Intel Threading Building Blocks

Two Mouths: Feeding Two from the Same Task in a Pipeline

Imagine you have a couple of pipelines and you want to merge results from different pipelines in a task.

Additionally, suppose the output of at least one task feeds more than a single task. Normally, a task can feed only one other task in a pipeline. The solution is to use a dummy task inside a task to link in the data for the task that feeds two dependents.

Figure 11-16 shows a diagram for a desired pipeline that contains a split such that the output of one stage is used by two other stages. This is a pretty typical-looking set of parallel pipelines such as you might find in any variety of signal-processing or multimedia applications. The temptation is to give up after the third step and do the merges back into the main program, but that would give up the parallelism available during the merges. Likewise, we do not want to execute the middle pipeline (ending with FILTER B) twice.

Figure 11-16. A pipeline with lots of parallelism, but a twist at the end

To focus on the problem we’re solving, we need only look at the last two phases, as shown in Figure 11-17. The actual solution is to create a DUMMY task solely for the sake of synchronization, as shown in Figure 11-18. Example 11-38 shows the code of the worker tasks and dummy task, and Example 11-39 shows the main program that spawns the three worker tasks. Keep in mind that because ...