high performance computing on graphics processing units: hgpu.org

A variety of multiprocessor architectures has proliferated even for off-the-shelf computing platforms. To make use of these platforms, traditional implementation frameworks focus on implementing Digital Signal Processing (DSP) applications using special platform features to achieve high performance. However, due to the fast evolution of the underlying architectures, solution redevelopment is error prone and re-usability of existing solutions and libraries is limited. In this thesis, we facilitate an efficient migration of DSP systems to multiprocessor platforms while systematically leveraging previous investment in optimized library kernels using dataflow design frameworks. We make these library elements, which are typically tailored to specialized architectures, more amenable to extensive analysis and optimization using an efficient and systematic process. In this thesis we provide techniques to allow such migration through four basic contributions: 1) We propose and develop a framework to explore efficient utilization of Single Instruction Multiple Data (SIMD) cores and accelerators available in heterogeneous multiprocessor platforms consisting of General Purpose Processors (GPPs) and Graphics Processing Units (GPUs). We also propose new scheduling techniques by applying extensive block processing in conjunction with appropriate task mapping and task ordering methods that match efficiently with the underlying architecture. The approach gives the developer the ability to prototype a GPU-accelerated application and explore its design space efficiently and effectively. 2) We introduce the concept of Partial Expansion Graphs (PEGs) as an implementation model and associated class of scheduling strategies. PEGs are designed to help realize DSP systems in terms of forms and granularities of parallelism that are well matched to the given applications and targeted platforms. PEGs also facilitate derivation of both static and dynamic scheduling techniques, depending on the amount of variability in task execution times and other operating conditions. We show how to implement efficient PEG-based scheduling methods using real time operating systems, and to re-use pre-optimized libraries of DSP components within such implementations. 3) We develop new algorithms for scheduling and mapping systems implemented using PEGs. Collectively, these algorithms operate in three steps. First, the amount of data parallelism in the application graph is tuned systematically over many iterations to profit from the available cores in the target platform. Then a mapping algorithm that uses graph analysis is developed to distribute data and task parallel instances over different cores while trying to balance the load of all processing units to make use of pipeline parallelism. Finally, we use a novel technique for performance evaluation by implementing the scheduler and a customizable solution on the programmable platform. This allows accurate fitness functions to be measured and used to drive runtime adaptation of schedules. 4) In addition to providing scheduling techniques for the mentioned applications and platforms, we also show how to integrate the resulting solution in the underlying environment. This is achieved by leveraging existing libraries and applying the GPP-GPU scheduling framework to augment a popular existing Software Defined Radio (SDR) development environment — GNU Radio — with a dataflow foundation and a stand-alone GPU-accelerated library. We also show how to realize the PEG model on real time operating system libraries, such as the Texas Instruments DSP/BIOS. A code generator that accepts a manual system designer solution as well as automatically configured solutions is provided to complete the design flow starting from application model to running system.