Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Computer Engineering


Gowdy, John N

Committee Member

Woodard , Damon L

Committee Member

Birchfield , Stanley T

Committee Member

Ligon III , Walter B


The move to more parallel computing architectures places more responsibility on the programmer to achieve greater performance. The programmer must now have a greater understanding of the underlying architecture and the inherent algorithmic parallelism. Using parallel computing architectures for exploiting algorithmic parallelism can be a complex task. This dissertation demonstrates various techniques for using parallel computing architectures to exploit algorithmic parallelism. Specifically, three pattern recognition (PR) approaches are examined for acceleration across multiple parallel computing architectures, namely field programmable gate arrays (FPGAs) and general purpose graphical processing units (GPGPUs).
Phase-only filter correlation for fingerprint identification was studied as the first PR approach. This approach's sensitivity to angular rotations, scaling, and missing data was surveyed. Additionally, a novel FPGA implementation of this algorithm was created using fixed point computations, deep pipelining, and four computation phases. Communication and computation were overlapped to efficiently process large fingerprint galleries. The FPGA implementation showed approximately a 47 times speedup over a central processing unit (CPU) implementation with negligible impact on precision.
For the second PR approach, a spiking neural network (SNN) algorithm for a character recognition application was examined. A novel FPGA implementation of the approach was developed incorporating a scalable modular SNN processing element (PE) to efficiently perform neural computations. The modular SNN PE incorporated streaming memory, fixed point computation, and deep pipelining. This design showed speedups of approximately 3.3 and 8.5 times over CPU implementations for 624 and 9,264 sized neural networks, respectively. Results indicate that the PE design could scale to process larger sized networks easily.
Finally for the third PR approach, cellular simultaneous recurrent networks (CSRNs) were investigated for GPGPU acceleration. Particularly, the applications of maze traversal and face recognition were studied. Novel GPGPU implementations were developed employing varying quantities of task-level, data-level, and instruction-level parallelism to achieve efficient runtime performance. Furthermore, the performance of the face recognition application was examined across a heterogeneous cluster of multi-core and GPGPU architectures. A combination of multi-core processors and GPGPUs achieved roughly a 996 times speedup over a single-core CPU implementation.
From examining these PR approaches for acceleration, this dissertation presents useful techniques and insight applicable to other algorithms to improve performance when designing a parallel implementation.