CAREER: Automated Software Understanding for Retargeting Embedded Image Processing Software for Data Parallel Execution

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

This research applies software understanding and reengineering techniques to the problem of retargeting embedded image processing software to emerging parallel architectures that are capable of efficiently exploiting the inherent data parallelism in these applications. Image and video processing software is found in a wide range of embedded devices. There is a significant demand for low-power, compact imaging devices, e.g., for smart cameras, wearable computers, and autonomous vehicle vision systems. Migrating image and video processing programs to new or extended hardware platforms that support data parallel execution will greatly enhance the image processing capabilities and sophistication of embedded devices. This research seeks to automatically expose the inherent data parallelism in these applications to fuel a new generation of high performance, high efficiency embedded processing systems. Software understanding techniques are being used to provide a deep model of what computation is being performed so that a broader set of parallelization transformations can be accomplished and more opportunities for parallelization can be identified. This work also integrates software reengineering more fully into the education of computer engineers and scientists, and help students learn parallel processing concepts. It will support interactive application of parallelizing transformations and allow students to sharpen their discovery and problem solving skills