Acceleration of deep convolutional neural networks on multiprocessor system-on-chip

In this master thesis some of the most promising existing frameworks and implementations of deep convolutional neural networks on multiprocessor system-on-chips (MPSoCs) are researched and evaluated. The thesis’ starting point was a previousthesis which evaluated possible deep learning models and frameworks for object detection on infra-red images conducted in the spring of 2018. In order to fit an existing deep convolutional neural network (DCNN) on a Multiple-Processor-System on Chip it needs modifications. Most DCNNs are trained on Graphic processing units (GPUs) with a bit width of 32 bit. This is not optimal for a platform with hard memory constraints such as the MPSoC which means it needs to be shortened. The optimal bit width depends on the network structure and requirements in terms of throughput and accuracy although most of the currently available object detection networks drop significantly when reduced below 6 bits width. After reducing the bit width, the network needs to be quantized and pruned for better memory usage. After quantization it can be implemented using one of many existing frameworks. This thesis focuses on Xilinx CHaiDNN and DNNWeaver V2 though it touches a little on revision, HLS4ML and DNNWeaver V1 as well. In conclusion the implementation of two network models on Xilinx Zynq UltraScale+ ZCU102 using CHaiDNN were evaluated. Conversion of existing network were done and quantization tested though not fully working. The results were a two to six times more power efficient implementation in comparison to GPU inference