Efficient FPGA implementation and power modelling of image and signal processing IP cores

Abstract

Field Programmable Gate Arrays (FPGAs) are the technology of choice in a number ofimage and signal processing application areas such as consumer electronics, instrumentation, medical data processing and avionics due to their reasonable energy consumption, high performance, security, low design-turnaround time and reconfigurability. Low power FPGA devices are also emerging as competitive solutions for mobile and thermally constrained platforms. Most computationally intensive image and signal processing algorithms also consume a lot of power leading to a number of issues including reduced mobility, reliability concerns and increased design cost among others. Power dissipation has become one of the most important challenges, particularly for FPGAs. Addressing this problem requires optimisation and awareness at all levels in the design flow. The key achievements of the work presented in this thesis are summarised here. Behavioural level optimisation strategies have been used for implementing matrix product and inner product through the use of mathematical techniques such as Distributed Arithmetic (DA) and its variations including offset binary coding, sparse factorisation and novel vector level transformations. Applications to test the impact of these algorithmic and arithmetic transformations include the fast Hadamard/Walsh transforms and Gaussian mixture models. Complete design space exploration has been performed on these cores, and where appropriate, they have been shown to clearly outperform comparable existing implementations. At the architectural level, strategies such as parallelism, pipelining and systolisation have been successfully applied for the design and optimisation of a number of cores including colour space conversion, finite Radon transform, finite ridgelet transform and circular convolution. A pioneering study into the influence of supply voltage scaling for FPGA based designs, used in conjunction with performance enhancing strategies such as parallelism and pipelining has been performed. Initial results are very promising and indicated significant potential for future research in this area. A key contribution of this work includes the development of a novel high level power macromodelling technique for design space exploration and characterisation of custom IP cores for FPGAs, called Functional Level Power Analysis and Modelling (FLPAM). FLPAM is scalable, platform independent and compares favourably with existing approaches. A hybrid, top-down design flow paradigm integrating FLPAM with commercially available design tools for systematic optimisation of IP cores has also been developed.