Experience with automatic mapping of sensor-based applications

An important goal of the Fx compiler project at Carnegie Mellon is to support the use of high level programming languages like High Performance Fortran for applications in image processing, signal processing and computer vision. Such applications typically require fast processing of streams of relatively small data sets. Fx introduced the notion of task parallelism within a data parallel language to support the simultaneous processing of multiple data sets in parallel or pipelined fashion. This introduced the problem of finding a good mapping for such applications and a set of automatic techniques were developed to find the best mappings for different performance goals. This paper reports on experience with automatic techniques for mapping pipelined applications. Results are presented from an image processing pipeline, a multibaseline stereo program and a narrowband tracking radar programs. We show that automatic mapping techniques are not perfect but can be used effectively if a variety of practical constraints ore given proper attention.