| 1 | /* |
| 2 | Copyright (c) 2005-2019 Intel Corporation |
| 3 | |
| 4 | Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | you may not use this file except in compliance with the License. |
| 6 | You may obtain a copy of the License at |
| 7 | |
| 8 | http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | |
| 10 | Unless required by applicable law or agreed to in writing, software |
| 11 | distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | See the License for the specific language governing permissions and |
| 14 | limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #include "harness_task.h" |
| 18 | #include "tbb/atomic.h" |
| 19 | #include "tbb/tbb_thread.h" |
| 20 | #include "tbb/task_scheduler_init.h" |
| 21 | #include <cstdlib> |
| 22 | |
| 23 | //------------------------------------------------------------------------ |
| 24 | // Test for task::spawn_children and task_list |
| 25 | //------------------------------------------------------------------------ |
| 26 | |
| 27 | class UnboundedlyRecursiveOnUnboundedStealingTask : public tbb::task { |
| 28 | typedef UnboundedlyRecursiveOnUnboundedStealingTask this_type; |
| 29 | |
| 30 | this_type *m_Parent; |
| 31 | const int m_Depth; |
| 32 | volatile bool m_GoAhead; |
| 33 | |
| 34 | // Well, virtually unboundedly, for any practical purpose |
| 35 | static const int max_depth = 1000000; |
| 36 | |
| 37 | public: |
| 38 | UnboundedlyRecursiveOnUnboundedStealingTask( this_type *parent_ = NULL, int depth_ = max_depth ) |
| 39 | : m_Parent(parent_) |
| 40 | , m_Depth(depth_) |
| 41 | , m_GoAhead(true) |
| 42 | {} |
| 43 | |
| 44 | tbb::task* execute() __TBB_override { |
| 45 | // Using large padding array speeds up reaching stealing limit |
| 46 | const int paddingSize = 16 * 1024; |
| 47 | volatile char padding[paddingSize]; |
| 48 | if( !m_Parent || (m_Depth > 0 && m_Parent->m_GoAhead) ) { |
| 49 | if ( m_Parent ) { |
| 50 | // We are stolen, let our parent start waiting for us |
| 51 | m_Parent->m_GoAhead = false; |
| 52 | } |
| 53 | tbb::task &t = *new( allocate_child() ) this_type(this, m_Depth - 1); |
| 54 | set_ref_count( 2 ); |
| 55 | spawn( t ); |
| 56 | // Give a willing thief a chance to steal |
| 57 | for( int i = 0; i < 1000000 && m_GoAhead; ++i ) { |
| 58 | ++padding[i % paddingSize]; |
| 59 | __TBB_Yield(); |
| 60 | } |
| 61 | // If our child has not been stolen yet, then prohibit it siring ones |
| 62 | // of its own (when this thread executes it inside the next wait_for_all) |
| 63 | m_GoAhead = false; |
| 64 | wait_for_all(); |
| 65 | } |
| 66 | return NULL; |
| 67 | } |
| 68 | }; // UnboundedlyRecursiveOnUnboundedStealingTask |
| 69 | |
| 70 | tbb::atomic<int> Count; |
| 71 | |
| 72 | class RecursiveTask: public tbb::task { |
| 73 | const int m_ChildCount; |
| 74 | const int m_Depth; |
| 75 | //! Spawn tasks in list. Exact method depends upon m_Depth&bit_mask. |
| 76 | void SpawnList( tbb::task_list& list, int bit_mask ) { |
| 77 | if( m_Depth&bit_mask ) { |
| 78 | // Take address to check that signature of spawn(task_list&) is static. |
| 79 | void (*s)(tbb::task_list&) = &tbb::task::spawn; |
| 80 | (*s)(list); |
| 81 | ASSERT( list.empty(), NULL ); |
| 82 | wait_for_all(); |
| 83 | } else { |
| 84 | spawn_and_wait_for_all(list); |
| 85 | ASSERT( list.empty(), NULL ); |
| 86 | } |
| 87 | } |
| 88 | public: |
| 89 | RecursiveTask( int child_count, int depth_ ) : m_ChildCount(child_count), m_Depth(depth_) {} |
| 90 | tbb::task* execute() __TBB_override { |
| 91 | ++Count; |
| 92 | if( m_Depth>0 ) { |
| 93 | tbb::task_list list; |
| 94 | ASSERT( list.empty(), NULL ); |
| 95 | for( int k=0; k<m_ChildCount; ++k ) { |
| 96 | list.push_back( *new( allocate_child() ) RecursiveTask(m_ChildCount/2,m_Depth-1 ) ); |
| 97 | ASSERT( !list.empty(), NULL ); |
| 98 | } |
| 99 | set_ref_count( m_ChildCount+1 ); |
| 100 | SpawnList( list, 1 ); |
| 101 | // Now try reusing this as the parent. |
| 102 | set_ref_count(2); |
| 103 | list.push_back( *new ( allocate_child() ) tbb::empty_task() ); |
| 104 | SpawnList( list, 2 ); |
| 105 | } |
| 106 | return NULL; |
| 107 | } |
| 108 | }; |
| 109 | |
| 110 | //! Compute what Count should be after RecursiveTask(child_count,depth) runs. |
| 111 | static int Expected( int child_count, int depth ) { |
| 112 | return depth<=0 ? 1 : 1+child_count*Expected(child_count/2,depth-1); |
| 113 | } |
| 114 | |
| 115 | void TestStealLimit( int nthread ) { |
| 116 | #if __TBB_DEFINE_MIC |
| 117 | REMARK( "skipping steal limiting heuristics for %d threads\n" , nthread ); |
| 118 | #else// !_TBB_DEFINE_MIC |
| 119 | REMARK( "testing steal limiting heuristics for %d threads\n" , nthread ); |
| 120 | tbb::task_scheduler_init init(nthread); |
| 121 | tbb::task &t = *new( tbb::task::allocate_root() ) UnboundedlyRecursiveOnUnboundedStealingTask(); |
| 122 | tbb::task::spawn_root_and_wait(t); |
| 123 | #endif// _TBB_DEFINE_MIC |
| 124 | } |
| 125 | |
| 126 | //! Test task::spawn( task_list& ) |
| 127 | void TestSpawnChildren( int nthread ) { |
| 128 | REMARK("testing task::spawn(task_list&) for %d threads\n" ,nthread); |
| 129 | tbb::task_scheduler_init init(nthread); |
| 130 | for( int j=0; j<50; ++j ) { |
| 131 | Count = 0; |
| 132 | RecursiveTask& p = *new( tbb::task::allocate_root() ) RecursiveTask(j,4); |
| 133 | tbb::task::spawn_root_and_wait(p); |
| 134 | int expected = Expected(j,4); |
| 135 | ASSERT( Count==expected, NULL ); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | //! Test task::spawn_root_and_wait( task_list& ) |
| 140 | void TestSpawnRootList( int nthread ) { |
| 141 | REMARK("testing task::spawn_root_and_wait(task_list&) for %d threads\n" ,nthread); |
| 142 | tbb::task_scheduler_init init(nthread); |
| 143 | for( int j=0; j<5; ++j ) |
| 144 | for( int k=0; k<10; ++k ) { |
| 145 | Count = 0; |
| 146 | tbb::task_list list; |
| 147 | for( int i=0; i<k; ++i ) |
| 148 | list.push_back( *new( tbb::task::allocate_root() ) RecursiveTask(j,4) ); |
| 149 | tbb::task::spawn_root_and_wait(list); |
| 150 | int expected = k*Expected(j,4); |
| 151 | ASSERT( Count==expected, NULL ); |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | //------------------------------------------------------------------------ |
| 156 | // Test for task::recycle_as_safe_continuation |
| 157 | //------------------------------------------------------------------------ |
| 158 | |
| 159 | void TestSafeContinuation( int nthread ) { |
| 160 | REMARK("testing task::recycle_as_safe_continuation for %d threads\n" ,nthread); |
| 161 | tbb::task_scheduler_init init(nthread); |
| 162 | for( int j=8; j<33; ++j ) { |
| 163 | TaskGenerator& p = *new( tbb::task::allocate_root() ) TaskGenerator(j,5); |
| 164 | tbb::task::spawn_root_and_wait(p); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | //------------------------------------------------------------------------ |
| 169 | // Test affinity interface |
| 170 | //------------------------------------------------------------------------ |
| 171 | tbb::atomic<int> TotalCount; |
| 172 | |
| 173 | struct AffinityTask: public tbb::task { |
| 174 | const affinity_id expected_affinity_id; |
| 175 | bool noted; |
| 176 | /** Computing affinities is NOT supported by TBB, and may disappear in the future. |
| 177 | It is done here for sake of unit testing. */ |
| 178 | AffinityTask( int expected_affinity_id_ ) : |
| 179 | expected_affinity_id(affinity_id(expected_affinity_id_)), |
| 180 | noted(false) |
| 181 | { |
| 182 | set_affinity(expected_affinity_id); |
| 183 | ASSERT( 0u-expected_affinity_id>0u, "affinity_id not an unsigned integral type?" ); |
| 184 | ASSERT( affinity()==expected_affinity_id, NULL ); |
| 185 | } |
| 186 | tbb::task* execute() __TBB_override { |
| 187 | ++TotalCount; |
| 188 | return NULL; |
| 189 | } |
| 190 | void note_affinity( affinity_id id ) __TBB_override { |
| 191 | // There is no guarantee in TBB that a task runs on its affinity thread. |
| 192 | // However, the current implementation does accidentally guarantee it |
| 193 | // under certain conditions, such as the conditions here. |
| 194 | // We exploit those conditions for sake of unit testing. |
| 195 | ASSERT( id!=expected_affinity_id, NULL ); |
| 196 | ASSERT( !noted, "note_affinity_id called twice!" ); |
| 197 | ASSERT ( &self() == (tbb::task*)this, "Wrong innermost running task" ); |
| 198 | noted = true; |
| 199 | } |
| 200 | }; |
| 201 | |
| 202 | /** Note: This test assumes a lot about the internal implementation of affinity. |
| 203 | Do NOT use this as an example of good programming practice with TBB */ |
| 204 | void TestAffinity( int nthread ) { |
| 205 | TotalCount = 0; |
| 206 | int n = tbb::task_scheduler_init::default_num_threads(); |
| 207 | if( n>nthread ) |
| 208 | n = nthread; |
| 209 | tbb::task_scheduler_init init(n); |
| 210 | tbb::empty_task* t = new( tbb::task::allocate_root() ) tbb::empty_task; |
| 211 | tbb::task::affinity_id affinity_id = t->affinity(); |
| 212 | ASSERT( affinity_id==0, NULL ); |
| 213 | // Set ref_count for n-1 children, plus 1 for the wait. |
| 214 | t->set_ref_count(n); |
| 215 | // Spawn n-1 affinitized children. |
| 216 | for( int i=1; i<n; ++i ) |
| 217 | tbb::task::spawn( *new(t->allocate_child()) AffinityTask(i) ); |
| 218 | if( n>1 ) { |
| 219 | // Keep master from stealing |
| 220 | while( TotalCount!=n-1 ) |
| 221 | __TBB_Yield(); |
| 222 | } |
| 223 | // Wait for the children |
| 224 | t->wait_for_all(); |
| 225 | int k = 0; |
| 226 | GetTaskPtr(k)->destroy(*t); |
| 227 | ASSERT(k==1,NULL); |
| 228 | } |
| 229 | |
| 230 | struct NoteAffinityTask: public tbb::task { |
| 231 | bool noted; |
| 232 | NoteAffinityTask( int id ) : noted(false) |
| 233 | { |
| 234 | set_affinity(affinity_id(id)); |
| 235 | } |
| 236 | ~NoteAffinityTask () { |
| 237 | ASSERT (noted, "note_affinity has not been called" ); |
| 238 | } |
| 239 | tbb::task* execute() __TBB_override { |
| 240 | return NULL; |
| 241 | } |
| 242 | void note_affinity( affinity_id /*id*/ ) __TBB_override { |
| 243 | noted = true; |
| 244 | ASSERT ( &self() == (tbb::task*)this, "Wrong innermost running task" ); |
| 245 | } |
| 246 | }; |
| 247 | |
| 248 | // This test checks one of the paths inside the scheduler by affinitizing the child task |
| 249 | // to non-existent thread so that it is proxied in the local task pool but not retrieved |
| 250 | // by another thread. |
| 251 | // If no workers requested, the extra slot #2 is allocated for a worker thread to serve |
| 252 | // "enqueued" tasks. In this test, it is used only for the affinity purpose. |
| 253 | void TestNoteAffinityContext() { |
| 254 | tbb::task_scheduler_init init(1); |
| 255 | tbb::empty_task* t = new( tbb::task::allocate_root() ) tbb::empty_task; |
| 256 | t->set_ref_count(2); |
| 257 | // This master in the absence of workers will have an affinity id of 1. |
| 258 | // So use another number to make the task get proxied. |
| 259 | tbb::task::spawn( *new(t->allocate_child()) NoteAffinityTask(2) ); |
| 260 | t->wait_for_all(); |
| 261 | tbb::task::destroy(*t); |
| 262 | } |
| 263 | |
| 264 | //------------------------------------------------------------------------ |
| 265 | // Test that recovery actions work correctly for task::allocate_* methods |
| 266 | // when a task's constructor throws an exception. |
| 267 | //------------------------------------------------------------------------ |
| 268 | |
| 269 | #if TBB_USE_EXCEPTIONS |
| 270 | static int TestUnconstructibleTaskCount; |
| 271 | |
| 272 | struct ConstructionFailure { |
| 273 | }; |
| 274 | |
| 275 | #if __TBB_MSVC_UNREACHABLE_CODE_IGNORED |
| 276 | // Suppress pointless "unreachable code" warning. |
| 277 | #pragma warning (push) |
| 278 | #pragma warning (disable: 4702) |
| 279 | #endif |
| 280 | |
| 281 | //! Task that cannot be constructed. |
| 282 | template<size_t N> |
| 283 | struct UnconstructibleTask: public tbb::empty_task { |
| 284 | char space[N]; |
| 285 | UnconstructibleTask() { |
| 286 | throw ConstructionFailure(); |
| 287 | } |
| 288 | }; |
| 289 | |
| 290 | #if __TBB_MSVC_UNREACHABLE_CODE_IGNORED |
| 291 | #pragma warning (pop) |
| 292 | #endif |
| 293 | |
| 294 | #define TRY_BAD_CONSTRUCTION(x) \ |
| 295 | { \ |
| 296 | try { \ |
| 297 | new(x) UnconstructibleTask<N>; \ |
| 298 | } catch( const ConstructionFailure& ) { \ |
| 299 | ASSERT( parent()==original_parent, NULL ); \ |
| 300 | ASSERT( ref_count()==original_ref_count, "incorrectly changed ref_count" );\ |
| 301 | ++TestUnconstructibleTaskCount; \ |
| 302 | } \ |
| 303 | } |
| 304 | |
| 305 | template<size_t N> |
| 306 | struct RootTaskForTestUnconstructibleTask: public tbb::task { |
| 307 | tbb::task* execute() __TBB_override { |
| 308 | tbb::task* original_parent = parent(); |
| 309 | ASSERT( original_parent!=NULL, NULL ); |
| 310 | int original_ref_count = ref_count(); |
| 311 | TRY_BAD_CONSTRUCTION( allocate_root() ); |
| 312 | TRY_BAD_CONSTRUCTION( allocate_child() ); |
| 313 | TRY_BAD_CONSTRUCTION( allocate_continuation() ); |
| 314 | TRY_BAD_CONSTRUCTION( allocate_additional_child_of(*this) ); |
| 315 | return NULL; |
| 316 | } |
| 317 | }; |
| 318 | |
| 319 | template<size_t N> |
| 320 | void TestUnconstructibleTask() { |
| 321 | TestUnconstructibleTaskCount = 0; |
| 322 | tbb::task_scheduler_init init; |
| 323 | tbb::task* t = new( tbb::task::allocate_root() ) RootTaskForTestUnconstructibleTask<N>; |
| 324 | tbb::task::spawn_root_and_wait(*t); |
| 325 | ASSERT( TestUnconstructibleTaskCount==4, NULL ); |
| 326 | } |
| 327 | #endif /* TBB_USE_EXCEPTIONS */ |
| 328 | |
| 329 | //------------------------------------------------------------------------ |
| 330 | // Test for alignment problems with task objects. |
| 331 | //------------------------------------------------------------------------ |
| 332 | |
| 333 | #if _MSC_VER && !defined(__INTEL_COMPILER) |
| 334 | // Workaround for pointless warning "structure was padded due to __declspec(align()) |
| 335 | #pragma warning (push) |
| 336 | #pragma warning (disable: 4324) |
| 337 | #endif |
| 338 | |
| 339 | //! Task with members of type T. |
| 340 | /** The task recursively creates tasks. */ |
| 341 | template<typename T> |
| 342 | class TaskWithMember: public tbb::task { |
| 343 | T x; |
| 344 | T y; |
| 345 | unsigned char count; |
| 346 | tbb::task* execute() __TBB_override { |
| 347 | x = y; |
| 348 | if( count>0 ) { |
| 349 | set_ref_count(2); |
| 350 | tbb::task* t = new( allocate_child() ) TaskWithMember<T>(count-1); |
| 351 | spawn_and_wait_for_all(*t); |
| 352 | } |
| 353 | return NULL; |
| 354 | } |
| 355 | public: |
| 356 | TaskWithMember( unsigned char n ) : count(n) {} |
| 357 | }; |
| 358 | |
| 359 | #if _MSC_VER && !defined(__INTEL_COMPILER) |
| 360 | #pragma warning (pop) |
| 361 | #endif |
| 362 | |
| 363 | template<typename T> |
| 364 | void TestAlignmentOfOneClass() { |
| 365 | typedef TaskWithMember<T> task_type; |
| 366 | tbb::task* t = new( tbb::task::allocate_root() ) task_type(10); |
| 367 | tbb::task::spawn_root_and_wait(*t); |
| 368 | } |
| 369 | |
| 370 | #include "harness_m128.h" |
| 371 | |
| 372 | void TestAlignment() { |
| 373 | REMARK("testing alignment\n" ); |
| 374 | tbb::task_scheduler_init init; |
| 375 | // Try types that have variety of alignments |
| 376 | TestAlignmentOfOneClass<char>(); |
| 377 | TestAlignmentOfOneClass<short>(); |
| 378 | TestAlignmentOfOneClass<int>(); |
| 379 | TestAlignmentOfOneClass<long>(); |
| 380 | TestAlignmentOfOneClass<void*>(); |
| 381 | TestAlignmentOfOneClass<float>(); |
| 382 | TestAlignmentOfOneClass<double>(); |
| 383 | #if HAVE_m128 |
| 384 | TestAlignmentOfOneClass<__m128>(); |
| 385 | #endif |
| 386 | #if HAVE_m256 |
| 387 | if (have_AVX()) TestAlignmentOfOneClass<__m256>(); |
| 388 | #endif |
| 389 | } |
| 390 | |
| 391 | //------------------------------------------------------------------------ |
| 392 | // Test for recursing on left while spawning on right |
| 393 | //------------------------------------------------------------------------ |
| 394 | |
| 395 | int Fib( int n ); |
| 396 | |
| 397 | struct RightFibTask: public tbb::task { |
| 398 | int* y; |
| 399 | const int n; |
| 400 | RightFibTask( int* y_, int n_ ) : y(y_), n(n_) {} |
| 401 | task* execute() __TBB_override { |
| 402 | *y = Fib(n-1); |
| 403 | return 0; |
| 404 | } |
| 405 | }; |
| 406 | |
| 407 | int Fib( int n ) { |
| 408 | if( n<2 ) { |
| 409 | return n; |
| 410 | } else { |
| 411 | // y actually does not need to be initialized. It is initialized solely to suppress |
| 412 | // a gratuitous warning "potentially uninitialized local variable". |
| 413 | int y=-1; |
| 414 | tbb::task* root_task = new( tbb::task::allocate_root() ) tbb::empty_task; |
| 415 | root_task->set_ref_count(2); |
| 416 | tbb::task::spawn( *new( root_task->allocate_child() ) RightFibTask(&y,n) ); |
| 417 | int x = Fib(n-2); |
| 418 | root_task->wait_for_all(); |
| 419 | tbb::task::destroy(*root_task); |
| 420 | return y+x; |
| 421 | } |
| 422 | } |
| 423 | |
| 424 | void TestLeftRecursion( int p ) { |
| 425 | REMARK("testing non-spawned roots for %d threads\n" ,p); |
| 426 | tbb::task_scheduler_init init(p); |
| 427 | int sum = 0; |
| 428 | for( int i=0; i<100; ++i ) |
| 429 | sum +=Fib(10); |
| 430 | ASSERT( sum==5500, NULL ); |
| 431 | } |
| 432 | |
| 433 | //------------------------------------------------------------------------ |
| 434 | // Test for computing with DAG of tasks. |
| 435 | //------------------------------------------------------------------------ |
| 436 | |
| 437 | class DagTask: public tbb::task { |
| 438 | typedef unsigned long long number_t; |
| 439 | const int i, j; |
| 440 | number_t sum_from_left, sum_from_above; |
| 441 | void check_sum( number_t sum ) { |
| 442 | number_t expected_sum = 1; |
| 443 | for( int k=i+1; k<=i+j; ++k ) |
| 444 | expected_sum *= k; |
| 445 | for( int k=1; k<=j; ++k ) |
| 446 | expected_sum /= k; |
| 447 | ASSERT(sum==expected_sum, NULL); |
| 448 | } |
| 449 | public: |
| 450 | DagTask *successor_to_below, *successor_to_right; |
| 451 | DagTask( int i_, int j_ ) : i(i_), j(j_), sum_from_left(0), sum_from_above(0) {} |
| 452 | task* execute() __TBB_override { |
| 453 | ASSERT( ref_count()==0, NULL ); |
| 454 | number_t sum = i==0 && j==0 ? 1 : sum_from_left+sum_from_above; |
| 455 | check_sum(sum); |
| 456 | ++execution_count; |
| 457 | if( DagTask* t = successor_to_right ) { |
| 458 | t->sum_from_left = sum; |
| 459 | if( t->decrement_ref_count()==0 ) |
| 460 | // Test using spawn to evaluate DAG |
| 461 | spawn( *t ); |
| 462 | } |
| 463 | if( DagTask* t = successor_to_below ) { |
| 464 | t->sum_from_above = sum; |
| 465 | if( t->add_ref_count(-1)==0 ) |
| 466 | // Test using bypass to evaluate DAG |
| 467 | return t; |
| 468 | } |
| 469 | return NULL; |
| 470 | } |
| 471 | ~DagTask() {++destruction_count;} |
| 472 | static tbb::atomic<int> execution_count; |
| 473 | static tbb::atomic<int> destruction_count; |
| 474 | }; |
| 475 | |
| 476 | tbb::atomic<int> DagTask::execution_count; |
| 477 | tbb::atomic<int> DagTask::destruction_count; |
| 478 | |
| 479 | void TestDag( int p ) { |
| 480 | REMARK("testing evaluation of DAG for %d threads\n" ,p); |
| 481 | tbb::task_scheduler_init init(p); |
| 482 | DagTask::execution_count=0; |
| 483 | DagTask::destruction_count=0; |
| 484 | const int n = 10; |
| 485 | DagTask* a[n][n]; |
| 486 | for( int i=0; i<n; ++i ) |
| 487 | for( int j=0; j<n; ++j ) |
| 488 | a[i][j] = new( tbb::task::allocate_root() ) DagTask(i,j); |
| 489 | for( int i=0; i<n; ++i ) |
| 490 | for( int j=0; j<n; ++j ) { |
| 491 | a[i][j]->successor_to_below = i+1<n ? a[i+1][j] : NULL; |
| 492 | a[i][j]->successor_to_right = j+1<n ? a[i][j+1] : NULL; |
| 493 | a[i][j]->set_ref_count((i>0)+(j>0)); |
| 494 | } |
| 495 | a[n-1][n-1]->increment_ref_count(); |
| 496 | a[n-1][n-1]->spawn_and_wait_for_all(*a[0][0]); |
| 497 | ASSERT( DagTask::execution_count == n*n - 1, NULL ); |
| 498 | tbb::task::destroy(*a[n-1][n-1]); |
| 499 | ASSERT( DagTask::destruction_count > n*n - p, NULL ); |
| 500 | while ( DagTask::destruction_count != n*n ) |
| 501 | __TBB_Yield(); |
| 502 | } |
| 503 | |
| 504 | #include "harness_barrier.h" |
| 505 | |
| 506 | class RelaxedOwnershipTask: public tbb::task { |
| 507 | tbb::task &m_taskToSpawn, |
| 508 | &m_taskToDestroy, |
| 509 | &m_taskToExecute; |
| 510 | static Harness::SpinBarrier m_barrier; |
| 511 | |
| 512 | tbb::task* execute () __TBB_override { |
| 513 | tbb::task &p = *parent(); |
| 514 | tbb::task &r = *new( allocate_root() ) tbb::empty_task; |
| 515 | r.set_ref_count( 1 ); |
| 516 | m_barrier.wait(); |
| 517 | p.spawn( *new(p.allocate_child()) tbb::empty_task ); |
| 518 | p.spawn( *new(task::allocate_additional_child_of(p)) tbb::empty_task ); |
| 519 | p.spawn( m_taskToSpawn ); |
| 520 | p.destroy( m_taskToDestroy ); |
| 521 | r.spawn_and_wait_for_all( m_taskToExecute ); |
| 522 | p.destroy( r ); |
| 523 | return NULL; |
| 524 | } |
| 525 | public: |
| 526 | RelaxedOwnershipTask ( tbb::task& toSpawn, tbb::task& toDestroy, tbb::task& toExecute ) |
| 527 | : m_taskToSpawn(toSpawn) |
| 528 | , m_taskToDestroy(toDestroy) |
| 529 | , m_taskToExecute(toExecute) |
| 530 | {} |
| 531 | static void SetBarrier ( int numThreads ) { m_barrier.initialize( numThreads ); } |
| 532 | }; |
| 533 | |
| 534 | Harness::SpinBarrier RelaxedOwnershipTask::m_barrier; |
| 535 | |
| 536 | void TestRelaxedOwnership( int p ) { |
| 537 | if ( p < 2 ) |
| 538 | return; |
| 539 | |
| 540 | if( unsigned(p)>tbb::tbb_thread::hardware_concurrency() ) |
| 541 | return; |
| 542 | |
| 543 | REMARK("testing tasks exercising relaxed ownership freedom for %d threads\n" , p); |
| 544 | tbb::task_scheduler_init init(p); |
| 545 | RelaxedOwnershipTask::SetBarrier(p); |
| 546 | tbb::task &r = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 547 | tbb::task_list tl; |
| 548 | for ( int i = 0; i < p; ++i ) { |
| 549 | tbb::task &tS = *new( r.allocate_child() ) tbb::empty_task, |
| 550 | &tD = *new( r.allocate_child() ) tbb::empty_task, |
| 551 | &tE = *new( r.allocate_child() ) tbb::empty_task; |
| 552 | tl.push_back( *new( r.allocate_child() ) RelaxedOwnershipTask(tS, tD, tE) ); |
| 553 | } |
| 554 | r.set_ref_count( 5 * p + 1 ); |
| 555 | int k=0; |
| 556 | GetTaskPtr(k)->spawn( tl ); |
| 557 | ASSERT(k==1,NULL); |
| 558 | r.wait_for_all(); |
| 559 | r.destroy( r ); |
| 560 | } |
| 561 | |
| 562 | //------------------------------------------------------------------------ |
| 563 | // Test for running TBB scheduler on user-created thread. |
| 564 | //------------------------------------------------------------------------ |
| 565 | |
| 566 | void RunSchedulerInstanceOnUserThread( int n_child ) { |
| 567 | tbb::task* e = new( tbb::task::allocate_root() ) tbb::empty_task; |
| 568 | e->set_ref_count(1+n_child); |
| 569 | for( int i=0; i<n_child; ++i ) |
| 570 | tbb::task::spawn( *new(e->allocate_child()) tbb::empty_task ); |
| 571 | e->wait_for_all(); |
| 572 | e->destroy(*e); |
| 573 | } |
| 574 | |
| 575 | void TestUserThread( int p ) { |
| 576 | tbb::task_scheduler_init init(p); |
| 577 | // Try with both 0 and 1 children. Only the latter scenario permits stealing. |
| 578 | for( int n_child=0; n_child<2; ++n_child ) { |
| 579 | tbb::tbb_thread t( RunSchedulerInstanceOnUserThread, n_child ); |
| 580 | t.join(); |
| 581 | } |
| 582 | } |
| 583 | |
| 584 | class TaskWithChildToSteal : public tbb::task { |
| 585 | const int m_Depth; |
| 586 | volatile bool m_GoAhead; |
| 587 | |
| 588 | public: |
| 589 | TaskWithChildToSteal( int depth_ ) |
| 590 | : m_Depth(depth_) |
| 591 | , m_GoAhead(false) |
| 592 | {} |
| 593 | |
| 594 | tbb::task* execute() __TBB_override { |
| 595 | m_GoAhead = true; |
| 596 | if ( m_Depth > 0 ) { |
| 597 | TaskWithChildToSteal &t = *new( allocate_child() ) TaskWithChildToSteal(m_Depth - 1); |
| 598 | t.SpawnAndWaitOnParent(); |
| 599 | } |
| 600 | else |
| 601 | Harness::Sleep(50); // The last task in chain sleeps for 50 ms |
| 602 | return NULL; |
| 603 | } |
| 604 | |
| 605 | void SpawnAndWaitOnParent() { |
| 606 | parent()->set_ref_count( 2 ); |
| 607 | parent()->spawn( *this ); |
| 608 | while (!this->m_GoAhead ) |
| 609 | __TBB_Yield(); |
| 610 | parent()->wait_for_all(); |
| 611 | } |
| 612 | }; // TaskWithChildToSteal |
| 613 | |
| 614 | // Success criterion of this test is not hanging |
| 615 | void TestDispatchLoopResponsiveness() { |
| 616 | REMARK("testing that dispatch loops do not go into eternal sleep when all remaining children are stolen\n" ); |
| 617 | // Recursion depth values test the following sorts of dispatch loops |
| 618 | // 0 - master's outermost |
| 619 | // 1 - worker's nested |
| 620 | // 2 - master's nested |
| 621 | tbb::task_scheduler_init init(2); |
| 622 | tbb::task &r = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 623 | for ( int depth = 0; depth < 3; ++depth ) { |
| 624 | TaskWithChildToSteal &t = *new( r.allocate_child() ) TaskWithChildToSteal(depth); |
| 625 | t.SpawnAndWaitOnParent(); |
| 626 | } |
| 627 | r.destroy(r); |
| 628 | } |
| 629 | |
| 630 | void TestWaitDiscriminativenessWithoutStealing() { |
| 631 | REMARK( "testing that task::wait_for_all is specific to the root it is called on (no workers)\n" ); |
| 632 | // The test relies on the strict LIFO scheduling order in the absence of workers |
| 633 | tbb::task_scheduler_init init(1); |
| 634 | tbb::task &r1 = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 635 | tbb::task &r2 = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 636 | const int NumChildren = 10; |
| 637 | r1.set_ref_count( NumChildren + 1 ); |
| 638 | r2.set_ref_count( NumChildren + 1 ); |
| 639 | for( int i=0; i < NumChildren; ++i ) { |
| 640 | tbb::empty_task &t1 = *new( r1.allocate_child() ) tbb::empty_task; |
| 641 | tbb::empty_task &t2 = *new( r2.allocate_child() ) tbb::empty_task; |
| 642 | tbb::task::spawn(t1); |
| 643 | tbb::task::spawn(t2); |
| 644 | } |
| 645 | r2.wait_for_all(); |
| 646 | ASSERT( r2.ref_count() <= 1, "Not all children of r2 executed" ); |
| 647 | ASSERT( r1.ref_count() > 1, "All children of r1 prematurely executed" ); |
| 648 | r1.wait_for_all(); |
| 649 | ASSERT( r1.ref_count() <= 1, "Not all children of r1 executed" ); |
| 650 | r1.destroy(r1); |
| 651 | r2.destroy(r2); |
| 652 | } |
| 653 | |
| 654 | |
| 655 | using tbb::internal::spin_wait_until_eq; |
| 656 | |
| 657 | //! Deterministic emulation of a long running task |
| 658 | class LongRunningTask : public tbb::task { |
| 659 | volatile bool& m_CanProceed; |
| 660 | |
| 661 | tbb::task* execute() __TBB_override { |
| 662 | spin_wait_until_eq( m_CanProceed, true ); |
| 663 | return NULL; |
| 664 | } |
| 665 | public: |
| 666 | LongRunningTask ( volatile bool& canProceed ) : m_CanProceed(canProceed) {} |
| 667 | }; |
| 668 | |
| 669 | void TestWaitDiscriminativenessWithStealing() { |
| 670 | if( tbb::tbb_thread::hardware_concurrency() < 2 ) |
| 671 | return; |
| 672 | REMARK( "testing that task::wait_for_all is specific to the root it is called on (one worker)\n" ); |
| 673 | volatile bool canProceed = false; |
| 674 | tbb::task_scheduler_init init(2); |
| 675 | tbb::task &r1 = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 676 | tbb::task &r2 = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 677 | r1.set_ref_count( 2 ); |
| 678 | r2.set_ref_count( 2 ); |
| 679 | tbb::task& t1 = *new( r1.allocate_child() ) tbb::empty_task; |
| 680 | tbb::task& t2 = *new( r2.allocate_child() ) LongRunningTask(canProceed); |
| 681 | tbb::task::spawn(t2); |
| 682 | tbb::task::spawn(t1); |
| 683 | r1.wait_for_all(); |
| 684 | ASSERT( r1.ref_count() <= 1, "Not all children of r1 executed" ); |
| 685 | ASSERT( r2.ref_count() == 2, "All children of r2 prematurely executed" ); |
| 686 | canProceed = true; |
| 687 | r2.wait_for_all(); |
| 688 | ASSERT( r2.ref_count() <= 1, "Not all children of r2 executed" ); |
| 689 | r1.destroy(r1); |
| 690 | r2.destroy(r2); |
| 691 | } |
| 692 | |
| 693 | struct MasterBody : NoAssign, Harness::NoAfterlife { |
| 694 | static Harness::SpinBarrier my_barrier; |
| 695 | |
| 696 | class BarrenButLongTask : public tbb::task { |
| 697 | volatile bool& m_Started; |
| 698 | volatile bool& m_CanProceed; |
| 699 | |
| 700 | tbb::task* execute() __TBB_override { |
| 701 | m_Started = true; |
| 702 | spin_wait_until_eq( m_CanProceed, true ); |
| 703 | volatile int k = 0; |
| 704 | for ( int i = 0; i < 1000000; ++i ) ++k; |
| 705 | return NULL; |
| 706 | } |
| 707 | public: |
| 708 | BarrenButLongTask ( volatile bool& started, volatile bool& can_proceed ) |
| 709 | : m_Started(started), m_CanProceed(can_proceed) |
| 710 | {} |
| 711 | }; |
| 712 | |
| 713 | class BinaryRecursiveTask : public tbb::task { |
| 714 | int m_Depth; |
| 715 | |
| 716 | tbb::task* execute() __TBB_override { |
| 717 | if( !m_Depth ) |
| 718 | return NULL; |
| 719 | set_ref_count(3); |
| 720 | spawn( *new( allocate_child() ) BinaryRecursiveTask(m_Depth - 1) ); |
| 721 | spawn( *new( allocate_child() ) BinaryRecursiveTask(m_Depth - 1) ); |
| 722 | wait_for_all(); |
| 723 | return NULL; |
| 724 | } |
| 725 | |
| 726 | void note_affinity( affinity_id ) __TBB_override { |
| 727 | ASSERT( false, "These tasks cannot be stolen" ); |
| 728 | } |
| 729 | public: |
| 730 | BinaryRecursiveTask ( int depth_ ) : m_Depth(depth_) {} |
| 731 | }; |
| 732 | |
| 733 | void operator() ( int id ) const { |
| 734 | if ( id ) { |
| 735 | tbb::task_scheduler_init init(2); |
| 736 | volatile bool child_started = false, |
| 737 | can_proceed = false; |
| 738 | tbb::task& r = *new( tbb::task::allocate_root() ) tbb::empty_task; |
| 739 | r.set_ref_count(2); |
| 740 | r.spawn( *new(r.allocate_child()) BarrenButLongTask(child_started, can_proceed) ); |
| 741 | spin_wait_until_eq( child_started, true ); |
| 742 | my_barrier.wait(); |
| 743 | can_proceed = true; |
| 744 | r.wait_for_all(); |
| 745 | r.destroy(r); |
| 746 | } |
| 747 | else { |
| 748 | my_barrier.wait(); |
| 749 | tbb::task_scheduler_init init(1); |
| 750 | Count = 0; |
| 751 | int depth = 16; |
| 752 | BinaryRecursiveTask& r = *new( tbb::task::allocate_root() ) BinaryRecursiveTask(depth); |
| 753 | tbb::task::spawn_root_and_wait(r); |
| 754 | } |
| 755 | } |
| 756 | public: |
| 757 | MasterBody ( int num_masters ) { my_barrier.initialize(num_masters); } |
| 758 | }; |
| 759 | |
| 760 | Harness::SpinBarrier MasterBody::my_barrier; |
| 761 | |
| 762 | /** Ensures that tasks spawned by a master thread or one of the workers servicing |
| 763 | it cannot be stolen by another master thread. **/ |
| 764 | void TestMastersIsolation ( int p ) { |
| 765 | // The test requires at least 3-way parallelism to work correctly |
| 766 | if ( p > 2 && tbb::task_scheduler_init::default_num_threads() >= p ) { |
| 767 | tbb::task_scheduler_init init(p); |
| 768 | NativeParallelFor( p, MasterBody(p) ); |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | struct waitable_task : tbb::task { |
| 773 | tbb::task* execute() __TBB_override { |
| 774 | recycle_as_safe_continuation(); // do not destroy the task after execution |
| 775 | set_parent(this); // decrement its own ref_count after completion |
| 776 | __TBB_Yield(); |
| 777 | return NULL; |
| 778 | } |
| 779 | }; |
| 780 | void TestWaitableTask() { |
| 781 | waitable_task &wt = *new( tbb::task::allocate_root() ) waitable_task; |
| 782 | for( int i = 0; i < 100000; i++ ) { |
| 783 | wt.set_ref_count(2); // prepare for waiting on it |
| 784 | wt.spawn(wt); |
| 785 | if( i&1 ) __TBB_Yield(); |
| 786 | wt.wait_for_all(); |
| 787 | } |
| 788 | wt.set_parent(NULL); // prevents assertions and atomics in task::destroy |
| 789 | tbb::task::destroy(wt); |
| 790 | } |
| 791 | |
| 792 | #if __TBB_PREVIEW_CRITICAL_TASKS |
| 793 | #include <stdexcept> |
| 794 | #include <vector> |
| 795 | #include <map> |
| 796 | #include "tbb/parallel_for.h" |
| 797 | |
| 798 | namespace CriticalTaskSupport { |
| 799 | |
| 800 | using tbb::task; |
| 801 | task* g_root_task = NULL; |
| 802 | |
| 803 | // markers to capture execution profile (declaration order is important) |
| 804 | enum task_marker_t { |
| 805 | no_task, regular_task, isolated_regular_task, |
| 806 | outer_critical_task, nested_critical_task, critical_from_isolated_task, bypassed_critical_task |
| 807 | }; |
| 808 | enum bypassed_critical_task_stage_t { not_bypassed, bypassed, executed }; |
| 809 | |
| 810 | typedef std::vector< std::vector<task_marker_t> > task_map_t; |
| 811 | task_map_t g_execution_profile; |
| 812 | |
| 813 | const int g_per_thread_regular_tasks_num = 5; |
| 814 | const int g_isolated_regular_task_num = 3; |
| 815 | tbb::atomic<bool> g_is_critical_task_submitted; |
| 816 | size_t g_bypassed_critical_task_index = size_t(-1); |
| 817 | task* g_bypassed_task_pointer = NULL; |
| 818 | int g_bypassed_task_creator = -1; |
| 819 | tbb::atomic<bypassed_critical_task_stage_t> g_bypassed_critical_task_stage; |
| 820 | tbb::task_arena g_arena; |
| 821 | Harness::SpinBarrier g_spin_barrier; |
| 822 | |
| 823 | struct parallel_for_body { |
| 824 | parallel_for_body(task_marker_t task_marker, bool submit_critical = false) |
| 825 | : my_task_marker(task_marker), my_submit_critical(submit_critical) {} |
| 826 | void operator()( int i ) const; |
| 827 | private: |
| 828 | task_marker_t my_task_marker; |
| 829 | bool my_submit_critical; |
| 830 | }; |
| 831 | |
| 832 | struct IsolatedFunctor { |
| 833 | void operator()() const { |
| 834 | parallel_for_body body(isolated_regular_task, /*submit_critical=*/ true); |
| 835 | tbb::parallel_for( 0, g_isolated_regular_task_num, body, tbb::simple_partitioner() ); |
| 836 | } |
| 837 | }; |
| 838 | |
| 839 | struct CriticalTaskBody : public task { |
| 840 | CriticalTaskBody(task_marker_t task_marker) : my_task_mark(task_marker) {} |
| 841 | task* execute() __TBB_override { |
| 842 | task* ret_task = NULL; |
| 843 | task* nested_task = NULL; |
| 844 | int thread_idx = tbb::this_task_arena::current_thread_index(); |
| 845 | g_execution_profile[thread_idx].push_back(my_task_mark); |
| 846 | switch( my_task_mark ) { |
| 847 | case outer_critical_task: |
| 848 | g_spin_barrier.wait(); // allow each thread to take its own critical task |
| 849 | // prefill queue with critical tasks |
| 850 | nested_task = new( task::allocate_additional_child_of(*g_root_task) ) |
| 851 | CriticalTaskBody(nested_critical_task); |
| 852 | enqueue( *nested_task, tbb::priority_t(tbb::internal::priority_critical) ); |
| 853 | if( not_bypassed == |
| 854 | g_bypassed_critical_task_stage.compare_and_swap(bypassed, not_bypassed) ) { |
| 855 | |
| 856 | // first, should process all the work from isolated region |
| 857 | tbb::this_task_arena::isolate( IsolatedFunctor() ); |
| 858 | |
| 859 | CriticalTaskBody* bypassed_task = |
| 860 | new( task::allocate_additional_child_of(*g_root_task) ) |
| 861 | CriticalTaskBody(bypassed_critical_task); |
| 862 | g_bypassed_task_pointer = bypassed_task; |
| 863 | g_bypassed_critical_task_index = g_execution_profile[thread_idx].size() + 1; |
| 864 | g_bypassed_task_creator = thread_idx; |
| 865 | tbb::internal::make_critical(*bypassed_task); |
| 866 | ret_task = bypassed_task; |
| 867 | } |
| 868 | g_spin_barrier.wait(); // allow thread to execute isolated region |
| 869 | break; |
| 870 | case nested_critical_task: |
| 871 | // wait until bypassed critical task has been executed |
| 872 | g_spin_barrier.wait(); |
| 873 | break; |
| 874 | case bypassed_critical_task: |
| 875 | ASSERT( bypassed == g_bypassed_critical_task_stage, "Unexpected bypassed critical task" ); |
| 876 | g_bypassed_critical_task_stage = executed; |
| 877 | ASSERT( thread_idx == g_bypassed_task_creator, |
| 878 | "Bypassed critical task is not being executed by the thread that bypassed it." ); |
| 879 | ASSERT( g_bypassed_task_pointer == this, "This is not bypassed task." ); |
| 880 | ASSERT( g_bypassed_critical_task_index == g_execution_profile[thread_idx].size(), |
| 881 | "Bypassed critical task was not selected as the next task." ); |
| 882 | break; |
| 883 | case critical_from_isolated_task: |
| 884 | break; |
| 885 | default: |
| 886 | ASSERT( false, "Incorrect critical task id." ); |
| 887 | } |
| 888 | return ret_task; |
| 889 | } |
| 890 | private: |
| 891 | task_marker_t my_task_mark; |
| 892 | }; |
| 893 | |
| 894 | void parallel_for_body::operator()( int i ) const { |
| 895 | int thread_idx = tbb::this_task_arena::current_thread_index(); |
| 896 | g_execution_profile[thread_idx].push_back(my_task_marker); |
| 897 | if( my_submit_critical && i == 0 ) { |
| 898 | task* isolated_task = new( task::allocate_additional_child_of(*g_root_task) ) |
| 899 | CriticalTaskBody(critical_from_isolated_task); |
| 900 | task::enqueue( *isolated_task, tbb::priority_t(tbb::internal::priority_critical) ); |
| 901 | } |
| 902 | } |
| 903 | |
| 904 | struct TaskBody: public task { |
| 905 | TaskBody() {} |
| 906 | TaskBody(task_marker_t /*mark*/) {} |
| 907 | task* execute() __TBB_override { |
| 908 | int thread_idx = tbb::this_task_arena::current_thread_index(); |
| 909 | g_execution_profile[thread_idx].push_back(regular_task); |
| 910 | if( !g_is_critical_task_submitted ) { |
| 911 | g_spin_barrier.wait(); // allow each thread to take its own task. |
| 912 | // prefill task pools with regular tasks |
| 913 | int half = g_per_thread_regular_tasks_num / 2; |
| 914 | for( int i = 0; i < half; ++i ) { |
| 915 | task& t = *new( task::allocate_additional_child_of(*g_root_task) ) |
| 916 | TaskBody; |
| 917 | spawn(t); |
| 918 | } |
| 919 | { |
| 920 | // prefill with critical tasks |
| 921 | task& t = *new( task::allocate_additional_child_of(*g_root_task) ) |
| 922 | CriticalTaskBody(outer_critical_task); |
| 923 | tbb::internal::make_critical(t); |
| 924 | tbb::task::spawn(t); |
| 925 | } |
| 926 | // prefill task pools with regular tasks |
| 927 | for( int i = half; i < g_per_thread_regular_tasks_num; ++i ) { |
| 928 | task& t = *new( task::allocate_additional_child_of(*g_root_task) ) |
| 929 | TaskBody; |
| 930 | spawn(t); |
| 931 | } |
| 932 | g_is_critical_task_submitted.store<tbb::relaxed>(true); |
| 933 | g_spin_barrier.wait(); |
| 934 | } |
| 935 | return NULL; |
| 936 | } |
| 937 | }; |
| 938 | |
| 939 | template<typename TaskType, void(*submit_task)(task&)> |
| 940 | struct WorkCreator { |
| 941 | WorkCreator(task*& root_task, size_t num_tasks, size_t num_critical_tasks = 0, |
| 942 | tbb::task_group_context* ctx = NULL) |
| 943 | : my_root_task(root_task), my_num_tasks(num_tasks), my_num_critical_tasks(num_critical_tasks), |
| 944 | my_context(ctx) {} |
| 945 | void operator()() const { |
| 946 | ASSERT( my_root_task == NULL, "Incorrect test set up." ); |
| 947 | task* root_task = NULL; |
| 948 | if( my_context ) |
| 949 | root_task = new( task::allocate_root(*my_context) ) TaskType(regular_task); |
| 950 | else |
| 951 | root_task = new( task::allocate_root() ) TaskType(regular_task); |
| 952 | root_task->increment_ref_count(); |
| 953 | for( size_t i = 0; i < my_num_tasks; ++i ) { |
| 954 | task& t = *new( task::allocate_additional_child_of(*root_task) ) TaskType(regular_task); |
| 955 | submit_task(t); |
| 956 | } |
| 957 | for( size_t i = 0; i < my_num_critical_tasks; ++i ) { |
| 958 | task& t = *new( task::allocate_additional_child_of(*root_task) ) |
| 959 | TaskType( outer_critical_task ); |
| 960 | tbb::task::enqueue( t, tbb::priority_t(tbb::internal::priority_critical) ); |
| 961 | } |
| 962 | my_root_task = root_task; |
| 963 | } |
| 964 | private: |
| 965 | task*& my_root_task; |
| 966 | size_t my_num_tasks; |
| 967 | size_t my_num_critical_tasks; |
| 968 | tbb::task_group_context* my_context; |
| 969 | }; |
| 970 | |
| 971 | struct WorkAwaiter { |
| 972 | WorkAwaiter(task*& root_task) : my_root_task(root_task) {} |
| 973 | void operator()() const { |
| 974 | while( !my_root_task ) __TBB_Yield(); // waiting on a tree construction |
| 975 | my_root_task->wait_for_all(); |
| 976 | task::destroy(*my_root_task); |
| 977 | my_root_task = NULL; |
| 978 | } |
| 979 | private: |
| 980 | task*& my_root_task; |
| 981 | }; |
| 982 | |
| 983 | void TestSchedulerTaskSelectionWhenSpawn() { |
| 984 | REMARK( "\tPreferring critical tasks among spawned\n" ); |
| 985 | typedef std::multimap<task_marker_t, task_marker_t> state_machine_t; |
| 986 | typedef state_machine_t::iterator states_it; |
| 987 | task_marker_t from_to_pairs[] = { |
| 988 | // from regular |
| 989 | regular_task, regular_task, |
| 990 | regular_task, outer_critical_task, |
| 991 | // from outermost critical |
| 992 | outer_critical_task, isolated_regular_task, |
| 993 | outer_critical_task, critical_from_isolated_task, |
| 994 | outer_critical_task, nested_critical_task, |
| 995 | // from isolated regular |
| 996 | isolated_regular_task, isolated_regular_task, |
| 997 | isolated_regular_task, critical_from_isolated_task, |
| 998 | isolated_regular_task, bypassed_critical_task, |
| 999 | // from critical that was enqueued from isolated region |
| 1000 | critical_from_isolated_task, isolated_regular_task, |
| 1001 | critical_from_isolated_task, nested_critical_task, |
| 1002 | critical_from_isolated_task, regular_task, |
| 1003 | critical_from_isolated_task, bypassed_critical_task, |
| 1004 | // from bypassed critical |
| 1005 | bypassed_critical_task, nested_critical_task, |
| 1006 | bypassed_critical_task, critical_from_isolated_task, |
| 1007 | // from nested critical |
| 1008 | nested_critical_task, critical_from_isolated_task, |
| 1009 | nested_critical_task, regular_task |
| 1010 | }; |
| 1011 | |
| 1012 | state_machine_t allowed_transitions; |
| 1013 | for( size_t i = 0; i < sizeof(from_to_pairs) / sizeof(from_to_pairs[0]); i += 2 ) |
| 1014 | allowed_transitions.insert( std::make_pair( from_to_pairs[i], from_to_pairs[i+1] ) ); |
| 1015 | |
| 1016 | for( int num_threads = MinThread; num_threads <= MaxThread; ++num_threads ) { |
| 1017 | for( int repeat = 0; repeat < 10; ++repeat ) { |
| 1018 | // test initialization |
| 1019 | g_bypassed_critical_task_stage = not_bypassed; |
| 1020 | g_is_critical_task_submitted = false; |
| 1021 | g_bypassed_critical_task_index = size_t(-1); |
| 1022 | g_bypassed_task_creator = -1; |
| 1023 | g_bypassed_task_pointer = NULL; |
| 1024 | g_execution_profile.resize(num_threads); |
| 1025 | g_spin_barrier.initialize(num_threads); |
| 1026 | g_arena.initialize(num_threads); |
| 1027 | |
| 1028 | // test execution |
| 1029 | g_arena.execute( |
| 1030 | WorkCreator<TaskBody, task::spawn>(g_root_task, /*num_tasks=*/size_t(num_threads)) ); |
| 1031 | g_arena.execute( WorkAwaiter(g_root_task) ); |
| 1032 | |
| 1033 | // checking how execution went |
| 1034 | int critical_task_count = 0; |
| 1035 | for( int thread = 0; thread < num_threads; ++thread ) { |
| 1036 | bool started_critical_region = false; |
| 1037 | bool pass_through_critical_region = false; |
| 1038 | size_t thread_task_num = g_execution_profile[thread].size(); |
| 1039 | for( size_t task_index = 0; task_index < thread_task_num; ++task_index ) { |
| 1040 | const task_marker_t& executed_task = g_execution_profile[thread][task_index]; |
| 1041 | |
| 1042 | if( pass_through_critical_region ) { |
| 1043 | ASSERT( executed_task < outer_critical_task, |
| 1044 | "Thread did not process all the critical work at once." ); |
| 1045 | } else if( isolated_regular_task <= executed_task && |
| 1046 | executed_task <= bypassed_critical_task) { |
| 1047 | started_critical_region = true; |
| 1048 | if( isolated_regular_task < executed_task ) |
| 1049 | ++critical_task_count; |
| 1050 | if( bypassed_critical_task == executed_task ) { |
| 1051 | size_t expected_bypass_task_min_index = |
| 1052 | /* number of regular task before critical region */1 + |
| 1053 | /* number of outermost critical tasks before isolated region */ 1 + |
| 1054 | g_isolated_regular_task_num; |
| 1055 | size_t expected_bypass_task_max_index = expected_bypass_task_min_index + |
| 1056 | /* number of critical tasks inside isolated region */ 1; |
| 1057 | ASSERT( expected_bypass_task_min_index <= task_index && |
| 1058 | task_index <= expected_bypass_task_max_index, |
| 1059 | "Bypassed critical task has been executed in wrong order" ); |
| 1060 | } |
| 1061 | } else if( started_critical_region ) { |
| 1062 | pass_through_critical_region = true; |
| 1063 | started_critical_region = false; |
| 1064 | } |
| 1065 | |
| 1066 | if( thread_task_num - 1 == task_index ) |
| 1067 | continue; // no transition check for the last executed task |
| 1068 | const task_marker_t& next_task = g_execution_profile[thread][task_index + 1]; |
| 1069 | std::pair<states_it, states_it> range = |
| 1070 | allowed_transitions.equal_range( executed_task ); |
| 1071 | bool is_choosen_task_allowed = false; |
| 1072 | for (states_it it = range.first; it != range.second; ++it) { |
| 1073 | is_choosen_task_allowed |= next_task == it->second; |
| 1074 | } |
| 1075 | ASSERT( is_choosen_task_allowed, "Thread chose incorrect task for execution." ); |
| 1076 | } |
| 1077 | } |
| 1078 | ASSERT( critical_task_count == 2 * num_threads + 2, "Wrong number of critical tasks" ); |
| 1079 | ASSERT( g_bypassed_critical_task_stage == executed, "Was bypassed critical task executed?" ); |
| 1080 | |
| 1081 | // test deinitialization |
| 1082 | g_execution_profile.clear(); |
| 1083 | g_arena.terminate(); |
| 1084 | } |
| 1085 | } |
| 1086 | } |
| 1087 | |
| 1088 | struct TaskTypeExecutionMarker : public task { |
| 1089 | TaskTypeExecutionMarker( task_marker_t mark ) : my_mark( mark ) {} |
| 1090 | task* execute() __TBB_override { |
| 1091 | g_execution_profile[tbb::this_task_arena::current_thread_index()].push_back( my_mark ); |
| 1092 | return NULL; |
| 1093 | } |
| 1094 | private: |
| 1095 | task_marker_t my_mark; |
| 1096 | }; |
| 1097 | |
| 1098 | struct RegularTaskMarkChecker { |
| 1099 | bool operator()(const task_marker_t& m) { return regular_task == m; } |
| 1100 | }; |
| 1101 | |
| 1102 | void TestSchedulerTaskSelectionWhenEnqueue() { |
| 1103 | REMARK( "\tPreferring critical tasks among enqueued\n" ); |
| 1104 | g_execution_profile.clear(); |
| 1105 | // creating two profiles because of enforced concurrency |
| 1106 | g_execution_profile.resize(2); |
| 1107 | g_root_task = NULL; |
| 1108 | unsigned task_num = 99; |
| 1109 | unsigned num_critical_tasks = 1; |
| 1110 | g_arena.initialize( /*num_threads=*/1, /*reserved_for_masters=*/0 ); |
| 1111 | g_arena.enqueue( |
| 1112 | WorkCreator<TaskTypeExecutionMarker, task::enqueue>( |
| 1113 | g_root_task, task_num, num_critical_tasks) |
| 1114 | ); |
| 1115 | WorkAwaiter awaiter(g_root_task); awaiter(); // waiting outside arena |
| 1116 | g_arena.terminate(); |
| 1117 | |
| 1118 | unsigned idx = !g_execution_profile[1].empty(); |
| 1119 | ASSERT( g_execution_profile[!idx].empty(), "" ); |
| 1120 | |
| 1121 | ASSERT( g_execution_profile[idx].size() == task_num + num_critical_tasks, |
| 1122 | "Incorrect number of tasks executed" ); |
| 1123 | ASSERT( g_execution_profile[idx][0] == outer_critical_task, |
| 1124 | "Critical task was executed in wrong order." ); |
| 1125 | bool all_regular = true; |
| 1126 | for( std::vector<task_marker_t>::const_iterator it = g_execution_profile[idx].begin() + 1; |
| 1127 | it != g_execution_profile[idx].end(); ++it ) |
| 1128 | all_regular &= regular_task == *it; |
| 1129 | ASSERT( all_regular, "Critical task was executed in wrong order." ); |
| 1130 | } |
| 1131 | |
| 1132 | enum ways_to_cancel_t { |
| 1133 | by_explicit_call = 0, |
| 1134 | by_exception, |
| 1135 | no_cancellation |
| 1136 | }; |
| 1137 | |
| 1138 | tbb::atomic<size_t> g_num_executed_from_cancelled_context; |
| 1139 | tbb::atomic<size_t> g_num_executed_from_working_context; |
| 1140 | int g_cancelling_task_id = -1; |
| 1141 | |
| 1142 | #if _MSC_VER && !__INTEL_COMPILER |
| 1143 | #pragma warning (push) |
| 1144 | #pragma warning (disable: 4127) /* suppress conditional expression is constant */ |
| 1145 | #endif |
| 1146 | |
| 1147 | template<bool cancelled_group> |
| 1148 | struct ATask : public task { |
| 1149 | ATask( task_marker_t /*mark*/ ) : my_cancellation_method( no_cancellation ) {} |
| 1150 | ATask( ways_to_cancel_t cancellation_method ) : my_cancellation_method( cancellation_method ) {} |
| 1151 | task* execute() __TBB_override { |
| 1152 | while( ! g_is_critical_task_submitted ) __TBB_Yield(); |
| 1153 | // scheduler should take critical task as the next task for execution. |
| 1154 | bypassed_critical_task_stage_t previous_critical_task_stage = |
| 1155 | g_bypassed_critical_task_stage.compare_and_swap(bypassed, not_bypassed); |
| 1156 | while( |
| 1157 | cancelled_group // Only tasks from cancelled group wait |
| 1158 | && !this->is_cancelled() // for their group to be cancelled |
| 1159 | && !tbb::internal::is_critical(*this) // allowing thread that took critical task |
| 1160 | && bypassed == previous_critical_task_stage // to proceed and cancel the whole group. |
| 1161 | ) __TBB_Yield(); |
| 1162 | if( cancelled_group ) |
| 1163 | ++g_num_executed_from_cancelled_context; |
| 1164 | else |
| 1165 | ++g_num_executed_from_working_context; |
| 1166 | switch( my_cancellation_method ) { |
| 1167 | case by_explicit_call: |
| 1168 | g_cancelling_task_id = int(g_num_executed_from_cancelled_context); |
| 1169 | self().cancel_group_execution(); |
| 1170 | break; |
| 1171 | case by_exception: |
| 1172 | g_cancelling_task_id = int(g_num_executed_from_cancelled_context); |
| 1173 | throw std::runtime_error("Exception data" ); |
| 1174 | break; |
| 1175 | case no_cancellation: break; |
| 1176 | default: |
| 1177 | ASSERT( false, "Should not be here!" ); |
| 1178 | break; |
| 1179 | } |
| 1180 | return NULL; |
| 1181 | } |
| 1182 | private: |
| 1183 | ways_to_cancel_t my_cancellation_method; |
| 1184 | }; |
| 1185 | |
| 1186 | #if _MSC_VER && !__INTEL_COMPILER |
| 1187 | #pragma warning (pop) |
| 1188 | #endif |
| 1189 | |
| 1190 | template<void(*submit_task)(task&)> |
| 1191 | struct SubmitTaskFunctor { |
| 1192 | SubmitTaskFunctor( task& t ) : my_task( t ) {} |
| 1193 | void operator()() const { |
| 1194 | submit_task(my_task); |
| 1195 | } |
| 1196 | private: |
| 1197 | task& my_task; |
| 1198 | }; |
| 1199 | |
| 1200 | void TestCancellation(bool cancel_by_exception) { |
| 1201 | g_is_critical_task_submitted = false; |
| 1202 | g_bypassed_critical_task_stage = not_bypassed; |
| 1203 | tbb::task_group_context context_to_leave_working; |
| 1204 | tbb::task_group_context context_to_cancel; |
| 1205 | task* root_task_of_to_be_cancelled_context = NULL; |
| 1206 | task* root_task_of_working_to_completion_context = NULL; |
| 1207 | size_t task_num = 64; |
| 1208 | size_t task_num_for_cancelled_context = 2 * MaxThread; |
| 1209 | g_num_executed_from_cancelled_context = g_num_executed_from_working_context = 0; |
| 1210 | g_cancelling_task_id = -1; |
| 1211 | g_arena.initialize( MaxThread ); // leaving one slot to be occupied by master to submit the work |
| 1212 | g_arena.execute( |
| 1213 | WorkCreator<ATask</*cancelled_group=*/true>, task::spawn> |
| 1214 | (root_task_of_to_be_cancelled_context, task_num_for_cancelled_context, |
| 1215 | /*num_critical_tasks=*/0, &context_to_cancel) |
| 1216 | ); |
| 1217 | g_arena.execute( |
| 1218 | WorkCreator<ATask</*cancelled_group=*/false>, task::spawn> |
| 1219 | (root_task_of_working_to_completion_context, task_num, /*num_critical_tasks=*/1, |
| 1220 | &context_to_leave_working) |
| 1221 | ); |
| 1222 | ways_to_cancel_t cancellation_method = ways_to_cancel_t( cancel_by_exception ); |
| 1223 | task& terminating_task = *new( task::allocate_additional_child_of(*root_task_of_to_be_cancelled_context) ) |
| 1224 | ATask</*cancelled_group=*/true>( cancellation_method ); |
| 1225 | tbb::internal::make_critical( terminating_task ); // stop the work as soon as possible! |
| 1226 | g_arena.enqueue( SubmitTaskFunctor<task::enqueue>(terminating_task), |
| 1227 | tbb::priority_t(tbb::internal::priority_critical) ); |
| 1228 | g_is_critical_task_submitted = true; |
| 1229 | try { |
| 1230 | g_arena.execute( WorkAwaiter(root_task_of_to_be_cancelled_context) ); |
| 1231 | } catch( const std::runtime_error& e ) { |
| 1232 | ASSERT( cancel_by_exception, "Exception was not expected!" ); |
| 1233 | ASSERT( std::string(e.what()) == "Exception data" , "Unexpected exception data!" ); |
| 1234 | } catch( const tbb::captured_exception& e ) { |
| 1235 | ASSERT( cancel_by_exception, "Exception was not expected!" ); |
| 1236 | ASSERT( std::string(e.what()) == "Exception data" , "Unexpected exception data!" ); |
| 1237 | } catch( ... ) { |
| 1238 | ASSERT( false, "Failed to catch specific exception" ); |
| 1239 | } |
| 1240 | g_arena.execute( WorkAwaiter(root_task_of_working_to_completion_context) ); |
| 1241 | g_arena.terminate(); |
| 1242 | |
| 1243 | if( !cancel_by_exception ) { |
| 1244 | ASSERT( context_to_cancel.is_group_execution_cancelled(), "Execution must be cancelled" ); |
| 1245 | } |
| 1246 | ASSERT( !context_to_leave_working.is_group_execution_cancelled(), |
| 1247 | "Execution must NOT be cancelled" ); |
| 1248 | |
| 1249 | ASSERT( g_num_executed_from_working_context == task_num + /*one critical*/1, |
| 1250 | "Incorrect number of tasks executed!" ); |
| 1251 | ASSERT( g_num_executed_from_cancelled_context < task_num_for_cancelled_context, |
| 1252 | "Number of executed tasks from the cancelled context should be less than submitted!" ); |
| 1253 | ASSERT( 0 < g_cancelling_task_id && g_cancelling_task_id < MaxThread + 1, |
| 1254 | "Critical task was executed in wrong order." ); |
| 1255 | } |
| 1256 | |
| 1257 | void TestCancellationSupport(bool cancel_by_exception) { |
| 1258 | const char* test_type[] = { "by explicit call to cancel" , "by throwing an exception" }; |
| 1259 | REMARK( "\tCancellation support %s\n" , test_type[!!cancel_by_exception] ); |
| 1260 | TestCancellation( cancel_by_exception ); |
| 1261 | } |
| 1262 | |
| 1263 | namespace NestedArenaCase { |
| 1264 | |
| 1265 | static const size_t g_num_critical_tasks = 10; |
| 1266 | static const size_t g_num_critical_nested = 5; |
| 1267 | |
| 1268 | struct CriticalTask : public task { |
| 1269 | CriticalTask(task_marker_t /*mark*/) {} |
| 1270 | task* execute() __TBB_override { |
| 1271 | ++g_num_executed_from_working_context; |
| 1272 | task* nested_root = NULL; |
| 1273 | if( !g_is_critical_task_submitted ) { |
| 1274 | g_is_critical_task_submitted = true; |
| 1275 | g_arena.execute( |
| 1276 | WorkCreator<CriticalTask, task::spawn>(nested_root, /*num_tasks=*/size_t(0), |
| 1277 | g_num_critical_nested) ); |
| 1278 | g_arena.execute( WorkAwaiter(nested_root) ); |
| 1279 | } |
| 1280 | return NULL; |
| 1281 | } |
| 1282 | }; |
| 1283 | |
| 1284 | void TestInNestedArena(tbb::task_arena& outer_arena) { |
| 1285 | g_root_task = NULL; |
| 1286 | g_is_critical_task_submitted = false; |
| 1287 | g_num_executed_from_working_context = 0; |
| 1288 | g_arena.initialize( 1 ); |
| 1289 | outer_arena.execute( |
| 1290 | WorkCreator<CriticalTask, task::spawn>( |
| 1291 | g_root_task, /*num_tasks=*/size_t(0), g_num_critical_tasks) ); |
| 1292 | outer_arena.execute( WorkAwaiter(g_root_task) ); |
| 1293 | ASSERT( g_num_executed_from_working_context == g_num_critical_tasks + g_num_critical_nested, |
| 1294 | "Mismatch in number of critical tasks executed in nested and outer arenas." ); |
| 1295 | g_arena.terminate(); |
| 1296 | } |
| 1297 | |
| 1298 | void test() { |
| 1299 | REMARK( "\tWork in nested arenas\n" ); |
| 1300 | TestInNestedArena( g_arena ); |
| 1301 | |
| 1302 | tbb::task_arena a( 1 ); |
| 1303 | TestInNestedArena( a ); |
| 1304 | } |
| 1305 | } // namespace NestedArenaCase |
| 1306 | |
| 1307 | void test() { |
| 1308 | REMARK("Testing support for critical tasks\n" ); |
| 1309 | TestSchedulerTaskSelectionWhenSpawn(); |
| 1310 | TestSchedulerTaskSelectionWhenEnqueue(); |
| 1311 | TestCancellationSupport(/*cancel_by_exception=*/false); |
| 1312 | TestCancellationSupport(/*cancel_by_exception=*/true); |
| 1313 | NestedArenaCase::test(); |
| 1314 | } |
| 1315 | } // namespace CriticalTaskSupport |
| 1316 | #endif /* __TBB_PREVIEW_CRITICAL_TASKS */ |
| 1317 | |
| 1318 | int TestMain () { |
| 1319 | #if TBB_USE_EXCEPTIONS |
| 1320 | TestUnconstructibleTask<1>(); |
| 1321 | TestUnconstructibleTask<10000>(); |
| 1322 | #endif |
| 1323 | TestAlignment(); |
| 1324 | TestNoteAffinityContext(); |
| 1325 | TestDispatchLoopResponsiveness(); |
| 1326 | TestWaitDiscriminativenessWithoutStealing(); |
| 1327 | TestWaitDiscriminativenessWithStealing(); |
| 1328 | for( int p=MinThread; p<=MaxThread; ++p ) { |
| 1329 | TestSpawnChildren( p ); |
| 1330 | TestSpawnRootList( p ); |
| 1331 | TestSafeContinuation( p ); |
| 1332 | TestLeftRecursion( p ); |
| 1333 | TestDag( p ); |
| 1334 | TestAffinity( p ); |
| 1335 | TestUserThread( p ); |
| 1336 | TestStealLimit( p ); |
| 1337 | TestRelaxedOwnership( p ); |
| 1338 | TestMastersIsolation( p ); |
| 1339 | } |
| 1340 | TestWaitableTask(); |
| 1341 | #if __TBB_PREVIEW_CRITICAL_TASKS |
| 1342 | CriticalTaskSupport::test(); |
| 1343 | #endif |
| 1344 | return Harness::Done; |
| 1345 | } |
| 1346 | |