H-SIMD machine : configurable parallel computing for data-intensive applications
Department of Electrical and Computer Engineering
Doctor of Philosophy
Gerbessiotis, Alexandros V.
FPGA Data parallel
This dissertation presents a hierarchical single-instruction multiple-data (H-SLMD) configurable computing architecture to facilitate the efficient execution of data-intensive applications on field-programmable gate arrays (FPGAs). H-SIMD targets data-intensive applications for FPGA-based system designs. The H-SIMD machine is associated with a hierarchical instruction set architecture (HISA) which is developed for each application. The main objectives of this work are to facilitate ease of program development and high performance through ease of scheduling operations and overlapping communications with computations.
The H-SIMD machine is composed of the host, FPGA and nano-processor layers. They execute host SIMD instructions (HSIs), FPGA SIMD instructions (FSIs) and nano-processor instructions (NPLs), respectively. A distinction between communication and computation instructions is intended for all the HISA layers. The H-SIMD machine also employs a memory switching scheme to bridge the omnipresent large bandwidth gaps in configurable systems. To showcase the proposed high-performance approach, the conditions to fully overlap communications with computations are investigated for important applications. The building blocks in the H-SLMD machine, such as high-performance and area-efficient register files, are presented in detail. The H-SLMD machine hierarchy is implemented on a host Dell workstation and the Annapolis Wildstar II FPGA board. Significant speedups have been achieved for matrix multiplication (MM), 2-dimensional discrete cosine transform (2D DCT) and 2-dimensional fast Fourier transform (2D FFT) which are used widely in science and engineering.
In another FPGA-based programming paradigm, a high-level language (here ANSI C) can be used to program the FPGAs in a mode similar to that of the H-SIMD machine in terms of trying to minimize the effect of overheads. More specifically, a multi-threaded overlapping scheme is proposed to reduce as much as possible, or even completely hide, runtime FPGA reconfiguration overheads. Nevertheless, although the HLL-enabled reconfigurable machine allows software developers to customize FPGA functions easily, special architecture techniques are needed to achieve high-performance without significant penalty on area and clock frequency. Two important high-performance applications, matrix multiplication and image edge detection, are tested on the SRC-6 reconfigurable machine. The implemented algorithms are able to exploit the available data parallelism with independent functional units and application-specific cache support. Relevant performance and design tradeoffs are analyzed.
njit-etd2006-089 (102 pages ~ 4,789 KB pdf)
Please complete this Feedback Form to inform us about your experience using this website. It will assist us in better serving your information needs in the future. Thank You!
Created January 23, 2007