Visible to Intel only — GUID: GUID-87BF02EA-A960-4F4B-849D-3CEC9F66421A
Package Contents
Parallelizing Simple Loops
Parallelizing Complex Loops
Parallelizing Data Flow and Dependence Graphs
Work Isolation
Exceptions and Cancellation
Containers
Mutual Exclusion
Timing
Memory Allocation
The Task Scheduler
Design Patterns
Migrating from Threading Building Blocks (TBB)
Constrained APIs
Invoke a Callable Object
Appendix A Costs of Time Slicing
Appendix B Mixing With Other Threading Packages
References
parallel_for_each Body semantics and requirements
parallel_sort ranges interface extension
TBB_malloc_replacement_log Function
Parallel Reduction for rvalues
Type-specified message keys for join_node
Scalable Memory Pools
Helper Functions for Expressing Graphs
concurrent_lru_cache
task_group extensions
The customizing mutex type for concurrent_hash_map
Visible to Intel only — GUID: GUID-87BF02EA-A960-4F4B-849D-3CEC9F66421A
Flow Graph Basics: Mapping Nodes to Tasks
The following figure shows the timeline for one possible execution of the two node graph example in the previous section. The bodies of n and m will be referred to as λn and λm, respectively. The three calls to try_put spawn three tasks; each one applies the lambda expression, λn, on one of the three input messages. Because n has unlimited concurrency, these tasks can execute concurrently if there are enough threads available. The call to g.wait_for_all() blocks until there are no tasks executing in the graph. As with other wait_for_all functions in oneTBB, the thread that calls wait_for_all is not spinning idly during this time, but instead can join in executing other tasks from the work pool.
Execution Timeline of a Two Node Graph
As each task from n finishes, it puts its output to m, since m is a successor of n. Unlike node n, m has been constructed with a concurrency limit of 1 and therefore does not spawn all tasks immediately. Instead, it sequentially spawns tasks to execute its body, λm, on the messages in the order that they arrive. When all tasks are complete, the call to wait_for_all returns.
NOTE:
All execution in the flow graph happens asynchronously. The calls to try_put return control to the calling thread quickly, after either immediately spawning a task or buffering the message being passed. Likewise, the body tasks execute the lambda expressions and then put the result to any successor nodes. Only the call to wait_for_all blocks, as it should, and even in this case the calling thread may be used to execute tasks from the oneTBB work pool while it is waiting.
The above timeline shows the sequence when there are enough threads to execute all of the tasks that can be executed in parallel. If there are fewer threads, some spawned tasks will need to wait until a thread is available to execute them.