high performance computing on graphics processing units: hgpu.org

Hardware accelerators are a fundamental part of modern high performance computing (HPC) systems due to their performance capabilities. The two most commonly used accelerators are GPUs and FPGAs. Despite the easier programmability and better memory performance of GPUs, generally FPGAs perform equally well for computationally challenging applications while dramatically reducing the energy consumption. Furthermore, with the availability of high level synthesis (HLS), the use of FPGAs has become easier. This makes them an excellent candidate for modern HPC systems. This dissertation describes my research work done in the field of electronic design automation with the major focus on optimizing memory intensive applications modeled using high level language for FPGAs. This work can be split into two parts, one dealing with manual memory optimization while other advocates the use of automated algorithms to select and optimize the best application-specific cache layout. The first part covers the manual optimization of a realistic smart city application. The application implements two image processing algorithms in OpenCL language which computes velocity and density of vehicles on urban streets in real time. Several different implementations of these memory hungry algorithms are considered. The results show that using suitable optimizations and HLS optimization directives, FPGAs can produce results with performance similar to a GPU with an order of magnitude less energy consumption. The second part of the dissertation starts by observing that custom data caches implemented on FPGAs are only useful if their layout is in accordance to their data access pattern. In this work, we present a tool, PEDAL (Pattern Evinced Determination of Appropriate Layout), that can automatically tune the custom data caches based on analyzing address traces. PEDAL uses artificial intelligence algorithms to detect the pattern of each array and then design the optimal cache for that pattern. The comparison of the results of PEDAL with the exhaustive search of cache configurations and cache designed through a state-of-the-art algorithm from the literature proves that it can produce better configurations in less time.