 663 | (*__kmp_acquire_user_lock_with_checks_)(lck, gtid); \ |
| 664 | } |
| 665 | |
| 666 | #else |
| 667 | static inline int __kmp_acquire_user_lock_with_checks(kmp_user_lock_p lck, |
| 668 | kmp_int32 gtid) { |
| 669 | KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL)if (!(__kmp_acquire_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_acquire_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 669); }; |
| 670 | return (*__kmp_acquire_user_lock_with_checks_)(lck, gtid); |
| 671 | } |
| 672 | #endif |
| 673 | |
| 674 | extern int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 675 | kmp_int32 gtid); |
| 676 | |
| 677 | #if KMP_OS_LINUX1 && \ |
| 678 | (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM || KMP_ARCH_AARCH640) |
| 679 | |
| 680 | #include "kmp_i18n.h" /* AC: KMP_FATAL definition */ |
| 681 | extern int __kmp_env_consistency_check; /* AC: copy from kmp.h here */ |
| 682 | static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck, |
| 683 | kmp_int32 gtid) { |
| 684 | if (__kmp_user_lock_kind == lk_tas) { |
| 685 | if (__kmp_env_consistency_check) { |
| 686 | char const *const func = "omp_test_lock"; |
| 687 | if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZEsizeof(int)) && |
| 688 | lck->tas.lk.depth_locked != -1) { |
| 689 | KMP_FATAL(LockNestableUsedAsSimple, func)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_LockNestableUsedAsSimple , func), __kmp_msg_null); |
| 690 | } |
| 691 | } |
| 692 | return ((lck->tas.lk.poll == 0) && |
| 693 | __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)); |
| 694 | } else { |
| 695 | KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL)if (!(__kmp_test_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_test_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 695); }; |
| 696 | return (*__kmp_test_user_lock_with_checks_)(lck, gtid); |
| 697 | } |
| 698 | } |
| 699 | #else |
| 700 | static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck, |
| 701 | kmp_int32 gtid) { |
| 702 | KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL)if (!(__kmp_test_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_test_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 702); }; |
| 703 | return (*__kmp_test_user_lock_with_checks_)(lck, gtid); |
| 704 | } |
| 705 | #endif |
| 706 | |
| 707 | extern int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 708 | kmp_int32 gtid); |
| 709 | |
| 710 | static inline void __kmp_release_user_lock_with_checks(kmp_user_lock_p lck, |
| 711 | kmp_int32 gtid) { |
| 712 | KMP_DEBUG_ASSERT(__kmp_release_user_lock_with_checks_ != NULL)if (!(__kmp_release_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_release_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 712); }; |
| 713 | (*__kmp_release_user_lock_with_checks_)(lck, gtid); |
| 714 | } |
| 715 | |
| 716 | extern void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck); |
| 717 | |
| 718 | static inline void __kmp_init_user_lock_with_checks(kmp_user_lock_p lck) { |
| 719 | KMP_DEBUG_ASSERT(__kmp_init_user_lock_with_checks_ != NULL)if (!(__kmp_init_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_init_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 719); }; |
| 720 | (*__kmp_init_user_lock_with_checks_)(lck); |
| 721 | } |
| 722 | |
| 723 | // We need a non-checking version of destroy lock for when the RTL is |
| 724 | // doing the cleanup as it can't always tell if the lock is nested or not. |
| 725 | extern void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck); |
| 726 | |
| 727 | static inline void __kmp_destroy_user_lock(kmp_user_lock_p lck) { |
| 728 | KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_ != NULL)if (!(__kmp_destroy_user_lock_ != __null)) { __kmp_debug_assert ("__kmp_destroy_user_lock_ != __null", "openmp/runtime/src/kmp_lock.h" , 728); }; |
| 729 | (*__kmp_destroy_user_lock_)(lck); |
| 730 | } |
| 731 | |
| 732 | extern void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck); |
| 733 | |
| 734 | static inline void __kmp_destroy_user_lock_with_checks(kmp_user_lock_p lck) { |
| 735 | KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_with_checks_ != NULL)if (!(__kmp_destroy_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_destroy_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 735); }; |
| 736 | (*__kmp_destroy_user_lock_with_checks_)(lck); |
| 737 | } |
| 738 | |
| 739 | extern int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 740 | kmp_int32 gtid); |
| 741 | |
| 742 | #if KMP_OS_LINUX1 && (KMP_ARCH_X860 || KMP_ARCH_X86_641) |
| 743 | |
| 744 | #define __kmp_acquire_nested_user_lock_with_checks(lck, gtid, depth) \ |
| 745 | if (__kmp_user_lock_kind == lk_tas) { \ |
| 746 | if (__kmp_env_consistency_check) { \ |
| 747 | char const *const func = "omp_set_nest_lock"; \ |
| 748 | if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZEsizeof(void *)) && \ |
| 749 | lck->tas.lk.depth_locked == -1) { \ |
| 750 | KMP_FATAL(LockSimpleUsedAsNestable, func)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_LockSimpleUsedAsNestable , func), __kmp_msg_null); \ |
| 751 | } \ |
| 752 | } \ |
| 753 | if (lck->tas.lk.poll - 1 == gtid) { \ |
| 754 | lck->tas.lk.depth_locked += 1; \ |
| 755 | *depth = KMP_LOCK_ACQUIRED_NEXT0; \ |
| 756 | } else { \ |
| 757 | if ((lck->tas.lk.poll != 0) || \ |
| 758 | !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) { \ |
| 759 | kmp_uint32 spins; \ |
| 760 | kmp_uint64 time; \ |
| 761 | KMP_FSYNC_PREPARE(lck)(!__kmp_itt_fsync_prepare_ptr__3_0) ? (void)0 : __kmp_itt_fsync_prepare_ptr__3_0 ((void *)(lck)); \ |
| 762 | KMP_INIT_YIELD(spins){ (spins) = __kmp_yield_init; }; \ |
| 763 | KMP_INIT_BACKOFF(time){ (time) = __kmp_pause_init; }; \ |
| 764 | do { \ |
| 765 | KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time){ if (__kmp_tpause_enabled) { if (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))) { __kmp_tpause(0, (time) ); } else { __kmp_tpause(__kmp_tpause_hint, (time)); } (time) = (time << 1 | 1) & ((kmp_uint64)0xFFFF); } else { __kmp_x86_pause(); if ((((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (((__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)))))) { __kmp_yield(); } else if (__kmp_use_yield == 1) { (spins) -= 2; if (!(spins)) { __kmp_yield(); (spins) = __kmp_yield_next; } } } }; \ |
| 766 | } while ( \ |
| 767 | (lck->tas.lk.poll != 0) || \ |
| 768 | !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)); \ |
| 769 | } \ |
| 770 | lck->tas.lk.depth_locked = 1; \ |
| 771 | *depth = KMP_LOCK_ACQUIRED_FIRST1; \ |
| 772 | } \ |
| 773 | KMP_FSYNC_ACQUIRED(lck)(!__kmp_itt_fsync_acquired_ptr__3_0) ? (void)0 : __kmp_itt_fsync_acquired_ptr__3_0 ((void *)(lck)); \ |
| 774 | } else { \ |
| 775 | KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL)if (!(__kmp_acquire_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert("__kmp_acquire_nested_user_lock_with_checks_ != __null" , "openmp/runtime/src/kmp_lock.h", 775); }; \ |
| 776 | *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid); \ |
| 777 | } |
| 778 | |
| 779 | #else |
| 780 | static inline void |
| 781 | __kmp_acquire_nested_user_lock_with_checks(kmp_user_lock_p lck, kmp_int32 gtid, |
| 782 | int *depth) { |
| 783 | KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL)if (!(__kmp_acquire_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert("__kmp_acquire_nested_user_lock_with_checks_ != __null" , "openmp/runtime/src/kmp_lock.h", 783); }; |
| 784 | *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid); |
| 785 | } |
| 786 | #endif |
| 787 | |
| 788 | extern int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 789 | kmp_int32 gtid); |
| 790 | |
| 791 | #if KMP_OS_LINUX1 && (KMP_ARCH_X860 || KMP_ARCH_X86_641) |
| 792 | static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck, |
| 793 | kmp_int32 gtid) { |
| 794 | if (__kmp_user_lock_kind == lk_tas) { |
| 795 | int retval; |
| 796 | if (__kmp_env_consistency_check) { |
| 797 | char const *const func = "omp_test_nest_lock"; |
| 798 | if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZEsizeof(void *)) && |
| 799 | lck->tas.lk.depth_locked == -1) { |
| 800 | KMP_FATAL(LockSimpleUsedAsNestable, func)__kmp_fatal(__kmp_msg_format(kmp_i18n_msg_LockSimpleUsedAsNestable , func), __kmp_msg_null); |
| 801 | } |
| 802 | } |
| 803 | KMP_DEBUG_ASSERT(gtid >= 0)if (!(gtid >= 0)) { __kmp_debug_assert("gtid >= 0", "openmp/runtime/src/kmp_lock.h" , 803); }; |
| 804 | if (lck->tas.lk.poll - 1 == |
| 805 | gtid) { /* __kmp_get_tas_lock_owner( lck ) == gtid */ |
| 806 | return ++lck->tas.lk.depth_locked; /* same owner, depth increased */ |
| 807 | } |
| 808 | retval = ((lck->tas.lk.poll == 0) && |
| 809 | __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)); |
| 810 | if (retval) { |
| 811 | KMP_MB(); |
| 812 | lck->tas.lk.depth_locked = 1; |
| 813 | } |
| 814 | return retval; |
| 815 | } else { |
| 816 | KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL)if (!(__kmp_test_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_test_nested_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 816); }; |
| 817 | return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid); |
| 818 | } |
| 819 | } |
| 820 | #else |
| 821 | static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck, |
| 822 | kmp_int32 gtid) { |
| 823 | KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL)if (!(__kmp_test_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_test_nested_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 823); }; |
| 824 | return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid); |
| 825 | } |
| 826 | #endif |
| 827 | |
| 828 | extern int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck, |
| 829 | kmp_int32 gtid); |
| 830 | |
| 831 | static inline int |
| 832 | __kmp_release_nested_user_lock_with_checks(kmp_user_lock_p lck, |
| 833 | kmp_int32 gtid) { |
| 834 | KMP_DEBUG_ASSERT(__kmp_release_nested_user_lock_with_checks_ != NULL)if (!(__kmp_release_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert("__kmp_release_nested_user_lock_with_checks_ != __null" , "openmp/runtime/src/kmp_lock.h", 834); }; |
| 835 | return (*__kmp_release_nested_user_lock_with_checks_)(lck, gtid); |
| 836 | } |
| 837 | |
| 838 | extern void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck); |
| 839 | |
| 840 | static inline void |
| 841 | __kmp_init_nested_user_lock_with_checks(kmp_user_lock_p lck) { |
| 842 | KMP_DEBUG_ASSERT(__kmp_init_nested_user_lock_with_checks_ != NULL)if (!(__kmp_init_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert ("__kmp_init_nested_user_lock_with_checks_ != __null", "openmp/runtime/src/kmp_lock.h" , 842); }; |
| 843 | (*__kmp_init_nested_user_lock_with_checks_)(lck); |
| 844 | } |
| 845 | |
| 846 | extern void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck); |
| 847 | |
| 848 | static inline void |
| 849 | __kmp_destroy_nested_user_lock_with_checks(kmp_user_lock_p lck) { |
| 850 | KMP_DEBUG_ASSERT(__kmp_destroy_nested_user_lock_with_checks_ != NULL)if (!(__kmp_destroy_nested_user_lock_with_checks_ != __null)) { __kmp_debug_assert("__kmp_destroy_nested_user_lock_with_checks_ != __null" , "openmp/runtime/src/kmp_lock.h", 850); }; |
| 851 | (*__kmp_destroy_nested_user_lock_with_checks_)(lck); |
| 852 | } |
| 853 | |
| 854 | // user lock functions which do not necessarily exist for all lock kinds. |
| 855 | // |
| 856 | // The "set" functions usually have wrapper routines that check for a NULL set |
| 857 | // function pointer and call it if non-NULL. |
| 858 | // |
| 859 | // In some cases, it makes sense to have a "get" wrapper function check for a |
| 860 | // NULL get function pointer and return NULL / invalid value / error code if |
| 861 | // the function pointer is NULL. |
| 862 | // |
| 863 | // In other cases, the calling code really should differentiate between an |
| 864 | // unimplemented function and one that is implemented but returning NULL / |
| 865 | // invalid value. If this is the case, no get function wrapper exists. |
| 866 | |
| 867 | extern int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck); |
| 868 | |
| 869 | // no set function; fields set during local allocation |
| 870 | |
| 871 | extern const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck); |
| 872 | |
| 873 | static inline const ident_t *__kmp_get_user_lock_location(kmp_user_lock_p lck) { |
| 874 | if (__kmp_get_user_lock_location_ != NULL__null) { |
| 875 | return (*__kmp_get_user_lock_location_)(lck); |
| 876 | } else { |
| 877 | return NULL__null; |
| 878 | } |
| 879 | } |
| 880 | |
| 881 | extern void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck, |
| 882 | const ident_t *loc); |
| 883 | |
| 884 | static inline void __kmp_set_user_lock_location(kmp_user_lock_p lck, |
| 885 | const ident_t *loc) { |
| 886 | if (__kmp_set_user_lock_location_ != NULL__null) { |
| 887 | (*__kmp_set_user_lock_location_)(lck, loc); |
| 888 | } |
| 889 | } |
| 890 | |
| 891 | extern kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck); |
| 892 | |
| 893 | extern void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck, |
| 894 | kmp_lock_flags_t flags); |
| 895 | |
| 896 | static inline void __kmp_set_user_lock_flags(kmp_user_lock_p lck, |
| 897 | kmp_lock_flags_t flags) { |
| 898 | if (__kmp_set_user_lock_flags_ != NULL__null) { |
| 899 | (*__kmp_set_user_lock_flags_)(lck, flags); |
| 900 | } |
| 901 | } |
| 902 | |
| 903 | // The function which sets up all of the vtbl pointers for kmp_user_lock_t. |
| 904 | extern void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind); |
| 905 | |
| 906 | // Macros for binding user lock functions. |
| 907 | #define KMP_BIND_USER_LOCK_TEMPLATE(nest, kind, suffix) \ |
| 908 | { \ |
| 909 | __kmp_acquire##nest##user_lock_with_checks_ = (int (*)( \ |
| 910 | kmp_user_lock_p, kmp_int32))__kmp_acquire##nest##kind##_##suffix; \ |
| 911 | __kmp_release##nest##user_lock_with_checks_ = (int (*)( \ |
| 912 | kmp_user_lock_p, kmp_int32))__kmp_release##nest##kind##_##suffix; \ |
| 913 | __kmp_test##nest##user_lock_with_checks_ = (int (*)( \ |
| 914 | kmp_user_lock_p, kmp_int32))__kmp_test##nest##kind##_##suffix; \ |
| 915 | __kmp_init##nest##user_lock_with_checks_ = \ |
| 916 | (void (*)(kmp_user_lock_p))__kmp_init##nest##kind##_##suffix; \ |
| 917 | __kmp_destroy##nest##user_lock_with_checks_ = \ |
| 918 | (void (*)(kmp_user_lock_p))__kmp_destroy##nest##kind##_##suffix; \ |
| 919 | } |
| 920 | |
| 921 | #define KMP_BIND_USER_LOCK(kind) KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock) |
| 922 | #define KMP_BIND_USER_LOCK_WITH_CHECKS(kind) \ |
| 923 | KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock_with_checks) |
| 924 | #define KMP_BIND_NESTED_USER_LOCK(kind) \ |
| 925 | KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock) |
| 926 | #define KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(kind) \ |
| 927 | KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock_with_checks) |
| 928 | |
| 929 | // User lock table & lock allocation |
| 930 | /* On 64-bit Linux* OS (and OS X*) GNU compiler allocates only 4 bytems memory |
| 931 | for lock variable, which is not enough to store a pointer, so we have to use |
| 932 | lock indexes instead of pointers and maintain lock table to map indexes to |
| 933 | pointers. |
| 934 | |
| 935 | |
| 936 | Note: The first element of the table is not a pointer to lock! It is a |
| 937 | pointer to previously allocated table (or NULL if it is the first table). |
| 938 | |
| 939 | Usage: |
| 940 | |
| 941 | if ( OMP_LOCK_T_SIZE < sizeof( <lock> ) ) { // or OMP_NEST_LOCK_T_SIZE |
| 942 | Lock table is fully utilized. User locks are indexes, so table is used on |
| 943 | user lock operation. |
| 944 | Note: it may be the case (lin_32) that we don't need to use a lock |
| 945 | table for regular locks, but do need the table for nested locks. |
| 946 | } |
| 947 | else { |
| 948 | Lock table initialized but not actually used. |
| 949 | } |
| 950 | */ |
| 951 | |
| 952 | struct kmp_lock_table { |
| 953 | kmp_lock_index_t used; // Number of used elements |
| 954 | kmp_lock_index_t allocated; // Number of allocated elements |
| 955 | kmp_user_lock_p *table; // Lock table. |
| 956 | }; |
| 957 | |
| 958 | typedef struct kmp_lock_table kmp_lock_table_t; |
| 959 | |
| 960 | extern kmp_lock_table_t __kmp_user_lock_table; |
| 961 | extern kmp_user_lock_p __kmp_lock_pool; |
| 962 | |
| 963 | struct kmp_block_of_locks { |
| 964 | struct kmp_block_of_locks *next_block; |
| 965 | void *locks; |
| 966 | }; |
| 967 | |
| 968 | typedef struct kmp_block_of_locks kmp_block_of_locks_t; |
| 969 | |
| 970 | extern kmp_block_of_locks_t *__kmp_lock_blocks; |
| 971 | extern int __kmp_num_locks_in_block; |
| 972 | |
| 973 | extern kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, |
| 974 | kmp_int32 gtid, |
| 975 | kmp_lock_flags_t flags); |
| 976 | extern void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid, |
| 977 | kmp_user_lock_p lck); |
| 978 | extern kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, |
| 979 | char const *func); |
| 980 | extern void __kmp_cleanup_user_locks(); |
| 981 | |
| 982 | #define KMP_CHECK_USER_LOCK_INIT() \ |
| 983 | { \ |
| 984 | if (!TCR_4(__kmp_init_user_locks)(__kmp_init_user_locks)) { \ |
| 985 | __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); \ |
| 986 | if (!TCR_4(__kmp_init_user_locks)(__kmp_init_user_locks)) { \ |
| 987 | TCW_4(__kmp_init_user_locks, TRUE)(__kmp_init_user_locks) = ((!0)); \ |
| 988 | } \ |
| 989 | __kmp_release_bootstrap_lock(&__kmp_initz_lock); \ |
| 990 | } \ |
| 991 | } |
| 992 | |
| 993 | #endif // KMP_USE_DYNAMIC_LOCK |
| 994 | |
| 995 | #undef KMP_PAD |
| 996 | #undef KMP_GTID_DNE(-2) |
| 997 | |
| 998 | #if KMP_USE_DYNAMIC_LOCK1 |
| 999 | // KMP_USE_DYNAMIC_LOCK enables dynamic dispatch of lock functions without |
| 1000 | // breaking the current compatibility. Essential functionality of this new code |
| 1001 | // is dynamic dispatch, but it also implements (or enables implementation of) |
| 1002 | // hinted user lock and critical section which will be part of OMP 4.5 soon. |
| 1003 | // |
| 1004 | // Lock type can be decided at creation time (i.e., lock initialization), and |
| 1005 | // subsequent lock function call on the created lock object requires type |
| 1006 | // extraction and call through jump table using the extracted type. This type |
| 1007 | // information is stored in two different ways depending on the size of the lock |
| 1008 | // object, and we differentiate lock types by this size requirement - direct and |
| 1009 | // indirect locks. |
| 1010 | // |
| 1011 | // Direct locks: |
| 1012 | // A direct lock object fits into the space created by the compiler for an |
| 1013 | // omp_lock_t object, and TAS/Futex lock falls into this category. We use low |
| 1014 | // one byte of the lock object as the storage for the lock type, and appropriate |
| 1015 | // bit operation is required to access the data meaningful to the lock |
| 1016 | // algorithms. Also, to differentiate direct lock from indirect lock, 1 is |
| 1017 | // written to LSB of the lock object. The newly introduced "hle" lock is also a |
| 1018 | // direct lock. |
| 1019 | // |
| 1020 | // Indirect locks: |
| 1021 | // An indirect lock object requires more space than the compiler-generated |
| 1022 | // space, and it should be allocated from heap. Depending on the size of the |
| 1023 | // compiler-generated space for the lock (i.e., size of omp_lock_t), this |
| 1024 | // omp_lock_t object stores either the address of the heap-allocated indirect |
| 1025 | // lock (void * fits in the object) or an index to the indirect lock table entry |
| 1026 | // that holds the address. Ticket/Queuing/DRDPA/Adaptive lock falls into this |
| 1027 | // category, and the newly introduced "rtm" lock is also an indirect lock which |
| 1028 | // was implemented on top of the Queuing lock. When the omp_lock_t object holds |
| 1029 | // an index (not lock address), 0 is written to LSB to differentiate the lock |
| 1030 | // from a direct lock, and the remaining part is the actual index to the |
| 1031 | // indirect lock table. |
| 1032 | |
| 1033 | #include <stdint.h> // for uintptr_t |
| 1034 | |
| 1035 | // Shortcuts |
| 1036 | #define KMP_USE_INLINED_TAS(1 && (0 || 1 || KMP_ARCH_ARM)) && 1 \ |
| 1037 | (KMP_OS_LINUX1 && (KMP_ARCH_X860 || KMP_ARCH_X86_641 || KMP_ARCH_ARM)) && 1 |
| 1038 | #define KMP_USE_INLINED_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) && 0 KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) && 0 |
| 1039 | |
| 1040 | // List of lock definitions; all nested locks are indirect locks. |
| 1041 | // hle lock is xchg lock prefixed with XACQUIRE/XRELEASE. |
| 1042 | // All nested locks are indirect lock types. |
| 1043 | #if KMP_USE_TSX(0 || 1) && !0 |
| 1044 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 1045 | #define KMP_FOREACH_D_LOCK(m, a)m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a) m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a) |
| 1046 | #define KMP_FOREACH_I_LOCK(m, a)m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing , a) m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a) m (nested_queuing, a) m(nested_drdpa, a) \ |
| 1047 | m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a) \ |
| 1048 | m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a) \ |
| 1049 | m(nested_queuing, a) m(nested_drdpa, a) |
| 1050 | #else |
| 1051 | #define KMP_FOREACH_D_LOCK(m, a)m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a) m(tas, a) m(hle, a) m(rtm_spin, a) |
| 1052 | #define KMP_FOREACH_I_LOCK(m, a)m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing , a) m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a) m (nested_queuing, a) m(nested_drdpa, a) \ |
| 1053 | m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a) \ |
| 1054 | m(nested_tas, a) m(nested_ticket, a) m(nested_queuing, a) \ |
| 1055 | m(nested_drdpa, a) |
| 1056 | #endif // KMP_USE_FUTEX |
| 1057 | #define KMP_LAST_D_LOCKlockseq_rtm_spin lockseq_rtm_spin |
| 1058 | #else |
| 1059 | #if KMP_USE_FUTEX(1 && (0 || 1 || KMP_ARCH_ARM || 0)) |
| 1060 | #define KMP_FOREACH_D_LOCK(m, a)m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a) m(tas, a) m(futex, a) |
| 1061 | #define KMP_FOREACH_I_LOCK(m, a)m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing , a) m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a) m (nested_queuing, a) m(nested_drdpa, a) \ |
| 1062 | m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_futex, a) \ |
| 1063 | m(nested_ticket, a) m(nested_queuing, a) m(nested_drdpa, a) |
| 1064 | #define KMP_LAST_D_LOCKlockseq_rtm_spin lockseq_futex |
| 1065 | #else |
| 1066 | #define KMP_FOREACH_D_LOCK(m, a)m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a) m(tas, a) |
| 1067 | #define KMP_FOREACH_I_LOCK(m, a)m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing , a) m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a) m (nested_queuing, a) m(nested_drdpa, a) \ |
| 1068 | m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_ticket, a) \ |
| 1069 | m(nested_queuing, a) m(nested_drdpa, a) |
| 1070 | #define KMP_LAST_D_LOCKlockseq_rtm_spin lockseq_tas |
| 1071 | #endif // KMP_USE_FUTEX |
| 1072 | #endif // KMP_USE_TSX |
| 1073 | |
| 1074 | // Information used in dynamic dispatch |
| 1075 | #define KMP_LOCK_SHIFT8 \ |
| 1076 | 8 // number of low bits to be used as tag for direct locks |
| 1077 | #define KMP_FIRST_D_LOCKlockseq_tas lockseq_tas |
| 1078 | #define KMP_FIRST_I_LOCKlockseq_ticket lockseq_ticket |
| 1079 | #define KMP_LAST_I_LOCKlockseq_nested_drdpa lockseq_nested_drdpa |
| 1080 | #define KMP_NUM_I_LOCKS(locktag_nested_drdpa + 1) \ |
| 1081 | (locktag_nested_drdpa + 1) // number of indirect lock types |
| 1082 | |
| 1083 | // Base type for dynamic locks. |
| 1084 | typedef kmp_uint32 kmp_dyna_lock_t; |
| 1085 | |
| 1086 | // Lock sequence that enumerates all lock kinds. Always make this enumeration |
| 1087 | // consistent with kmp_lockseq_t in the include directory. |
| 1088 | typedef enum { |
| 1089 | lockseq_indirect = 0, |
| 1090 | #define expand_seq(l, a) lockseq_##l, |
| 1091 | KMP_FOREACH_D_LOCK(expand_seq, 0)expand_seq(tas, 0) expand_seq(futex, 0) expand_seq(hle, 0) expand_seq (rtm_spin, 0) KMP_FOREACH_I_LOCK(expand_seq, 0)expand_seq(ticket, 0) expand_seq(queuing, 0) expand_seq(adaptive , 0) expand_seq(drdpa, 0) expand_seq(rtm_queuing, 0) expand_seq (nested_tas, 0) expand_seq(nested_futex, 0) expand_seq(nested_ticket , 0) expand_seq(nested_queuing, 0) expand_seq(nested_drdpa, 0 ) |
| 1092 | #undef expand_seq |
| 1093 | } kmp_dyna_lockseq_t; |
| 1094 | |
| 1095 | // Enumerates indirect lock tags. |
| 1096 | typedef enum { |
| 1097 | #define expand_tag(l, a) locktag_##l, |
| 1098 | KMP_FOREACH_I_LOCK(expand_tag, 0)expand_tag(ticket, 0) expand_tag(queuing, 0) expand_tag(adaptive , 0) expand_tag(drdpa, 0) expand_tag(rtm_queuing, 0) expand_tag (nested_tas, 0) expand_tag(nested_futex, 0) expand_tag(nested_ticket , 0) expand_tag(nested_queuing, 0) expand_tag(nested_drdpa, 0 ) |
| 1099 | #undef expand_tag |
| 1100 | } kmp_indirect_locktag_t; |
| 1101 | |
| 1102 | // Utility macros that extract information from lock sequences. |
| 1103 | #define KMP_IS_D_LOCK(seq)((seq) >= lockseq_tas && (seq) <= lockseq_rtm_spin ) \ |
| 1104 | ((seq) >= KMP_FIRST_D_LOCKlockseq_tas && (seq) <= KMP_LAST_D_LOCKlockseq_rtm_spin) |
| 1105 | #define KMP_IS_I_LOCK(seq)((seq) >= lockseq_ticket && (seq) <= lockseq_nested_drdpa ) \ |
| 1106 | ((seq) >= KMP_FIRST_I_LOCKlockseq_ticket && (seq) <= KMP_LAST_I_LOCKlockseq_nested_drdpa) |
| 1107 | #define KMP_GET_I_TAG(seq)(kmp_indirect_locktag_t)((seq)-lockseq_ticket) (kmp_indirect_locktag_t)((seq)-KMP_FIRST_I_LOCKlockseq_ticket) |
| 1108 | #define KMP_GET_D_TAG(seq)((seq) << 1 | 1) ((seq) << 1 | 1) |
| 1109 | |
| 1110 | // Enumerates direct lock tags starting from indirect tag. |
| 1111 | typedef enum { |
| 1112 | #define expand_tag(l, a) locktag_##l = KMP_GET_D_TAG(lockseq_##l)((lockseq_##l) << 1 | 1), |
| 1113 | KMP_FOREACH_D_LOCK(expand_tag, 0)expand_tag(tas, 0) expand_tag(futex, 0) expand_tag(hle, 0) expand_tag (rtm_spin, 0) |
| 1114 | #undef expand_tag |
| 1115 | } kmp_direct_locktag_t; |
| 1116 | |
| 1117 | // Indirect lock type |
| 1118 | typedef struct { |
| 1119 | kmp_user_lock_p lock; |
| 1120 | kmp_indirect_locktag_t type; |
| 1121 | } kmp_indirect_lock_t; |
| 1122 | |
| 1123 | // Function tables for direct locks. Set/unset/test differentiate functions |
| 1124 | // with/without consistency checking. |
| 1125 | extern void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t); |
| 1126 | extern void (**__kmp_direct_destroy)(kmp_dyna_lock_t *); |
| 1127 | extern int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32); |
| 1128 | extern int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32); |
| 1129 | extern int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32); |
| 1130 | |
| 1131 | // Function tables for indirect locks. Set/unset/test differentiate functions |
| 1132 | // with/without consistency checking. |
| 1133 | extern void (*__kmp_indirect_init[])(kmp_user_lock_p); |
| 1134 | extern void (**__kmp_indirect_destroy)(kmp_user_lock_p); |
| 1135 | extern int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32); |
| 1136 | extern int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32); |
| 1137 | extern int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32); |
| 1138 | |
| 1139 | // Extracts direct lock tag from a user lock pointer |
| 1140 | #define KMP_EXTRACT_D_TAG(l)(*((kmp_dyna_lock_t *)(l)) & ((1 << 8) - 1) & - (*((kmp_dyna_lock_t *)(l)) & 1)) \ |
| 1141 | (*((kmp_dyna_lock_t *)(l)) & ((1 << KMP_LOCK_SHIFT8) - 1) & \ |
| 1142 | -(*((kmp_dyna_lock_t *)(l)) & 1)) |
| 1143 | |
| 1144 | // Extracts indirect lock index from a user lock pointer |
| 1145 | #define KMP_EXTRACT_I_INDEX(l)(*(kmp_lock_index_t *)(l) >> 1) (*(kmp_lock_index_t *)(l) >> 1) |
| 1146 | |
| 1147 | // Returns function pointer to the direct lock function with l (kmp_dyna_lock_t |
| 1148 | // *) and op (operation type). |
| 1149 | #define KMP_D_LOCK_FUNC(l, op)__kmp_direct_op[(*((kmp_dyna_lock_t *)(l)) & ((1 << 8) - 1) & -(*((kmp_dyna_lock_t *)(l)) & 1))] __kmp_direct_##op[KMP_EXTRACT_D_TAG(l)(*((kmp_dyna_lock_t *)(l)) & ((1 << 8) - 1) & - (*((kmp_dyna_lock_t *)(l)) & 1))] |
| 1150 | |
| 1151 | // Returns function pointer to the indirect lock function with l |
| 1152 | // (kmp_indirect_lock_t *) and op (operation type). |
| 1153 | #define KMP_I_LOCK_FUNC(l, op)__kmp_indirect_op[((kmp_indirect_lock_t *)(l))->type] \ |
| 1154 | __kmp_indirect_##op[((kmp_indirect_lock_t *)(l))->type] |
| 1155 | |
| 1156 | // Initializes a direct lock with the given lock pointer and lock sequence. |
| 1157 | #define KMP_INIT_D_LOCK(l, seq)__kmp_direct_init[((seq) << 1 | 1)]((kmp_dyna_lock_t *) l, seq) \ |
| 1158 | __kmp_direct_init[KMP_GET_D_TAG(seq)((seq) << 1 | 1)]((kmp_dyna_lock_t *)l, seq) |
| 1159 | |
| 1160 | // Initializes an indirect lock with the given lock pointer and lock sequence. |
| 1161 | #define KMP_INIT_I_LOCK(l, seq)__kmp_direct_init[0]((kmp_dyna_lock_t *)(l), seq) \ |
| 1162 | __kmp_direct_init[0]((kmp_dyna_lock_t *)(l), seq) |
| 1163 | |
| 1164 | // Returns "free" lock value for the given lock type. |
| 1165 | #define KMP_LOCK_FREE(type)(locktag_type) (locktag_##type) |
| 1166 | |
| 1167 | // Returns "busy" lock value for the given lock teyp. |
| 1168 | #define KMP_LOCK_BUSY(v, type)((v) << 8 | locktag_type) ((v) << KMP_LOCK_SHIFT8 | locktag_##type) |
| 1169 | |
| 1170 | // Returns lock value after removing (shifting) lock tag. |
| 1171 | #define KMP_LOCK_STRIP(v)((v) >> 8) ((v) >> KMP_LOCK_SHIFT8) |
| 1172 | |
| 1173 | // Initializes global states and data structures for managing dynamic user |
| 1174 | // locks. |
| 1175 | extern void __kmp_init_dynamic_user_locks(); |
| 1176 | |
| 1177 | // Allocates and returns an indirect lock with the given indirect lock tag. |
| 1178 | extern kmp_indirect_lock_t * |
| 1179 | __kmp_allocate_indirect_lock(void **, kmp_int32, kmp_indirect_locktag_t); |
| 1180 | |
| 1181 | // Cleans up global states and data structures for managing dynamic user locks. |
| 1182 | extern void __kmp_cleanup_indirect_user_locks(); |
| 1183 | |
| 1184 | // Default user lock sequence when not using hinted locks. |
| 1185 | extern kmp_dyna_lockseq_t __kmp_user_lock_seq; |
| 1186 | |
| 1187 | // Jump table for "set lock location", available only for indirect locks. |
| 1188 | extern void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS(locktag_nested_drdpa + 1)])(kmp_user_lock_p, |
| 1189 | const ident_t *); |
| 1190 | #define KMP_SET_I_LOCK_LOCATION(lck, loc){ if (__kmp_indirect_set_location[(lck)->type] != __null) __kmp_indirect_set_location [(lck)->type]((lck)->lock, loc); } \ |
| 1191 | { \ |
| 1192 | if (__kmp_indirect_set_location[(lck)->type] != NULL__null) \ |
| 1193 | __kmp_indirect_set_location[(lck)->type]((lck)->lock, loc); \ |
| 1194 | } |
| 1195 | |
| 1196 | // Jump table for "set lock flags", available only for indirect locks. |
| 1197 | extern void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS(locktag_nested_drdpa + 1)])(kmp_user_lock_p, |
| 1198 | kmp_lock_flags_t); |
| 1199 | #define KMP_SET_I_LOCK_FLAGS(lck, flag){ if (__kmp_indirect_set_flags[(lck)->type] != __null) __kmp_indirect_set_flags [(lck)->type]((lck)->lock, flag); } \ |
| 1200 | { \ |
| 1201 | if (__kmp_indirect_set_flags[(lck)->type] != NULL__null) \ |
| 1202 | __kmp_indirect_set_flags[(lck)->type]((lck)->lock, flag); \ |
| 1203 | } |
| 1204 | |
| 1205 | // Jump table for "get lock location", available only for indirect locks. |
| 1206 | extern const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS(locktag_nested_drdpa + 1)])( |
| 1207 | kmp_user_lock_p); |
| 1208 | #define KMP_GET_I_LOCK_LOCATION(lck)(__kmp_indirect_get_location[(lck)->type] != __null ? __kmp_indirect_get_location [(lck)->type]((lck)->lock) : __null) \ |
| 1209 | (__kmp_indirect_get_location[(lck)->type] != NULL__null \ |
| 1210 | ? __kmp_indirect_get_location[(lck)->type]((lck)->lock) \ |
| 1211 | : NULL__null) |
| 1212 | |
| 1213 | // Jump table for "get lock flags", available only for indirect locks. |
| 1214 | extern kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS(locktag_nested_drdpa + 1)])( |
| 1215 | kmp_user_lock_p); |
| 1216 | #define KMP_GET_I_LOCK_FLAGS(lck)(__kmp_indirect_get_flags[(lck)->type] != __null ? __kmp_indirect_get_flags [(lck)->type]((lck)->lock) : __null) \ |
| 1217 | (__kmp_indirect_get_flags[(lck)->type] != NULL__null \ |
| 1218 | ? __kmp_indirect_get_flags[(lck)->type]((lck)->lock) \ |
| 1219 | : NULL__null) |
| 1220 | |
| 1221 | // number of kmp_indirect_lock_t objects to be allocated together |
| 1222 | #define KMP_I_LOCK_CHUNK1024 1024 |
| 1223 | // Keep at a power of 2 since it is used in multiplication & division |
| 1224 | KMP_BUILD_ASSERT(KMP_I_LOCK_CHUNK % 2 == 0)static_assert(1024 % 2 == 0, "Build condition error"); |
| 1225 | // number of row entries in the initial lock table |
| 1226 | #define KMP_I_LOCK_TABLE_INIT_NROW_PTRS8 8 |
| 1227 | |
| 1228 | // Lock table for indirect locks. |
| 1229 | typedef struct kmp_indirect_lock_table { |
| 1230 | kmp_indirect_lock_t **table; // blocks of indirect locks allocated |
| 1231 | kmp_uint32 nrow_ptrs; // number *table pointer entries in table |
| 1232 | kmp_lock_index_t next; // index to the next lock to be allocated |
| 1233 | struct kmp_indirect_lock_table *next_table; |
| 1234 | } kmp_indirect_lock_table_t; |
| 1235 | |
| 1236 | extern kmp_indirect_lock_table_t __kmp_i_lock_table; |
| 1237 | |
| 1238 | // Returns the indirect lock associated with the given index. |
| 1239 | // Returns nullptr if no lock at given index |
| 1240 | static inline kmp_indirect_lock_t *__kmp_get_i_lock(kmp_lock_index_t idx) { |
| 1241 | kmp_indirect_lock_table_t *lock_table = &__kmp_i_lock_table; |
| 1242 | while (lock_table) { |
| 1243 | kmp_lock_index_t max_locks = lock_table->nrow_ptrs * KMP_I_LOCK_CHUNK1024; |
| 1244 | if (idx < max_locks) { |
| 1245 | kmp_lock_index_t row = idx / KMP_I_LOCK_CHUNK1024; |
| 1246 | kmp_lock_index_t col = idx % KMP_I_LOCK_CHUNK1024; |
| 1247 | if (!lock_table->table[row] || idx >= lock_table->next) |
| 1248 | break; |
| 1249 | return &lock_table->table[row][col]; |
| 1250 | } |
| 1251 | idx -= max_locks; |
| 1252 | lock_table = lock_table->next_table; |
| 1253 | } |
| 1254 | return nullptr; |
| 1255 | } |
| 1256 | |
| 1257 | // Number of locks in a lock block, which is fixed to "1" now. |
| 1258 | // TODO: No lock block implementation now. If we do support, we need to manage |
| 1259 | // lock block data structure for each indirect lock type. |
| 1260 | extern int __kmp_num_locks_in_block; |
| 1261 | |
| 1262 | // Fast lock table lookup without consistency checking |
| 1263 | #define KMP_LOOKUP_I_LOCK(l)((sizeof(int) < sizeof(void *)) ? __kmp_get_i_lock((*(kmp_lock_index_t *)(l) >> 1)) : *((kmp_indirect_lock_t **)(l))) \ |
| 1264 | ((OMP_LOCK_T_SIZEsizeof(int) < sizeof(void *)) \ |
| 1265 | ? __kmp_get_i_lock(KMP_EXTRACT_I_INDEX(l)(*(kmp_lock_index_t *)(l) >> 1)) \ |
| 1266 | : *((kmp_indirect_lock_t **)(l))) |
| 1267 | |
| 1268 | // Used once in kmp_error.cpp |
| 1269 | extern kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p, kmp_uint32); |
| 1270 | |
| 1271 | #else // KMP_USE_DYNAMIC_LOCK |
| 1272 | |
| 1273 | #define KMP_LOCK_BUSY(v, type)((v) << 8 | locktag_type) (v) |
| 1274 | #define KMP_LOCK_FREE(type)(locktag_type) 0 |
| 1275 | #define KMP_LOCK_STRIP(v)((v) >> 8) (v) |
| 1276 | |
| 1277 | #endif // KMP_USE_DYNAMIC_LOCK |
| 1278 | |
| 1279 | // data structure for using backoff within spin locks. |
| 1280 | typedef struct { |
| 1281 | kmp_uint32 step; // current step |
| 1282 | kmp_uint32 max_backoff; // upper bound of outer delay loop |
| 1283 | kmp_uint32 min_tick; // size of inner delay loop in ticks (machine-dependent) |
| 1284 | } kmp_backoff_t; |
| 1285 | |
| 1286 | // Runtime's default backoff parameters |
| 1287 | extern kmp_backoff_t __kmp_spin_backoff_params; |
| 1288 | |
| 1289 | // Backoff function |
| 1290 | extern void __kmp_spin_backoff(kmp_backoff_t *); |
| 1291 | |
| 1292 | #ifdef __cplusplus201703L |
| 1293 | } // extern "C" |
| 1294 | #endif // __cplusplus |
| 1295 | |
| 1296 | #endif /* KMP_LOCK_H */ |