Hardware Acceleration in the Context of Motion Control for Autonomous Systems

Abstract: State estimation filters are computationally intensive blocks used to calculate uncertain/unknown state values from noisy/not available sensor inputs in any autonomous systems. The inputs to the actuators depend on these filter’s output and thus the scheduling of filter has to be at very small time intervals. The aim of this thesis is to investigate the possibility of using hardware accelerators to perform this computation. To make a comparative study, 3 filters that predicts 4, 8 and 16 state information was developed and implemented in Arm real time and application purpose CPU, NVIDIA Quadro and Turing GPU, and Xilinx FPGA programmable logic. The execution, memory transfer time, and the total developement time to realise the logic in CPU, GPU and FPGA is discussed. The CUDA developement environment was used for the GPU implementation and Vivado HLS with SDSoc environment was used for the FPGA implementation. The thesis concludes that a hardware accelerator is needed if the filter estimates 16 or more state information even if the processor is entirely dedicated for the computation of filter logic. Otherwise, for a 4 and 8 state filter the processor shows similar performance as an accelerator. However, in a real time environment the processor is the brain of the system, so it has to give instructions to many other functions parallelly. In such an environment, the instruction and data caches of the processor will be disturbed and there will be a fluctuation in the execution time of the filter for every iteration. For this, the best and worst case processor timings are calculated and discussed.