Parallelization of While-Loops in Nested Loop Programs for Real-time Multiprocessor Systems

Many applications with stream processing behavior contain one or more loops with an unknown number of iterations. These loops have to be parallelized in order to utilize the maximum capacity of an embedded multiprocessor platform and thus increase the total throughput. This thesis presents a method to automatically extract a parallel task graph based on function level parallelism from a sequential nested loop program with while-loops. The task graph does not rely on a fixed order of the execution of the tasks. The notion of a single assignment section is introduced, which is used in the parallelization approach. A single assignment section is less restrictive when programming the sequential nested loop programs than when demanding single assignment code. Instead of the complete program, only in a single assignment section single assignment must hold. This makes the parallelization of all while-loops possible. Communication between the extracted tasks is done via circular buffers. In a circular buffer reading and writing is only allowed inside a window. For each circular buffer a choice is made between a buffer with sliding or overlapping windows depending on the costs. In a circular buffer with sliding windows the read and write windows can not overlap while this is allowed in a circular buffer with overlapping windows. An analysis on the overhead of the circular buffer is performed and a comparison between circular buffers with sliding and overlapping windows is made. Synchronization for the used circular buffers is inserted into the task graph to ensure the same functional behavior as the sequential nested loop program. Sufficient buffer capacities for a deadlock free execution of the parallel task graph are determined using data flow analysis. For each while-loop a cyclo-static data flow model is derived based on the models for circular buffers with sliding and overlapping windows. The model for overlapping windows is also presented in this thesis. The models for the while-loops are combined such that buffer capacities can be determined for the complete nested loop program. An extension to cyclo-static data flow models is proposed such that the combining of the while-loops can be modeled. A DVB-T radio receiver where the user can switch channels after an undetermined amount of time illustrates the parallelization approach. By means of execution traces it is shown that a higher throughput can be achieved when the presented parallelization approach is used. An extension to circular buffers with sliding or overlapping windows is presented