Date of Award
Doctor of Philosophy (PhD)
Electrical & Computer Engineering
Vijayan K. Asari
Shirshak K. Dhali
The increasing capacity and capabilities of FPGA devices in recent years provide an attractive option for performance-hungry applications in the image and video processing domain. FPGA devices are often used as implementation platforms for image and video processing algorithms for real-time applications due to their programmable structure that can exploit inherent spatial and temporal parallelism. While performance and area remain as two main design criteria, power consumption has become an important design goal especially for mobile devices. Reduction in power consumption can be achieved by reducing the supply voltage, capacitances, clock frequency and switching activities in a circuit. Switching activities can be reduced by architectural optimization of the processing cores such as adders, multipliers, multiply and accumulators (MACS), etc. This dissertation research focuses on reducing the switching activities in digital circuits by considering data dependencies in bit level, word level and block level neighborhoods in a video frame.
The bit level data neighborhood dependency consideration for power reduction is illustrated in the design of pipelined array, Booth and log-based multipliers. For an array multiplier, operands of the multipliers are partitioned into higher and lower parts so that the probability of the higher order parts being zero or one increases. The gating technique for the pipelined approach deactivates part(s) of the multiplier when the above special values are detected. For the Booth multiplier, the partitioning and gating technique is integrated into the Booth recoding scheme. In addition, a delay correction strategy is developed for the Booth multiplier to reduce the switching activities of the sign extension part in the partial products. A novel architecture design for the computation of log and inverse-log functions for the reduction of power consumption in arithmetic circuits is also presented. This also utilizes the proposed partitioning and gating technique for further dynamic power reduction by reducing the switching activities.
The word level and block level data dependencies for reducing the dynamic power consumption are illustrated by presenting the design of a 2-D convolution architecture. Here the similarities of the neighboring pixels in window-based operations of image and video processing algorithms are considered for reduced switching activities. A partitioning and detection mechanism is developed to deactivate the parallel architecture for window-based operations if higher order parts of the pixel values are the same. A neighborhood dependent approach (NDA) is incorporated with different window buffering schemes. Consideration of the symmetry property in filter kernels is also applied with the NDA method for further reduction of switching activities.
The proposed design methodologies are implemented and evaluated in a FPGA environment. It is observed that the dynamic power consumption in FPGA-based circuit implementations is significantly reduced in bit level, data level and block level architectures when compared to state-of-the-art design techniques. A specific application for the design of a real-time video processing system incorporating the proposed design methodologies for low power consumption is also presented. An image enhancement application is considered and the proposed partitioning and gating, and NDA methods are utilized in the design of the enhancement system. Experimental results show that the proposed multi-level power aware methodology achieves considerable power reduction. Research work is progressing In utilizing the data dependencies in subsequent frames in a video stream for the reduction of circuit switching activities and thereby the dynamic power consumption.
Ngo, Hau T..
"Power-Aware Design Methodologies for FPGA-Based Implementation of Video Processing Systems"
(2007). Doctor of Philosophy (PhD), Dissertation, Electrical & Computer Engineering, Old Dominion University, DOI: 10.25777/j6kw-q685