high performance computing on graphics processing units: hgpu.org

In this dissertation, we address the problem of runtime adaptation of the application to its execution environment. A typical example is changing theprocessing element on which a computation is executed, considering the available processing elements in the system. This is done based on the information and instrumentation provided by the compiler and taking into account the status of the environment. The work focuses on heterogeneous multicore embedded architectures. We address three aspects of application optimizations: hardware software mapping, memory allocation and parallel execution. For each aspect, an algorithm is developed and, using a suitable application, it is tested on the hardware platform. The programming paradigm on which this work is based is the Molen programming paradigm, extended and adapted for our specific platform and operating environment. The hardware software mapping algorithm objective is to choose at runtime, on which processing element it is more efficient to execute a function. For the memory allocation we propose an algorithm, that using compile-time gathered information and the current execution environment, decides on the best allocation for memory, at runtime. For dealing with parallel applications we developed an algorithm that selects the best trade-off between area and speedup and decides on the number of concurrent units that execute. The experiments were performed on an embedded multicore heterogeneous platform, namely the hArtes Hardware Platform (hHP). This platform contains an ARM processor as General Purpose Processor (GPP), an Atmel Magic Diopsis Digital Signal Processor (DSP) and a Xilinx Virtex4 Field Programmable Gate Array (FPGA). The applications used to validate the algorithms are real life applications from the multimedia field: a video encoder/decoder and a wavefield synthesis application. The mapping algorithms obtains improvements between 5% and 43%. We showed this is an adaptable algorithm, that will adapt the execution in case the execution overhead increases. The memory allocation algorithm obtained a speedup of 18% on the selected application. For this algorithm we show that the solution is within 14% of the optimal solution, computed using Integer Linear Programming (ILP). The scenario based selection of parallel computations, is between 21% to 92% better than existing solutions.