Serialized multitasking code generation from dataflow specification

Abstract

This paper is concerned about multitasking embedded software development from the system specification to the final implementation including design space exploration(DSE). In the proposed framework, dataflow model is used for task specification. Multitasking software is generated for the performance evaluation of architecture candidates during the DSE process. Since the same code is also used for the final implementation, it is highly desirable to make it portable and efficient. In this paper, we propose a serialized multitasking code generation technique from dataflow specification to run the multitasking application without OS on any target processor. The code serialization also reduces runtime overhead of task switching as previous works have reported. By separating run-time scheduler generation from task code generation, various scheduling policies can be explored. Experiments with DiVX application confirm the viability of the proposed technique.