Copyrights to these papers may be held by the publishers. The download files are preprints. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.
Peter A. Milder (PhD. thesis, Electrical and Computer Engineering, Carnegie Mellon University, 2010)
A Mathematical Approach for Compiling and Optimizing Hardware Implementations of DSP Transforms
Preprint (4.4 MB)
Linear signal transforms (such as the discrete Fourier transform) are frequently used in digital signal processing and related fields. Algorithms for computing linear transforms are well-understood and typically have a high degree of regularity and parallelism, making them well-suited for implementation as sequential datapaths on field-programmable gate arrays or application-specific integrated circuits. Nonetheless, transforms are difficult to implement due to the large number of algorithmic options available and ways that algorithms can be mapped to sequential datapaths. Further, the best choices depend heavily on the resource budget and performance goals of the target application. Thus, it is difficult for a designer to determine which set of options will best meet a given set of requirements. This thesis proposes the Spiral hardware generation framework, a hardware compilation and optimization tool based on a mathematical formula language. A formula written in this language specifies a particular transform algorithm executed using a particular sequential datapath. This language allows high-level representation of sequential hardware structures that reuse datapath elements multiple times in the computation of a transform. The language accomplishes this by providing a formal connection between structure in the algorithm and sequential reuse in the datapath. This proposed language drives a hardware compilation system that takes as input a problem specification with directives that define characteristics of the desired datapath; the system then automatically generates an algorithm, maps the algorithm to a datapath, and outputs synthesizable register transfer level Verilog. This thesis evaluates the generated designs when synthesized for field-programmable gate array or application-specific integrated circuit. Its evaluations consider designs across multiple transforms, datatypes, and design goals, and its results show that Spiral is able to automatically providea wide tradeoff between cost (e.g., area, power) and performance. This tradeoff space compares well with existing benchmarks, but allows the designer much more flexibility to find the design best suited to his or her needs.Keywords: IP cores for FPGA/ASIC, Discrete/fast Fourier transform, Streaming permutations, OFDM, Discrete/fast cosine transforms