Year: 2004
Type: dissertation
Abstract: Digital devices and their underlying DSP technology are widespread in modern society.Consequently, optimizing DSP applications is necessary to satisfy tight power, space, and bandwidth constraints.As code length and complexity increases for these applications, the value of time-intensive, manual expert analysis has decreased.In its place, robust, compiler-generated optimizations have become necessary.We target linear state-space sections of applications to analyze and optimize, and use the programming language StreamIt to implement and test our ideas.StreamIt enables its users to create filters and connect them in.asimple, structured manner.A linear state-space filter may have a set of state variables that it uses from execution to execution.Such a filter has the property that its outputs are a linear combination of its inputs and states, and the states are updated by a linear combination of the inputs and states.Examples of such filters that use states include IIR filters and linear difference equations.We automate the following steps in the StreamIt compiler.We extract the representation for each filter that is linear state-space.We combine representations of filters to form a single linear state-space filter.We remove unnecessary states and reduce the number of computations needed per filter execution.Lastly, we convert the optimized filter back to StreamIt code.We have compiled a number of StreamIt benchmark applications using our analysis and optimizations.Our data indicates state-space replacement works as well as linear replacement (a filter is linear if its outputs are an affine combination of its inputs) for stateless filters, and improves performance by as much as 66% for combinations of filters with state.
Author: Sitij Agrawal
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2126895182