| File: | projects/openmp/runtime/src/kmp_tasking.cpp |
| Warning: | line 1430, column 65 The right operand of '-' is a garbage value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* | |||
| 2 | * kmp_tasking.cpp -- OpenMP 3.0 tasking support. | |||
| 3 | */ | |||
| 4 | ||||
| 5 | //===----------------------------------------------------------------------===// | |||
| 6 | // | |||
| 7 | // The LLVM Compiler Infrastructure | |||
| 8 | // | |||
| 9 | // This file is dual licensed under the MIT and the University of Illinois Open | |||
| 10 | // Source Licenses. See LICENSE.txt for details. | |||
| 11 | // | |||
| 12 | //===----------------------------------------------------------------------===// | |||
| 13 | ||||
| 14 | #include "kmp.h" | |||
| 15 | #include "kmp_i18n.h" | |||
| 16 | #include "kmp_itt.h" | |||
| 17 | #include "kmp_stats.h" | |||
| 18 | #include "kmp_wait_release.h" | |||
| 19 | ||||
| 20 | #if OMPT_SUPPORT1 | |||
| 21 | #include "ompt-specific.h" | |||
| 22 | #endif | |||
| 23 | ||||
| 24 | #include "tsan_annotations.h" | |||
| 25 | ||||
| 26 | /* forward declaration */ | |||
| 27 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
| 28 | kmp_info_t *this_thr); | |||
| 29 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
| 30 | kmp_thread_data_t *thread_data); | |||
| 31 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
| 32 | kmp_task_team_t *task_team); | |||
| 33 | ||||
| 34 | #ifdef OMP_45_ENABLED(50 >= 45) | |||
| 35 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask); | |||
| 36 | #endif | |||
| 37 | ||||
| 38 | #ifdef BUILD_TIED_TASK_STACK | |||
| 39 | ||||
| 40 | // __kmp_trace_task_stack: print the tied tasks from the task stack in order | |||
| 41 | // from top do bottom | |||
| 42 | // | |||
| 43 | // gtid: global thread identifier for thread containing stack | |||
| 44 | // thread_data: thread data for task team thread containing stack | |||
| 45 | // threshold: value above which the trace statement triggers | |||
| 46 | // location: string identifying call site of this function (for trace) | |||
| 47 | static void __kmp_trace_task_stack(kmp_int32 gtid, | |||
| 48 | kmp_thread_data_t *thread_data, | |||
| 49 | int threshold, char *location) { | |||
| 50 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 51 | kmp_taskdata_t **stack_top = task_stack->ts_top; | |||
| 52 | kmp_int32 entries = task_stack->ts_entries; | |||
| 53 | kmp_taskdata_t *tied_task; | |||
| 54 | ||||
| 55 | KA_TRACE(if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
| 56 | threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
| 57 | ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, "if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
| 58 | "first_block = %p, stack_top = %p \n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); } | |||
| 59 | location, gtid, entries, task_stack->ts_first_block, stack_top))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, " "first_block = %p, stack_top = %p \n", location, gtid, entries , task_stack->ts_first_block, stack_top); }; | |||
| 60 | ||||
| 61 | KMP_DEBUG_ASSERT(stack_top != NULL)((stack_top != __null) ? 0 : __kmp_debug_assert("stack_top != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 61)); | |||
| 62 | KMP_DEBUG_ASSERT(entries > 0)((entries > 0) ? 0 : __kmp_debug_assert("entries > 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 62)); | |||
| 63 | ||||
| 64 | while (entries != 0) { | |||
| 65 | KMP_DEBUG_ASSERT(stack_top != &task_stack->ts_first_block.sb_block[0])((stack_top != &task_stack->ts_first_block.sb_block[0] ) ? 0 : __kmp_debug_assert("stack_top != &task_stack->ts_first_block.sb_block[0]" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 65)); | |||
| 66 | // fix up ts_top if we need to pop from previous block | |||
| 67 | if (entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 68 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(stack_top); | |||
| 69 | ||||
| 70 | stack_block = stack_block->sb_prev; | |||
| 71 | stack_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
| 72 | } | |||
| 73 | ||||
| 74 | // finish bookkeeping | |||
| 75 | stack_top--; | |||
| 76 | entries--; | |||
| 77 | ||||
| 78 | tied_task = *stack_top; | |||
| 79 | ||||
| 80 | KMP_DEBUG_ASSERT(tied_task != NULL)((tied_task != __null) ? 0 : __kmp_debug_assert("tied_task != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 80)); | |||
| 81 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 81)); | |||
| 82 | ||||
| 83 | KA_TRACE(threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
| 84 | ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, "if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
| 85 | "stack_top=%p, tied_task=%p\n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); } | |||
| 86 | location, gtid, entries, stack_top, tied_task))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(%s): gtid=%d, entry=%d, " "stack_top=%p, tied_task=%p\n", location, gtid, entries, stack_top , tied_task); }; | |||
| 87 | } | |||
| 88 | KMP_DEBUG_ASSERT(stack_top == &task_stack->ts_first_block.sb_block[0])((stack_top == &task_stack->ts_first_block.sb_block[0] ) ? 0 : __kmp_debug_assert("stack_top == &task_stack->ts_first_block.sb_block[0]" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 88)); | |||
| 89 | ||||
| 90 | KA_TRACE(threshold,if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); } | |||
| 91 | ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n",if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); } | |||
| 92 | location, gtid))if (kmp_a_debug >= threshold) { __kmp_debug_printf ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n" , location, gtid); }; | |||
| 93 | } | |||
| 94 | ||||
| 95 | // __kmp_init_task_stack: initialize the task stack for the first time | |||
| 96 | // after a thread_data structure is created. | |||
| 97 | // It should not be necessary to do this again (assuming the stack works). | |||
| 98 | // | |||
| 99 | // gtid: global thread identifier of calling thread | |||
| 100 | // thread_data: thread data for task team thread containing stack | |||
| 101 | static void __kmp_init_task_stack(kmp_int32 gtid, | |||
| 102 | kmp_thread_data_t *thread_data) { | |||
| 103 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 104 | kmp_stack_block_t *first_block; | |||
| 105 | ||||
| 106 | // set up the first block of the stack | |||
| 107 | first_block = &task_stack->ts_first_block; | |||
| 108 | task_stack->ts_top = (kmp_taskdata_t **)first_block; | |||
| 109 | memset((void *)first_block, '\0', | |||
| 110 | TASK_STACK_BLOCK_SIZE * sizeof(kmp_taskdata_t *)); | |||
| 111 | ||||
| 112 | // initialize the stack to be empty | |||
| 113 | task_stack->ts_entries = TASK_STACK_EMPTY; | |||
| 114 | first_block->sb_next = NULL__null; | |||
| 115 | first_block->sb_prev = NULL__null; | |||
| 116 | } | |||
| 117 | ||||
| 118 | // __kmp_free_task_stack: free the task stack when thread_data is destroyed. | |||
| 119 | // | |||
| 120 | // gtid: global thread identifier for calling thread | |||
| 121 | // thread_data: thread info for thread containing stack | |||
| 122 | static void __kmp_free_task_stack(kmp_int32 gtid, | |||
| 123 | kmp_thread_data_t *thread_data) { | |||
| 124 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 125 | kmp_stack_block_t *stack_block = &task_stack->ts_first_block; | |||
| 126 | ||||
| 127 | KMP_DEBUG_ASSERT(task_stack->ts_entries == TASK_STACK_EMPTY)((task_stack->ts_entries == TASK_STACK_EMPTY) ? 0 : __kmp_debug_assert ("task_stack->ts_entries == TASK_STACK_EMPTY", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 127)); | |||
| 128 | // free from the second block of the stack | |||
| 129 | while (stack_block != NULL__null) { | |||
| 130 | kmp_stack_block_t *next_block = (stack_block) ? stack_block->sb_next : NULL__null; | |||
| 131 | ||||
| 132 | stack_block->sb_next = NULL__null; | |||
| 133 | stack_block->sb_prev = NULL__null; | |||
| 134 | if (stack_block != &task_stack->ts_first_block) { | |||
| 135 | __kmp_thread_free(thread,___kmp_thread_free((thread), (stack_block), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 136) | |||
| 136 | stack_block)___kmp_thread_free((thread), (stack_block), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 136); // free the block, if not the first | |||
| 137 | } | |||
| 138 | stack_block = next_block; | |||
| 139 | } | |||
| 140 | // initialize the stack to be empty | |||
| 141 | task_stack->ts_entries = 0; | |||
| 142 | task_stack->ts_top = NULL__null; | |||
| 143 | } | |||
| 144 | ||||
| 145 | // __kmp_push_task_stack: Push the tied task onto the task stack. | |||
| 146 | // Grow the stack if necessary by allocating another block. | |||
| 147 | // | |||
| 148 | // gtid: global thread identifier for calling thread | |||
| 149 | // thread: thread info for thread containing stack | |||
| 150 | // tied_task: the task to push on the stack | |||
| 151 | static void __kmp_push_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
| 152 | kmp_taskdata_t *tied_task) { | |||
| 153 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
| 154 | kmp_thread_data_t *thread_data = | |||
| 155 | &thread->th.th_task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 155)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
| 156 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 157 | ||||
| 158 | if (tied_task->td_flags.team_serial || tied_task->td_flags.tasking_ser) { | |||
| 159 | return; // Don't push anything on stack if team or team tasks are serialized | |||
| 160 | } | |||
| 161 | ||||
| 162 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 162)); | |||
| 163 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)((task_stack->ts_top != __null) ? 0 : __kmp_debug_assert("task_stack->ts_top != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 163)); | |||
| 164 | ||||
| 165 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); } | |||
| 166 | ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); } | |||
| 167 | gtid, thread, tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n" , gtid, thread, tied_task); }; | |||
| 168 | // Store entry | |||
| 169 | *(task_stack->ts_top) = tied_task; | |||
| 170 | ||||
| 171 | // Do bookkeeping for next push | |||
| 172 | task_stack->ts_top++; | |||
| 173 | task_stack->ts_entries++; | |||
| 174 | ||||
| 175 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 176 | // Find beginning of this task block | |||
| 177 | kmp_stack_block_t *stack_block = | |||
| 178 | (kmp_stack_block_t *)(task_stack->ts_top - TASK_STACK_BLOCK_SIZE); | |||
| 179 | ||||
| 180 | // Check if we already have a block | |||
| 181 | if (stack_block->sb_next != | |||
| 182 | NULL__null) { // reset ts_top to beginning of next block | |||
| 183 | task_stack->ts_top = &stack_block->sb_next->sb_block[0]; | |||
| 184 | } else { // Alloc new block and link it up | |||
| 185 | kmp_stack_block_t *new_block = (kmp_stack_block_t *)__kmp_thread_calloc( | |||
| 186 | thread, sizeof(kmp_stack_block_t)); | |||
| 187 | ||||
| 188 | task_stack->ts_top = &new_block->sb_block[0]; | |||
| 189 | stack_block->sb_next = new_block; | |||
| 190 | new_block->sb_prev = stack_block; | |||
| 191 | new_block->sb_next = NULL__null; | |||
| 192 | ||||
| 193 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
| 194 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
| 195 | ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n",if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); } | |||
| 196 | gtid, tied_task, new_block))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n" , gtid, tied_task, new_block); }; | |||
| 197 | } | |||
| 198 | } | |||
| 199 | KA_TRACE(20, ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); } | |||
| 200 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
| 201 | } | |||
| 202 | ||||
| 203 | // __kmp_pop_task_stack: Pop the tied task from the task stack. Don't return | |||
| 204 | // the task, just check to make sure it matches the ending task passed in. | |||
| 205 | // | |||
| 206 | // gtid: global thread identifier for the calling thread | |||
| 207 | // thread: thread info structure containing stack | |||
| 208 | // tied_task: the task popped off the stack | |||
| 209 | // ending_task: the task that is ending (should match popped task) | |||
| 210 | static void __kmp_pop_task_stack(kmp_int32 gtid, kmp_info_t *thread, | |||
| 211 | kmp_taskdata_t *ending_task) { | |||
| 212 | // GEH - need to consider what to do if tt_threads_data not allocated yet | |||
| 213 | kmp_thread_data_t *thread_data = | |||
| 214 | &thread->th.th_task_team->tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 214)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
| 215 | kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks; | |||
| 216 | kmp_taskdata_t *tied_task; | |||
| 217 | ||||
| 218 | if (ending_task->td_flags.team_serial || ending_task->td_flags.tasking_ser) { | |||
| 219 | // Don't pop anything from stack if team or team tasks are serialized | |||
| 220 | return; | |||
| 221 | } | |||
| 222 | ||||
| 223 | KMP_DEBUG_ASSERT(task_stack->ts_top != NULL)((task_stack->ts_top != __null) ? 0 : __kmp_debug_assert("task_stack->ts_top != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 223)); | |||
| 224 | KMP_DEBUG_ASSERT(task_stack->ts_entries > 0)((task_stack->ts_entries > 0) ? 0 : __kmp_debug_assert( "task_stack->ts_entries > 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 224)); | |||
| 225 | ||||
| 226 | KA_TRACE(20, ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n" , gtid, thread); } | |||
| 227 | thread))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n" , gtid, thread); }; | |||
| 228 | ||||
| 229 | // fix up ts_top if we need to pop from previous block | |||
| 230 | if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) { | |||
| 231 | kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(task_stack->ts_top); | |||
| 232 | ||||
| 233 | stack_block = stack_block->sb_prev; | |||
| 234 | task_stack->ts_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE]; | |||
| 235 | } | |||
| 236 | ||||
| 237 | // finish bookkeeping | |||
| 238 | task_stack->ts_top--; | |||
| 239 | task_stack->ts_entries--; | |||
| 240 | ||||
| 241 | tied_task = *(task_stack->ts_top); | |||
| 242 | ||||
| 243 | KMP_DEBUG_ASSERT(tied_task != NULL)((tied_task != __null) ? 0 : __kmp_debug_assert("tied_task != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 243)); | |||
| 244 | KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED)((tied_task->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("tied_task->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 244)); | |||
| 245 | KMP_DEBUG_ASSERT(tied_task == ending_task)((tied_task == ending_task) ? 0 : __kmp_debug_assert("tied_task == ending_task" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 245)); // If we built the stack correctly | |||
| 246 | ||||
| 247 | KA_TRACE(20, ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n", gtid,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); } | |||
| 248 | tied_task))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n" , gtid, tied_task); }; | |||
| 249 | return; | |||
| 250 | } | |||
| 251 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 252 | ||||
| 253 | // __kmp_push_task: Add a task to the thread's deque | |||
| 254 | static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) { | |||
| 255 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 256 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 257 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 258 | kmp_int32 tid = __kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 258)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid); | |||
| 259 | kmp_thread_data_t *thread_data; | |||
| 260 | ||||
| 261 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d trying to push task %p.\n" , gtid, taskdata); } | |||
| 262 | ("__kmp_push_task: T#%d trying to push task %p.\n", gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d trying to push task %p.\n" , gtid, taskdata); }; | |||
| 263 | ||||
| 264 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
| 265 | // untied task needs to increment counter so that the task structure is not | |||
| 266 | // freed prematurely | |||
| 267 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 268 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
| 269 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
| 270 | ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); } | |||
| 271 | gtid, counter, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n" , gtid, counter, taskdata); }; | |||
| 272 | } | |||
| 273 | ||||
| 274 | // The first check avoids building task_team thread data if serialized | |||
| 275 | if (taskdata->td_flags.task_serial) { | |||
| 276 | KA_TRACE(20, ("__kmp_push_task: T#%d team serialized; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 277 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 278 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d team serialized; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
| 279 | return TASK_NOT_PUSHED1; | |||
| 280 | } | |||
| 281 | ||||
| 282 | // Now that serialized tasks have returned, we can assume that we are not in | |||
| 283 | // immediate exec mode | |||
| 284 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 284)); | |||
| 285 | if (!KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
| 286 | __kmp_enable_tasking(task_team, thread); | |||
| 287 | } | |||
| 288 | KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE)(((task_team->tt.tt_found_tasks) == (!0)) ? 0 : __kmp_debug_assert ("(task_team->tt.tt_found_tasks) == (!0)", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 288)); | |||
| 289 | KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL)((((void *)(task_team->tt.tt_threads_data)) != __null) ? 0 : __kmp_debug_assert("((void *)(task_team->tt.tt_threads_data)) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 289)); | |||
| 290 | ||||
| 291 | // Find tasking deque specific to encountering thread | |||
| 292 | thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 293 | ||||
| 294 | // No lock needed since only owner can allocate | |||
| 295 | if (thread_data->td.td_deque == NULL__null) { | |||
| 296 | __kmp_alloc_task_deque(thread, thread_data); | |||
| 297 | } | |||
| 298 | ||||
| 299 | // Check if deque is full | |||
| 300 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 301 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 302 | KA_TRACE(20, ("__kmp_push_task: T#%d deque is full; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 303 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 304 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
| 305 | return TASK_NOT_PUSHED1; | |||
| 306 | } | |||
| 307 | ||||
| 308 | // Lock the deque for the task push operation | |||
| 309 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 310 | ||||
| 311 | #if OMP_45_ENABLED(50 >= 45) | |||
| 312 | // Need to recheck as we can get a proxy task from a thread outside of OpenMP | |||
| 313 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 314 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 315 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 316 | KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; returning "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 317 | "TASK_NOT_PUSHED for task %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); } | |||
| 318 | gtid, taskdata))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d deque is full on 2nd check; returning " "TASK_NOT_PUSHED for task %p\n", gtid, taskdata); }; | |||
| 319 | return TASK_NOT_PUSHED1; | |||
| 320 | } | |||
| 321 | #else | |||
| 322 | // Must have room since no thread can add tasks but calling thread | |||
| 323 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <(((thread_data->td.td_deque_ntasks) < ((thread_data-> td).td_deque_size)) ? 0 : __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 324)) | |||
| 324 | TASK_DEQUE_SIZE(thread_data->td))(((thread_data->td.td_deque_ntasks) < ((thread_data-> td).td_deque_size)) ? 0 : __kmp_debug_assert("(thread_data->td.td_deque_ntasks) < ((thread_data->td).td_deque_size)" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 324)); | |||
| 325 | #endif | |||
| 326 | ||||
| 327 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = | |||
| 328 | taskdata; // Push taskdata | |||
| 329 | // Wrap index. | |||
| 330 | thread_data->td.td_deque_tail = | |||
| 331 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 332 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
| 333 | TCR_4(thread_data->td.td_deque_ntasks) + 1)(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1); // Adjust task count | |||
| 334 | ||||
| 335 | KA_TRACE(20, ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 336 | "task=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 337 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 338 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: " "task=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data ->td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); }; | |||
| 339 | ||||
| 340 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 341 | ||||
| 342 | return TASK_SUCCESSFULLY_PUSHED0; | |||
| 343 | } | |||
| 344 | ||||
| 345 | // __kmp_pop_current_task_from_thread: set up current task from called thread | |||
| 346 | // when team ends | |||
| 347 | // | |||
| 348 | // this_thr: thread structure to set current_task in. | |||
| 349 | void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) { | |||
| 350 | KF_TRACE(10, ("__kmp_pop_current_task_from_thread(enter): T#%d "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 351 | "this_thread=%p, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 352 | "curtask_parent=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 353 | 0, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 354 | this_thr->th.th_current_task->td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(enter): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); }; | |||
| 355 | ||||
| 356 | this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent; | |||
| 357 | ||||
| 358 | KF_TRACE(10, ("__kmp_pop_current_task_from_thread(exit): T#%d "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 359 | "this_thread=%p, curtask=%p, "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 360 | "curtask_parent=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 361 | 0, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); } | |||
| 362 | this_thr->th.th_current_task->td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_pop_current_task_from_thread(exit): T#%d " "this_thread=%p, curtask=%p, " "curtask_parent=%p\n", 0, this_thr , this_thr->th.th_current_task, this_thr->th.th_current_task ->td_parent); }; | |||
| 363 | } | |||
| 364 | ||||
| 365 | // __kmp_push_current_task_to_thread: set up current task in called thread for a | |||
| 366 | // new team | |||
| 367 | // | |||
| 368 | // this_thr: thread structure to set up | |||
| 369 | // team: team for implicit task data | |||
| 370 | // tid: thread within team to set up | |||
| 371 | void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team, | |||
| 372 | int tid) { | |||
| 373 | // current task of the thread is a parent of the new just created implicit | |||
| 374 | // tasks of new team | |||
| 375 | KF_TRACE(10, ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 376 | "curtask=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 377 | "parent_task=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 378 | tid, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 379 | team->t.t_implicit_task_taskdata[tid].td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); }; | |||
| 380 | ||||
| 381 | KMP_DEBUG_ASSERT(this_thr != NULL)((this_thr != __null) ? 0 : __kmp_debug_assert("this_thr != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 381)); | |||
| 382 | ||||
| 383 | if (tid == 0) { | |||
| 384 | if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) { | |||
| 385 | team->t.t_implicit_task_taskdata[0].td_parent = | |||
| 386 | this_thr->th.th_current_task; | |||
| 387 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0]; | |||
| 388 | } | |||
| 389 | } else { | |||
| 390 | team->t.t_implicit_task_taskdata[tid].td_parent = | |||
| 391 | team->t.t_implicit_task_taskdata[0].td_parent; | |||
| 392 | this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid]; | |||
| 393 | } | |||
| 394 | ||||
| 395 | KF_TRACE(10, ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 396 | "curtask=%p "if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 397 | "parent_task=%p\n",if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 398 | tid, this_thr, this_thr->th.th_current_task,if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); } | |||
| 399 | team->t.t_implicit_task_taskdata[tid].td_parent))if (kmp_f_debug >= 10) { __kmp_debug_printf ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p " "curtask=%p " "parent_task=%p\n", tid, this_thr, this_thr-> th.th_current_task, team->t.t_implicit_task_taskdata[tid]. td_parent); }; | |||
| 400 | } | |||
| 401 | ||||
| 402 | // __kmp_task_start: bookkeeping for a task starting execution | |||
| 403 | // | |||
| 404 | // GTID: global thread id of calling thread | |||
| 405 | // task: task starting execution | |||
| 406 | // current_task: task suspending | |||
| 407 | static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task, | |||
| 408 | kmp_taskdata_t *current_task) { | |||
| 409 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 410 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 411 | ||||
| 412 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); } | |||
| 413 | ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); } | |||
| 414 | gtid, taskdata, current_task))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n" , gtid, taskdata, current_task); }; | |||
| 415 | ||||
| 416 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 416)); | |||
| 417 | ||||
| 418 | // mark currently executing task as suspended | |||
| 419 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
| 420 | // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 ); | |||
| 421 | current_task->td_flags.executing = 0; | |||
| 422 | ||||
| 423 | // Add task to stack if tied | |||
| 424 | #ifdef BUILD_TIED_TASK_STACK | |||
| 425 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 426 | __kmp_push_task_stack(gtid, thread, taskdata); | |||
| 427 | } | |||
| 428 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 429 | ||||
| 430 | // mark starting task as executing and as current task | |||
| 431 | thread->th.th_current_task = taskdata; | |||
| 432 | ||||
| 433 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||((taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 434)) | |||
| 434 | taskdata->td_flags.tiedness == TASK_UNTIED)((taskdata->td_flags.started == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.started == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 434)); | |||
| 435 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||((taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 436)) | |||
| 436 | taskdata->td_flags.tiedness == TASK_UNTIED)((taskdata->td_flags.executing == 0 || taskdata->td_flags .tiedness == 0) ? 0 : __kmp_debug_assert("taskdata->td_flags.executing == 0 || taskdata->td_flags.tiedness == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 436)); | |||
| 437 | taskdata->td_flags.started = 1; | |||
| 438 | taskdata->td_flags.executing = 1; | |||
| 439 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 439)); | |||
| 440 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 440)); | |||
| 441 | ||||
| 442 | // GEH TODO: shouldn't we pass some sort of location identifier here? | |||
| 443 | // APT: yes, we will pass location here. | |||
| 444 | // need to store current thread state (in a thread or taskdata structure) | |||
| 445 | // before setting work_state, otherwise wrong state is set after end of task | |||
| 446 | ||||
| 447 | KA_TRACE(10, ("__kmp_task_start(exit): T#%d task=%p\n", gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_start(exit): T#%d task=%p\n" , gtid, taskdata); }; | |||
| 448 | ||||
| 449 | return; | |||
| 450 | } | |||
| 451 | ||||
| 452 | #if OMPT_SUPPORT1 | |||
| 453 | //------------------------------------------------------------------------------ | |||
| 454 | // __ompt_task_init: | |||
| 455 | // Initialize OMPT fields maintained by a task. This will only be called after | |||
| 456 | // ompt_start_tool, so we already know whether ompt is enabled or not. | |||
| 457 | ||||
| 458 | static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) { | |||
| 459 | // The calls to __ompt_task_init already have the ompt_enabled condition. | |||
| 460 | task->ompt_task_info.task_data.value = 0; | |||
| 461 | task->ompt_task_info.frame.exit_frame = NULL__null; | |||
| 462 | task->ompt_task_info.frame.enter_frame = NULL__null; | |||
| 463 | #if OMP_40_ENABLED(50 >= 40) | |||
| 464 | task->ompt_task_info.ndeps = 0; | |||
| 465 | task->ompt_task_info.deps = NULL__null; | |||
| 466 | #endif /* OMP_40_ENABLED */ | |||
| 467 | } | |||
| 468 | ||||
| 469 | // __ompt_task_start: | |||
| 470 | // Build and trigger task-begin event | |||
| 471 | static inline void __ompt_task_start(kmp_task_t *task, | |||
| 472 | kmp_taskdata_t *current_task, | |||
| 473 | kmp_int32 gtid) { | |||
| 474 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 475 | ompt_task_status_t status = ompt_task_others; | |||
| 476 | if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) { | |||
| 477 | status = ompt_task_yield; | |||
| 478 | __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0; | |||
| 479 | } | |||
| 480 | /* let OMPT know that we're about to run this task */ | |||
| 481 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
| 482 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
| 483 | &(current_task->ompt_task_info.task_data), status, | |||
| 484 | &(taskdata->ompt_task_info.task_data)); | |||
| 485 | } | |||
| 486 | taskdata->ompt_task_info.scheduling_parent = current_task; | |||
| 487 | } | |||
| 488 | ||||
| 489 | // __ompt_task_finish: | |||
| 490 | // Build and trigger final task-schedule event | |||
| 491 | static inline void __ompt_task_finish(kmp_task_t *task, | |||
| 492 | kmp_taskdata_t *resumed_task) { | |||
| 493 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 494 | ompt_task_status_t status = ompt_task_complete; | |||
| 495 | if (taskdata->td_flags.tiedness == TASK_UNTIED0 && | |||
| 496 | KMP_TEST_THEN_ADD32(&(taskdata->td_untied_count), 0)__sync_fetch_and_add((volatile kmp_int32 *)(&(taskdata-> td_untied_count)), (kmp_int32)(0)) > 1) | |||
| 497 | status = ompt_task_others; | |||
| 498 | if (__kmp_omp_cancellation && taskdata->td_taskgroup && | |||
| 499 | taskdata->td_taskgroup->cancel_request == cancel_taskgroup) { | |||
| 500 | status = ompt_task_cancel; | |||
| 501 | } | |||
| 502 | ||||
| 503 | /* let OMPT know that we're returning to the callee task */ | |||
| 504 | if (ompt_enabled.ompt_callback_task_schedule) { | |||
| 505 | ompt_callbacks.ompt_callback(ompt_callback_task_schedule)ompt_callback_task_schedule_callback( | |||
| 506 | &(taskdata->ompt_task_info.task_data), status, | |||
| 507 | &((resumed_task ? resumed_task | |||
| 508 | : (taskdata->ompt_task_info.scheduling_parent | |||
| 509 | ? taskdata->ompt_task_info.scheduling_parent | |||
| 510 | : taskdata->td_parent)) | |||
| 511 | ->ompt_task_info.task_data)); | |||
| 512 | } | |||
| 513 | } | |||
| 514 | #endif | |||
| 515 | ||||
| 516 | template <bool ompt> | |||
| 517 | static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid, | |||
| 518 | kmp_task_t *task, | |||
| 519 | void *frame_address, | |||
| 520 | void *return_address) { | |||
| 521 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 522 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 523 | ||||
| 524 | 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); } | |||
| 525 | "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); } | |||
| 526 | 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); }; | |||
| 527 | ||||
| 528 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
| 529 | // untied task needs to increment counter so that the task structure is not | |||
| 530 | // freed prematurely | |||
| 531 | kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 532 | 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); } | |||
| 533 | "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); } | |||
| 534 | 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); }; | |||
| 535 | } | |||
| 536 | ||||
| 537 | taskdata->td_flags.task_serial = | |||
| 538 | 1; // Execute this task immediately, not deferred. | |||
| 539 | __kmp_task_start(gtid, task, current_task); | |||
| 540 | ||||
| 541 | #if OMPT_SUPPORT1 | |||
| 542 | if (ompt) { | |||
| 543 | if (current_task->ompt_task_info.frame.enter_frame == NULL__null) { | |||
| 544 | current_task->ompt_task_info.frame.enter_frame = | |||
| 545 | taskdata->ompt_task_info.frame.exit_frame = frame_address; | |||
| 546 | } | |||
| 547 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 548 | ompt_task_info_t *parent_info = &(current_task->ompt_task_info); | |||
| 549 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 550 | &(parent_info->task_data), &(parent_info->frame), | |||
| 551 | &(taskdata->ompt_task_info.task_data), | |||
| 552 | 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, | |||
| 553 | return_address); | |||
| 554 | } | |||
| 555 | __ompt_task_start(task, current_task, gtid); | |||
| 556 | } | |||
| 557 | #endif // OMPT_SUPPORT | |||
| 558 | ||||
| 559 | 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); } | |||
| 560 | 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); }; | |||
| 561 | } | |||
| 562 | ||||
| 563 | #if OMPT_SUPPORT1 | |||
| 564 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 565 | static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 566 | kmp_task_t *task, | |||
| 567 | void *frame_address, | |||
| 568 | void *return_address) { | |||
| 569 | __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address, | |||
| 570 | return_address); | |||
| 571 | } | |||
| 572 | #endif // OMPT_SUPPORT | |||
| 573 | ||||
| 574 | // __kmpc_omp_task_begin_if0: report that a given serialized task has started | |||
| 575 | // execution | |||
| 576 | // | |||
| 577 | // loc_ref: source location information; points to beginning of task block. | |||
| 578 | // gtid: global thread number. | |||
| 579 | // task: task thunk for the started task. | |||
| 580 | void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
| 581 | kmp_task_t *task) { | |||
| 582 | #if OMPT_SUPPORT1 | |||
| 583 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 584 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 585 | __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task, | |||
| 586 | OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1), | |||
| 587 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 588 | return; | |||
| 589 | } | |||
| 590 | #endif | |||
| 591 | __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL__null, NULL__null); | |||
| 592 | } | |||
| 593 | ||||
| 594 | #ifdef TASK_UNUSED | |||
| 595 | // __kmpc_omp_task_begin: report that a given task has started execution | |||
| 596 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
| 597 | void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) { | |||
| 598 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 599 | ||||
| 600 | 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 ); } | |||
| 601 | 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 ); } | |||
| 602 | ("__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 ); } | |||
| 603 | 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 ); }; | |||
| 604 | ||||
| 605 | __kmp_task_start(gtid, task, current_task); | |||
| 606 | ||||
| 607 | 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)); } | |||
| 608 | 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)); }; | |||
| 609 | return; | |||
| 610 | } | |||
| 611 | #endif // TASK_UNUSED | |||
| 612 | ||||
| 613 | // __kmp_free_task: free the current task space and the space for shareds | |||
| 614 | // | |||
| 615 | // gtid: Global thread ID of calling thread | |||
| 616 | // taskdata: task to free | |||
| 617 | // thread: thread data structure of caller | |||
| 618 | static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata, | |||
| 619 | kmp_info_t *thread) { | |||
| 620 | 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); } | |||
| 621 | taskdata))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_free_task: T#%d freeing data from task %p\n" , gtid, taskdata); }; | |||
| 622 | ||||
| 623 | // Check to make sure all flags and counters have the correct values | |||
| 624 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 624)); | |||
| 625 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0)((taskdata->td_flags.executing == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.executing == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 625)); | |||
| 626 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1)((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 626)); | |||
| 627 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 627)); | |||
| 628 | KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||((taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1) ? 0 : __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 629)) | |||
| 629 | taskdata->td_flags.task_serial == 1)((taskdata->td_allocated_child_tasks == 0 || taskdata-> td_flags.task_serial == 1) ? 0 : __kmp_debug_assert("taskdata->td_allocated_child_tasks == 0 || taskdata->td_flags.task_serial == 1" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 629)); | |||
| 630 | KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0)((taskdata->td_incomplete_child_tasks == 0) ? 0 : __kmp_debug_assert ("taskdata->td_incomplete_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 630)); | |||
| 631 | ||||
| 632 | taskdata->td_flags.freed = 1; | |||
| 633 | ANNOTATE_HAPPENS_BEFORE(taskdata); | |||
| 634 | // deallocate the taskdata and shared variable blocks associated with this task | |||
| 635 | #if USE_FAST_MEMORY3 | |||
| 636 | __kmp_fast_free(thread, taskdata)___kmp_fast_free((thread), (taskdata), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 636); | |||
| 637 | #else /* ! USE_FAST_MEMORY */ | |||
| 638 | __kmp_thread_free(thread, taskdata)___kmp_thread_free((thread), (taskdata), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 638); | |||
| 639 | #endif | |||
| 640 | ||||
| 641 | 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); }; | |||
| 642 | } | |||
| 643 | ||||
| 644 | // __kmp_free_task_and_ancestors: free the current task and ancestors without | |||
| 645 | // children | |||
| 646 | // | |||
| 647 | // gtid: Global thread ID of calling thread | |||
| 648 | // taskdata: task to free | |||
| 649 | // thread: thread data structure of caller | |||
| 650 | static void __kmp_free_task_and_ancestors(kmp_int32 gtid, | |||
| 651 | kmp_taskdata_t *taskdata, | |||
| 652 | kmp_info_t *thread) { | |||
| 653 | #if OMP_45_ENABLED(50 >= 45) | |||
| 654 | // Proxy tasks must always be allowed to free their parents | |||
| 655 | // because they can be run in background even in serial mode. | |||
| 656 | kmp_int32 team_serial = | |||
| 657 | (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) && | |||
| 658 | !taskdata->td_flags.proxy; | |||
| 659 | #else | |||
| 660 | kmp_int32 team_serial = | |||
| 661 | taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser; | |||
| 662 | #endif | |||
| 663 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 663)); | |||
| 664 | ||||
| 665 | 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; | |||
| 666 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 666)); | |||
| 667 | ||||
| 668 | // Now, go up the ancestor tree to see if any ancestors can now be freed. | |||
| 669 | while (children == 0) { | |||
| 670 | kmp_taskdata_t *parent_taskdata = taskdata->td_parent; | |||
| 671 | ||||
| 672 | 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); } | |||
| 673 | "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); } | |||
| 674 | 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); }; | |||
| 675 | ||||
| 676 | // --- Deallocate my ancestor task --- | |||
| 677 | __kmp_free_task(gtid, taskdata, thread); | |||
| 678 | ||||
| 679 | taskdata = parent_taskdata; | |||
| 680 | ||||
| 681 | // Stop checking ancestors at implicit task instead of walking up ancestor | |||
| 682 | // tree to avoid premature deallocation of ancestors. | |||
| 683 | if (team_serial || taskdata->td_flags.tasktype == TASK_IMPLICIT0) | |||
| 684 | return; | |||
| 685 | ||||
| 686 | // Predecrement simulated by "- 1" calculation | |||
| 687 | children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks)(&taskdata->td_allocated_child_tasks)->fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
| 688 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 688)); | |||
| 689 | } | |||
| 690 | ||||
| 691 | 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); } | |||
| 692 | 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); } | |||
| 693 | "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); } | |||
| 694 | 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); }; | |||
| 695 | } | |||
| 696 | ||||
| 697 | // __kmp_task_finish: bookkeeping to do when a task finishes execution | |||
| 698 | // | |||
| 699 | // gtid: global thread ID for calling thread | |||
| 700 | // task: task to be finished | |||
| 701 | // resumed_task: task to be resumed. (may be NULL if task is serialized) | |||
| 702 | static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task, | |||
| 703 | kmp_taskdata_t *resumed_task) { | |||
| 704 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 705 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 706 | kmp_task_team_t *task_team = | |||
| 707 | thread->th.th_task_team; // might be NULL for serial teams... | |||
| 708 | kmp_int32 children = 0; | |||
| 709 | ||||
| 710 | 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); } | |||
| 711 | "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); } | |||
| 712 | 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); }; | |||
| 713 | ||||
| 714 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 714)); | |||
| 715 | ||||
| 716 | // Pop task from stack if tied | |||
| 717 | #ifdef BUILD_TIED_TASK_STACK | |||
| 718 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 719 | __kmp_pop_task_stack(gtid, thread, taskdata); | |||
| 720 | } | |||
| 721 | #endif /* BUILD_TIED_TASK_STACK */ | |||
| 722 | ||||
| 723 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
| 724 | // untied task needs to check the counter so that the task structure is not | |||
| 725 | // freed prematurely | |||
| 726 | kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count)(&taskdata->td_untied_count)->fetch_sub(1, std::memory_order_acq_rel ) - 1; | |||
| 727 | 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); } | |||
| 728 | 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); } | |||
| 729 | ("__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); } | |||
| 730 | 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); }; | |||
| 731 | if (counter > 0) { | |||
| 732 | // untied task is not done, to be continued possibly by other thread, do | |||
| 733 | // not free it now | |||
| 734 | if (resumed_task == NULL__null) { | |||
| 735 | KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial)((taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("taskdata->td_flags.task_serial", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 735)); | |||
| 736 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
| 737 | // task is the parent | |||
| 738 | } | |||
| 739 | thread->th.th_current_task = resumed_task; // restore current_task | |||
| 740 | resumed_task->td_flags.executing = 1; // resume previous task | |||
| 741 | 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); } | |||
| 742 | "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); } | |||
| 743 | 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); }; | |||
| 744 | return; | |||
| 745 | } | |||
| 746 | } | |||
| 747 | ||||
| 748 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 748)); | |||
| 749 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
| 750 | KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1)((taskdata->td_flags.started == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.started == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 750)); | |||
| 751 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 751)); | |||
| 752 | ||||
| 753 | // Only need to keep track of count if team parallel and tasking not | |||
| 754 | // serialized | |||
| 755 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) { | |||
| 756 | // Predecrement simulated by "- 1" calculation | |||
| 757 | children = | |||
| 758 | KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks)(&taskdata->td_parent->td_incomplete_child_tasks)-> fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
| 759 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 759)); | |||
| 760 | #if OMP_40_ENABLED(50 >= 40) | |||
| 761 | if (taskdata->td_taskgroup) | |||
| 762 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
| 763 | #if OMP_45_ENABLED(50 >= 45) | |||
| 764 | } | |||
| 765 | // if we found proxy tasks there could exist a dependency chain | |||
| 766 | // with the proxy task as origin | |||
| 767 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) || | |||
| 768 | (task_team && task_team->tt.tt_found_proxy_tasks)) { | |||
| 769 | #endif | |||
| 770 | __kmp_release_deps(gtid, taskdata); | |||
| 771 | #endif | |||
| 772 | } | |||
| 773 | ||||
| 774 | // td_flags.executing must be marked as 0 after __kmp_release_deps has been | |||
| 775 | // called. Othertwise, if a task is executed immediately from the release_deps | |||
| 776 | // code, the flag will be reset to 1 again by this same function | |||
| 777 | KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1)((taskdata->td_flags.executing == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.executing == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 777)); | |||
| 778 | taskdata->td_flags.executing = 0; // suspend the finishing task | |||
| 779 | ||||
| 780 | 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); } | |||
| 781 | 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); } | |||
| 782 | 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); }; | |||
| 783 | ||||
| 784 | #if OMP_40_ENABLED(50 >= 40) | |||
| 785 | /* If the tasks' destructor thunk flag has been set, we need to invoke the | |||
| 786 | destructor thunk that has been generated by the compiler. The code is | |||
| 787 | placed here, since at this point other tasks might have been released | |||
| 788 | hence overlapping the destructor invokations with some other work in the | |||
| 789 | released tasks. The OpenMP spec is not specific on when the destructors | |||
| 790 | are invoked, so we should be free to choose. */ | |||
| 791 | if (taskdata->td_flags.destructors_thunk) { | |||
| 792 | kmp_routine_entry_t destr_thunk = task->data1.destructors; | |||
| 793 | KMP_ASSERT(destr_thunk)((destr_thunk) ? 0 : __kmp_debug_assert("destr_thunk", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 793)); | |||
| 794 | destr_thunk(gtid, task); | |||
| 795 | } | |||
| 796 | #endif // OMP_40_ENABLED | |||
| 797 | ||||
| 798 | // bookkeeping for resuming task: | |||
| 799 | // GEH - note tasking_ser => task_serial | |||
| 800 | KMP_DEBUG_ASSERT((((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 802)) | |||
| 801 | (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==(((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 802)) | |||
| 802 | taskdata->td_flags.task_serial)(((taskdata->td_flags.tasking_ser || taskdata->td_flags .task_serial) == taskdata->td_flags.task_serial) ? 0 : __kmp_debug_assert ("(taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) == taskdata->td_flags.task_serial" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 802)); | |||
| 803 | if (taskdata->td_flags.task_serial) { | |||
| 804 | if (resumed_task == NULL__null) { | |||
| 805 | resumed_task = taskdata->td_parent; // In a serialized task, the resumed | |||
| 806 | // task is the parent | |||
| 807 | } | |||
| 808 | } else { | |||
| 809 | KMP_DEBUG_ASSERT(resumed_task !=((resumed_task != __null) ? 0 : __kmp_debug_assert("resumed_task != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 810)) | |||
| 810 | NULL)((resumed_task != __null) ? 0 : __kmp_debug_assert("resumed_task != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 810)); // verify that resumed task is passed as arguemnt | |||
| 811 | } | |||
| 812 | ||||
| 813 | // Free this task and then ancestor tasks if they have no children. | |||
| 814 | // Restore th_current_task first as suggested by John: | |||
| 815 | // johnmc: if an asynchronous inquiry peers into the runtime system | |||
| 816 | // it doesn't see the freed task as the current task. | |||
| 817 | thread->th.th_current_task = resumed_task; | |||
| 818 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
| 819 | ||||
| 820 | // TODO: GEH - make sure root team implicit task is initialized properly. | |||
| 821 | // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 ); | |||
| 822 | resumed_task->td_flags.executing = 1; // resume previous task | |||
| 823 | ||||
| 824 | 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); } | |||
| 825 | 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); } | |||
| 826 | 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); }; | |||
| 827 | ||||
| 828 | return; | |||
| 829 | } | |||
| 830 | ||||
| 831 | template <bool ompt> | |||
| 832 | static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref, | |||
| 833 | kmp_int32 gtid, | |||
| 834 | kmp_task_t *task) { | |||
| 835 | 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)); } | |||
| 836 | 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)); }; | |||
| 837 | // this routine will provide task to resume | |||
| 838 | __kmp_task_finish(gtid, task, NULL__null); | |||
| 839 | ||||
| 840 | 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)); } | |||
| 841 | 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)); }; | |||
| 842 | ||||
| 843 | #if OMPT_SUPPORT1 | |||
| 844 | if (ompt) { | |||
| 845 | __ompt_task_finish(task, NULL__null); | |||
| 846 | omp_frame_t *ompt_frame; | |||
| 847 | __ompt_get_task_info_internal(0, NULL__null, NULL__null, &ompt_frame, NULL__null, NULL__null); | |||
| 848 | ompt_frame->enter_frame = NULL__null; | |||
| 849 | } | |||
| 850 | #endif | |||
| 851 | ||||
| 852 | return; | |||
| 853 | } | |||
| 854 | ||||
| 855 | #if OMPT_SUPPORT1 | |||
| 856 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 857 | void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 858 | kmp_task_t *task) { | |||
| 859 | __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task); | |||
| 860 | } | |||
| 861 | #endif // OMPT_SUPPORT | |||
| 862 | ||||
| 863 | // __kmpc_omp_task_complete_if0: report that a task has completed execution | |||
| 864 | // | |||
| 865 | // loc_ref: source location information; points to end of task block. | |||
| 866 | // gtid: global thread number. | |||
| 867 | // task: task thunk for the completed task. | |||
| 868 | void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, | |||
| 869 | kmp_task_t *task) { | |||
| 870 | #if OMPT_SUPPORT1 | |||
| 871 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 872 | __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task); | |||
| 873 | return; | |||
| 874 | } | |||
| 875 | #endif | |||
| 876 | __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task); | |||
| 877 | } | |||
| 878 | ||||
| 879 | #ifdef TASK_UNUSED | |||
| 880 | // __kmpc_omp_task_complete: report that a task has completed execution | |||
| 881 | // NEVER GENERATED BY COMPILER, DEPRECATED!!! | |||
| 882 | void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, | |||
| 883 | kmp_task_t *task) { | |||
| 884 | 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)); } | |||
| 885 | 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)); }; | |||
| 886 | ||||
| 887 | __kmp_task_finish(gtid, task, NULL__null); // Not sure how to find task to resume | |||
| 888 | ||||
| 889 | 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)); } | |||
| 890 | 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)); }; | |||
| 891 | return; | |||
| 892 | } | |||
| 893 | #endif // TASK_UNUSED | |||
| 894 | ||||
| 895 | // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit | |||
| 896 | // task for a given thread | |||
| 897 | // | |||
| 898 | // loc_ref: reference to source location of parallel region | |||
| 899 | // this_thr: thread data structure corresponding to implicit task | |||
| 900 | // team: team for this_thr | |||
| 901 | // tid: thread id of given thread within team | |||
| 902 | // set_curr_task: TRUE if need to push current task to thread | |||
| 903 | // NOTE: Routine does not set up the implicit task ICVS. This is assumed to | |||
| 904 | // have already been done elsewhere. | |||
| 905 | // TODO: Get better loc_ref. Value passed in may be NULL | |||
| 906 | void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, | |||
| 907 | kmp_team_t *team, int tid, int set_curr_task) { | |||
| 908 | kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid]; | |||
| 909 | ||||
| 910 | 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"); } | |||
| 911 | 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"); } | |||
| 912 | ("__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"); } | |||
| 913 | 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"); }; | |||
| 914 | ||||
| 915 | task->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 916 | task->td_team = team; | |||
| 917 | // task->td_parent = NULL; // fix for CQ230101 (broken parent task info | |||
| 918 | // in debugger) | |||
| 919 | task->td_ident = loc_ref; | |||
| 920 | task->td_taskwait_ident = NULL__null; | |||
| 921 | task->td_taskwait_counter = 0; | |||
| 922 | task->td_taskwait_thread = 0; | |||
| 923 | ||||
| 924 | task->td_flags.tiedness = TASK_TIED1; | |||
| 925 | task->td_flags.tasktype = TASK_IMPLICIT0; | |||
| 926 | #if OMP_45_ENABLED(50 >= 45) | |||
| 927 | task->td_flags.proxy = TASK_FULL0; | |||
| 928 | #endif | |||
| 929 | ||||
| 930 | // All implicit tasks are executed immediately, not deferred | |||
| 931 | task->td_flags.task_serial = 1; | |||
| 932 | task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
| 933 | task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
| 934 | ||||
| 935 | task->td_flags.started = 1; | |||
| 936 | task->td_flags.executing = 1; | |||
| 937 | task->td_flags.complete = 0; | |||
| 938 | task->td_flags.freed = 0; | |||
| 939 | ||||
| 940 | #if OMP_40_ENABLED(50 >= 40) | |||
| 941 | task->td_depnode = NULL__null; | |||
| 942 | #endif | |||
| 943 | task->td_last_tied = task; | |||
| 944 | ||||
| 945 | if (set_curr_task) { // only do this init first time thread is created | |||
| 946 | KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0)(&task->td_incomplete_child_tasks)->store(0, std::memory_order_release ); | |||
| 947 | // Not used: don't need to deallocate implicit task | |||
| 948 | KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0)(&task->td_allocated_child_tasks)->store(0, std::memory_order_release ); | |||
| 949 | #if OMP_40_ENABLED(50 >= 40) | |||
| 950 | task->td_taskgroup = NULL__null; // An implicit task does not have taskgroup | |||
| 951 | task->td_dephash = NULL__null; | |||
| 952 | #endif | |||
| 953 | __kmp_push_current_task_to_thread(this_thr, team, tid); | |||
| 954 | } else { | |||
| 955 | KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0)((task->td_incomplete_child_tasks == 0) ? 0 : __kmp_debug_assert ("task->td_incomplete_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 955)); | |||
| 956 | KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0)((task->td_allocated_child_tasks == 0) ? 0 : __kmp_debug_assert ("task->td_allocated_child_tasks == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 956)); | |||
| 957 | } | |||
| 958 | ||||
| 959 | #if OMPT_SUPPORT1 | |||
| 960 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 961 | __ompt_task_init(task, tid); | |||
| 962 | #endif | |||
| 963 | ||||
| 964 | 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); } | |||
| 965 | 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); }; | |||
| 966 | } | |||
| 967 | ||||
| 968 | // __kmp_finish_implicit_task: Release resources associated to implicit tasks | |||
| 969 | // at the end of parallel regions. Some resources are kept for reuse in the next | |||
| 970 | // parallel region. | |||
| 971 | // | |||
| 972 | // thread: thread data structure corresponding to implicit task | |||
| 973 | void __kmp_finish_implicit_task(kmp_info_t *thread) { | |||
| 974 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
| 975 | if (task->td_dephash) | |||
| 976 | __kmp_dephash_free_entries(thread, task->td_dephash); | |||
| 977 | } | |||
| 978 | ||||
| 979 | // __kmp_free_implicit_task: Release resources associated to implicit tasks | |||
| 980 | // when these are destroyed regions | |||
| 981 | // | |||
| 982 | // thread: thread data structure corresponding to implicit task | |||
| 983 | void __kmp_free_implicit_task(kmp_info_t *thread) { | |||
| 984 | kmp_taskdata_t *task = thread->th.th_current_task; | |||
| 985 | if (task && task->td_dephash) { | |||
| 986 | __kmp_dephash_free(thread, task->td_dephash); | |||
| 987 | task->td_dephash = NULL__null; | |||
| 988 | } | |||
| 989 | } | |||
| 990 | ||||
| 991 | // Round up a size to a power of two specified by val: Used to insert padding | |||
| 992 | // between structures co-allocated using a single malloc() call | |||
| 993 | static size_t __kmp_round_up_to_val(size_t size, size_t val) { | |||
| 994 | if (size & (val - 1)) { | |||
| 995 | size &= ~(val - 1); | |||
| 996 | if (size <= KMP_SIZE_T_MAX(0xFFFFFFFFFFFFFFFF) - val) { | |||
| 997 | size += val; // Round up if there is no overflow. | |||
| 998 | } | |||
| 999 | } | |||
| 1000 | return size; | |||
| 1001 | } // __kmp_round_up_to_va | |||
| 1002 | ||||
| 1003 | // __kmp_task_alloc: Allocate the taskdata and task data structures for a task | |||
| 1004 | // | |||
| 1005 | // loc_ref: source location information | |||
| 1006 | // gtid: global thread number. | |||
| 1007 | // flags: include tiedness & task type (explicit vs. implicit) of the ''new'' | |||
| 1008 | // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine. | |||
| 1009 | // sizeof_kmp_task_t: Size in bytes of kmp_task_t data structure including | |||
| 1010 | // private vars accessed in task. | |||
| 1011 | // sizeof_shareds: Size in bytes of array of pointers to shared vars accessed | |||
| 1012 | // in task. | |||
| 1013 | // task_entry: Pointer to task code entry point generated by compiler. | |||
| 1014 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
| 1015 | kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1016 | kmp_tasking_flags_t *flags, | |||
| 1017 | size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
| 1018 | kmp_routine_entry_t task_entry) { | |||
| 1019 | kmp_task_t *task; | |||
| 1020 | kmp_taskdata_t *taskdata; | |||
| 1021 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1022 | kmp_team_t *team = thread->th.th_team; | |||
| 1023 | kmp_taskdata_t *parent_task = thread->th.th_current_task; | |||
| 1024 | size_t shareds_offset; | |||
| 1025 | ||||
| 1026 | 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 ); } | |||
| 1027 | "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 ); } | |||
| 1028 | 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 ); } | |||
| 1029 | 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 ); }; | |||
| 1030 | ||||
| 1031 | if (parent_task->td_flags.final) { | |||
| 1032 | if (flags->merged_if0) { | |||
| 1033 | } | |||
| 1034 | flags->final = 1; | |||
| 1035 | } | |||
| 1036 | if (flags->tiedness == TASK_UNTIED0 && !team->t.t_serialized) { | |||
| 1037 | // Untied task encountered causes the TSC algorithm to check entire deque of | |||
| 1038 | // the victim thread. If no untied task encountered, then checking the head | |||
| 1039 | // of the deque should be enough. | |||
| 1040 | 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); | |||
| 1041 | } | |||
| 1042 | ||||
| 1043 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1044 | if (flags->proxy == TASK_PROXY1) { | |||
| 1045 | flags->tiedness = TASK_UNTIED0; | |||
| 1046 | flags->merged_if0 = 1; | |||
| 1047 | ||||
| 1048 | /* are we running in a sequential parallel or tskm_immediate_exec... we need | |||
| 1049 | tasking support enabled */ | |||
| 1050 | if ((thread->th.th_task_team) == NULL__null) { | |||
| 1051 | /* This should only happen if the team is serialized | |||
| 1052 | setup a task team and propagate it to the thread */ | |||
| 1053 | KMP_DEBUG_ASSERT(team->t.t_serialized)((team->t.t_serialized) ? 0 : __kmp_debug_assert("team->t.t_serialized" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1053)); | |||
| 1054 | 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); } | |||
| 1055 | ("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); } | |||
| 1056 | gtid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d creating task team in __kmp_task_alloc for proxy task\n" , gtid); }; | |||
| 1057 | __kmp_task_team_setup( | |||
| 1058 | thread, team, | |||
| 1059 | 1); // 1 indicates setup the current team regardless of nthreads | |||
| 1060 | thread->th.th_task_team = team->t.t_task_team[thread->th.th_task_state]; | |||
| 1061 | } | |||
| 1062 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 1063 | ||||
| 1064 | /* tasking must be enabled now as the task might not be pushed */ | |||
| 1065 | if (!KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
| 1066 | KA_TRACE(if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
| 1067 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n" , gtid); } | |||
| 1068 | ("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); }; | |||
| 1069 | __kmp_enable_tasking(task_team, thread); | |||
| 1070 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
| 1071 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 1072 | // No lock needed since only owner can allocate | |||
| 1073 | if (thread_data->td.td_deque == NULL__null) { | |||
| 1074 | __kmp_alloc_task_deque(thread, thread_data); | |||
| 1075 | } | |||
| 1076 | } | |||
| 1077 | ||||
| 1078 | if (task_team->tt.tt_found_proxy_tasks == FALSE0) | |||
| 1079 | TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE)(task_team->tt.tt_found_proxy_tasks) = ((!0)); | |||
| 1080 | } | |||
| 1081 | #endif | |||
| 1082 | ||||
| 1083 | // Calculate shared structure offset including padding after kmp_task_t struct | |||
| 1084 | // to align pointers in shared struct | |||
| 1085 | shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t; | |||
| 1086 | shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *)); | |||
| 1087 | ||||
| 1088 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
| 1089 | 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); } | |||
| 1090 | shareds_offset))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d First malloc size: %ld\n" , gtid, shareds_offset); }; | |||
| 1091 | 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); } | |||
| 1092 | sizeof_shareds))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_alloc: T#%d Second malloc size: %ld\n" , gtid, sizeof_shareds); }; | |||
| 1093 | ||||
| 1094 | // Avoid double allocation here by combining shareds with taskdata | |||
| 1095 | #if USE_FAST_MEMORY3 | |||
| 1096 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, shareds_offset +___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1097) | |||
| 1097 | sizeof_shareds)___kmp_fast_allocate((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1097); | |||
| 1098 | #else /* ! USE_FAST_MEMORY */ | |||
| 1099 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, shareds_offset +___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1100) | |||
| 1100 | sizeof_shareds)___kmp_thread_malloc((thread), (shareds_offset + sizeof_shareds ), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1100); | |||
| 1101 | #endif /* USE_FAST_MEMORY */ | |||
| 1102 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
| 1103 | ||||
| 1104 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 1105 | ||||
| 1106 | // Make sure task & taskdata are aligned appropriately | |||
| 1107 | #if KMP_ARCH_X860 || KMP_ARCH_PPC64(0 || 0) || !KMP_HAVE_QUAD0 | |||
| 1108 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0)(((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0 ) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1108)); | |||
| 1109 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0)(((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1109)); | |||
| 1110 | #else | |||
| 1111 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0)(((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1111)); | |||
| 1112 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0)(((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1112)); | |||
| 1113 | #endif | |||
| 1114 | if (sizeof_shareds > 0) { | |||
| 1115 | // Avoid double allocation here by combining shareds with taskdata | |||
| 1116 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
| 1117 | // Make sure shareds struct is aligned to pointer size | |||
| 1118 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1119)) | |||
| 1119 | 0)(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1119)); | |||
| 1120 | } else { | |||
| 1121 | task->shareds = NULL__null; | |||
| 1122 | } | |||
| 1123 | task->routine = task_entry; | |||
| 1124 | task->part_id = 0; // AC: Always start with 0 part id | |||
| 1125 | ||||
| 1126 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 1127 | taskdata->td_team = team; | |||
| 1128 | taskdata->td_alloc_thread = thread; | |||
| 1129 | taskdata->td_parent = parent_task; | |||
| 1130 | taskdata->td_level = parent_task->td_level + 1; // increment nesting level | |||
| 1131 | KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0)(&taskdata->td_untied_count)->store(0, std::memory_order_relaxed ); | |||
| 1132 | taskdata->td_ident = loc_ref; | |||
| 1133 | taskdata->td_taskwait_ident = NULL__null; | |||
| 1134 | taskdata->td_taskwait_counter = 0; | |||
| 1135 | taskdata->td_taskwait_thread = 0; | |||
| 1136 | KMP_DEBUG_ASSERT(taskdata->td_parent != NULL)((taskdata->td_parent != __null) ? 0 : __kmp_debug_assert( "taskdata->td_parent != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1136)); | |||
| 1137 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1138 | // avoid copying icvs for proxy tasks | |||
| 1139 | if (flags->proxy == TASK_FULL0) | |||
| 1140 | #endif | |||
| 1141 | copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs); | |||
| 1142 | ||||
| 1143 | taskdata->td_flags.tiedness = flags->tiedness; | |||
| 1144 | taskdata->td_flags.final = flags->final; | |||
| 1145 | taskdata->td_flags.merged_if0 = flags->merged_if0; | |||
| 1146 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1147 | taskdata->td_flags.destructors_thunk = flags->destructors_thunk; | |||
| 1148 | #endif // OMP_40_ENABLED | |||
| 1149 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1150 | taskdata->td_flags.proxy = flags->proxy; | |||
| 1151 | taskdata->td_task_team = thread->th.th_task_team; | |||
| 1152 | taskdata->td_size_alloc = shareds_offset + sizeof_shareds; | |||
| 1153 | #endif | |||
| 1154 | taskdata->td_flags.tasktype = TASK_EXPLICIT1; | |||
| 1155 | ||||
| 1156 | // GEH - TODO: fix this to copy parent task's value of tasking_ser flag | |||
| 1157 | taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec); | |||
| 1158 | ||||
| 1159 | // GEH - TODO: fix this to copy parent task's value of team_serial flag | |||
| 1160 | taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0; | |||
| 1161 | ||||
| 1162 | // GEH - Note we serialize the task if the team is serialized to make sure | |||
| 1163 | // implicit parallel region tasks are not left until program termination to | |||
| 1164 | // execute. Also, it helps locality to execute immediately. | |||
| 1165 | ||||
| 1166 | taskdata->td_flags.task_serial = | |||
| 1167 | (parent_task->td_flags.final || taskdata->td_flags.team_serial || | |||
| 1168 | taskdata->td_flags.tasking_ser); | |||
| 1169 | ||||
| 1170 | taskdata->td_flags.started = 0; | |||
| 1171 | taskdata->td_flags.executing = 0; | |||
| 1172 | taskdata->td_flags.complete = 0; | |||
| 1173 | taskdata->td_flags.freed = 0; | |||
| 1174 | ||||
| 1175 | taskdata->td_flags.native = flags->native; | |||
| 1176 | ||||
| 1177 | KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0)(&taskdata->td_incomplete_child_tasks)->store(0, std ::memory_order_relaxed); | |||
| 1178 | // start at one because counts current task and children | |||
| 1179 | KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1)(&taskdata->td_allocated_child_tasks)->store(1, std ::memory_order_relaxed); | |||
| 1180 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1181 | taskdata->td_taskgroup = | |||
| 1182 | parent_task->td_taskgroup; // task inherits taskgroup from the parent task | |||
| 1183 | taskdata->td_dephash = NULL__null; | |||
| 1184 | taskdata->td_depnode = NULL__null; | |||
| 1185 | #endif | |||
| 1186 | if (flags->tiedness == TASK_UNTIED0) | |||
| 1187 | taskdata->td_last_tied = NULL__null; // will be set when the task is scheduled | |||
| 1188 | else | |||
| 1189 | taskdata->td_last_tied = taskdata; | |||
| 1190 | ||||
| 1191 | // Only need to keep track of child task counts if team parallel and tasking not | |||
| 1192 | // serialized or if it is a proxy task | |||
| 1193 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1194 | if (flags->proxy == TASK_PROXY1 || | |||
| 1195 | !(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) | |||
| 1196 | #else | |||
| 1197 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) | |||
| 1198 | #endif | |||
| 1199 | { | |||
| 1200 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
| 1201 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1202 | if (parent_task->td_taskgroup) | |||
| 1203 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
| 1204 | #endif | |||
| 1205 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
| 1206 | // implicit not deallocated | |||
| 1207 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) { | |||
| 1208 | 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); | |||
| 1209 | } | |||
| 1210 | } | |||
| 1211 | ||||
| 1212 | 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); } | |||
| 1213 | 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); }; | |||
| 1214 | ANNOTATE_HAPPENS_BEFORE(task); | |||
| 1215 | ||||
| 1216 | #if OMPT_SUPPORT1 | |||
| 1217 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1218 | __ompt_task_init(taskdata, gtid); | |||
| 1219 | #endif | |||
| 1220 | ||||
| 1221 | return task; | |||
| 1222 | } | |||
| 1223 | ||||
| 1224 | kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1225 | kmp_int32 flags, size_t sizeof_kmp_task_t, | |||
| 1226 | size_t sizeof_shareds, | |||
| 1227 | kmp_routine_entry_t task_entry) { | |||
| 1228 | kmp_task_t *retval; | |||
| 1229 | kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; | |||
| 1230 | ||||
| 1231 | input_flags->native = FALSE0; | |||
| 1232 | // __kmp_task_alloc() sets up all other runtime flags | |||
| 1233 | ||||
| 1234 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1235 | KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1236 | "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1237 | gtid, loc_ref, input_flags->tiedness ? "tied " : "untied",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1238 | input_flags->proxy ? "proxy" : "", sizeof_kmp_task_t,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); } | |||
| 1239 | sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", input_flags ->proxy ? "proxy" : "", sizeof_kmp_task_t, sizeof_shareds, task_entry); }; | |||
| 1240 | #else | |||
| 1241 | KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
| 1242 | "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
| 1243 | gtid, loc_ref, input_flags->tiedness ? "tied " : "untied",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); } | |||
| 1244 | sizeof_kmp_task_t, sizeof_shareds, task_entry))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) " "sizeof_task=%ld sizeof_shared=%ld entry=%p\n", gtid, loc_ref , input_flags->tiedness ? "tied " : "untied", sizeof_kmp_task_t , sizeof_shareds, task_entry); }; | |||
| 1245 | #endif | |||
| 1246 | ||||
| 1247 | retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t, | |||
| 1248 | sizeof_shareds, task_entry); | |||
| 1249 | ||||
| 1250 | 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); }; | |||
| 1251 | ||||
| 1252 | return retval; | |||
| 1253 | } | |||
| 1254 | ||||
| 1255 | // __kmp_invoke_task: invoke the specified task | |||
| 1256 | // | |||
| 1257 | // gtid: global thread ID of caller | |||
| 1258 | // task: the task to invoke | |||
| 1259 | // current_task: the task to resume after task invokation | |||
| 1260 | static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task, | |||
| 1261 | kmp_taskdata_t *current_task) { | |||
| 1262 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 1263 | kmp_uint64 cur_time; | |||
| 1264 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1265 | int discard = 0 /* false */; | |||
| 1266 | #endif | |||
| 1267 | 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); } | |||
| 1268 | 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); } | |||
| 1269 | 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); }; | |||
| 1270 | KMP_DEBUG_ASSERT(task)((task) ? 0 : __kmp_debug_assert("task", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1270)); | |||
| 1271 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1272 | if (taskdata->td_flags.proxy == TASK_PROXY1 && | |||
| 1273 | taskdata->td_flags.complete == 1) { | |||
| 1274 | // This is a proxy task that was already completed but it needs to run | |||
| 1275 | // its bottom-half finish | |||
| 1276 | 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); } | |||
| 1277 | 30,if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n" , gtid, taskdata); } | |||
| 1278 | ("__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); } | |||
| 1279 | 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); }; | |||
| 1280 | ||||
| 1281 | __kmp_bottom_half_finish_proxy(gtid, task); | |||
| 1282 | ||||
| 1283 | 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 ); } | |||
| 1284 | "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 ); } | |||
| 1285 | 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 ); }; | |||
| 1286 | ||||
| 1287 | return; | |||
| 1288 | } | |||
| 1289 | #endif | |||
| 1290 | ||||
| 1291 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 1292 | if (__kmp_forkjoin_frames_mode == 3) { | |||
| 1293 | // Get the current time stamp to measure task execution time to correct | |||
| 1294 | // barrier imbalance time | |||
| 1295 | cur_time = __itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0(); | |||
| 1296 | } | |||
| 1297 | #endif | |||
| 1298 | ||||
| 1299 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1300 | // Proxy tasks are not handled by the runtime | |||
| 1301 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
| 1302 | #endif | |||
| 1303 | ANNOTATE_HAPPENS_AFTER(task); | |||
| 1304 | __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded | |||
| 1305 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1306 | } | |||
| 1307 | #endif | |||
| 1308 | ||||
| 1309 | #if OMPT_SUPPORT1 | |||
| 1310 | ompt_thread_info_t oldInfo; | |||
| 1311 | kmp_info_t *thread; | |||
| 1312 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1313 | // Store the threads states and restore them after the task | |||
| 1314 | thread = __kmp_threads[gtid]; | |||
| 1315 | oldInfo = thread->th.ompt_thread_info; | |||
| 1316 | thread->th.ompt_thread_info.wait_id = 0; | |||
| 1317 | thread->th.ompt_thread_info.state = (thread->th.th_team_serialized) | |||
| 1318 | ? omp_state_work_serial | |||
| 1319 | : omp_state_work_parallel; | |||
| 1320 | taskdata->ompt_task_info.frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0)__builtin_frame_address(0); | |||
| 1321 | } | |||
| 1322 | #endif | |||
| 1323 | ||||
| 1324 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1325 | // TODO: cancel tasks if the parallel region has also been cancelled | |||
| 1326 | // TODO: check if this sequence can be hoisted above __kmp_task_start | |||
| 1327 | // if cancellation has been enabled for this run ... | |||
| 1328 | if (__kmp_omp_cancellation) { | |||
| 1329 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
| 1330 | kmp_team_t *this_team = this_thr->th.th_team; | |||
| 1331 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
| 1332 | if ((taskgroup && taskgroup->cancel_request) || | |||
| 1333 | (this_team->t.t_cancel_request == cancel_parallel)) { | |||
| 1334 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1335 | ompt_data_t *task_data; | |||
| 1336 | if (UNLIKELY(ompt_enabled.ompt_callback_cancel)__builtin_expect(!!(ompt_enabled.ompt_callback_cancel), 0)) { | |||
| 1337 | __ompt_get_task_info_internal(0, NULL__null, &task_data, NULL__null, NULL__null, NULL__null); | |||
| 1338 | ompt_callbacks.ompt_callback(ompt_callback_cancel)ompt_callback_cancel_callback( | |||
| 1339 | task_data, | |||
| 1340 | ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup | |||
| 1341 | : ompt_cancel_parallel) | | |||
| 1342 | ompt_cancel_discarded_task, | |||
| 1343 | NULL__null); | |||
| 1344 | } | |||
| 1345 | #endif | |||
| 1346 | KMP_COUNT_BLOCK(TASK_cancelled)((void)0); | |||
| 1347 | // this task belongs to a task group and we need to cancel it | |||
| 1348 | discard = 1 /* true */; | |||
| 1349 | } | |||
| 1350 | } | |||
| 1351 | ||||
| 1352 | // Invoke the task routine and pass in relevant data. | |||
| 1353 | // Thunks generated by gcc take a different argument list. | |||
| 1354 | if (!discard) { | |||
| 1355 | if (taskdata->td_flags.tiedness == TASK_UNTIED0) { | |||
| 1356 | taskdata->td_last_tied = current_task->td_last_tied; | |||
| 1357 | KMP_DEBUG_ASSERT(taskdata->td_last_tied)((taskdata->td_last_tied) ? 0 : __kmp_debug_assert("taskdata->td_last_tied" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1357)); | |||
| 1358 | } | |||
| 1359 | #if KMP_STATS_ENABLED0 | |||
| 1360 | KMP_COUNT_BLOCK(TASK_executed)((void)0); | |||
| 1361 | switch (KMP_GET_THREAD_STATE()((void)0)) { | |||
| 1362 | case FORK_JOIN_BARRIER: | |||
| 1363 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar)((void)0); | |||
| 1364 | break; | |||
| 1365 | case PLAIN_BARRIER: | |||
| 1366 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar)((void)0); | |||
| 1367 | break; | |||
| 1368 | case TASKYIELD: | |||
| 1369 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield)((void)0); | |||
| 1370 | break; | |||
| 1371 | case TASKWAIT: | |||
| 1372 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait)((void)0); | |||
| 1373 | break; | |||
| 1374 | case TASKGROUP: | |||
| 1375 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup)((void)0); | |||
| 1376 | break; | |||
| 1377 | default: | |||
| 1378 | KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate)((void)0); | |||
| 1379 | break; | |||
| 1380 | } | |||
| 1381 | #endif // KMP_STATS_ENABLED | |||
| 1382 | #endif // OMP_40_ENABLED | |||
| 1383 | ||||
| 1384 | // OMPT task begin | |||
| 1385 | #if OMPT_SUPPORT1 | |||
| 1386 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1387 | __ompt_task_start(task, current_task, gtid); | |||
| 1388 | #endif | |||
| 1389 | ||||
| 1390 | #ifdef KMP_GOMP_COMPAT | |||
| 1391 | if (taskdata->td_flags.native) { | |||
| 1392 | ((void (*)(void *))(*(task->routine)))(task->shareds); | |||
| 1393 | } else | |||
| 1394 | #endif /* KMP_GOMP_COMPAT */ | |||
| 1395 | { | |||
| 1396 | (*(task->routine))(gtid, task); | |||
| 1397 | } | |||
| 1398 | KMP_POP_PARTITIONED_TIMER()((void)0); | |||
| 1399 | ||||
| 1400 | #if OMPT_SUPPORT1 | |||
| 1401 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1402 | __ompt_task_finish(task, current_task); | |||
| 1403 | #endif | |||
| 1404 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1405 | } | |||
| 1406 | #endif // OMP_40_ENABLED | |||
| 1407 | ||||
| 1408 | #if OMPT_SUPPORT1 | |||
| 1409 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1410 | thread->th.ompt_thread_info = oldInfo; | |||
| 1411 | taskdata->ompt_task_info.frame.exit_frame = NULL__null; | |||
| 1412 | } | |||
| 1413 | #endif | |||
| 1414 | ||||
| 1415 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1416 | // Proxy tasks are not handled by the runtime | |||
| 1417 | if (taskdata->td_flags.proxy != TASK_PROXY1) { | |||
| 1418 | #endif | |||
| 1419 | ANNOTATE_HAPPENS_BEFORE(taskdata->td_parent); | |||
| 1420 | __kmp_task_finish(gtid, task, current_task); // OMPT only if not discarded | |||
| 1421 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1422 | } | |||
| 1423 | #endif | |||
| 1424 | ||||
| 1425 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 1426 | // Barrier imbalance - correct arrive time after the task finished | |||
| 1427 | if (__kmp_forkjoin_frames_mode == 3) { | |||
| 1428 | kmp_info_t *this_thr = __kmp_threads[gtid]; | |||
| 1429 | if (this_thr->th.th_bar_arrive_time) { | |||
| 1430 | this_thr->th.th_bar_arrive_time += (__itt_get_timestamp(!__kmp_itt_get_timestamp_ptr__3_0) ? 0 : __kmp_itt_get_timestamp_ptr__3_0() - cur_time); | |||
| ||||
| 1431 | } | |||
| 1432 | } | |||
| 1433 | #endif | |||
| 1434 | 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); } | |||
| 1435 | 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); } | |||
| 1436 | ("__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); } | |||
| 1437 | 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); }; | |||
| 1438 | return; | |||
| 1439 | } | |||
| 1440 | ||||
| 1441 | // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution | |||
| 1442 | // | |||
| 1443 | // loc_ref: location of original task pragma (ignored) | |||
| 1444 | // gtid: Global Thread ID of encountering thread | |||
| 1445 | // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task'' | |||
| 1446 | // Returns: | |||
| 1447 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1448 | // be resumed later. | |||
| 1449 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1450 | // resumed later. | |||
| 1451 | kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1452 | kmp_task_t *new_task) { | |||
| 1453 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1454 | ||||
| 1455 | 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); } | |||
| 1456 | 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); }; | |||
| 1457 | ||||
| 1458 | #if OMPT_SUPPORT1 | |||
| 1459 | kmp_taskdata_t *parent; | |||
| 1460 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1461 | parent = new_taskdata->td_parent; | |||
| 1462 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 1463 | ompt_data_t task_data = ompt_data_none; | |||
| 1464 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 1465 | parent ? &(parent->ompt_task_info.task_data) : &task_data, | |||
| 1466 | parent ? &(parent->ompt_task_info.frame) : NULL__null, | |||
| 1467 | &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0, | |||
| 1468 | OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 1469 | } | |||
| 1470 | } | |||
| 1471 | #endif | |||
| 1472 | ||||
| 1473 | /* Should we execute the new task or queue it? For now, let's just always try | |||
| 1474 | to queue it. If the queue fills up, then we'll execute it. */ | |||
| 1475 | ||||
| 1476 | if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
| 1477 | { // Execute this task immediately | |||
| 1478 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 1479 | new_taskdata->td_flags.task_serial = 1; | |||
| 1480 | __kmp_invoke_task(gtid, new_task, current_task); | |||
| 1481 | } | |||
| 1482 | ||||
| 1483 | 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); } | |||
| 1484 | 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); } | |||
| 1485 | ("__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); } | |||
| 1486 | "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); } | |||
| 1487 | 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); }; | |||
| 1488 | ||||
| 1489 | ANNOTATE_HAPPENS_BEFORE(new_task); | |||
| 1490 | #if OMPT_SUPPORT1 | |||
| 1491 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1492 | parent->ompt_task_info.frame.enter_frame = NULL__null; | |||
| 1493 | } | |||
| 1494 | #endif | |||
| 1495 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1496 | } | |||
| 1497 | ||||
| 1498 | // __kmp_omp_task: Schedule a non-thread-switchable task for execution | |||
| 1499 | // | |||
| 1500 | // gtid: Global Thread ID of encountering thread | |||
| 1501 | // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc() | |||
| 1502 | // serialize_immediate: if TRUE then if the task is executed immediately its | |||
| 1503 | // execution will be serialized | |||
| 1504 | // Returns: | |||
| 1505 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1506 | // be resumed later. | |||
| 1507 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1508 | // resumed later. | |||
| 1509 | kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, | |||
| 1510 | bool serialize_immediate) { | |||
| 1511 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1512 | ||||
| 1513 | /* Should we execute the new task or queue it? For now, let's just always try to | |||
| 1514 | queue it. If the queue fills up, then we'll execute it. */ | |||
| 1515 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1516 | if (new_taskdata->td_flags.proxy == TASK_PROXY1 || | |||
| 1517 | __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
| 1518 | #else | |||
| 1519 | if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED1) // if cannot defer | |||
| 1520 | #endif | |||
| 1521 | { // Execute this task immediately | |||
| 1522 | kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task; | |||
| 1523 | if (serialize_immediate) | |||
| 1524 | new_taskdata->td_flags.task_serial = 1; | |||
| 1525 | __kmp_invoke_task(gtid, new_task, current_task); | |||
| 1526 | } | |||
| 1527 | ||||
| 1528 | ANNOTATE_HAPPENS_BEFORE(new_task); | |||
| 1529 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1530 | } | |||
| 1531 | ||||
| 1532 | // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a | |||
| 1533 | // non-thread-switchable task from the parent thread only! | |||
| 1534 | // | |||
| 1535 | // loc_ref: location of original task pragma (ignored) | |||
| 1536 | // gtid: Global Thread ID of encountering thread | |||
| 1537 | // new_task: non-thread-switchable task thunk allocated by | |||
| 1538 | // __kmp_omp_task_alloc() | |||
| 1539 | // Returns: | |||
| 1540 | // TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to | |||
| 1541 | // be resumed later. | |||
| 1542 | // TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be | |||
| 1543 | // resumed later. | |||
| 1544 | kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1545 | kmp_task_t *new_task) { | |||
| 1546 | kmp_int32 res; | |||
| 1547 | KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK)((void)0); | |||
| 1548 | ||||
| 1549 | #if KMP_DEBUG1 || OMPT_SUPPORT1 | |||
| 1550 | kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task)(((kmp_taskdata_t *)new_task) - 1); | |||
| 1551 | #endif | |||
| 1552 | 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); } | |||
| 1553 | 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); }; | |||
| 1554 | ||||
| 1555 | #if OMPT_SUPPORT1 | |||
| 1556 | kmp_taskdata_t *parent = NULL__null; | |||
| 1557 | if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)__builtin_expect(!!(ompt_enabled.enabled && !new_taskdata ->td_flags.started), 0)) { | |||
| 1558 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 1559 | parent = new_taskdata->td_parent; | |||
| 1560 | if (!parent->ompt_task_info.frame.enter_frame) | |||
| 1561 | parent->ompt_task_info.frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1); | |||
| 1562 | if (ompt_enabled.ompt_callback_task_create) { | |||
| 1563 | ompt_data_t task_data = ompt_data_none; | |||
| 1564 | ompt_callbacks.ompt_callback(ompt_callback_task_create)ompt_callback_task_create_callback( | |||
| 1565 | parent ? &(parent->ompt_task_info.task_data) : &task_data, | |||
| 1566 | parent ? &(parent->ompt_task_info.frame) : NULL__null, | |||
| 1567 | &(new_taskdata->ompt_task_info.task_data), | |||
| 1568 | 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, | |||
| 1569 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 1570 | } | |||
| 1571 | } | |||
| 1572 | #endif | |||
| 1573 | ||||
| 1574 | res = __kmp_omp_task(gtid, new_task, true); | |||
| 1575 | ||||
| 1576 | 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 ); } | |||
| 1577 | "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 ); } | |||
| 1578 | 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 ); }; | |||
| 1579 | #if OMPT_SUPPORT1 | |||
| 1580 | if (UNLIKELY(ompt_enabled.enabled && parent != NULL)__builtin_expect(!!(ompt_enabled.enabled && parent != __null), 0)) { | |||
| 1581 | parent->ompt_task_info.frame.enter_frame = NULL__null; | |||
| 1582 | } | |||
| 1583 | #endif | |||
| 1584 | return res; | |||
| 1585 | } | |||
| 1586 | ||||
| 1587 | template <bool ompt> | |||
| 1588 | static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1589 | void *frame_address, | |||
| 1590 | void *return_address) { | |||
| 1591 | kmp_taskdata_t *taskdata; | |||
| 1592 | kmp_info_t *thread; | |||
| 1593 | int thread_finished = FALSE0; | |||
| 1594 | KMP_SET_THREAD_STATE_BLOCK(TASKWAIT)((void)0); | |||
| 1595 | ||||
| 1596 | 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); }; | |||
| 1597 | ||||
| 1598 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
| 1599 | thread = __kmp_threads[gtid]; | |||
| 1600 | taskdata = thread->th.th_current_task; | |||
| 1601 | ||||
| 1602 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1603 | ompt_data_t *my_task_data; | |||
| 1604 | ompt_data_t *my_parallel_data; | |||
| 1605 | ||||
| 1606 | if (ompt) { | |||
| 1607 | my_task_data = &(taskdata->ompt_task_info.task_data); | |||
| 1608 | my_parallel_data = OMPT_CUR_TEAM_DATA(thread)(&(thread->th.th_team->t.ompt_team_info.parallel_data )); | |||
| 1609 | ||||
| 1610 | taskdata->ompt_task_info.frame.enter_frame = frame_address; | |||
| 1611 | ||||
| 1612 | if (ompt_enabled.ompt_callback_sync_region) { | |||
| 1613 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 1614 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
| 1615 | my_task_data, return_address); | |||
| 1616 | } | |||
| 1617 | ||||
| 1618 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
| 1619 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 1620 | ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data, | |||
| 1621 | my_task_data, return_address); | |||
| 1622 | } | |||
| 1623 | } | |||
| 1624 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 1625 | ||||
| 1626 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
| 1627 | // taskwait. | |||
| 1628 | #if USE_ITT_BUILD1 | |||
| 1629 | // Note: These values are used by ITT events as well. | |||
| 1630 | #endif /* USE_ITT_BUILD */ | |||
| 1631 | taskdata->td_taskwait_counter += 1; | |||
| 1632 | taskdata->td_taskwait_ident = loc_ref; | |||
| 1633 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 1634 | ||||
| 1635 | #if USE_ITT_BUILD1 | |||
| 1636 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
| 1637 | if (itt_sync_obj != NULL__null) | |||
| 1638 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
| 1639 | #endif /* USE_ITT_BUILD */ | |||
| 1640 | ||||
| 1641 | bool must_wait = | |||
| 1642 | !taskdata->td_flags.team_serial && !taskdata->td_flags.final; | |||
| 1643 | ||||
| 1644 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1645 | must_wait = must_wait || (thread->th.th_task_team != NULL__null && | |||
| 1646 | thread->th.th_task_team->tt.tt_found_proxy_tasks); | |||
| 1647 | #endif | |||
| 1648 | if (must_wait) { | |||
| 1649 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)) | |||
| 1650 | &(taskdata->td_incomplete_child_tasks))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskdata->td_incomplete_child_tasks)), | |||
| 1651 | 0U); | |||
| 1652 | while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) != 0) { | |||
| 1653 | flag.execute_tasks(thread, gtid, FALSE0, | |||
| 1654 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 1655 | __kmp_task_stealing_constraint); | |||
| 1656 | } | |||
| 1657 | } | |||
| 1658 | #if USE_ITT_BUILD1 | |||
| 1659 | if (itt_sync_obj != NULL__null) | |||
| 1660 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
| 1661 | #endif /* USE_ITT_BUILD */ | |||
| 1662 | ||||
| 1663 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
| 1664 | // negated. | |||
| 1665 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
| 1666 | ||||
| 1667 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1668 | if (ompt) { | |||
| 1669 | if (ompt_enabled.ompt_callback_sync_region_wait) { | |||
| 1670 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 1671 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
| 1672 | my_task_data, return_address); | |||
| 1673 | } | |||
| 1674 | if (ompt_enabled.ompt_callback_sync_region) { | |||
| 1675 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 1676 | ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data, | |||
| 1677 | my_task_data, return_address); | |||
| 1678 | } | |||
| 1679 | taskdata->ompt_task_info.frame.enter_frame = NULL__null; | |||
| 1680 | } | |||
| 1681 | #endif // OMPT_SUPPORT && OMPT_OPTIONAL | |||
| 1682 | ||||
| 1683 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
| 1684 | } | |||
| 1685 | ||||
| 1686 | 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); } | |||
| 1687 | "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); } | |||
| 1688 | 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); }; | |||
| 1689 | ||||
| 1690 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1691 | } | |||
| 1692 | ||||
| 1693 | #if OMPT_SUPPORT1 | |||
| 1694 | OMPT_NOINLINE__attribute__((noinline)) | |||
| 1695 | static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid, | |||
| 1696 | void *frame_address, | |||
| 1697 | void *return_address) { | |||
| 1698 | return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address, | |||
| 1699 | return_address); | |||
| 1700 | } | |||
| 1701 | #endif // OMPT_SUPPORT | |||
| 1702 | ||||
| 1703 | // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are | |||
| 1704 | // complete | |||
| 1705 | kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) { | |||
| 1706 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1707 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 1708 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 1709 | return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(1)__builtin_frame_address(1), | |||
| 1710 | OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid)); | |||
| 1711 | } | |||
| 1712 | #endif | |||
| 1713 | return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL__null, NULL__null); | |||
| 1714 | } | |||
| 1715 | ||||
| 1716 | // __kmpc_omp_taskyield: switch to a different task | |||
| 1717 | kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) { | |||
| 1718 | kmp_taskdata_t *taskdata; | |||
| 1719 | kmp_info_t *thread; | |||
| 1720 | int thread_finished = FALSE0; | |||
| 1721 | ||||
| 1722 | KMP_COUNT_BLOCK(OMP_TASKYIELD)((void)0); | |||
| 1723 | KMP_SET_THREAD_STATE_BLOCK(TASKYIELD)((void)0); | |||
| 1724 | ||||
| 1725 | 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); } | |||
| 1726 | 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); }; | |||
| 1727 | ||||
| 1728 | if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) { | |||
| 1729 | thread = __kmp_threads[gtid]; | |||
| 1730 | taskdata = thread->th.th_current_task; | |||
| 1731 | // Should we model this as a task wait or not? | |||
| 1732 | // Debugger: The taskwait is active. Store location and thread encountered the | |||
| 1733 | // taskwait. | |||
| 1734 | #if USE_ITT_BUILD1 | |||
| 1735 | // Note: These values are used by ITT events as well. | |||
| 1736 | #endif /* USE_ITT_BUILD */ | |||
| 1737 | taskdata->td_taskwait_counter += 1; | |||
| 1738 | taskdata->td_taskwait_ident = loc_ref; | |||
| 1739 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 1740 | ||||
| 1741 | #if USE_ITT_BUILD1 | |||
| 1742 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
| 1743 | if (itt_sync_obj != NULL__null) | |||
| 1744 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
| 1745 | #endif /* USE_ITT_BUILD */ | |||
| 1746 | if (!taskdata->td_flags.team_serial) { | |||
| 1747 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 1748 | if (task_team != NULL__null) { | |||
| 1749 | if (KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
| 1750 | #if OMPT_SUPPORT1 | |||
| 1751 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1752 | thread->th.ompt_thread_info.ompt_task_yielded = 1; | |||
| 1753 | #endif | |||
| 1754 | __kmp_execute_tasks_32( | |||
| 1755 | thread, gtid, NULL__null, FALSE0, | |||
| 1756 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 1757 | __kmp_task_stealing_constraint); | |||
| 1758 | #if OMPT_SUPPORT1 | |||
| 1759 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 1760 | thread->th.ompt_thread_info.ompt_task_yielded = 0; | |||
| 1761 | #endif | |||
| 1762 | } | |||
| 1763 | } | |||
| 1764 | } | |||
| 1765 | #if USE_ITT_BUILD1 | |||
| 1766 | if (itt_sync_obj != NULL__null) | |||
| 1767 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
| 1768 | #endif /* USE_ITT_BUILD */ | |||
| 1769 | ||||
| 1770 | // Debugger: The taskwait is completed. Location remains, but thread is | |||
| 1771 | // negated. | |||
| 1772 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; | |||
| 1773 | } | |||
| 1774 | ||||
| 1775 | 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); } | |||
| 1776 | "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); } | |||
| 1777 | 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); }; | |||
| 1778 | ||||
| 1779 | return TASK_CURRENT_NOT_QUEUED0; | |||
| 1780 | } | |||
| 1781 | ||||
| 1782 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
| 1783 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1784 | // Task Reduction implementation | |||
| 1785 | ||||
| 1786 | typedef struct kmp_task_red_flags { | |||
| 1787 | unsigned lazy_priv : 1; // hint: (1) use lazy allocation (big objects) | |||
| 1788 | unsigned reserved31 : 31; | |||
| 1789 | } kmp_task_red_flags_t; | |||
| 1790 | ||||
| 1791 | // internal structure for reduction data item related info | |||
| 1792 | typedef struct kmp_task_red_data { | |||
| 1793 | void *reduce_shar; // shared reduction item | |||
| 1794 | size_t reduce_size; // size of data item | |||
| 1795 | void *reduce_priv; // thread specific data | |||
| 1796 | void *reduce_pend; // end of private data for comparison op | |||
| 1797 | void *reduce_init; // data initialization routine | |||
| 1798 | void *reduce_fini; // data finalization routine | |||
| 1799 | void *reduce_comb; // data combiner routine | |||
| 1800 | kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
| 1801 | } kmp_task_red_data_t; | |||
| 1802 | ||||
| 1803 | // structure sent us by compiler - one per reduction item | |||
| 1804 | typedef struct kmp_task_red_input { | |||
| 1805 | void *reduce_shar; // shared reduction item | |||
| 1806 | size_t reduce_size; // size of data item | |||
| 1807 | void *reduce_init; // data initialization routine | |||
| 1808 | void *reduce_fini; // data finalization routine | |||
| 1809 | void *reduce_comb; // data combiner routine | |||
| 1810 | kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
| 1811 | } kmp_task_red_input_t; | |||
| 1812 | ||||
| 1813 | /*! | |||
| 1814 | @ingroup TASKING | |||
| 1815 | @param gtid Global thread ID | |||
| 1816 | @param num Number of data items to reduce | |||
| 1817 | @param data Array of data for reduction | |||
| 1818 | @return The taskgroup identifier | |||
| 1819 | ||||
| 1820 | Initialize task reduction for the taskgroup. | |||
| 1821 | */ | |||
| 1822 | void *__kmpc_task_reduction_init(int gtid, int num, void *data) { | |||
| 1823 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1824 | kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup; | |||
| 1825 | kmp_int32 nth = thread->th.th_team_nproc; | |||
| 1826 | kmp_task_red_input_t *input = (kmp_task_red_input_t *)data; | |||
| 1827 | kmp_task_red_data_t *arr; | |||
| 1828 | ||||
| 1829 | // check input data just in case | |||
| 1830 | KMP_ASSERT(tg != NULL)((tg != __null) ? 0 : __kmp_debug_assert("tg != NULL", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1830)); | |||
| 1831 | KMP_ASSERT(data != NULL)((data != __null) ? 0 : __kmp_debug_assert("data != NULL", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1831)); | |||
| 1832 | KMP_ASSERT(num > 0)((num > 0) ? 0 : __kmp_debug_assert("num > 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1832)); | |||
| 1833 | if (nth == 1) { | |||
| 1834 | 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); } | |||
| 1835 | gtid, tg))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n" , gtid, tg); }; | |||
| 1836 | return (void *)tg; | |||
| 1837 | } | |||
| 1838 | 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); } | |||
| 1839 | 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); }; | |||
| 1840 | arr = (kmp_task_red_data_t *)__kmp_thread_malloc(___kmp_thread_malloc((thread), (num * sizeof(kmp_task_red_data_t )), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1841) | |||
| 1841 | thread, num * sizeof(kmp_task_red_data_t))___kmp_thread_malloc((thread), (num * sizeof(kmp_task_red_data_t )), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1841); | |||
| 1842 | for (int i = 0; i < num; ++i) { | |||
| 1843 | void (*f_init)(void *) = (void (*)(void *))(input[i].reduce_init); | |||
| 1844 | size_t size = input[i].reduce_size - 1; | |||
| 1845 | // round the size up to cache line per thread-specific item | |||
| 1846 | size += CACHE_LINE64 - size % CACHE_LINE64; | |||
| 1847 | KMP_ASSERT(input[i].reduce_comb != NULL)((input[i].reduce_comb != __null) ? 0 : __kmp_debug_assert("input[i].reduce_comb != NULL" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1847)); // combiner is mandatory | |||
| 1848 | arr[i].reduce_shar = input[i].reduce_shar; | |||
| 1849 | arr[i].reduce_size = size; | |||
| 1850 | arr[i].reduce_init = input[i].reduce_init; | |||
| 1851 | arr[i].reduce_fini = input[i].reduce_fini; | |||
| 1852 | arr[i].reduce_comb = input[i].reduce_comb; | |||
| 1853 | arr[i].flags = input[i].flags; | |||
| 1854 | if (!input[i].flags.lazy_priv) { | |||
| 1855 | // allocate cache-line aligned block and fill it with zeros | |||
| 1856 | arr[i].reduce_priv = __kmp_allocate(nth * size)___kmp_allocate((nth * size), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1856); | |||
| 1857 | arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size; | |||
| 1858 | if (f_init != NULL__null) { | |||
| 1859 | // initialize thread-specific items | |||
| 1860 | for (int j = 0; j < nth; ++j) { | |||
| 1861 | f_init((char *)(arr[i].reduce_priv) + j * size); | |||
| 1862 | } | |||
| 1863 | } | |||
| 1864 | } else { | |||
| 1865 | // only allocate space for pointers now, | |||
| 1866 | // objects will be lazily allocated/initialized once requested | |||
| 1867 | arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *))___kmp_allocate((nth * sizeof(void *)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1867); | |||
| 1868 | } | |||
| 1869 | } | |||
| 1870 | tg->reduce_data = (void *)arr; | |||
| 1871 | tg->reduce_num_data = num; | |||
| 1872 | return (void *)tg; | |||
| 1873 | } | |||
| 1874 | ||||
| 1875 | /*! | |||
| 1876 | @ingroup TASKING | |||
| 1877 | @param gtid Global thread ID | |||
| 1878 | @param tskgrp The taskgroup ID (optional) | |||
| 1879 | @param data Shared location of the item | |||
| 1880 | @return The pointer to per-thread data | |||
| 1881 | ||||
| 1882 | Get thread-specific location of data item | |||
| 1883 | */ | |||
| 1884 | void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) { | |||
| 1885 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1886 | kmp_int32 nth = thread->th.th_team_nproc; | |||
| 1887 | if (nth == 1) | |||
| 1888 | return data; // nothing to do | |||
| 1889 | ||||
| 1890 | kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp; | |||
| 1891 | if (tg == NULL__null) | |||
| 1892 | tg = thread->th.th_current_task->td_taskgroup; | |||
| 1893 | KMP_ASSERT(tg != NULL)((tg != __null) ? 0 : __kmp_debug_assert("tg != NULL", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1893)); | |||
| 1894 | kmp_task_red_data_t *arr = (kmp_task_red_data_t *)(tg->reduce_data); | |||
| 1895 | kmp_int32 num = tg->reduce_num_data; | |||
| 1896 | kmp_int32 tid = thread->th.th_info.ds.ds_tid; | |||
| 1897 | ||||
| 1898 | KMP_ASSERT(data != NULL)((data != __null) ? 0 : __kmp_debug_assert("data != NULL", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1898)); | |||
| 1899 | while (tg != NULL__null) { | |||
| 1900 | for (int i = 0; i < num; ++i) { | |||
| 1901 | if (!arr[i].flags.lazy_priv) { | |||
| 1902 | if (data == arr[i].reduce_shar || | |||
| 1903 | (data >= arr[i].reduce_priv && data < arr[i].reduce_pend)) | |||
| 1904 | return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size; | |||
| 1905 | } else { | |||
| 1906 | // check shared location first | |||
| 1907 | void **p_priv = (void **)(arr[i].reduce_priv); | |||
| 1908 | if (data == arr[i].reduce_shar) | |||
| 1909 | goto found; | |||
| 1910 | // check if we get some thread specific location as parameter | |||
| 1911 | for (int j = 0; j < nth; ++j) | |||
| 1912 | if (data == p_priv[j]) | |||
| 1913 | goto found; | |||
| 1914 | continue; // not found, continue search | |||
| 1915 | found: | |||
| 1916 | if (p_priv[tid] == NULL__null) { | |||
| 1917 | // allocate thread specific object lazily | |||
| 1918 | void (*f_init)(void *) = (void (*)(void *))(arr[i].reduce_init); | |||
| 1919 | p_priv[tid] = __kmp_allocate(arr[i].reduce_size)___kmp_allocate((arr[i].reduce_size), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1919); | |||
| 1920 | if (f_init != NULL__null) { | |||
| 1921 | f_init(p_priv[tid]); | |||
| 1922 | } | |||
| 1923 | } | |||
| 1924 | return p_priv[tid]; | |||
| 1925 | } | |||
| 1926 | } | |||
| 1927 | tg = tg->parent; | |||
| 1928 | arr = (kmp_task_red_data_t *)(tg->reduce_data); | |||
| 1929 | num = tg->reduce_num_data; | |||
| 1930 | } | |||
| 1931 | KMP_ASSERT2(0, "Unknown task reduction item")((0) ? 0 : __kmp_debug_assert(("Unknown task reduction item") , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1931)); | |||
| 1932 | return NULL__null; // ERROR, this line never executed | |||
| 1933 | } | |||
| 1934 | ||||
| 1935 | // Finalize task reduction. | |||
| 1936 | // Called from __kmpc_end_taskgroup() | |||
| 1937 | static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) { | |||
| 1938 | kmp_int32 nth = th->th.th_team_nproc; | |||
| 1939 | KMP_DEBUG_ASSERT(nth > 1)((nth > 1) ? 0 : __kmp_debug_assert("nth > 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1939)); // should not be called if nth == 1 | |||
| 1940 | kmp_task_red_data_t *arr = (kmp_task_red_data_t *)tg->reduce_data; | |||
| 1941 | kmp_int32 num = tg->reduce_num_data; | |||
| 1942 | for (int i = 0; i < num; ++i) { | |||
| 1943 | void *sh_data = arr[i].reduce_shar; | |||
| 1944 | void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini); | |||
| 1945 | void (*f_comb)(void *, void *) = | |||
| 1946 | (void (*)(void *, void *))(arr[i].reduce_comb); | |||
| 1947 | if (!arr[i].flags.lazy_priv) { | |||
| 1948 | void *pr_data = arr[i].reduce_priv; | |||
| 1949 | size_t size = arr[i].reduce_size; | |||
| 1950 | for (int j = 0; j < nth; ++j) { | |||
| 1951 | void *priv_data = (char *)pr_data + j * size; | |||
| 1952 | f_comb(sh_data, priv_data); // combine results | |||
| 1953 | if (f_fini) | |||
| 1954 | f_fini(priv_data); // finalize if needed | |||
| 1955 | } | |||
| 1956 | } else { | |||
| 1957 | void **pr_data = (void **)(arr[i].reduce_priv); | |||
| 1958 | for (int j = 0; j < nth; ++j) { | |||
| 1959 | if (pr_data[j] != NULL__null) { | |||
| 1960 | f_comb(sh_data, pr_data[j]); // combine results | |||
| 1961 | if (f_fini) | |||
| 1962 | f_fini(pr_data[j]); // finalize if needed | |||
| 1963 | __kmp_free(pr_data[j])___kmp_free((pr_data[j]), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1963); | |||
| 1964 | } | |||
| 1965 | } | |||
| 1966 | } | |||
| 1967 | __kmp_free(arr[i].reduce_priv)___kmp_free((arr[i].reduce_priv), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1967); | |||
| 1968 | } | |||
| 1969 | __kmp_thread_free(th, arr)___kmp_thread_free((th), (arr), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1969); | |||
| 1970 | tg->reduce_data = NULL__null; | |||
| 1971 | tg->reduce_num_data = 0; | |||
| 1972 | } | |||
| 1973 | #endif | |||
| 1974 | ||||
| 1975 | #if OMP_40_ENABLED(50 >= 40) | |||
| 1976 | // __kmpc_taskgroup: Start a new taskgroup | |||
| 1977 | void __kmpc_taskgroup(ident_t *loc, int gtid) { | |||
| 1978 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 1979 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 1980 | kmp_taskgroup_t *tg_new = | |||
| 1981 | (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t))___kmp_thread_malloc((thread), (sizeof(kmp_taskgroup_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 1981); | |||
| 1982 | 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); }; | |||
| 1983 | KMP_ATOMIC_ST_RLX(&tg_new->count, 0)(&tg_new->count)->store(0, std::memory_order_relaxed ); | |||
| 1984 | KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq)(&tg_new->cancel_request)->store(cancel_noreq, std:: memory_order_relaxed); | |||
| 1985 | tg_new->parent = taskdata->td_taskgroup; | |||
| 1986 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
| 1987 | #if OMP_45_ENABLED(50 >= 45) | |||
| 1988 | tg_new->reduce_data = NULL__null; | |||
| 1989 | tg_new->reduce_num_data = 0; | |||
| 1990 | #endif | |||
| 1991 | taskdata->td_taskgroup = tg_new; | |||
| 1992 | ||||
| 1993 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 1994 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
| 1995 | void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
| 1996 | if (!codeptr) | |||
| 1997 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
| 1998 | kmp_team_t *team = thread->th.th_team; | |||
| 1999 | ompt_data_t my_task_data = taskdata->ompt_task_info.task_data; | |||
| 2000 | // FIXME: I think this is wrong for lwt! | |||
| 2001 | ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
| 2002 | ||||
| 2003 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2004 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
| 2005 | &(my_task_data), codeptr); | |||
| 2006 | } | |||
| 2007 | #endif | |||
| 2008 | } | |||
| 2009 | ||||
| 2010 | // __kmpc_end_taskgroup: Wait until all tasks generated by the current task | |||
| 2011 | // and its descendants are complete | |||
| 2012 | void __kmpc_end_taskgroup(ident_t *loc, int gtid) { | |||
| 2013 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 2014 | kmp_taskdata_t *taskdata = thread->th.th_current_task; | |||
| 2015 | kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup; | |||
| 2016 | int thread_finished = FALSE0; | |||
| 2017 | ||||
| 2018 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2019 | kmp_team_t *team; | |||
| 2020 | ompt_data_t my_task_data; | |||
| 2021 | ompt_data_t my_parallel_data; | |||
| 2022 | void *codeptr; | |||
| 2023 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) { | |||
| 2024 | team = thread->th.th_team; | |||
| 2025 | my_task_data = taskdata->ompt_task_info.task_data; | |||
| 2026 | // FIXME: I think this is wrong for lwt! | |||
| 2027 | my_parallel_data = team->t.ompt_team_info.parallel_data; | |||
| 2028 | codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid)__ompt_load_return_address(gtid); | |||
| 2029 | if (!codeptr) | |||
| 2030 | codeptr = OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0); | |||
| 2031 | } | |||
| 2032 | #endif | |||
| 2033 | ||||
| 2034 | 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); }; | |||
| 2035 | KMP_DEBUG_ASSERT(taskgroup != NULL)((taskgroup != __null) ? 0 : __kmp_debug_assert("taskgroup != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2035)); | |||
| 2036 | KMP_SET_THREAD_STATE_BLOCK(TASKGROUP)((void)0); | |||
| 2037 | ||||
| 2038 | if (__kmp_tasking_mode != tskm_immediate_exec) { | |||
| 2039 | // mark task as waiting not on a barrier | |||
| 2040 | taskdata->td_taskwait_counter += 1; | |||
| 2041 | taskdata->td_taskwait_ident = loc; | |||
| 2042 | taskdata->td_taskwait_thread = gtid + 1; | |||
| 2043 | #if USE_ITT_BUILD1 | |||
| 2044 | // For ITT the taskgroup wait is similar to taskwait until we need to | |||
| 2045 | // distinguish them | |||
| 2046 | void *itt_sync_obj = __kmp_itt_taskwait_object(gtid); | |||
| 2047 | if (itt_sync_obj != NULL__null) | |||
| 2048 | __kmp_itt_taskwait_starting(gtid, itt_sync_obj); | |||
| 2049 | #endif /* USE_ITT_BUILD */ | |||
| 2050 | ||||
| 2051 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2052 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
| 2053 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2054 | ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data), | |||
| 2055 | &(my_task_data), codeptr); | |||
| 2056 | } | |||
| 2057 | #endif | |||
| 2058 | ||||
| 2059 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2060 | if (!taskdata->td_flags.team_serial || | |||
| 2061 | (thread->th.th_task_team != NULL__null && | |||
| 2062 | thread->th.th_task_team->tt.tt_found_proxy_tasks)) | |||
| 2063 | #else | |||
| 2064 | if (!taskdata->td_flags.team_serial) | |||
| 2065 | #endif | |||
| 2066 | { | |||
| 2067 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count))reinterpret_cast<std::atomic<kmp_uint32> *>(& (taskgroup->count)), | |||
| 2068 | 0U); | |||
| 2069 | while (KMP_ATOMIC_LD_ACQ(&taskgroup->count)(&taskgroup->count)->load(std::memory_order_acquire ) != 0) { | |||
| 2070 | flag.execute_tasks(thread, gtid, FALSE0, | |||
| 2071 | &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, | |||
| 2072 | __kmp_task_stealing_constraint); | |||
| 2073 | } | |||
| 2074 | } | |||
| 2075 | taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting | |||
| 2076 | ||||
| 2077 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2078 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region_wait ), 0)) { | |||
| 2079 | ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)ompt_callback_sync_region_wait_callback( | |||
| 2080 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
| 2081 | &(my_task_data), codeptr); | |||
| 2082 | } | |||
| 2083 | #endif | |||
| 2084 | ||||
| 2085 | #if USE_ITT_BUILD1 | |||
| 2086 | if (itt_sync_obj != NULL__null) | |||
| 2087 | __kmp_itt_taskwait_finished(gtid, itt_sync_obj); | |||
| 2088 | #endif /* USE_ITT_BUILD */ | |||
| 2089 | } | |||
| 2090 | KMP_DEBUG_ASSERT(taskgroup->count == 0)((taskgroup->count == 0) ? 0 : __kmp_debug_assert("taskgroup->count == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2090)); | |||
| 2091 | ||||
| 2092 | // TODO: change to OMP_50_ENABLED, need to change build tools for this to work | |||
| 2093 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2094 | if (taskgroup->reduce_data != NULL__null) // need to reduce? | |||
| 2095 | __kmp_task_reduction_fini(thread, taskgroup); | |||
| 2096 | #endif | |||
| 2097 | // Restore parent taskgroup for the current task | |||
| 2098 | taskdata->td_taskgroup = taskgroup->parent; | |||
| 2099 | __kmp_thread_free(thread, taskgroup)___kmp_thread_free((thread), (taskgroup), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2099); | |||
| 2100 | ||||
| 2101 | 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); } | |||
| 2102 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n" , gtid, taskdata); }; | |||
| 2103 | ANNOTATE_HAPPENS_AFTER(taskdata); | |||
| 2104 | ||||
| 2105 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 2106 | if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)__builtin_expect(!!(ompt_enabled.ompt_callback_sync_region), 0 )) { | |||
| 2107 | ompt_callbacks.ompt_callback(ompt_callback_sync_region)ompt_callback_sync_region_callback( | |||
| 2108 | ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data), | |||
| 2109 | &(my_task_data), codeptr); | |||
| 2110 | } | |||
| 2111 | #endif | |||
| 2112 | } | |||
| 2113 | #endif | |||
| 2114 | ||||
| 2115 | // __kmp_remove_my_task: remove a task from my own deque | |||
| 2116 | static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid, | |||
| 2117 | kmp_task_team_t *task_team, | |||
| 2118 | kmp_int32 is_constrained) { | |||
| 2119 | kmp_task_t *task; | |||
| 2120 | kmp_taskdata_t *taskdata; | |||
| 2121 | kmp_thread_data_t *thread_data; | |||
| 2122 | kmp_uint32 tail; | |||
| 2123 | ||||
| 2124 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2124)); | |||
| 2125 | KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=((task_team->tt.tt_threads_data != __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2126)) | |||
| 2126 | NULL)((task_team->tt.tt_threads_data != __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_threads_data != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2126)); // Caller should check this condition | |||
| 2127 | ||||
| 2128 | thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2128)), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)]; | |||
| 2129 | ||||
| 2130 | 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); } | |||
| 2131 | 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); } | |||
| 2132 | 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); }; | |||
| 2133 | ||||
| 2134 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
| 2135 | 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 ); } | |||
| 2136 | ("__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 ); } | |||
| 2137 | "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 ); } | |||
| 2138 | 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 ); } | |||
| 2139 | 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 ); }; | |||
| 2140 | return NULL__null; | |||
| 2141 | } | |||
| 2142 | ||||
| 2143 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2144 | ||||
| 2145 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) == 0) { | |||
| 2146 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2147 | 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 ); } | |||
| 2148 | ("__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 ); } | |||
| 2149 | "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 ); } | |||
| 2150 | 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 ); } | |||
| 2151 | 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 ); }; | |||
| 2152 | return NULL__null; | |||
| 2153 | } | |||
| 2154 | ||||
| 2155 | tail = (thread_data->td.td_deque_tail - 1) & | |||
| 2156 | TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); // Wrap index. | |||
| 2157 | taskdata = thread_data->td.td_deque[tail]; | |||
| 2158 | ||||
| 2159 | if (is_constrained && (taskdata->td_flags.tiedness == TASK_TIED1)) { | |||
| 2160 | // we need to check if the candidate obeys task scheduling constraint (TSC) | |||
| 2161 | // only descendant of all deferred tied tasks can be scheduled, checking | |||
| 2162 | // the last one is enough, as it in turn is the descendant of all others | |||
| 2163 | kmp_taskdata_t *current = thread->th.th_current_task->td_last_tied; | |||
| 2164 | KMP_DEBUG_ASSERT(current != NULL)((current != __null) ? 0 : __kmp_debug_assert("current != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2164)); | |||
| 2165 | // check if last tied task is not suspended on barrier | |||
| 2166 | if (current->td_flags.tasktype == TASK_EXPLICIT1 || | |||
| 2167 | current->td_taskwait_thread > 0) { // <= 0 on barrier | |||
| 2168 | kmp_int32 level = current->td_level; | |||
| 2169 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
| 2170 | while (parent != current && parent->td_level > level) { | |||
| 2171 | parent = parent->td_parent; // check generation up to the level of the | |||
| 2172 | // current task | |||
| 2173 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2173)); | |||
| 2174 | } | |||
| 2175 | if (parent != current) { | |||
| 2176 | // The TSC does not allow to steal victim task | |||
| 2177 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2178 | KA_TRACE(10, ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
| 2179 | "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 ); } | |||
| 2180 | 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 ); } | |||
| 2181 | thread_data->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: " "ntasks=%d head=%u tail=%u\n", gtid, thread_data->td.td_deque_ntasks , thread_data->td.td_deque_head, thread_data->td.td_deque_tail ); } | |||
| 2182 | 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 ); }; | |||
| 2183 | return NULL__null; | |||
| 2184 | } | |||
| 2185 | } | |||
| 2186 | } | |||
| 2187 | ||||
| 2188 | thread_data->td.td_deque_tail = tail; | |||
| 2189 | 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); | |||
| 2190 | ||||
| 2191 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2192 | ||||
| 2193 | KA_TRACE(10, ("__kmp_remove_my_task(exit #2): T#%d task %p removed: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2194 | "ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2195 | gtid, taskdata, thread_data->td.td_deque_ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); } | |||
| 2196 | thread_data->td.td_deque_head, thread_data->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_remove_my_task(exit #2): T#%d task %p removed: " "ntasks=%d head=%u tail=%u\n", gtid, taskdata, thread_data-> td.td_deque_ntasks, thread_data->td.td_deque_head, thread_data ->td.td_deque_tail); }; | |||
| 2197 | ||||
| 2198 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 2199 | return task; | |||
| 2200 | } | |||
| 2201 | ||||
| 2202 | // __kmp_steal_task: remove a task from another thread's deque | |||
| 2203 | // Assume that calling thread has already checked existence of | |||
| 2204 | // task_team thread_data before calling this routine. | |||
| 2205 | static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid, | |||
| 2206 | kmp_task_team_t *task_team, | |||
| 2207 | std::atomic<kmp_int32> *unfinished_threads, | |||
| 2208 | int *thread_finished, | |||
| 2209 | kmp_int32 is_constrained) { | |||
| 2210 | kmp_task_t *task; | |||
| 2211 | kmp_taskdata_t *taskdata; | |||
| 2212 | kmp_taskdata_t *current; | |||
| 2213 | kmp_thread_data_t *victim_td, *threads_data; | |||
| 2214 | kmp_int32 level, target; | |||
| 2215 | kmp_int32 victim_tid; | |||
| 2216 | ||||
| 2217 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2217)); | |||
| 2218 | ||||
| 2219 | threads_data = task_team->tt.tt_threads_data; | |||
| 2220 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2220)); // Caller should check this condition | |||
| 2221 | ||||
| 2222 | victim_tid = victim_thr->th.th_info.ds.ds_tid; | |||
| 2223 | victim_td = &threads_data[victim_tid]; | |||
| 2224 | ||||
| 2225 | KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2227)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2226 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2227)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2227 | gtid, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2227)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2228 | victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2227)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2229 | victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(enter): T#%d try to steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2227)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 2230 | ||||
| 2231 | if (TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks) == 0) { | |||
| 2232 | KA_TRACE(10, ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2234)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2233 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2234)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2234 | gtid, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2234)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2235 | victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2234)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); } | |||
| 2236 | victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2234)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head , victim_td->td.td_deque_tail); }; | |||
| 2237 | return NULL__null; | |||
| 2238 | } | |||
| 2239 | ||||
| 2240 | __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 2241 | ||||
| 2242 | int ntasks = TCR_4(victim_td->td.td_deque_ntasks)(victim_td->td.td_deque_ntasks); | |||
| 2243 | // Check again after we acquire the lock | |||
| 2244 | if (ntasks == 0) { | |||
| 2245 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 2246 | KA_TRACE(10, ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2248)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2247 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2248)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2248 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2248)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2249 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, ((((victim_thr ) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2248)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
| 2250 | return NULL__null; | |||
| 2251 | } | |||
| 2252 | ||||
| 2253 | KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL)((victim_td->td.td_deque != __null) ? 0 : __kmp_debug_assert ("victim_td->td.td_deque != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2253)); | |||
| 2254 | ||||
| 2255 | taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head]; | |||
| 2256 | if (is_constrained && (taskdata->td_flags.tiedness == TASK_TIED1)) { | |||
| 2257 | // we need to check if the candidate obeys task scheduling constraint (TSC) | |||
| 2258 | // only descendant of all deferred tied tasks can be scheduled, checking | |||
| 2259 | // the last one is enough, as it in turn is the descendant of all others | |||
| 2260 | current = __kmp_threads[gtid]->th.th_current_task->td_last_tied; | |||
| 2261 | KMP_DEBUG_ASSERT(current != NULL)((current != __null) ? 0 : __kmp_debug_assert("current != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2261)); | |||
| 2262 | // check if last tied task is not suspended on barrier | |||
| 2263 | if (current->td_flags.tasktype == TASK_EXPLICIT1 || | |||
| 2264 | current->td_taskwait_thread > 0) { // <= 0 on barrier | |||
| 2265 | level = current->td_level; | |||
| 2266 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
| 2267 | while (parent != current && parent->td_level > level) { | |||
| 2268 | parent = parent->td_parent; // check generation up to the level of the | |||
| 2269 | // current task | |||
| 2270 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2270)); | |||
| 2271 | } | |||
| 2272 | if (parent != current) { | |||
| 2273 | if (!task_team->tt.tt_untied_task_encountered) { | |||
| 2274 | // The TSC does not allow to steal victim task | |||
| 2275 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 2276 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2279)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2277 | ("__kmp_steal_task(exit #3): T#%d could not steal from "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2279)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2278 | "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2279)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2279 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2279)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2280 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #3): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2279)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
| 2281 | return NULL__null; | |||
| 2282 | } | |||
| 2283 | taskdata = NULL__null; // will check other tasks in victim's deque | |||
| 2284 | } | |||
| 2285 | } | |||
| 2286 | } | |||
| 2287 | if (taskdata != NULL__null) { | |||
| 2288 | // Bump head pointer and Wrap. | |||
| 2289 | victim_td->td.td_deque_head = | |||
| 2290 | (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 2291 | } else { | |||
| 2292 | int i; | |||
| 2293 | // walk through victim's deque trying to steal any task | |||
| 2294 | target = victim_td->td.td_deque_head; | |||
| 2295 | for (i = 1; i < ntasks; ++i) { | |||
| 2296 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 2297 | taskdata = victim_td->td.td_deque[target]; | |||
| 2298 | if (taskdata->td_flags.tiedness == TASK_TIED1) { | |||
| 2299 | // check if the candidate obeys the TSC | |||
| 2300 | kmp_taskdata_t *parent = taskdata->td_parent; | |||
| 2301 | // check generation up to the level of the current task | |||
| 2302 | while (parent != current && parent->td_level > level) { | |||
| 2303 | parent = parent->td_parent; | |||
| 2304 | KMP_DEBUG_ASSERT(parent != NULL)((parent != __null) ? 0 : __kmp_debug_assert("parent != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2304)); | |||
| 2305 | } | |||
| 2306 | if (parent != current) { | |||
| 2307 | // The TSC does not allow to steal the candidate | |||
| 2308 | taskdata = NULL__null; | |||
| 2309 | continue; | |||
| 2310 | } else { | |||
| 2311 | // found victim tied task | |||
| 2312 | break; | |||
| 2313 | } | |||
| 2314 | } else { | |||
| 2315 | // found victim untied task | |||
| 2316 | break; | |||
| 2317 | } | |||
| 2318 | } | |||
| 2319 | if (taskdata == NULL__null) { | |||
| 2320 | // No appropriate candidate to steal found | |||
| 2321 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 2322 | KA_TRACE(10, ("__kmp_steal_task(exit #4): T#%d could not steal from "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2324)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2323 | "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2324)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2324 | gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2324)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2325 | victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #4): T#%d could not steal from " "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n", gtid, (((( victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2324)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
| 2326 | return NULL__null; | |||
| 2327 | } | |||
| 2328 | int prev = target; | |||
| 2329 | for (i = i + 1; i < ntasks; ++i) { | |||
| 2330 | // shift remaining tasks in the deque left by 1 | |||
| 2331 | target = (target + 1) & TASK_DEQUE_MASK(victim_td->td)((victim_td->td).td_deque_size - 1); | |||
| 2332 | victim_td->td.td_deque[prev] = victim_td->td.td_deque[target]; | |||
| 2333 | prev = target; | |||
| 2334 | } | |||
| 2335 | KMP_DEBUG_ASSERT(victim_td->td.td_deque_tail ==((victim_td->td.td_deque_tail == ((target + 1) & ((victim_td ->td).td_deque_size - 1))) ? 0 : __kmp_debug_assert("victim_td->td.td_deque_tail == ((target + 1) & ((victim_td->td).td_deque_size - 1))" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2336)) | |||
| 2336 | ((target + 1) & TASK_DEQUE_MASK(victim_td->td)))((victim_td->td.td_deque_tail == ((target + 1) & ((victim_td ->td).td_deque_size - 1))) ? 0 : __kmp_debug_assert("victim_td->td.td_deque_tail == ((target + 1) & ((victim_td->td).td_deque_size - 1))" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2336)); | |||
| 2337 | victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped)) | |||
| 2338 | } | |||
| 2339 | if (*thread_finished) { | |||
| 2340 | // We need to un-mark this victim as a finished victim. This must be done | |||
| 2341 | // before releasing the lock, or else other threads (starting with the | |||
| 2342 | // master victim) might be prematurely released from the barrier!!! | |||
| 2343 | kmp_int32 count; | |||
| 2344 | ||||
| 2345 | count = KMP_ATOMIC_INC(unfinished_threads)(unfinished_threads)->fetch_add(1, std::memory_order_acq_rel ); | |||
| 2346 | ||||
| 2347 | 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); } | |||
| 2348 | 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); } | |||
| 2349 | ("__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); } | |||
| 2350 | 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); }; | |||
| 2351 | ||||
| 2352 | *thread_finished = FALSE0; | |||
| 2353 | } | |||
| 2354 | TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1)(victim_td->td.td_deque_ntasks) = (ntasks - 1); | |||
| 2355 | ||||
| 2356 | __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock); | |||
| 2357 | ||||
| 2358 | KMP_COUNT_BLOCK(TASK_stolen)((void)0); | |||
| 2359 | KA_TRACE(10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2362)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2360 | ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2362)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2361 | "task_team=%p ntasks=%d head=%u tail=%u\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2362)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2362 | gtid, taskdata, __kmp_gtid_from_thread(victim_thr), task_team,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2362)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); } | |||
| 2363 | ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: " "task_team=%p ntasks=%d head=%u tail=%u\n", gtid, taskdata, ( (((victim_thr) != __null) ? 0 : __kmp_debug_assert("(victim_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2362)), (victim_thr)->th.th_info.ds.ds_gtid), task_team, ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail ); }; | |||
| 2364 | ||||
| 2365 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 2366 | return task; | |||
| 2367 | } | |||
| 2368 | ||||
| 2369 | // __kmp_execute_tasks_template: Choose and execute tasks until either the | |||
| 2370 | // condition is statisfied (return true) or there are none left (return false). | |||
| 2371 | // | |||
| 2372 | // final_spin is TRUE if this is the spin at the release barrier. | |||
| 2373 | // thread_finished indicates whether the thread is finished executing all | |||
| 2374 | // the tasks it has on its deque, and is at the release barrier. | |||
| 2375 | // spinner is the location on which to spin. | |||
| 2376 | // spinner == NULL means only execute a single task and return. | |||
| 2377 | // checker is the value to check to terminate the spin. | |||
| 2378 | template <class C> | |||
| 2379 | static inline int __kmp_execute_tasks_template( | |||
| 2380 | kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin, | |||
| 2381 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 2382 | kmp_int32 is_constrained) { | |||
| 2383 | kmp_task_team_t *task_team = thread->th.th_task_team; | |||
| 2384 | kmp_thread_data_t *threads_data; | |||
| 2385 | kmp_task_t *task; | |||
| 2386 | kmp_info_t *other_thread; | |||
| 2387 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 2388 | std::atomic<kmp_int32> *unfinished_threads; | |||
| 2389 | kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0, | |||
| 2390 | tid = thread->th.th_info.ds.ds_tid; | |||
| 2391 | ||||
| 2392 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2392)); | |||
| 2393 | KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid])((thread == __kmp_threads[gtid]) ? 0 : __kmp_debug_assert("thread == __kmp_threads[gtid]" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2393)); | |||
| 2394 | ||||
| 2395 | if (task_team == NULL__null) | |||
| 2396 | return FALSE0; | |||
| 2397 | ||||
| 2398 | 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) ; } | |||
| 2399 | "*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) ; } | |||
| 2400 | 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) ; }; | |||
| 2401 | ||||
| 2402 | thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP0; | |||
| 2403 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
| 2404 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2404)); | |||
| 2405 | ||||
| 2406 | nthreads = task_team->tt.tt_nproc; | |||
| 2407 | unfinished_threads = &(task_team->tt.tt_unfinished_threads); | |||
| 2408 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2409 | KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks)((nthreads > 1 || task_team->tt.tt_found_proxy_tasks) ? 0 : __kmp_debug_assert("nthreads > 1 || task_team->tt.tt_found_proxy_tasks" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2409)); | |||
| 2410 | #else | |||
| 2411 | KMP_DEBUG_ASSERT(nthreads > 1)((nthreads > 1) ? 0 : __kmp_debug_assert("nthreads > 1" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2411)); | |||
| 2412 | #endif | |||
| 2413 | KMP_DEBUG_ASSERT(*unfinished_threads >= 0)((*unfinished_threads >= 0) ? 0 : __kmp_debug_assert("*unfinished_threads >= 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2413)); | |||
| 2414 | ||||
| 2415 | while (1) { // Outer loop keeps trying to find tasks in case of single thread | |||
| 2416 | // getting tasks from target constructs | |||
| 2417 | while (1) { // Inner loop to find a task and execute it | |||
| 2418 | task = NULL__null; | |||
| 2419 | if (use_own_tasks) { // check on own queue first | |||
| 2420 | task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained); | |||
| 2421 | } | |||
| 2422 | if ((task == NULL__null) && (nthreads > 1)) { // Steal a task | |||
| 2423 | int asleep = 1; | |||
| 2424 | use_own_tasks = 0; | |||
| 2425 | // Try to steal from the last place I stole from successfully. | |||
| 2426 | if (victim_tid == -2) { // haven't stolen anything yet | |||
| 2427 | victim_tid = threads_data[tid].td.td_deque_last_stolen; | |||
| 2428 | if (victim_tid != | |||
| 2429 | -1) // if we have a last stolen from victim, get the thread | |||
| 2430 | other_thread = threads_data[victim_tid].td.td_thr; | |||
| 2431 | } | |||
| 2432 | if (victim_tid != -1) { // found last victim | |||
| 2433 | asleep = 0; | |||
| 2434 | } else if (!new_victim) { // no recent steals and we haven't already | |||
| 2435 | // used a new victim; select a random thread | |||
| 2436 | do { // Find a different thread to steal work from. | |||
| 2437 | // Pick a random thread. Initial plan was to cycle through all the | |||
| 2438 | // threads, and only return if we tried to steal from every thread, | |||
| 2439 | // and failed. Arch says that's not such a great idea. | |||
| 2440 | victim_tid = __kmp_get_random(thread) % (nthreads - 1); | |||
| 2441 | if (victim_tid >= tid) { | |||
| 2442 | ++victim_tid; // Adjusts random distribution to exclude self | |||
| 2443 | } | |||
| 2444 | // Found a potential victim | |||
| 2445 | other_thread = threads_data[victim_tid].td.td_thr; | |||
| 2446 | // There is a slight chance that __kmp_enable_tasking() did not wake | |||
| 2447 | // up all threads waiting at the barrier. If victim is sleeping, | |||
| 2448 | // then wake it up. Since we were going to pay the cache miss | |||
| 2449 | // penalty for referencing another thread's kmp_info_t struct | |||
| 2450 | // anyway, | |||
| 2451 | // the check shouldn't cost too much performance at this point. In | |||
| 2452 | // extra barrier mode, tasks do not sleep at the separate tasking | |||
| 2453 | // barrier, so this isn't a problem. | |||
| 2454 | asleep = 0; | |||
| 2455 | if ((__kmp_tasking_mode == tskm_task_teams) && | |||
| 2456 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) && | |||
| 2457 | (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc))((void *)(const_cast<void *>(other_thread->th.th_sleep_loc ))) != | |||
| 2458 | NULL__null)) { | |||
| 2459 | asleep = 1; | |||
| 2460 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(other_thread)((((other_thread) != __null) ? 0 : __kmp_debug_assert("(other_thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2460)), (other_thread)->th.th_info.ds.ds_gtid), | |||
| 2461 | other_thread->th.th_sleep_loc); | |||
| 2462 | // A sleeping thread should not have any tasks on it's queue. | |||
| 2463 | // There is a slight possibility that it resumes, steals a task | |||
| 2464 | // from another thread, which spawns more tasks, all in the time | |||
| 2465 | // that it takes this thread to check => don't write an assertion | |||
| 2466 | // that the victim's queue is empty. Try stealing from a | |||
| 2467 | // different thread. | |||
| 2468 | } | |||
| 2469 | } while (asleep); | |||
| 2470 | } | |||
| 2471 | ||||
| 2472 | if (!asleep) { | |||
| 2473 | // We have a victim to try to steal from | |||
| 2474 | task = __kmp_steal_task(other_thread, gtid, task_team, | |||
| 2475 | unfinished_threads, thread_finished, | |||
| 2476 | is_constrained); | |||
| 2477 | } | |||
| 2478 | if (task != NULL__null) { // set last stolen to victim | |||
| 2479 | if (threads_data[tid].td.td_deque_last_stolen != victim_tid) { | |||
| 2480 | threads_data[tid].td.td_deque_last_stolen = victim_tid; | |||
| 2481 | // The pre-refactored code did not try more than 1 successful new | |||
| 2482 | // vicitm, unless the last one generated more local tasks; | |||
| 2483 | // new_victim keeps track of this | |||
| 2484 | new_victim = 1; | |||
| 2485 | } | |||
| 2486 | } else { // No tasks found; unset last_stolen | |||
| 2487 | 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); | |||
| 2488 | victim_tid = -2; // no successful victim found | |||
| 2489 | } | |||
| 2490 | } | |||
| 2491 | ||||
| 2492 | if (task == NULL__null) // break out of tasking loop | |||
| 2493 | break; | |||
| 2494 | ||||
| 2495 | // Found a task; execute it | |||
| 2496 | #if USE_ITT_BUILD1 && USE_ITT_NOTIFY1 | |||
| 2497 | if (__itt_sync_create_ptr__kmp_itt_sync_create_ptr__3_0 || KMP_ITT_DEBUG0) { | |||
| 2498 | if (itt_sync_obj == NULL__null) { // we are at fork barrier where we could not | |||
| 2499 | // get the object reliably | |||
| 2500 | itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier); | |||
| 2501 | } | |||
| 2502 | __kmp_itt_task_starting(itt_sync_obj); | |||
| 2503 | } | |||
| 2504 | #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */ | |||
| 2505 | __kmp_invoke_task(gtid, task, current_task); | |||
| 2506 | #if USE_ITT_BUILD1 | |||
| 2507 | if (itt_sync_obj != NULL__null) | |||
| 2508 | __kmp_itt_task_finished(itt_sync_obj); | |||
| 2509 | #endif /* USE_ITT_BUILD */ | |||
| 2510 | // If this thread is only partway through the barrier and the condition is | |||
| 2511 | // met, then return now, so that the barrier gather/release pattern can | |||
| 2512 | // proceed. If this thread is in the last spin loop in the barrier, | |||
| 2513 | // waiting to be released, we know that the termination condition will not | |||
| 2514 | // be satisified, so don't waste any cycles checking it. | |||
| 2515 | if (flag == NULL__null || (!final_spin && flag->done_check())) { | |||
| 2516 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 2517 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 2518 | ("__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); } | |||
| 2519 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
| 2520 | return TRUE(!0); | |||
| 2521 | } | |||
| 2522 | if (thread->th.th_task_team == NULL__null) { | |||
| 2523 | break; | |||
| 2524 | } | |||
| 2525 | // Yield before executing next task | |||
| 2526 | KMP_YIELD(__kmp_library == library_throughput){ __kmp_x86_pause(); __kmp_yield((__kmp_library == library_throughput )); }; | |||
| 2527 | // If execution of a stolen task results in more tasks being placed on our | |||
| 2528 | // run queue, reset use_own_tasks | |||
| 2529 | if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks)(threads_data[tid].td.td_deque_ntasks) != 0) { | |||
| 2530 | 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); } | |||
| 2531 | "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); } | |||
| 2532 | gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d stolen task spawned " "other tasks, restart\n", gtid); }; | |||
| 2533 | use_own_tasks = 1; | |||
| 2534 | new_victim = 0; | |||
| 2535 | } | |||
| 2536 | } | |||
| 2537 | ||||
| 2538 | // The task source has been exhausted. If in final spin loop of barrier, check | |||
| 2539 | // if termination condition is satisfied. | |||
| 2540 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2541 | // The work queue may be empty but there might be proxy tasks still | |||
| 2542 | // executing | |||
| 2543 | if (final_spin && | |||
| 2544 | KMP_ATOMIC_LD_ACQ(¤t_task->td_incomplete_child_tasks)(¤t_task->td_incomplete_child_tasks)->load(std ::memory_order_acquire) == 0) | |||
| 2545 | #else | |||
| 2546 | if (final_spin) | |||
| 2547 | #endif | |||
| 2548 | { | |||
| 2549 | // First, decrement the #unfinished threads, if that has not already been | |||
| 2550 | // done. This decrement might be to the spin location, and result in the | |||
| 2551 | // termination condition being satisfied. | |||
| 2552 | if (!*thread_finished) { | |||
| 2553 | kmp_int32 count; | |||
| 2554 | ||||
| 2555 | count = KMP_ATOMIC_DEC(unfinished_threads)(unfinished_threads)->fetch_sub(1, std::memory_order_acq_rel ) - 1; | |||
| 2556 | 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 ); } | |||
| 2557 | "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 ); } | |||
| 2558 | 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 ); }; | |||
| 2559 | *thread_finished = TRUE(!0); | |||
| 2560 | } | |||
| 2561 | ||||
| 2562 | // It is now unsafe to reference thread->th.th_team !!! | |||
| 2563 | // Decrementing task_team->tt.tt_unfinished_threads can allow the master | |||
| 2564 | // thread to pass through the barrier, where it might reset each thread's | |||
| 2565 | // th.th_team field for the next parallel region. If we can steal more | |||
| 2566 | // work, we know that this has not happened yet. | |||
| 2567 | if (flag != NULL__null && flag->done_check()) { | |||
| 2568 | KA_TRACE(if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 2569 | 15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); } | |||
| 2570 | ("__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); } | |||
| 2571 | gtid))if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n" , gtid); }; | |||
| 2572 | return TRUE(!0); | |||
| 2573 | } | |||
| 2574 | } | |||
| 2575 | ||||
| 2576 | // If this thread's task team is NULL, master has recognized that there are | |||
| 2577 | // no more tasks; bail out | |||
| 2578 | if (thread->th.th_task_team == NULL__null) { | |||
| 2579 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d no more tasks\n" , gtid); } | |||
| 2580 | ("__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); }; | |||
| 2581 | return FALSE0; | |||
| 2582 | } | |||
| 2583 | ||||
| 2584 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2585 | // We could be getting tasks from target constructs; if this is the only | |||
| 2586 | // thread, keep trying to execute tasks from own queue | |||
| 2587 | if (nthreads == 1) | |||
| 2588 | use_own_tasks = 1; | |||
| 2589 | else | |||
| 2590 | #endif | |||
| 2591 | { | |||
| 2592 | KA_TRACE(15,if (kmp_a_debug >= 15) { __kmp_debug_printf ("__kmp_execute_tasks_template: T#%d can't find work\n" , gtid); } | |||
| 2593 | ("__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); }; | |||
| 2594 | return FALSE0; | |||
| 2595 | } | |||
| 2596 | } | |||
| 2597 | } | |||
| 2598 | ||||
| 2599 | int __kmp_execute_tasks_32( | |||
| 2600 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32 *flag, int final_spin, | |||
| 2601 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 2602 | kmp_int32 is_constrained) { | |||
| 2603 | return __kmp_execute_tasks_template( | |||
| 2604 | thread, gtid, flag, final_spin, | |||
| 2605 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 2606 | } | |||
| 2607 | ||||
| 2608 | int __kmp_execute_tasks_64( | |||
| 2609 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64 *flag, int final_spin, | |||
| 2610 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 2611 | kmp_int32 is_constrained) { | |||
| 2612 | return __kmp_execute_tasks_template( | |||
| 2613 | thread, gtid, flag, final_spin, | |||
| 2614 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 2615 | } | |||
| 2616 | ||||
| 2617 | int __kmp_execute_tasks_oncore( | |||
| 2618 | kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin, | |||
| 2619 | int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, | |||
| 2620 | kmp_int32 is_constrained) { | |||
| 2621 | return __kmp_execute_tasks_template( | |||
| 2622 | thread, gtid, flag, final_spin, | |||
| 2623 | thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj, is_constrained); | |||
| 2624 | } | |||
| 2625 | ||||
| 2626 | // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the | |||
| 2627 | // next barrier so they can assist in executing enqueued tasks. | |||
| 2628 | // First thread in allocates the task team atomically. | |||
| 2629 | static void __kmp_enable_tasking(kmp_task_team_t *task_team, | |||
| 2630 | kmp_info_t *this_thr) { | |||
| 2631 | kmp_thread_data_t *threads_data; | |||
| 2632 | int nthreads, i, is_init_thread; | |||
| 2633 | ||||
| 2634 | KA_TRACE(10, ("__kmp_enable_tasking(enter): T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(enter): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2635)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
| 2635 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(enter): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2635)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
| 2636 | ||||
| 2637 | KMP_DEBUG_ASSERT(task_team != NULL)((task_team != __null) ? 0 : __kmp_debug_assert("task_team != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2637)); | |||
| 2638 | KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL)((this_thr->th.th_team != __null) ? 0 : __kmp_debug_assert ("this_thr->th.th_team != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2638)); | |||
| 2639 | ||||
| 2640 | nthreads = task_team->tt.tt_nproc; | |||
| 2641 | KMP_DEBUG_ASSERT(nthreads > 0)((nthreads > 0) ? 0 : __kmp_debug_assert("nthreads > 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2641)); | |||
| 2642 | KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc)((nthreads == this_thr->th.th_team->t.t_nproc) ? 0 : __kmp_debug_assert ("nthreads == this_thr->th.th_team->t.t_nproc", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2642)); | |||
| 2643 | ||||
| 2644 | // Allocate or increase the size of threads_data if necessary | |||
| 2645 | is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team); | |||
| 2646 | ||||
| 2647 | if (!is_init_thread) { | |||
| 2648 | // Some other thread already set up the array. | |||
| 2649 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2652)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
| 2650 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2652)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
| 2651 | ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2652)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
| 2652 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2652)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
| 2653 | return; | |||
| 2654 | } | |||
| 2655 | threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data)((void *)(task_team->tt.tt_threads_data)); | |||
| 2656 | KMP_DEBUG_ASSERT(threads_data != NULL)((threads_data != __null) ? 0 : __kmp_debug_assert("threads_data != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2656)); | |||
| 2657 | ||||
| 2658 | if ((__kmp_tasking_mode == tskm_task_teams) && | |||
| 2659 | (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647))) { | |||
| 2660 | // Release any threads sleeping at the barrier, so that they can steal | |||
| 2661 | // tasks and execute them. In extra barrier mode, tasks do not sleep | |||
| 2662 | // at the separate tasking barrier, so this isn't a problem. | |||
| 2663 | for (i = 0; i < nthreads; i++) { | |||
| 2664 | volatile void *sleep_loc; | |||
| 2665 | kmp_info_t *thread = threads_data[i].td.td_thr; | |||
| 2666 | ||||
| 2667 | if (i == this_thr->th.th_info.ds.ds_tid) { | |||
| 2668 | continue; | |||
| 2669 | } | |||
| 2670 | // Since we haven't locked the thread's suspend mutex lock at this | |||
| 2671 | // point, there is a small window where a thread might be putting | |||
| 2672 | // itself to sleep, but hasn't set the th_sleep_loc field yet. | |||
| 2673 | // To work around this, __kmp_execute_tasks_template() periodically checks | |||
| 2674 | // see if other threads are sleeping (using the same random mechanism that | |||
| 2675 | // is used for task stealing) and awakens them if they are. | |||
| 2676 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
| 2677 | NULL__null) { | |||
| 2678 | KF_TRACE(50, ("__kmp_enable_tasking: T#%d waking up thread T#%d\n",if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2679)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2680)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 2679 | __kmp_gtid_from_thread(this_thr),if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2679)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2680)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 2680 | __kmp_gtid_from_thread(thread)))if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d waking up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2679)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2680)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
| 2681 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2681)), (thread)->th.th_info.ds.ds_gtid), sleep_loc); | |||
| 2682 | } else { | |||
| 2683 | KF_TRACE(50, ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n",if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2684)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2685)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 2684 | __kmp_gtid_from_thread(this_thr),if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2684)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2685)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 2685 | __kmp_gtid_from_thread(thread)))if (kmp_f_debug >= 50) { __kmp_debug_printf ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2684)), (this_thr)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2685)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
| 2686 | } | |||
| 2687 | } | |||
| 2688 | } | |||
| 2689 | ||||
| 2690 | KA_TRACE(10, ("__kmp_enable_tasking(exit): T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2691)), (this_thr)->th.th_info.ds.ds_gtid)); } | |||
| 2691 | __kmp_gtid_from_thread(this_thr)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_enable_tasking(exit): T#%d\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2691)), (this_thr)->th.th_info.ds.ds_gtid)); }; | |||
| 2692 | } | |||
| 2693 | ||||
| 2694 | /* // TODO: Check the comment consistency | |||
| 2695 | * Utility routines for "task teams". A task team (kmp_task_t) is kind of | |||
| 2696 | * like a shadow of the kmp_team_t data struct, with a different lifetime. | |||
| 2697 | * After a child * thread checks into a barrier and calls __kmp_release() from | |||
| 2698 | * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no | |||
| 2699 | * longer assume that the kmp_team_t structure is intact (at any moment, the | |||
| 2700 | * master thread may exit the barrier code and free the team data structure, | |||
| 2701 | * and return the threads to the thread pool). | |||
| 2702 | * | |||
| 2703 | * This does not work with the the tasking code, as the thread is still | |||
| 2704 | * expected to participate in the execution of any tasks that may have been | |||
| 2705 | * spawned my a member of the team, and the thread still needs access to all | |||
| 2706 | * to each thread in the team, so that it can steal work from it. | |||
| 2707 | * | |||
| 2708 | * Enter the existence of the kmp_task_team_t struct. It employs a reference | |||
| 2709 | * counting mechanims, and is allocated by the master thread before calling | |||
| 2710 | * __kmp_<barrier_kind>_release, and then is release by the last thread to | |||
| 2711 | * exit __kmp_<barrier_kind>_release at the next barrier. I.e. the lifetimes | |||
| 2712 | * of the kmp_task_team_t structs for consecutive barriers can overlap | |||
| 2713 | * (and will, unless the master thread is the last thread to exit the barrier | |||
| 2714 | * release phase, which is not typical). | |||
| 2715 | * | |||
| 2716 | * The existence of such a struct is useful outside the context of tasking, | |||
| 2717 | * but for now, I'm trying to keep it specific to the OMP_30_ENABLED macro, | |||
| 2718 | * so that any performance differences show up when comparing the 2.5 vs. 3.0 | |||
| 2719 | * libraries. | |||
| 2720 | * | |||
| 2721 | * We currently use the existence of the threads array as an indicator that | |||
| 2722 | * tasks were spawned since the last barrier. If the structure is to be | |||
| 2723 | * useful outside the context of tasking, then this will have to change, but | |||
| 2724 | * not settting the field minimizes the performance impact of tasking on | |||
| 2725 | * barriers, when no explicit tasks were spawned (pushed, actually). | |||
| 2726 | */ | |||
| 2727 | ||||
| 2728 | static kmp_task_team_t *__kmp_free_task_teams = | |||
| 2729 | NULL__null; // Free list for task_team data structures | |||
| 2730 | // Lock for task team data structures | |||
| 2731 | kmp_bootstrap_lock_t __kmp_task_team_lock = | |||
| 2732 | KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock){ { { true } , &((__kmp_task_team_lock)), __null, { 0U }, { 0U }, { 0 }, { -1 } } }; | |||
| 2733 | ||||
| 2734 | // __kmp_alloc_task_deque: | |||
| 2735 | // Allocates a task deque for a particular thread, and initialize the necessary | |||
| 2736 | // data structures relating to the deque. This only happens once per thread | |||
| 2737 | // per task team since task teams are recycled. No lock is needed during | |||
| 2738 | // allocation since each thread allocates its own deque. | |||
| 2739 | static void __kmp_alloc_task_deque(kmp_info_t *thread, | |||
| 2740 | kmp_thread_data_t *thread_data) { | |||
| 2741 | __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2742 | KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL)((thread_data->td.td_deque == __null) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque == __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2742)); | |||
| 2743 | ||||
| 2744 | // Initialize last stolen task field to "none" | |||
| 2745 | thread_data->td.td_deque_last_stolen = -1; | |||
| 2746 | ||||
| 2747 | KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0)(((thread_data->td.td_deque_ntasks) == 0) ? 0 : __kmp_debug_assert ("(thread_data->td.td_deque_ntasks) == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2747)); | |||
| 2748 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0)((thread_data->td.td_deque_head == 0) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque_head == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2748)); | |||
| 2749 | KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0)((thread_data->td.td_deque_tail == 0) ? 0 : __kmp_debug_assert ("thread_data->td.td_deque_tail == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2749)); | |||
| 2750 | ||||
| 2751 | KE_TRACE(if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2754)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
| 2752 | 10,if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2754)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
| 2753 | ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2754)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); } | |||
| 2754 | __kmp_gtid_from_thread(thread), INITIAL_TASK_DEQUE_SIZE, thread_data))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2754)), (thread)->th.th_info.ds.ds_gtid), (1 << 8) , thread_data); }; | |||
| 2755 | // Allocate space for task deque, and zero the deque | |||
| 2756 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
| 2757 | // kmp_reap_task_team( ). | |||
| 2758 | thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2759) | |||
| 2759 | INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *))___kmp_allocate(((1 << 8) * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2759); | |||
| 2760 | thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE(1 << 8); | |||
| 2761 | } | |||
| 2762 | ||||
| 2763 | // __kmp_realloc_task_deque: | |||
| 2764 | // Re-allocates a task deque for a particular thread, copies the content from | |||
| 2765 | // the old deque and adjusts the necessary data structures relating to the | |||
| 2766 | // deque. This operation must be done with a the deque_lock being held | |||
| 2767 | static void __kmp_realloc_task_deque(kmp_info_t *thread, | |||
| 2768 | kmp_thread_data_t *thread_data) { | |||
| 2769 | kmp_int32 size = TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size); | |||
| 2770 | kmp_int32 new_size = 2 * size; | |||
| 2771 | ||||
| 2772 | KE_TRACE(10, ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2774)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); } | |||
| 2773 | "%d] for thread_data %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2774)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); } | |||
| 2774 | __kmp_gtid_from_thread(thread), size, new_size, thread_data))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to " "%d] for thread_data %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2774)), (thread)->th.th_info.ds.ds_gtid), size, new_size , thread_data); }; | |||
| 2775 | ||||
| 2776 | kmp_taskdata_t **new_deque = | |||
| 2777 | (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *))___kmp_allocate((new_size * sizeof(kmp_taskdata_t *)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2777); | |||
| 2778 | ||||
| 2779 | int i, j; | |||
| 2780 | for (i = thread_data->td.td_deque_head, j = 0; j < size; | |||
| 2781 | i = (i + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1), j++) | |||
| 2782 | new_deque[j] = thread_data->td.td_deque[i]; | |||
| 2783 | ||||
| 2784 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2784); | |||
| 2785 | ||||
| 2786 | thread_data->td.td_deque_head = 0; | |||
| 2787 | thread_data->td.td_deque_tail = size; | |||
| 2788 | thread_data->td.td_deque = new_deque; | |||
| 2789 | thread_data->td.td_deque_size = new_size; | |||
| 2790 | } | |||
| 2791 | ||||
| 2792 | // __kmp_free_task_deque: | |||
| 2793 | // Deallocates a task deque for a particular thread. Happens at library | |||
| 2794 | // deallocation so don't need to reset all thread data fields. | |||
| 2795 | static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) { | |||
| 2796 | if (thread_data->td.td_deque != NULL__null) { | |||
| 2797 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2798 | TCW_4(thread_data->td.td_deque_ntasks, 0)(thread_data->td.td_deque_ntasks) = (0); | |||
| 2799 | __kmp_free(thread_data->td.td_deque)___kmp_free((thread_data->td.td_deque), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2799); | |||
| 2800 | thread_data->td.td_deque = NULL__null; | |||
| 2801 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 2802 | } | |||
| 2803 | ||||
| 2804 | #ifdef BUILD_TIED_TASK_STACK | |||
| 2805 | // GEH: Figure out what to do here for td_susp_tied_tasks | |||
| 2806 | if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) { | |||
| 2807 | __kmp_free_task_stack(__kmp_thread_from_gtid(gtid)((((gtid) >= 0) ? 0 : __kmp_debug_assert("(gtid) >= 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2807)), __kmp_threads[(gtid)]), thread_data); | |||
| 2808 | } | |||
| 2809 | #endif // BUILD_TIED_TASK_STACK | |||
| 2810 | } | |||
| 2811 | ||||
| 2812 | // __kmp_realloc_task_threads_data: | |||
| 2813 | // Allocates a threads_data array for a task team, either by allocating an | |||
| 2814 | // initial array or enlarging an existing array. Only the first thread to get | |||
| 2815 | // the lock allocs or enlarges the array and re-initializes the array eleemnts. | |||
| 2816 | // That thread returns "TRUE", the rest return "FALSE". | |||
| 2817 | // Assumes that the new array size is given by task_team -> tt.tt_nproc. | |||
| 2818 | // The current size is given by task_team -> tt.tt_max_threads. | |||
| 2819 | static int __kmp_realloc_task_threads_data(kmp_info_t *thread, | |||
| 2820 | kmp_task_team_t *task_team) { | |||
| 2821 | kmp_thread_data_t **threads_data_p; | |||
| 2822 | kmp_int32 nthreads, maxthreads; | |||
| 2823 | int is_init_thread = FALSE0; | |||
| 2824 | ||||
| 2825 | if (TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
| 2826 | // Already reallocated and initialized. | |||
| 2827 | return FALSE0; | |||
| 2828 | } | |||
| 2829 | ||||
| 2830 | threads_data_p = &task_team->tt.tt_threads_data; | |||
| 2831 | nthreads = task_team->tt.tt_nproc; | |||
| 2832 | maxthreads = task_team->tt.tt_max_threads; | |||
| 2833 | ||||
| 2834 | // All threads must lock when they encounter the first task of the implicit | |||
| 2835 | // task region to make sure threads_data fields are (re)initialized before | |||
| 2836 | // used. | |||
| 2837 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 2838 | ||||
| 2839 | if (!TCR_4(task_team->tt.tt_found_tasks)(task_team->tt.tt_found_tasks)) { | |||
| 2840 | // first thread to enable tasking | |||
| 2841 | kmp_team_t *team = thread->th.th_team; | |||
| 2842 | int i; | |||
| 2843 | ||||
| 2844 | is_init_thread = TRUE(!0); | |||
| 2845 | if (maxthreads < nthreads) { | |||
| 2846 | ||||
| 2847 | if (*threads_data_p != NULL__null) { | |||
| 2848 | kmp_thread_data_t *old_data = *threads_data_p; | |||
| 2849 | kmp_thread_data_t *new_data = NULL__null; | |||
| 2850 | ||||
| 2851 | KE_TRACE(if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2855)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
| 2852 | 10,if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2855)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
| 2853 | ("__kmp_realloc_task_threads_data: T#%d reallocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2855)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
| 2854 | "threads data for task_team %p, new_size = %d, old_size = %d\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2855)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); } | |||
| 2855 | __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d reallocating " "threads data for task_team %p, new_size = %d, old_size = %d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2855)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads , maxthreads); }; | |||
| 2856 | // Reallocate threads_data to have more elements than current array | |||
| 2857 | // Cannot use __kmp_thread_realloc() because threads not around for | |||
| 2858 | // kmp_reap_task_team( ). Note all new array entries are initialized | |||
| 2859 | // to zero by __kmp_allocate(). | |||
| 2860 | new_data = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2861) | |||
| 2861 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2861); | |||
| 2862 | // copy old data to new data | |||
| 2863 | 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)) | |||
| 2864 | (void *)old_data, maxthreads * sizeof(kmp_thread_data_t))memcpy((void *)new_data, (void *)old_data, maxthreads * sizeof (kmp_thread_data_t)); | |||
| 2865 | ||||
| 2866 | #ifdef BUILD_TIED_TASK_STACK | |||
| 2867 | // GEH: Figure out if this is the right thing to do | |||
| 2868 | for (i = maxthreads; i < nthreads; i++) { | |||
| 2869 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 2870 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2870)), (thread)->th.th_info.ds.ds_gtid), thread_data); | |||
| 2871 | } | |||
| 2872 | #endif // BUILD_TIED_TASK_STACK | |||
| 2873 | // Install the new data and free the old data | |||
| 2874 | (*threads_data_p) = new_data; | |||
| 2875 | __kmp_free(old_data)___kmp_free((old_data), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2875); | |||
| 2876 | } else { | |||
| 2877 | KE_TRACE(10, ("__kmp_realloc_task_threads_data: T#%d allocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2879)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); } | |||
| 2878 | "threads data for task_team %p, size = %d\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2879)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); } | |||
| 2879 | __kmp_gtid_from_thread(thread), task_team, nthreads))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_realloc_task_threads_data: T#%d allocating " "threads data for task_team %p, size = %d\n", ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2879)), (thread)->th.th_info.ds.ds_gtid), task_team, nthreads ); }; | |||
| 2880 | // Make the initial allocate for threads_data array, and zero entries | |||
| 2881 | // Cannot use __kmp_thread_calloc() because threads not around for | |||
| 2882 | // kmp_reap_task_team( ). | |||
| 2883 | ANNOTATE_IGNORE_WRITES_BEGIN(); | |||
| 2884 | *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2885) | |||
| 2885 | nthreads * sizeof(kmp_thread_data_t))___kmp_allocate((nthreads * sizeof(kmp_thread_data_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2885); | |||
| 2886 | ANNOTATE_IGNORE_WRITES_END(); | |||
| 2887 | #ifdef BUILD_TIED_TASK_STACK | |||
| 2888 | // GEH: Figure out if this is the right thing to do | |||
| 2889 | for (i = 0; i < nthreads; i++) { | |||
| 2890 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 2891 | __kmp_init_task_stack(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2891)), (thread)->th.th_info.ds.ds_gtid), thread_data); | |||
| 2892 | } | |||
| 2893 | #endif // BUILD_TIED_TASK_STACK | |||
| 2894 | } | |||
| 2895 | task_team->tt.tt_max_threads = nthreads; | |||
| 2896 | } else { | |||
| 2897 | // If array has (more than) enough elements, go ahead and use it | |||
| 2898 | KMP_DEBUG_ASSERT(*threads_data_p != NULL)((*threads_data_p != __null) ? 0 : __kmp_debug_assert("*threads_data_p != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2898)); | |||
| 2899 | } | |||
| 2900 | ||||
| 2901 | // initialize threads_data pointers back to thread_info structures | |||
| 2902 | for (i = 0; i < nthreads; i++) { | |||
| 2903 | kmp_thread_data_t *thread_data = &(*threads_data_p)[i]; | |||
| 2904 | thread_data->td.td_thr = team->t.t_threads[i]; | |||
| 2905 | ||||
| 2906 | if (thread_data->td.td_deque_last_stolen >= nthreads) { | |||
| 2907 | // The last stolen field survives across teams / barrier, and the number | |||
| 2908 | // of threads may have changed. It's possible (likely?) that a new | |||
| 2909 | // parallel region will exhibit the same behavior as previous region. | |||
| 2910 | thread_data->td.td_deque_last_stolen = -1; | |||
| 2911 | } | |||
| 2912 | } | |||
| 2913 | ||||
| 2914 | KMP_MB(); | |||
| 2915 | TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE)(task_team->tt.tt_found_tasks) = ((!0)); | |||
| 2916 | } | |||
| 2917 | ||||
| 2918 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 2919 | return is_init_thread; | |||
| 2920 | } | |||
| 2921 | ||||
| 2922 | // __kmp_free_task_threads_data: | |||
| 2923 | // Deallocates a threads_data array for a task team, including any attached | |||
| 2924 | // tasking deques. Only occurs at library shutdown. | |||
| 2925 | static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) { | |||
| 2926 | __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 2927 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
| 2928 | int i; | |||
| 2929 | for (i = 0; i < task_team->tt.tt_max_threads; i++) { | |||
| 2930 | __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]); | |||
| 2931 | } | |||
| 2932 | __kmp_free(task_team->tt.tt_threads_data)___kmp_free((task_team->tt.tt_threads_data), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2932); | |||
| 2933 | task_team->tt.tt_threads_data = NULL__null; | |||
| 2934 | } | |||
| 2935 | __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 2936 | } | |||
| 2937 | ||||
| 2938 | // __kmp_allocate_task_team: | |||
| 2939 | // Allocates a task team associated with a specific team, taking it from | |||
| 2940 | // the global task team free list if possible. Also initializes data | |||
| 2941 | // structures. | |||
| 2942 | static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread, | |||
| 2943 | kmp_team_t *team) { | |||
| 2944 | kmp_task_team_t *task_team = NULL__null; | |||
| 2945 | int nthreads; | |||
| 2946 | ||||
| 2947 | KA_TRACE(20, ("__kmp_allocate_task_team: T#%d entering; team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d entering; team = %p\n" , (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2948)), (thread)->th.th_info.ds.ds_gtid) : -1), team); } | |||
| 2948 | (thread ? __kmp_gtid_from_thread(thread) : -1), team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d entering; team = %p\n" , (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2948)), (thread)->th.th_info.ds.ds_gtid) : -1), team); }; | |||
| 2949 | ||||
| 2950 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
| 2951 | // Take a task team from the task team pool | |||
| 2952 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 2953 | if (__kmp_free_task_teams != NULL__null) { | |||
| 2954 | task_team = __kmp_free_task_teams; | |||
| 2955 | TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next)((__kmp_free_task_teams)) = ((task_team->tt.tt_next)); | |||
| 2956 | task_team->tt.tt_next = NULL__null; | |||
| 2957 | } | |||
| 2958 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 2959 | } | |||
| 2960 | ||||
| 2961 | if (task_team == NULL__null) { | |||
| 2962 | KE_TRACE(10, ("__kmp_allocate_task_team: T#%d allocating "if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2964)), (thread)->th.th_info.ds.ds_gtid), team); } | |||
| 2963 | "task team for team %p\n",if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2964)), (thread)->th.th_info.ds.ds_gtid), team); } | |||
| 2964 | __kmp_gtid_from_thread(thread), team))if (kmp_e_debug >= 10) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d allocating " "task team for team %p\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2964)), (thread)->th.th_info.ds.ds_gtid), team); }; | |||
| 2965 | // Allocate a new task team if one is not available. | |||
| 2966 | // Cannot use __kmp_thread_malloc() because threads not around for | |||
| 2967 | // kmp_reap_task_team( ). | |||
| 2968 | task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t))___kmp_allocate((sizeof(kmp_task_team_t)), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2968); | |||
| 2969 | __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock); | |||
| 2970 | // AC: __kmp_allocate zeroes returned memory | |||
| 2971 | // task_team -> tt.tt_threads_data = NULL; | |||
| 2972 | // task_team -> tt.tt_max_threads = 0; | |||
| 2973 | // task_team -> tt.tt_next = NULL; | |||
| 2974 | } | |||
| 2975 | ||||
| 2976 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
| 2977 | #if OMP_45_ENABLED(50 >= 45) | |||
| 2978 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 2979 | #endif | |||
| 2980 | task_team->tt.tt_nproc = nthreads = team->t.t_nproc; | |||
| 2981 | ||||
| 2982 | KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads)(&task_team->tt.tt_unfinished_threads)->store(nthreads , std::memory_order_release); | |||
| 2983 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
| 2984 | ||||
| 2985 | KA_TRACE(20, ("__kmp_allocate_task_team: T#%d exiting; task_team = %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2987)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
| 2986 | "unfinished_threads init'd to %d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2987)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
| 2987 | (thread ? __kmp_gtid_from_thread(thread) : -1), task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2987)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); } | |||
| 2988 | KMP_ATOMIC_LD_RLX(&task_team->tt.tt_unfinished_threads)))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_allocate_task_team: T#%d exiting; task_team = %p " "unfinished_threads init'd to %d\n", (thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2987)), (thread)->th.th_info.ds.ds_gtid) : -1), task_team , (&task_team->tt.tt_unfinished_threads)->load(std:: memory_order_relaxed)); }; | |||
| 2989 | return task_team; | |||
| 2990 | } | |||
| 2991 | ||||
| 2992 | // __kmp_free_task_team: | |||
| 2993 | // Frees the task team associated with a specific thread, and adds it | |||
| 2994 | // to the global task team free list. | |||
| 2995 | void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) { | |||
| 2996 | KA_TRACE(20, ("__kmp_free_task_team: T#%d task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_team: T#%d task_team = %p\n" , thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2997)), (thread)->th.th_info.ds.ds_gtid) : -1, task_team ); } | |||
| 2997 | thread ? __kmp_gtid_from_thread(thread) : -1, task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_free_task_team: T#%d task_team = %p\n" , thread ? ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 2997)), (thread)->th.th_info.ds.ds_gtid) : -1, task_team ); }; | |||
| 2998 | ||||
| 2999 | // Put task team back on free list | |||
| 3000 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3001 | ||||
| 3002 | KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL)((task_team->tt.tt_next == __null) ? 0 : __kmp_debug_assert ("task_team->tt.tt_next == __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3002)); | |||
| 3003 | task_team->tt.tt_next = __kmp_free_task_teams; | |||
| 3004 | TCW_PTR(__kmp_free_task_teams, task_team)((__kmp_free_task_teams)) = ((task_team)); | |||
| 3005 | ||||
| 3006 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3007 | } | |||
| 3008 | ||||
| 3009 | // __kmp_reap_task_teams: | |||
| 3010 | // Free all the task teams on the task team free list. | |||
| 3011 | // Should only be done during library shutdown. | |||
| 3012 | // Cannot do anything that needs a thread structure or gtid since they are | |||
| 3013 | // already gone. | |||
| 3014 | void __kmp_reap_task_teams(void) { | |||
| 3015 | kmp_task_team_t *task_team; | |||
| 3016 | ||||
| 3017 | if (TCR_PTR(__kmp_free_task_teams)((void *)(__kmp_free_task_teams)) != NULL__null) { | |||
| 3018 | // Free all task_teams on the free list | |||
| 3019 | __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3020 | while ((task_team = __kmp_free_task_teams) != NULL__null) { | |||
| 3021 | __kmp_free_task_teams = task_team->tt.tt_next; | |||
| 3022 | task_team->tt.tt_next = NULL__null; | |||
| 3023 | ||||
| 3024 | // Free threads_data if necessary | |||
| 3025 | if (task_team->tt.tt_threads_data != NULL__null) { | |||
| 3026 | __kmp_free_task_threads_data(task_team); | |||
| 3027 | } | |||
| 3028 | __kmp_free(task_team)___kmp_free((task_team), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3028); | |||
| 3029 | } | |||
| 3030 | __kmp_release_bootstrap_lock(&__kmp_task_team_lock); | |||
| 3031 | } | |||
| 3032 | } | |||
| 3033 | ||||
| 3034 | // __kmp_wait_to_unref_task_teams: | |||
| 3035 | // Some threads could still be in the fork barrier release code, possibly | |||
| 3036 | // trying to steal tasks. Wait for each thread to unreference its task team. | |||
| 3037 | void __kmp_wait_to_unref_task_teams(void) { | |||
| 3038 | kmp_info_t *thread; | |||
| 3039 | kmp_uint32 spins; | |||
| 3040 | int done; | |||
| 3041 | ||||
| 3042 | KMP_INIT_YIELD(spins){ (spins) = __kmp_yield_init; }; | |||
| 3043 | ||||
| 3044 | for (;;) { | |||
| 3045 | done = TRUE(!0); | |||
| 3046 | ||||
| 3047 | // TODO: GEH - this may be is wrong because some sync would be necessary | |||
| 3048 | // in case threads are added to the pool during the traversal. Need to | |||
| 3049 | // verify that lock for thread pool is held when calling this routine. | |||
| 3050 | for (thread = CCAST(kmp_info_t *, __kmp_thread_pool)const_cast<kmp_info_t *>(__kmp_thread_pool); thread != NULL__null; | |||
| 3051 | thread = thread->th.th_next_pool) { | |||
| 3052 | #if KMP_OS_WINDOWS0 | |||
| 3053 | DWORD exit_val; | |||
| 3054 | #endif | |||
| 3055 | if (TCR_PTR(thread->th.th_task_team)((void *)(thread->th.th_task_team)) == NULL__null) { | |||
| 3056 | KA_TRACE(10, ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3057)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3057 | __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3057)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
| 3058 | continue; | |||
| 3059 | } | |||
| 3060 | #if KMP_OS_WINDOWS0 | |||
| 3061 | // TODO: GEH - add this check for Linux* OS / OS X* as well? | |||
| 3062 | if (!__kmp_is_thread_alive(thread, &exit_val)) { | |||
| 3063 | thread->th.th_task_team = NULL__null; | |||
| 3064 | continue; | |||
| 3065 | } | |||
| 3066 | #endif | |||
| 3067 | ||||
| 3068 | done = FALSE0; // Because th_task_team pointer is not NULL for this thread | |||
| 3069 | ||||
| 3070 | KA_TRACE(10, ("__kmp_wait_to_unref_task_team: Waiting for T#%d to "if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3072)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3071 | "unreference task_team\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3072)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3072 | __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: Waiting for T#%d to " "unreference task_team\n", ((((thread) != __null) ? 0 : __kmp_debug_assert ("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3072)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
| 3073 | ||||
| 3074 | if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME(2147483647)) { | |||
| 3075 | volatile void *sleep_loc; | |||
| 3076 | // If the thread is sleeping, awaken it. | |||
| 3077 | if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))((void *)(const_cast<void *>(thread->th.th_sleep_loc )))) != | |||
| 3078 | NULL__null) { | |||
| 3079 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3080 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3081 | ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n",if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid)); } | |||
| 3082 | __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread)))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n" , ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid), ((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3082)), (thread)->th.th_info.ds.ds_gtid)); }; | |||
| 3083 | __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread)((((thread) != __null) ? 0 : __kmp_debug_assert("(thread) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3083)), (thread)->th.th_info.ds.ds_gtid), sleep_loc); | |||
| 3084 | } | |||
| 3085 | } | |||
| 3086 | } | |||
| 3087 | if (done) { | |||
| 3088 | break; | |||
| 3089 | } | |||
| 3090 | ||||
| 3091 | // If we are oversubscribed, or have waited a bit (and library mode is | |||
| 3092 | // throughput), yield. Pause is in the following code. | |||
| 3093 | KMP_YIELD(TCR_4(__kmp_nth) > __kmp_avail_proc){ __kmp_x86_pause(); __kmp_yield(((__kmp_nth) > __kmp_avail_proc )); }; | |||
| 3094 | KMP_YIELD_SPIN(spins){ __kmp_x86_pause(); (spins) -= 2; if (!(spins)) { __kmp_yield (1); (spins) = __kmp_yield_next; } }; // Yields only if KMP_LIBRARY=throughput | |||
| 3095 | } | |||
| 3096 | } | |||
| 3097 | ||||
| 3098 | // __kmp_task_team_setup: Create a task_team for the current team, but use | |||
| 3099 | // an already created, unused one if it already exists. | |||
| 3100 | void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) { | |||
| 3101 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3101)); | |||
| 3102 | ||||
| 3103 | // If this task_team hasn't been created yet, allocate it. It will be used in | |||
| 3104 | // the region after the next. | |||
| 3105 | // If it exists, it is the current task team and shouldn't be touched yet as | |||
| 3106 | // it may still be in use. | |||
| 3107 | if (team->t.t_task_team[this_thr->th.th_task_state] == NULL__null && | |||
| 3108 | (always || team->t.t_nproc > 1)) { | |||
| 3109 | team->t.t_task_team[this_thr->th.th_task_state] = | |||
| 3110 | __kmp_allocate_task_team(this_thr, team); | |||
| 3111 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created new task_team %p "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
| 3112 | "for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
| 3113 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
| 3114 | team->t.t_task_team[this_thr->th.th_task_state],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
| 3115 | ((team != NULL) ? team->t.t_id : -1),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); } | |||
| 3116 | this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created new task_team %p " "for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3113)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [this_thr->th.th_task_state], ((team != __null) ? team-> t.t_id : -1), this_thr->th.th_task_state); }; | |||
| 3117 | } | |||
| 3118 | ||||
| 3119 | // After threads exit the release, they will call sync, and then point to this | |||
| 3120 | // other task_team; make sure it is allocated and properly initialized. As | |||
| 3121 | // threads spin in the barrier release phase, they will continue to use the | |||
| 3122 | // previous task_team struct(above), until they receive the signal to stop | |||
| 3123 | // checking for tasks (they can't safely reference the kmp_team_t struct, | |||
| 3124 | // which could be reallocated by the master thread). No task teams are formed | |||
| 3125 | // for serialized teams. | |||
| 3126 | if (team->t.t_nproc > 1) { | |||
| 3127 | int other_team = 1 - this_thr->th.th_task_state; | |||
| 3128 | if (team->t.t_task_team[other_team] == NULL__null) { // setup other team as well | |||
| 3129 | team->t.t_task_team[other_team] = | |||
| 3130 | __kmp_allocate_task_team(this_thr, team); | |||
| 3131 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created second new "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3133)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3132 | "task_team %p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3133)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3133 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3133)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3134 | team->t.t_task_team[other_team],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3133)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3135 | ((team != NULL) ? team->t.t_id : -1), other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d created second new " "task_team %p for team %d at parity=%d\n", ((((this_thr) != __null ) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3133)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); }; | |||
| 3136 | } else { // Leave the old task team struct in place for the upcoming region; | |||
| 3137 | // adjust as needed | |||
| 3138 | kmp_task_team_t *task_team = team->t.t_task_team[other_team]; | |||
| 3139 | if (!task_team->tt.tt_active || | |||
| 3140 | team->t.t_nproc != task_team->tt.tt_nproc) { | |||
| 3141 | TCW_4(task_team->tt.tt_nproc, team->t.t_nproc)(task_team->tt.tt_nproc) = (team->t.t_nproc); | |||
| 3142 | TCW_4(task_team->tt.tt_found_tasks, FALSE)(task_team->tt.tt_found_tasks) = (0); | |||
| 3143 | #if OMP_45_ENABLED(50 >= 45) | |||
| 3144 | TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 3145 | #endif | |||
| 3146 | 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) | |||
| 3147 | team->t.t_nproc)(&task_team->tt.tt_unfinished_threads)->store(team-> t.t_nproc, std::memory_order_release); | |||
| 3148 | TCW_4(task_team->tt.tt_active, TRUE)(task_team->tt.tt_active) = ((!0)); | |||
| 3149 | } | |||
| 3150 | // if team size has changed, the first thread to enable tasking will | |||
| 3151 | // realloc threads_data if necessary | |||
| 3152 | KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d reset next task_team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3154)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3153 | "%p for team %d at parity=%d\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3154)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3154 | __kmp_gtid_from_thread(this_thr),if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3154)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3155 | team->t.t_task_team[other_team],if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3154)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); } | |||
| 3156 | ((team != NULL) ? team->t.t_id : -1), other_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_setup: Master T#%d reset next task_team " "%p for team %d at parity=%d\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3154)), (this_thr)->th.th_info.ds.ds_gtid), team->t.t_task_team [other_team], ((team != __null) ? team->t.t_id : -1), other_team ); }; | |||
| 3157 | } | |||
| 3158 | } | |||
| 3159 | } | |||
| 3160 | ||||
| 3161 | // __kmp_task_team_sync: Propagation of task team data from team to threads | |||
| 3162 | // which happens just after the release phase of a team barrier. This may be | |||
| 3163 | // called by any thread, but only for teams with # threads > 1. | |||
| 3164 | void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) { | |||
| 3165 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3165)); | |||
| 3166 | ||||
| 3167 | // Toggle the th_task_state field, to switch which task_team this thread | |||
| 3168 | // refers to | |||
| 3169 | this_thr->th.th_task_state = 1 - this_thr->th.th_task_state; | |||
| 3170 | // It is now safe to propagate the task team pointer from the team struct to | |||
| 3171 | // the current thread. | |||
| 3172 | 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])) | |||
| 3173 | 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])); | |||
| 3174 | KA_TRACE(20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3177)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
| 3175 | ("__kmp_task_team_sync: Thread T#%d task team switched to task_team "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3177)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
| 3176 | "%p from Team #%d (parity=%d)\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3177)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
| 3177 | __kmp_gtid_from_thread(this_thr), this_thr->th.th_task_team,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3177)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); } | |||
| 3178 | ((team != NULL) ? team->t.t_id : -1), this_thr->th.th_task_state))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_sync: Thread T#%d task team switched to task_team " "%p from Team #%d (parity=%d)\n", ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3177)), (this_thr)->th.th_info.ds.ds_gtid), this_thr-> th.th_task_team, ((team != __null) ? team->t.t_id : -1), this_thr ->th.th_task_state); }; | |||
| 3179 | } | |||
| 3180 | ||||
| 3181 | // __kmp_task_team_wait: Master thread waits for outstanding tasks after the | |||
| 3182 | // barrier gather phase. Only called by master thread if #threads in team > 1 or | |||
| 3183 | // if proxy tasks were created. | |||
| 3184 | // | |||
| 3185 | // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off | |||
| 3186 | // by passing in 0 optionally as the last argument. When wait is zero, master | |||
| 3187 | // thread does not wait for unfinished_threads to reach 0. | |||
| 3188 | void __kmp_task_team_wait( | |||
| 3189 | kmp_info_t *this_thr, | |||
| 3190 | kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), void *itt_sync_obj, int wait) { | |||
| 3191 | kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state]; | |||
| 3192 | ||||
| 3193 | KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec)((__kmp_tasking_mode != tskm_immediate_exec) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode != tskm_immediate_exec", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3193)); | |||
| 3194 | KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team)((task_team == this_thr->th.th_task_team) ? 0 : __kmp_debug_assert ("task_team == this_thr->th.th_task_team", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3194)); | |||
| 3195 | ||||
| 3196 | if ((task_team != NULL__null) && KMP_TASKING_ENABLED(task_team)(((task_team)->tt.tt_found_tasks) == (!0))) { | |||
| 3197 | if (wait) { | |||
| 3198 | KA_TRACE(20, ("__kmp_task_team_wait: Master T#%d waiting for all tasks "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3199 | "(for unfinished_threads to reach 0) on task_team = %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3200 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d waiting for all tasks " "(for unfinished_threads to reach 0) on task_team = %p\n", ( (((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3200)), (this_thr)->th.th_info.ds.ds_gtid), task_team); }; | |||
| 3201 | // Worker threads may have dropped through to release phase, but could | |||
| 3202 | // still be executing tasks. Wait here for tasks to complete. To avoid | |||
| 3203 | // memory contention, only master thread checks termination condition. | |||
| 3204 | kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads) | |||
| 3205 | &task_team->tt.tt_unfinished_threads)reinterpret_cast<std::atomic<kmp_uint32> *>(& task_team->tt.tt_unfinished_threads), | |||
| 3206 | 0U); | |||
| 3207 | flag.wait(this_thr, TRUE(!0) USE_ITT_BUILD_ARG(itt_sync_obj), itt_sync_obj); | |||
| 3208 | } | |||
| 3209 | // Deactivate the old task team, so that the worker threads will stop | |||
| 3210 | // referencing it while spinning. | |||
| 3211 | KA_TRACE(if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3215)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3212 | 20,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3215)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3213 | ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3215)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3214 | "setting active to false, setting local and team's pointer to NULL\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3215)), (this_thr)->th.th_info.ds.ds_gtid), task_team); } | |||
| 3215 | __kmp_gtid_from_thread(this_thr), task_team))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: " "setting active to false, setting local and team's pointer to NULL\n" , ((((this_thr) != __null) ? 0 : __kmp_debug_assert("(this_thr) != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3215)), (this_thr)->th.th_info.ds.ds_gtid), task_team); }; | |||
| 3216 | #if OMP_45_ENABLED(50 >= 45) | |||
| 3217 | KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||((task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)) ? 0 : __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3218)) | |||
| 3218 | task_team->tt.tt_found_proxy_tasks == TRUE)((task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)) ? 0 : __kmp_debug_assert("task_team->tt.tt_nproc > 1 || task_team->tt.tt_found_proxy_tasks == (!0)" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3218)); | |||
| 3219 | TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE)(task_team->tt.tt_found_proxy_tasks) = (0); | |||
| 3220 | #else | |||
| 3221 | KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1)((task_team->tt.tt_nproc > 1) ? 0 : __kmp_debug_assert( "task_team->tt.tt_nproc > 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3221)); | |||
| 3222 | #endif | |||
| 3223 | 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); | |||
| 3224 | TCW_SYNC_4(task_team->tt.tt_active, FALSE)(task_team->tt.tt_active) = (0); | |||
| 3225 | KMP_MB(); | |||
| 3226 | ||||
| 3227 | TCW_PTR(this_thr->th.th_task_team, NULL)((this_thr->th.th_task_team)) = ((__null)); | |||
| 3228 | } | |||
| 3229 | } | |||
| 3230 | ||||
| 3231 | // __kmp_tasking_barrier: | |||
| 3232 | // This routine may only called when __kmp_tasking_mode == tskm_extra_barrier. | |||
| 3233 | // Internal function to execute all tasks prior to a regular barrier or a join | |||
| 3234 | // barrier. It is a full barrier itself, which unfortunately turns regular | |||
| 3235 | // barriers into double barriers and join barriers into 1 1/2 barriers. | |||
| 3236 | void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) { | |||
| 3237 | 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 ) | |||
| 3238 | std::atomic<kmp_uint32> *,reinterpret_cast<std::atomic<kmp_uint32> *>(& team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads ) | |||
| 3239 | &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 ); | |||
| 3240 | int flag = FALSE0; | |||
| 3241 | KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier)((__kmp_tasking_mode == tskm_extra_barrier) ? 0 : __kmp_debug_assert ("__kmp_tasking_mode == tskm_extra_barrier", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3241)); | |||
| 3242 | ||||
| 3243 | #if USE_ITT_BUILD1 | |||
| 3244 | 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"); | |||
| 3245 | #endif /* USE_ITT_BUILD */ | |||
| 3246 | kmp_flag_32 spin_flag(spin, 0U); | |||
| 3247 | while (!spin_flag.execute_tasks(thread, gtid, TRUE(!0), | |||
| 3248 | &flag USE_ITT_BUILD_ARG(NULL), __null, 0)) { | |||
| 3249 | #if USE_ITT_BUILD1 | |||
| 3250 | // TODO: What about itt_sync_obj?? | |||
| 3251 | 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); | |||
| 3252 | #endif /* USE_ITT_BUILD */ | |||
| 3253 | ||||
| 3254 | if (TCR_4(__kmp_global.g.g_done)(__kmp_global.g.g_done)) { | |||
| 3255 | if (__kmp_global.g.g_abort) | |||
| 3256 | __kmp_abort_thread(); | |||
| 3257 | break; | |||
| 3258 | } | |||
| 3259 | KMP_YIELD(TRUE){ __kmp_x86_pause(); __kmp_yield(((!0))); }; // GH: We always yield here | |||
| 3260 | } | |||
| 3261 | #if USE_ITT_BUILD1 | |||
| 3262 | 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); | |||
| 3263 | #endif /* USE_ITT_BUILD */ | |||
| 3264 | } | |||
| 3265 | ||||
| 3266 | #if OMP_45_ENABLED(50 >= 45) | |||
| 3267 | ||||
| 3268 | // __kmp_give_task puts a task into a given thread queue if: | |||
| 3269 | // - the queue for that thread was created | |||
| 3270 | // - there's space in that queue | |||
| 3271 | // Because of this, __kmp_push_task needs to check if there's space after | |||
| 3272 | // getting the lock | |||
| 3273 | static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task, | |||
| 3274 | kmp_int32 pass) { | |||
| 3275 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 3276 | kmp_task_team_t *task_team = taskdata->td_task_team; | |||
| 3277 | ||||
| 3278 | 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); } | |||
| 3279 | taskdata, tid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_give_task: trying to give task %p to thread %d.\n" , taskdata, tid); }; | |||
| 3280 | ||||
| 3281 | // If task_team is NULL something went really bad... | |||
| 3282 | KMP_DEBUG_ASSERT(task_team != NULL)((task_team != __null) ? 0 : __kmp_debug_assert("task_team != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3282)); | |||
| 3283 | ||||
| 3284 | bool result = false; | |||
| 3285 | kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid]; | |||
| 3286 | ||||
| 3287 | if (thread_data->td.td_deque == NULL__null) { | |||
| 3288 | // There's no queue in this thread, go find another one | |||
| 3289 | // We're guaranteed that at least one thread has a queue | |||
| 3290 | 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); } | |||
| 3291 | ("__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); } | |||
| 3292 | 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); }; | |||
| 3293 | return result; | |||
| 3294 | } | |||
| 3295 | ||||
| 3296 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 3297 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 3298 | 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); } | |||
| 3299 | 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); } | |||
| 3300 | ("__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); } | |||
| 3301 | 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); }; | |||
| 3302 | ||||
| 3303 | // if this deque is bigger than the pass ratio give a chance to another | |||
| 3304 | // thread | |||
| 3305 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
| 3306 | return result; | |||
| 3307 | ||||
| 3308 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3309 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 3310 | ||||
| 3311 | } else { | |||
| 3312 | ||||
| 3313 | __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3314 | ||||
| 3315 | if (TCR_4(thread_data->td.td_deque_ntasks)(thread_data->td.td_deque_ntasks) >= | |||
| 3316 | TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size)) { | |||
| 3317 | 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); } | |||
| 3318 | "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); } | |||
| 3319 | 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); }; | |||
| 3320 | ||||
| 3321 | // if this deque is bigger than the pass ratio give a chance to another | |||
| 3322 | // thread | |||
| 3323 | if (TASK_DEQUE_SIZE(thread_data->td)((thread_data->td).td_deque_size) / INITIAL_TASK_DEQUE_SIZE(1 << 8) >= pass) | |||
| 3324 | goto release_and_exit; | |||
| 3325 | ||||
| 3326 | __kmp_realloc_task_deque(thread, thread_data); | |||
| 3327 | } | |||
| 3328 | } | |||
| 3329 | ||||
| 3330 | // lock is held here, and there is space in the deque | |||
| 3331 | ||||
| 3332 | thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata; | |||
| 3333 | // Wrap index. | |||
| 3334 | thread_data->td.td_deque_tail = | |||
| 3335 | (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td)((thread_data->td).td_deque_size - 1); | |||
| 3336 | TCW_4(thread_data->td.td_deque_ntasks,(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1) | |||
| 3337 | TCR_4(thread_data->td.td_deque_ntasks) + 1)(thread_data->td.td_deque_ntasks) = ((thread_data->td.td_deque_ntasks ) + 1); | |||
| 3338 | ||||
| 3339 | result = true; | |||
| 3340 | 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); } | |||
| 3341 | taskdata, tid))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_give_task: successfully gave task %p to thread %d.\n" , taskdata, tid); }; | |||
| 3342 | ||||
| 3343 | release_and_exit: | |||
| 3344 | __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock); | |||
| 3345 | ||||
| 3346 | return result; | |||
| 3347 | } | |||
| 3348 | ||||
| 3349 | /* The finish of the proxy tasks is divided in two pieces: | |||
| 3350 | - the top half is the one that can be done from a thread outside the team | |||
| 3351 | - the bottom half must be run from a them within the team | |||
| 3352 | ||||
| 3353 | In order to run the bottom half the task gets queued back into one of the | |||
| 3354 | threads of the team. Once the td_incomplete_child_task counter of the parent | |||
| 3355 | is decremented the threads can leave the barriers. So, the bottom half needs | |||
| 3356 | to be queued before the counter is decremented. The top half is therefore | |||
| 3357 | divided in two parts: | |||
| 3358 | - things that can be run before queuing the bottom half | |||
| 3359 | - things that must be run after queuing the bottom half | |||
| 3360 | ||||
| 3361 | This creates a second race as the bottom half can free the task before the | |||
| 3362 | second top half is executed. To avoid this we use the | |||
| 3363 | td_incomplete_child_task of the proxy task to synchronize the top and bottom | |||
| 3364 | half. */ | |||
| 3365 | static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
| 3366 | KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT)((taskdata->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3366)); | |||
| 3367 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3367)); | |||
| 3368 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0)((taskdata->td_flags.complete == 0) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3368)); | |||
| 3369 | KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0)((taskdata->td_flags.freed == 0) ? 0 : __kmp_debug_assert( "taskdata->td_flags.freed == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3369)); | |||
| 3370 | ||||
| 3371 | taskdata->td_flags.complete = 1; // mark the task as completed | |||
| 3372 | ||||
| 3373 | if (taskdata->td_taskgroup) | |||
| 3374 | KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count)(&taskdata->td_taskgroup->count)->fetch_sub(1, std ::memory_order_acq_rel); | |||
| 3375 | ||||
| 3376 | // Create an imaginary children for this task so the bottom half cannot | |||
| 3377 | // release the task before we have completed the second top half | |||
| 3378 | KMP_ATOMIC_INC(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->fetch_add(1 , std::memory_order_acq_rel); | |||
| 3379 | } | |||
| 3380 | ||||
| 3381 | static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) { | |||
| 3382 | kmp_int32 children = 0; | |||
| 3383 | ||||
| 3384 | // Predecrement simulated by "- 1" calculation | |||
| 3385 | children = | |||
| 3386 | KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks)(&taskdata->td_parent->td_incomplete_child_tasks)-> fetch_sub(1, std::memory_order_acq_rel) - 1; | |||
| 3387 | KMP_DEBUG_ASSERT(children >= 0)((children >= 0) ? 0 : __kmp_debug_assert("children >= 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3387)); | |||
| 3388 | ||||
| 3389 | // Remove the imaginary children | |||
| 3390 | KMP_ATOMIC_DEC(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->fetch_sub(1 , std::memory_order_acq_rel); | |||
| 3391 | } | |||
| 3392 | ||||
| 3393 | static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) { | |||
| 3394 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 3395 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 3396 | ||||
| 3397 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3397)); | |||
| 3398 | KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3399)) | |||
| 3399 | 1)((taskdata->td_flags.complete == 1) ? 0 : __kmp_debug_assert ("taskdata->td_flags.complete == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3399)); // top half must run before bottom half | |||
| 3400 | ||||
| 3401 | // We need to wait to make sure the top half is finished | |||
| 3402 | // Spinning here should be ok as this should happen quickly | |||
| 3403 | while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks)(&taskdata->td_incomplete_child_tasks)->load(std::memory_order_acquire ) > 0) | |||
| 3404 | ; | |||
| 3405 | ||||
| 3406 | __kmp_release_deps(gtid, taskdata); | |||
| 3407 | __kmp_free_task_and_ancestors(gtid, taskdata, thread); | |||
| 3408 | } | |||
| 3409 | ||||
| 3410 | /*! | |||
| 3411 | @ingroup TASKING | |||
| 3412 | @param gtid Global Thread ID of encountering thread | |||
| 3413 | @param ptask Task which execution is completed | |||
| 3414 | ||||
| 3415 | Execute the completation of a proxy task from a thread of that is part of the | |||
| 3416 | team. Run first and bottom halves directly. | |||
| 3417 | */ | |||
| 3418 | void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) { | |||
| 3419 | KMP_DEBUG_ASSERT(ptask != NULL)((ptask != __null) ? 0 : __kmp_debug_assert("ptask != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3419)); | |||
| 3420 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 3421 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); } | |||
| 3422 | 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); } | |||
| 3423 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
| 3424 | ||||
| 3425 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3425)); | |||
| 3426 | ||||
| 3427 | __kmp_first_top_half_finish_proxy(taskdata); | |||
| 3428 | __kmp_second_top_half_finish_proxy(taskdata); | |||
| 3429 | __kmp_bottom_half_finish_proxy(gtid, ptask); | |||
| 3430 | ||||
| 3431 | 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); } | |||
| 3432 | ("__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); } | |||
| 3433 | gtid, taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n" , gtid, taskdata); }; | |||
| 3434 | } | |||
| 3435 | ||||
| 3436 | /*! | |||
| 3437 | @ingroup TASKING | |||
| 3438 | @param ptask Task which execution is completed | |||
| 3439 | ||||
| 3440 | Execute the completation of a proxy task from a thread that could not belong to | |||
| 3441 | the team. | |||
| 3442 | */ | |||
| 3443 | void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) { | |||
| 3444 | KMP_DEBUG_ASSERT(ptask != NULL)((ptask != __null) ? 0 : __kmp_debug_assert("ptask != __null" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3444)); | |||
| 3445 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask)(((kmp_taskdata_t *)ptask) - 1); | |||
| 3446 | ||||
| 3447 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
| 3448 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); } | |||
| 3449 | ("__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); } | |||
| 3450 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n" , taskdata); }; | |||
| 3451 | ||||
| 3452 | KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY)((taskdata->td_flags.proxy == 1) ? 0 : __kmp_debug_assert( "taskdata->td_flags.proxy == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3452)); | |||
| 3453 | ||||
| 3454 | __kmp_first_top_half_finish_proxy(taskdata); | |||
| 3455 | ||||
| 3456 | // Enqueue task to complete bottom half completion from a thread within the | |||
| 3457 | // corresponding team | |||
| 3458 | kmp_team_t *team = taskdata->td_team; | |||
| 3459 | kmp_int32 nthreads = team->t.t_nproc; | |||
| 3460 | kmp_info_t *thread; | |||
| 3461 | ||||
| 3462 | // This should be similar to start_k = __kmp_get_random( thread ) % nthreads | |||
| 3463 | // but we cannot use __kmp_get_random here | |||
| 3464 | kmp_int32 start_k = 0; | |||
| 3465 | kmp_int32 pass = 1; | |||
| 3466 | kmp_int32 k = start_k; | |||
| 3467 | ||||
| 3468 | do { | |||
| 3469 | // For now we're just linearly trying to find a thread | |||
| 3470 | thread = team->t.t_threads[k]; | |||
| 3471 | k = (k + 1) % nthreads; | |||
| 3472 | ||||
| 3473 | // we did a full pass through all the threads | |||
| 3474 | if (k == start_k) | |||
| 3475 | pass = pass << 1; | |||
| 3476 | ||||
| 3477 | } while (!__kmp_give_task(thread, k, ptask, pass)); | |||
| 3478 | ||||
| 3479 | __kmp_second_top_half_finish_proxy(taskdata); | |||
| 3480 | ||||
| 3481 | KA_TRACE(if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
| 3482 | 10,if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); } | |||
| 3483 | ("__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); } | |||
| 3484 | taskdata))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n" , taskdata); }; | |||
| 3485 | } | |||
| 3486 | ||||
| 3487 | // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task | |||
| 3488 | // for taskloop | |||
| 3489 | // | |||
| 3490 | // thread: allocating thread | |||
| 3491 | // task_src: pointer to source task to be duplicated | |||
| 3492 | // returns: a pointer to the allocated kmp_task_t structure (task). | |||
| 3493 | kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) { | |||
| 3494 | kmp_task_t *task; | |||
| 3495 | kmp_taskdata_t *taskdata; | |||
| 3496 | kmp_taskdata_t *taskdata_src; | |||
| 3497 | kmp_taskdata_t *parent_task = thread->th.th_current_task; | |||
| 3498 | size_t shareds_offset; | |||
| 3499 | size_t task_size; | |||
| 3500 | ||||
| 3501 | 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); } | |||
| 3502 | task_src))if (kmp_a_debug >= 10) { __kmp_debug_printf ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n" , thread, task_src); }; | |||
| 3503 | taskdata_src = KMP_TASK_TO_TASKDATA(task_src)(((kmp_taskdata_t *)task_src) - 1); | |||
| 3504 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==((taskdata_src->td_flags.proxy == 0) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3505)) | |||
| 3505 | TASK_FULL)((taskdata_src->td_flags.proxy == 0) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.proxy == 0", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3505)); // it should not be proxy task | |||
| 3506 | KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT)((taskdata_src->td_flags.tasktype == 1) ? 0 : __kmp_debug_assert ("taskdata_src->td_flags.tasktype == 1", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3506)); | |||
| 3507 | task_size = taskdata_src->td_size_alloc; | |||
| 3508 | ||||
| 3509 | // Allocate a kmp_taskdata_t block and a kmp_task_t block. | |||
| 3510 | 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); } | |||
| 3511 | task_size))if (kmp_a_debug >= 30) { __kmp_debug_printf ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n" , thread, task_size); }; | |||
| 3512 | #if USE_FAST_MEMORY3 | |||
| 3513 | taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size)___kmp_fast_allocate((thread), (task_size), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3513); | |||
| 3514 | #else | |||
| 3515 | taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size)___kmp_thread_malloc((thread), (task_size), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3515); | |||
| 3516 | #endif /* USE_FAST_MEMORY */ | |||
| 3517 | KMP_MEMCPYmemcpy(taskdata, taskdata_src, task_size); | |||
| 3518 | ||||
| 3519 | task = KMP_TASKDATA_TO_TASK(taskdata)(kmp_task_t *)(taskdata + 1); | |||
| 3520 | ||||
| 3521 | // Initialize new task (only specific fields not affected by memcpy) | |||
| 3522 | taskdata->td_task_id = KMP_GEN_TASK_ID()(~0); | |||
| 3523 | if (task->shareds != NULL__null) { // need setup shareds pointer | |||
| 3524 | shareds_offset = (char *)task_src->shareds - (char *)taskdata_src; | |||
| 3525 | task->shareds = &((char *)taskdata)[shareds_offset]; | |||
| 3526 | KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3527)) | |||
| 3527 | 0)(((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1 )) == 0) ? 0 : __kmp_debug_assert("(((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) == 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3527)); | |||
| 3528 | } | |||
| 3529 | taskdata->td_alloc_thread = thread; | |||
| 3530 | taskdata->td_parent = parent_task; | |||
| 3531 | taskdata->td_taskgroup = | |||
| 3532 | parent_task | |||
| 3533 | ->td_taskgroup; // task inherits the taskgroup from the parent task | |||
| 3534 | ||||
| 3535 | // Only need to keep track of child task counts if team parallel and tasking | |||
| 3536 | // not serialized | |||
| 3537 | if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) { | |||
| 3538 | KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks)(&parent_task->td_incomplete_child_tasks)->fetch_add (1, std::memory_order_acq_rel); | |||
| 3539 | if (parent_task->td_taskgroup) | |||
| 3540 | KMP_ATOMIC_INC(&parent_task->td_taskgroup->count)(&parent_task->td_taskgroup->count)->fetch_add(1 , std::memory_order_acq_rel); | |||
| 3541 | // Only need to keep track of allocated child tasks for explicit tasks since | |||
| 3542 | // implicit not deallocated | |||
| 3543 | if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT1) | |||
| 3544 | 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); | |||
| 3545 | } | |||
| 3546 | ||||
| 3547 | 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); } | |||
| 3548 | ("__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); } | |||
| 3549 | 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); }; | |||
| 3550 | #if OMPT_SUPPORT1 | |||
| 3551 | if (UNLIKELY(ompt_enabled.enabled)__builtin_expect(!!(ompt_enabled.enabled), 0)) | |||
| 3552 | __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid); | |||
| 3553 | #endif | |||
| 3554 | return task; | |||
| 3555 | } | |||
| 3556 | ||||
| 3557 | // Routine optionally generated by the compiler for setting the lastprivate flag | |||
| 3558 | // and calling needed constructors for private/firstprivate objects | |||
| 3559 | // (used to form taskloop tasks from pattern task) | |||
| 3560 | // Parameters: dest task, src task, lastprivate flag. | |||
| 3561 | typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); | |||
| 3562 | ||||
| 3563 | KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8)static_assert(sizeof(long) == 4 || sizeof(long) == 8, "Build condition error" ); | |||
| 3564 | ||||
| 3565 | // class to encapsulate manipulating loop bounds in a taskloop task. | |||
| 3566 | // this abstracts away the Intel vs GOMP taskloop interface for setting/getting | |||
| 3567 | // the loop bound variables. | |||
| 3568 | class kmp_taskloop_bounds_t { | |||
| 3569 | kmp_task_t *task; | |||
| 3570 | const kmp_taskdata_t *taskdata; | |||
| 3571 | size_t lower_offset; | |||
| 3572 | size_t upper_offset; | |||
| 3573 | ||||
| 3574 | public: | |||
| 3575 | kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub) | |||
| 3576 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1)), | |||
| 3577 | lower_offset((char *)lb - (char *)task), | |||
| 3578 | upper_offset((char *)ub - (char *)task) { | |||
| 3579 | KMP_DEBUG_ASSERT((char *)lb > (char *)_task)(((char *)lb > (char *)_task) ? 0 : __kmp_debug_assert("(char *)lb > (char *)_task" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3579)); | |||
| 3580 | KMP_DEBUG_ASSERT((char *)ub > (char *)_task)(((char *)ub > (char *)_task) ? 0 : __kmp_debug_assert("(char *)ub > (char *)_task" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3580)); | |||
| 3581 | } | |||
| 3582 | kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds) | |||
| 3583 | : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)(((kmp_taskdata_t *)_task) - 1)), | |||
| 3584 | lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {} | |||
| 3585 | size_t get_lower_offset() const { return lower_offset; } | |||
| 3586 | size_t get_upper_offset() const { return upper_offset; } | |||
| 3587 | kmp_uint64 get_lb() const { | |||
| 3588 | kmp_int64 retval; | |||
| 3589 | #if defined(KMP_GOMP_COMPAT) | |||
| 3590 | // Intel task just returns the lower bound normally | |||
| 3591 | if (!taskdata->td_flags.native) { | |||
| 3592 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
| 3593 | } else { | |||
| 3594 | // GOMP task has to take into account the sizeof(long) | |||
| 3595 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 3596 | kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds); | |||
| 3597 | retval = (kmp_int64)*lb; | |||
| 3598 | } else { | |||
| 3599 | kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds); | |||
| 3600 | retval = (kmp_int64)*lb; | |||
| 3601 | } | |||
| 3602 | } | |||
| 3603 | #else | |||
| 3604 | retval = *(kmp_int64 *)((char *)task + lower_offset); | |||
| 3605 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 3606 | return retval; | |||
| 3607 | } | |||
| 3608 | kmp_uint64 get_ub() const { | |||
| 3609 | kmp_int64 retval; | |||
| 3610 | #if defined(KMP_GOMP_COMPAT) | |||
| 3611 | // Intel task just returns the upper bound normally | |||
| 3612 | if (!taskdata->td_flags.native) { | |||
| 3613 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
| 3614 | } else { | |||
| 3615 | // GOMP task has to take into account the sizeof(long) | |||
| 3616 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 3617 | kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds)reinterpret_cast<kmp_int32 *>(task->shareds) + 1; | |||
| 3618 | retval = (kmp_int64)*ub; | |||
| 3619 | } else { | |||
| 3620 | kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds)reinterpret_cast<kmp_int64 *>(task->shareds) + 1; | |||
| 3621 | retval = (kmp_int64)*ub; | |||
| 3622 | } | |||
| 3623 | } | |||
| 3624 | #else | |||
| 3625 | retval = *(kmp_int64 *)((char *)task + upper_offset); | |||
| 3626 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 3627 | return retval; | |||
| 3628 | } | |||
| 3629 | void set_lb(kmp_uint64 lb) { | |||
| 3630 | #if defined(KMP_GOMP_COMPAT) | |||
| 3631 | // Intel task just sets the lower bound normally | |||
| 3632 | if (!taskdata->td_flags.native) { | |||
| 3633 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
| 3634 | } else { | |||
| 3635 | // GOMP task has to take into account the sizeof(long) | |||
| 3636 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 3637 | kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds); | |||
| 3638 | *lower = (kmp_uint32)lb; | |||
| 3639 | } else { | |||
| 3640 | kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds); | |||
| 3641 | *lower = (kmp_uint64)lb; | |||
| 3642 | } | |||
| 3643 | } | |||
| 3644 | #else | |||
| 3645 | *(kmp_uint64 *)((char *)task + lower_offset) = lb; | |||
| 3646 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 3647 | } | |||
| 3648 | void set_ub(kmp_uint64 ub) { | |||
| 3649 | #if defined(KMP_GOMP_COMPAT) | |||
| 3650 | // Intel task just sets the upper bound normally | |||
| 3651 | if (!taskdata->td_flags.native) { | |||
| 3652 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
| 3653 | } else { | |||
| 3654 | // GOMP task has to take into account the sizeof(long) | |||
| 3655 | if (taskdata->td_size_loop_bounds == 4) { | |||
| 3656 | kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds)reinterpret_cast<kmp_uint32 *>(task->shareds) + 1; | |||
| 3657 | *upper = (kmp_uint32)ub; | |||
| 3658 | } else { | |||
| 3659 | kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds)reinterpret_cast<kmp_uint64 *>(task->shareds) + 1; | |||
| 3660 | *upper = (kmp_uint64)ub; | |||
| 3661 | } | |||
| 3662 | } | |||
| 3663 | #else | |||
| 3664 | *(kmp_uint64 *)((char *)task + upper_offset) = ub; | |||
| 3665 | #endif // defined(KMP_GOMP_COMPAT) | |||
| 3666 | } | |||
| 3667 | }; | |||
| 3668 | ||||
| 3669 | // __kmp_taskloop_linear: Start tasks of the taskloop linearly | |||
| 3670 | // | |||
| 3671 | // loc Source location information | |||
| 3672 | // gtid Global thread ID | |||
| 3673 | // task Pattern task, exposes the loop iteration range | |||
| 3674 | // lb Pointer to loop lower bound in task structure | |||
| 3675 | // ub Pointer to loop upper bound in task structure | |||
| 3676 | // st Loop stride | |||
| 3677 | // ub_glob Global upper bound (used for lastprivate check) | |||
| 3678 | // num_tasks Number of tasks to execute | |||
| 3679 | // grainsize Number of loop iterations per task | |||
| 3680 | // extras Number of chunks with grainsize+1 iterations | |||
| 3681 | // tc Iterations count | |||
| 3682 | // task_dup Tasks duplication routine | |||
| 3683 | void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task, | |||
| 3684 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 3685 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
| 3686 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
| 3687 | kmp_uint64 tc, void *task_dup) { | |||
| 3688 | KMP_COUNT_BLOCK(OMP_TASKLOOP)((void)0); | |||
| 3689 | KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling)((void)0); | |||
| 3690 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 3691 | // compiler provides global bounds here | |||
| 3692 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
| 3693 | kmp_uint64 lower = task_bounds.get_lb(); | |||
| 3694 | kmp_uint64 upper = task_bounds.get_ub(); | |||
| 3695 | kmp_uint64 i; | |||
| 3696 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 3697 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 3698 | kmp_task_t *next_task; | |||
| 3699 | kmp_int32 lastpriv = 0; | |||
| 3700 | ||||
| 3701 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3701)); | |||
| 3702 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3702)); | |||
| 3703 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3703)); | |||
| 3704 | KA_TRACE(20, ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
| 3705 | "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
| 3706 | gtid, num_tasks, grainsize, extras, lower, upper, ub_glob, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); } | |||
| 3707 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, " "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n", gtid, num_tasks , grainsize, extras, lower, upper, ub_glob, st, task_dup); }; | |||
| 3708 | ||||
| 3709 | // Launch num_tasks tasks, assign grainsize iterations each task | |||
| 3710 | for (i = 0; i < num_tasks; ++i) { | |||
| 3711 | kmp_uint64 chunk_minus_1; | |||
| 3712 | if (extras == 0) { | |||
| 3713 | chunk_minus_1 = grainsize - 1; | |||
| 3714 | } else { | |||
| 3715 | chunk_minus_1 = grainsize; | |||
| 3716 | --extras; // first extras iterations get bigger chunk (grainsize+1) | |||
| 3717 | } | |||
| 3718 | upper = lower + st * chunk_minus_1; | |||
| 3719 | if (i == num_tasks - 1) { | |||
| 3720 | // schedule the last task, set lastprivate flag if needed | |||
| 3721 | if (st == 1) { // most common case | |||
| 3722 | KMP_DEBUG_ASSERT(upper == *ub)((upper == *ub) ? 0 : __kmp_debug_assert("upper == *ub", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3722)); | |||
| 3723 | if (upper == ub_glob) | |||
| 3724 | lastpriv = 1; | |||
| 3725 | } else if (st > 0) { // positive loop stride | |||
| 3726 | KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper)(((kmp_uint64)st > *ub - upper) ? 0 : __kmp_debug_assert("(kmp_uint64)st > *ub - upper" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3726)); | |||
| 3727 | if ((kmp_uint64)st > ub_glob - upper) | |||
| 3728 | lastpriv = 1; | |||
| 3729 | } else { // negative loop stride | |||
| 3730 | KMP_DEBUG_ASSERT(upper + st < *ub)((upper + st < *ub) ? 0 : __kmp_debug_assert("upper + st < *ub" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3730)); | |||
| 3731 | if (upper - ub_glob < (kmp_uint64)(-st)) | |||
| 3732 | lastpriv = 1; | |||
| 3733 | } | |||
| 3734 | } | |||
| 3735 | next_task = __kmp_task_dup_alloc(thread, task); // allocate new task | |||
| 3736 | kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task)(((kmp_taskdata_t *)next_task) - 1); | |||
| 3737 | kmp_taskloop_bounds_t next_task_bounds = | |||
| 3738 | kmp_taskloop_bounds_t(next_task, task_bounds); | |||
| 3739 | ||||
| 3740 | // adjust task-specific bounds | |||
| 3741 | next_task_bounds.set_lb(lower); | |||
| 3742 | if (next_taskdata->td_flags.native) { | |||
| 3743 | next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1)); | |||
| 3744 | } else { | |||
| 3745 | next_task_bounds.set_ub(upper); | |||
| 3746 | } | |||
| 3747 | if (ptask_dup != NULL__null) // set lastprivate flag, construct fistprivates, etc. | |||
| 3748 | ptask_dup(next_task, task, lastpriv); | |||
| 3749 | 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()); } | |||
| 3750 | ("__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()); } | |||
| 3751 | "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()); } | |||
| 3752 | 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()); } | |||
| 3753 | 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()); } | |||
| 3754 | 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()); }; | |||
| 3755 | __kmp_omp_task(gtid, next_task, true); // schedule new task | |||
| 3756 | lower = upper + st; // adjust lower bound for the next iteration | |||
| 3757 | } | |||
| 3758 | // free the pattern task and exit | |||
| 3759 | __kmp_task_start(gtid, task, current_task); // make internal bookkeeping | |||
| 3760 | // do not execute the pattern task, just do internal bookkeeping | |||
| 3761 | __kmp_task_finish(gtid, task, current_task); | |||
| 3762 | } | |||
| 3763 | ||||
| 3764 | // Structure to keep taskloop parameters for auxiliary task | |||
| 3765 | // kept in the shareds of the task structure. | |||
| 3766 | typedef struct __taskloop_params { | |||
| 3767 | kmp_task_t *task; | |||
| 3768 | kmp_uint64 *lb; | |||
| 3769 | kmp_uint64 *ub; | |||
| 3770 | void *task_dup; | |||
| 3771 | kmp_int64 st; | |||
| 3772 | kmp_uint64 ub_glob; | |||
| 3773 | kmp_uint64 num_tasks; | |||
| 3774 | kmp_uint64 grainsize; | |||
| 3775 | kmp_uint64 extras; | |||
| 3776 | kmp_uint64 tc; | |||
| 3777 | kmp_uint64 num_t_min; | |||
| 3778 | } __taskloop_params_t; | |||
| 3779 | ||||
| 3780 | void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *, | |||
| 3781 | kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64, | |||
| 3782 | kmp_uint64, kmp_uint64, kmp_uint64, kmp_uint64, | |||
| 3783 | void *); | |||
| 3784 | ||||
| 3785 | // Execute part of the the taskloop submitted as a task. | |||
| 3786 | int __kmp_taskloop_task(int gtid, void *ptask) { | |||
| 3787 | __taskloop_params_t *p = | |||
| 3788 | (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds; | |||
| 3789 | kmp_task_t *task = p->task; | |||
| 3790 | kmp_uint64 *lb = p->lb; | |||
| 3791 | kmp_uint64 *ub = p->ub; | |||
| 3792 | void *task_dup = p->task_dup; | |||
| 3793 | // p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 3794 | kmp_int64 st = p->st; | |||
| 3795 | kmp_uint64 ub_glob = p->ub_glob; | |||
| 3796 | kmp_uint64 num_tasks = p->num_tasks; | |||
| 3797 | kmp_uint64 grainsize = p->grainsize; | |||
| 3798 | kmp_uint64 extras = p->extras; | |||
| 3799 | kmp_uint64 tc = p->tc; | |||
| 3800 | kmp_uint64 num_t_min = p->num_t_min; | |||
| 3801 | #if KMP_DEBUG1 | |||
| 3802 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 3803 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3803)); | |||
| 3804 | KA_TRACE(20, ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3805 | " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3806 | gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3807 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); }; | |||
| 3808 | #endif | |||
| 3809 | KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min)((num_tasks * 2 + 1 > num_t_min) ? 0 : __kmp_debug_assert( "num_tasks * 2 + 1 > num_t_min", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3809)); | |||
| 3810 | if (num_tasks > num_t_min) | |||
| 3811 | __kmp_taskloop_recur(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 3812 | grainsize, extras, tc, num_t_min, task_dup); | |||
| 3813 | else | |||
| 3814 | __kmp_taskloop_linear(NULL__null, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 3815 | grainsize, extras, tc, task_dup); | |||
| 3816 | ||||
| 3817 | 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); }; | |||
| 3818 | return 0; | |||
| 3819 | } | |||
| 3820 | ||||
| 3821 | // Schedule part of the the taskloop as a task, | |||
| 3822 | // execute the rest of the the taskloop. | |||
| 3823 | // | |||
| 3824 | // loc Source location information | |||
| 3825 | // gtid Global thread ID | |||
| 3826 | // task Pattern task, exposes the loop iteration range | |||
| 3827 | // lb Pointer to loop lower bound in task structure | |||
| 3828 | // ub Pointer to loop upper bound in task structure | |||
| 3829 | // st Loop stride | |||
| 3830 | // ub_glob Global upper bound (used for lastprivate check) | |||
| 3831 | // num_tasks Number of tasks to execute | |||
| 3832 | // grainsize Number of loop iterations per task | |||
| 3833 | // extras Number of chunks with grainsize+1 iterations | |||
| 3834 | // tc Iterations count | |||
| 3835 | // num_t_min Threashold to launch tasks recursively | |||
| 3836 | // task_dup Tasks duplication routine | |||
| 3837 | void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task, | |||
| 3838 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, | |||
| 3839 | kmp_uint64 ub_glob, kmp_uint64 num_tasks, | |||
| 3840 | kmp_uint64 grainsize, kmp_uint64 extras, | |||
| 3841 | kmp_uint64 tc, kmp_uint64 num_t_min, void *task_dup) { | |||
| 3842 | #if KMP_DEBUG1 | |||
| 3843 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 3844 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3844)); | |||
| 3845 | KMP_DEBUG_ASSERT(num_tasks > num_t_min)((num_tasks > num_t_min) ? 0 : __kmp_debug_assert("num_tasks > num_t_min" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3845)); | |||
| 3846 | KA_TRACE(20, ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3847 | " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3848 | gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); } | |||
| 3849 | task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize" " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n", gtid, taskdata , num_tasks, grainsize, extras, *lb, *ub, st, task_dup); }; | |||
| 3850 | #endif | |||
| 3851 | p_task_dup_t ptask_dup = (p_task_dup_t)task_dup; | |||
| 3852 | kmp_uint64 lower = *lb; | |||
| 3853 | kmp_uint64 upper = *ub; | |||
| 3854 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 3855 | // kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 3856 | kmp_task_t *next_task; | |||
| 3857 | kmp_int32 lastpriv = 0; | |||
| 3858 | size_t lower_offset = | |||
| 3859 | (char *)lb - (char *)task; // remember offset of lb in the task structure | |||
| 3860 | size_t upper_offset = | |||
| 3861 | (char *)ub - (char *)task; // remember offset of ub in the task structure | |||
| 3862 | ||||
| 3863 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3863)); | |||
| 3864 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3864)); | |||
| 3865 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3865)); | |||
| 3866 | ||||
| 3867 | // split the loop in two halves | |||
| 3868 | kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1; | |||
| 3869 | kmp_uint64 gr_size0 = grainsize; | |||
| 3870 | kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute | |||
| 3871 | kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task | |||
| 3872 | if (n_tsk0 <= extras) { | |||
| 3873 | gr_size0++; // integrate extras into grainsize | |||
| 3874 | ext0 = 0; // no extra iters in 1st half | |||
| 3875 | ext1 = extras - n_tsk0; // remaining extras | |||
| 3876 | tc0 = gr_size0 * n_tsk0; | |||
| 3877 | tc1 = tc - tc0; | |||
| 3878 | } else { // n_tsk0 > extras | |||
| 3879 | ext1 = 0; // no extra iters in 2nd half | |||
| 3880 | ext0 = extras; | |||
| 3881 | tc1 = grainsize * n_tsk1; | |||
| 3882 | tc0 = tc - tc1; | |||
| 3883 | } | |||
| 3884 | ub0 = lower + st * (tc0 - 1); | |||
| 3885 | lb1 = ub0 + st; | |||
| 3886 | ||||
| 3887 | // create pattern task for 2nd half of the loop | |||
| 3888 | next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task | |||
| 3889 | // adjust lower bound (upper bound is not changed) for the 2nd half | |||
| 3890 | *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1; | |||
| 3891 | if (ptask_dup != NULL__null) // construct fistprivates, etc. | |||
| 3892 | ptask_dup(next_task, task, 0); | |||
| 3893 | *ub = ub0; // adjust upper bound for the 1st half | |||
| 3894 | ||||
| 3895 | // create auxiliary task for 2nd half of the loop | |||
| 3896 | kmp_task_t *new_task = | |||
| 3897 | __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *), | |||
| 3898 | sizeof(__taskloop_params_t), &__kmp_taskloop_task); | |||
| 3899 | __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds; | |||
| 3900 | p->task = next_task; | |||
| 3901 | p->lb = (kmp_uint64 *)((char *)next_task + lower_offset); | |||
| 3902 | p->ub = (kmp_uint64 *)((char *)next_task + upper_offset); | |||
| 3903 | p->task_dup = task_dup; | |||
| 3904 | p->st = st; | |||
| 3905 | p->ub_glob = ub_glob; | |||
| 3906 | p->num_tasks = n_tsk1; | |||
| 3907 | p->grainsize = grainsize; | |||
| 3908 | p->extras = ext1; | |||
| 3909 | p->tc = tc1; | |||
| 3910 | p->num_t_min = num_t_min; | |||
| 3911 | __kmp_omp_task(gtid, new_task, true); // schedule new task | |||
| 3912 | ||||
| 3913 | // execute the 1st half of current subrange | |||
| 3914 | if (n_tsk0 > num_t_min) | |||
| 3915 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0, | |||
| 3916 | ext0, tc0, num_t_min, task_dup); | |||
| 3917 | else | |||
| 3918 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, | |||
| 3919 | gr_size0, ext0, tc0, task_dup); | |||
| 3920 | ||||
| 3921 | KA_TRACE(40, ("__kmpc_taskloop_recur(exit): T#%d\n", gtid))if (kmp_a_debug >= 40) { __kmp_debug_printf ("__kmpc_taskloop_recur(exit): T#%d\n" , gtid); }; | |||
| 3922 | } | |||
| 3923 | ||||
| 3924 | /*! | |||
| 3925 | @ingroup TASKING | |||
| 3926 | @param loc Source location information | |||
| 3927 | @param gtid Global thread ID | |||
| 3928 | @param task Task structure | |||
| 3929 | @param if_val Value of the if clause | |||
| 3930 | @param lb Pointer to loop lower bound in task structure | |||
| 3931 | @param ub Pointer to loop upper bound in task structure | |||
| 3932 | @param st Loop stride | |||
| 3933 | @param nogroup Flag, 1 if nogroup clause specified, 0 otherwise | |||
| 3934 | @param sched Schedule specified 0/1/2 for none/grainsize/num_tasks | |||
| 3935 | @param grainsize Schedule value if specified | |||
| 3936 | @param task_dup Tasks duplication routine | |||
| 3937 | ||||
| 3938 | Execute the taskloop construct. | |||
| 3939 | */ | |||
| 3940 | void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, | |||
| 3941 | kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, | |||
| 3942 | int sched, kmp_uint64 grainsize, void *task_dup) { | |||
| 3943 | kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task)(((kmp_taskdata_t *)task) - 1); | |||
| 3944 | KMP_DEBUG_ASSERT(task != NULL)((task != __null) ? 0 : __kmp_debug_assert("task != __null", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 3944)); | |||
| 3945 | ||||
| 3946 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 3947 | ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL__null); | |||
| 3948 | ompt_task_info_t *task_info = __ompt_get_task_info_object(0); | |||
| 3949 | if (ompt_enabled.ompt_callback_work) { | |||
| ||||
| 3950 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
| 3951 | ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data), | |||
| 3952 | &(task_info->task_data), 0, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 3953 | } | |||
| 3954 | #endif | |||
| 3955 | ||||
| 3956 | if (nogroup == 0) { | |||
| 3957 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 3958 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 3959 | #endif | |||
| 3960 | __kmpc_taskgroup(loc, gtid); | |||
| 3961 | } | |||
| 3962 | ||||
| 3963 | // ========================================================================= | |||
| 3964 | // calculate loop parameters | |||
| 3965 | kmp_taskloop_bounds_t task_bounds(task, lb, ub); | |||
| 3966 | kmp_uint64 tc; | |||
| 3967 | // compiler provides global bounds here | |||
| 3968 | kmp_uint64 lower = task_bounds.get_lb(); | |||
| 3969 | kmp_uint64 upper = task_bounds.get_ub(); | |||
| 3970 | kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag | |||
| 3971 | kmp_uint64 num_tasks = 0, extras = 0; | |||
| 3972 | kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks; | |||
| 3973 | kmp_info_t *thread = __kmp_threads[gtid]; | |||
| 3974 | kmp_taskdata_t *current_task = thread->th.th_current_task; | |||
| 3975 | ||||
| 3976 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); } | |||
| 3977 | "grain %llu(%d), dup %p\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); } | |||
| 3978 | gtid, taskdata, lower, upper, st, grainsize, sched, task_dup))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, " "grain %llu(%d), dup %p\n", gtid, taskdata, lower, upper, st , grainsize, sched, task_dup); }; | |||
| 3979 | ||||
| 3980 | // compute trip count | |||
| 3981 | if (st == 1) { // most common case | |||
| 3982 | tc = upper - lower + 1; | |||
| 3983 | } else if (st < 0) { | |||
| 3984 | tc = (lower - upper) / (-st) + 1; | |||
| 3985 | } else { // st > 0 | |||
| 3986 | tc = (upper - lower) / st + 1; | |||
| 3987 | } | |||
| 3988 | if (tc == 0) { | |||
| 3989 | KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d zero-trip loop\n", gtid))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop(exit): T#%d zero-trip loop\n" , gtid); }; | |||
| 3990 | // free the pattern task and exit | |||
| 3991 | __kmp_task_start(gtid, task, current_task); | |||
| 3992 | // do not execute anything for zero-trip loop | |||
| 3993 | __kmp_task_finish(gtid, task, current_task); | |||
| 3994 | return; | |||
| 3995 | } | |||
| 3996 | if (num_tasks_min == 0) | |||
| 3997 | // TODO: can we choose better default heuristic? | |||
| 3998 | num_tasks_min = | |||
| 3999 | 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))); | |||
| 4000 | ||||
| 4001 | // compute num_tasks/grainsize based on the input provided | |||
| 4002 | switch (sched) { | |||
| 4003 | case 0: // no schedule clause specified, we can choose the default | |||
| 4004 | // let's try to schedule (team_size*10) tasks | |||
| 4005 | grainsize = thread->th.th_team_nproc * 10; | |||
| 4006 | case 2: // num_tasks provided | |||
| 4007 | if (grainsize > tc) { | |||
| 4008 | num_tasks = tc; // too big num_tasks requested, adjust values | |||
| 4009 | grainsize = 1; | |||
| 4010 | extras = 0; | |||
| 4011 | } else { | |||
| 4012 | num_tasks = grainsize; | |||
| 4013 | grainsize = tc / num_tasks; | |||
| 4014 | extras = tc % num_tasks; | |||
| 4015 | } | |||
| 4016 | break; | |||
| 4017 | case 1: // grainsize provided | |||
| 4018 | if (grainsize > tc) { | |||
| 4019 | num_tasks = 1; // too big grainsize requested, adjust values | |||
| 4020 | grainsize = tc; | |||
| 4021 | extras = 0; | |||
| 4022 | } else { | |||
| 4023 | num_tasks = tc / grainsize; | |||
| 4024 | // adjust grainsize for balanced distribution of iterations | |||
| 4025 | grainsize = tc / num_tasks; | |||
| 4026 | extras = tc % num_tasks; | |||
| 4027 | } | |||
| 4028 | break; | |||
| 4029 | default: | |||
| 4030 | KMP_ASSERT2(0, "unknown scheduling of taskloop")((0) ? 0 : __kmp_debug_assert(("unknown scheduling of taskloop" ), "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 4030)); | |||
| 4031 | } | |||
| 4032 | KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras)((tc == num_tasks * grainsize + extras) ? 0 : __kmp_debug_assert ("tc == num_tasks * grainsize + extras", "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 4032)); | |||
| 4033 | KMP_DEBUG_ASSERT(num_tasks > extras)((num_tasks > extras) ? 0 : __kmp_debug_assert("num_tasks > extras" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 4033)); | |||
| 4034 | KMP_DEBUG_ASSERT(num_tasks > 0)((num_tasks > 0) ? 0 : __kmp_debug_assert("num_tasks > 0" , "/build/llvm-toolchain-snapshot-7~svn337957/projects/openmp/runtime/src/kmp_tasking.cpp" , 4034)); | |||
| 4035 | // ========================================================================= | |||
| 4036 | ||||
| 4037 | // check if clause value first | |||
| 4038 | // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native) | |||
| 4039 | if (if_val == 0) { // if(0) specified, mark task as serial | |||
| 4040 | taskdata->td_flags.task_serial = 1; | |||
| 4041 | taskdata->td_flags.tiedness = TASK_TIED1; // AC: serial task cannot be untied | |||
| 4042 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4043 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 4044 | #endif | |||
| 4045 | // always start serial tasks linearly | |||
| 4046 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 4047 | grainsize, extras, tc, task_dup); | |||
| 4048 | // !taskdata->td_flags.native => currently force linear spawning of tasks | |||
| 4049 | // for GOMP_taskloop | |||
| 4050 | } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) { | |||
| 4051 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
| 4052 | "(%lld), grain %llu, extras %llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
| 4053 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); }; | |||
| 4054 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4055 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 4056 | #endif | |||
| 4057 | __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 4058 | grainsize, extras, tc, num_tasks_min, task_dup); | |||
| 4059 | } else { | |||
| 4060 | KA_TRACE(20, ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu"if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
| 4061 | "(%lld), grain %llu, extras %llu\n",if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); } | |||
| 4062 | gtid, tc, num_tasks, num_tasks_min, grainsize, extras))if (kmp_a_debug >= 20) { __kmp_debug_printf ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu" "(%lld), grain %llu, extras %llu\n", gtid, tc, num_tasks, num_tasks_min , grainsize, extras); }; | |||
| 4063 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4064 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 4065 | #endif | |||
| 4066 | __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks, | |||
| 4067 | grainsize, extras, tc, task_dup); | |||
| 4068 | } | |||
| 4069 | ||||
| 4070 | if (nogroup == 0) { | |||
| 4071 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4072 | OMPT_STORE_RETURN_ADDRESS(gtid)if (ompt_enabled.enabled && gtid >= 0 && __kmp_threads [gtid] && !__kmp_threads[gtid]->th.ompt_thread_info .return_address) __kmp_threads[gtid]->th.ompt_thread_info. return_address = __builtin_return_address(0); | |||
| 4073 | #endif | |||
| 4074 | __kmpc_end_taskgroup(loc, gtid); | |||
| 4075 | } | |||
| 4076 | #if OMPT_SUPPORT1 && OMPT_OPTIONAL1 | |||
| 4077 | if (ompt_enabled.ompt_callback_work) { | |||
| 4078 | ompt_callbacks.ompt_callback(ompt_callback_work)ompt_callback_work_callback( | |||
| 4079 | ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data), | |||
| 4080 | &(task_info->task_data), 0, OMPT_GET_RETURN_ADDRESS(0)__builtin_return_address(0)); | |||
| 4081 | } | |||
| 4082 | #endif | |||
| 4083 | 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); }; | |||
| 4084 | } | |||
| 4085 | ||||
| 4086 | #endif |