Buffer Size Reduction through Control-Flow Decomposition

Abstract

Software synthesis from a data-flow model has been a very promising technique, especially for multimedia applications with contradicting requirements of high design complexity and fast time-to-market. In a dataflow model, buffer size is pessimistically determined through static analysis, thus results in large memory overhead even with optimization techniques such as buffer sharing and scheduling. So, reducing buffer size is one of the key issues of data-flow models. In this work, we propose a novel software synthesis technique to reduce buffer size through control-flow decomposition. We first traverse the control-flow within each actor of a data-flow graph and decompose it into a set of multiple execution paths. Then we transform the actor such that only one of the paths is executed at one invocation of the actor. The new actor may have to be invoked many times to complete the behavior of the original actor. By proper decomposition, we can make the new actor consume/produce much smaller amount of input/output data for each invocation, thereby reducing the input/output buffer size drastically. We automate the process of transformation and show the efficiency of the proposed approach through experiments with image/video multimedia applications.