Multidimensional Synchronous Dataflow for Modeling, Simulating, and Synthesizing Image, Video, and Spatial Array Processing Systems
Dr. Praveen K. Murthy
Department of Electrical Engineering and Computer Sciences
University of California at Berkeley
Berkeley, CA 94720-1772
http://ptolemy.eecs.berkeley. edu/~murthy
Wednesday, April 2nd, 5:00 PM, ENS 302
Many dataflow models use one-dimensional streams. While these models are ideally suited for one-dimensional digital signal processing (DSP) systems, much of the data parallelism can be lost if we model multidimensional DSP systems using models that have one-dimensional streams. Expressing multidimensional systems in this way is often awkward and counter-intuitive.
Multidimensional synchronous dataflow (MDSDF) is an extension of synchronous dataflow (SDF) to multiple dimensions. In this model, streams are generalized to arrays in multiple dimensions. We will show that MDSDF is well-suited for not only MD DSP systems, but also for expressing certain control flow structures concisely, in a way that SDF cannot.
We generalize the MDSDF model to model MD systems that use non-rectangular sampling schemes. The key challenge in designing this model is preserving static scheduling, which is accomplished by solving a set of augmented balance equations. We demonstrate the MDSDF model by capturing the parallelism in the conversion of 4/3 aspect ratio interlaced video to 16/9 aspect ratio interlaced video using multiple stages.
A list of digital signal processing seminars is available at from the ECE department Web pages under "Seminars". The Web address for the digital signal processing seminars is http://www.ece.utexas.edu/~bevans/dsp_seminars.html