Taking Spiral Beyond Digital Signal Processing

Tuesday Mar. 19, 2013
Hamerschlag Hall D-210
4:30-5.30 pm

Richard Veras (ECE, CMU)

Abstract

SPIRAL is a framework for automatically generating and tuning Digital Signal Pro-cessing (DSP) algorithms from formal descriptions to target microarchitectures. The power of SPIRAL resides in the layering of several Domain Specific Languages (DSL) which capture the necessary information to perform various classes of transformations on the DSP operation of interest. The first layer captures the data flow structure of the operation where a variety of transformations enable vectorization and thread level parallelization. This is then translated to an iterative space language which enables loop-level optimizations. After this, the operation is converted into a machine independent intermediate language which is further optimized and mapped to a target language like C. My current work is to expand this SPIRAL framework to include other domains including Stencil Operations, Dense Linear Algebra and Constraint Satisfaction Solvers. In my talk I will discuss how I have used the various layers of SPIRAL to capture non-DSP operations.

Bio

Richard Veras is a second year PhD Student in the Electrical and Computer Engineering Department at Carnegie Mellon University. He is a member of the SPIRAL group and is being advised by Dr. Franz Franchetti. He earned his bachelors in Computer Science and Mathematics at the University of Texas at Austin and his research interests are in Automatic Program Generation, and Scientific and High Performance Computing.

Back to the seminar page