| Graphics Processing Unit (GPU) performance is increasing faster than central processing unit (CPU) performance. This growth is driven by performance improvements that can be divided into the following three categories: algorithmic improvements, architectural improvements, and circuit-level improvements. In this dissertation I present techniques that improve the rendering performance of graphics hardware measured in speed, power consumption or image quality in each of these three areas.; At the algorithmic level, I introduce a method for using graphics hardware to rapidly and efficiently generate summed-area tables, which are data structures that hold pre-computed two-dimensional integrals of subsets of a given image, and present several novel rendering techniques that take advantage of summed-area tables to produce dynamic, high-quality images at interactive frame rates. These techniques improve the visual quality of images rendered on current commodity GPUs without requiring modifications to the underlying hardware or architecture.; At the architectural level, I propose modifications to the architecture of current GPUs that add conditional streaming capabilities. I describe a novel GPU-based ray-tracing algorithm that takes advantage of conditional output streams to reduce the memory bandwidth requirements by over an order of magnitude times when compared to previous techniques.; At the circuit level, I propose a compute-on-demand paradigm for the design of high-speed and energy-efficient graphics components. The goal of the compute-on-demand paradigm is to only perform computation at the bit-level when needed. The compute-on-demand paradigm exploits the data-dependent nature of computation, and thereby obtains speed and energy improvements by optimizing designs for the common case. This approach is illustrated with the design of a high-speed Z-comparator that is implemented using asynchronous logic. Asynchronous or "clockless" circuits were chosen for my implementations since they allow for data-dependent completion times and reduced power consumption by disabling inactive components. The resulting circuit-level implementation runs over 1.5 times faster while on dissipating 25% the energy of a comparable synchronous comparator for the average case.; Also at the circuit-level, I introduce a novel implementation of counterflow pipelining, which allows two streams of data to flow in opposite directions within the same pipeline without the need for complex arbitration. The advantages of this implementation are demonstrated by the design of a high-speed asynchronous Booth multiplier. While both the comparator and the multiplier are useful components of a graphics pipeline, the objective of this work was to propose the new design paradigm as a promising alternative to current graphics hardware design practices. |
