VLSI design and synthesis for a class of two-dimensional problems
This thesis presents a methodology to synthesize optimal VLSI designs by exploiting several global/local communication paradigms, which lead to solutions for a class of two dimensional problems (D{sup 2}), including signal/image processing, computer vision, and linear algebra. A two-dimensional problem (i.e. a D{sup 2} problem) is defined as one in which both input and output data can be expressed by a two dimensional array of size N x N. A class of D{sup 2} problems are categorized by the data dependency patterns of their algorithms. The class of problems was further divided into local, tree, butterfly (or FFT-type) as well as broadcast-type problems. The procedure to synthesize a D{sup 2} algorithm are reported. First, efficient generic algorithms are designed based on those communication paradigms. Unfortunately, most of these generic algorithms, in this dissertation are not shift invariant, which is the fundamental condition for mapping parallel algorithms to systolic architectures based on the systolic synthesis methodology. Then, this deficiency was remedied by adding virtual nodes and edges. This artifice induces new problems, such as requiring redundant hardware optimization. Next, the new problems were handled by applying graph optimization algorithms and logic minimization techniques. By applying these techniques, we can improve the PE utilization rate from 50% to 100% for the FFT algorithms under certain conditions. Also, we can reduce half of the memory size for the FFT-type algorithms which input and output data from the raster scan I/O device. The idea of combining techniques of the synthesis methodology with the improvement strategy based on the algebraic properties of the computation node are presented. Many tree and combination of local and tree algorithms can be synthesized to achieve the lower bound in terms of processor number and time complexity after applying the combining techniques.
- Research Organization:
- North Carolina State Univ., Raleigh, NC (USA)
- OSTI ID:
- 6596548
- Resource Relation:
- Other Information: Thesis (Ph. D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Testing and reconfiguration techniques for VLSI processor arrays
Real-time, large scale optimization of water network systems using a subdomain approach.
Related Subjects
COMPUTER ARCHITECTURE
DESIGN
ALGORITHMS
ARRAY PROCESSORS
COMMUNICATIONS
COMPUTER GRAPHICS
DECISION TREE ANALYSIS
INFORMATION SYSTEMS
OPTIMIZATION
PARALLEL PROCESSING
TIMING PROPERTIES
TWO-DIMENSIONAL CALCULATIONS
MATHEMATICAL LOGIC
PROGRAMMING
990200* - Mathematics & Computers
990300 - Information Handling