Year: 2022
Type: dissertation
Abstract: The ongoing deployment of 5G NR is to bring a completely new wave of technology and revolutionize wireless data transfer. The new standard will provide improved data rates, lower latency, better energy efficiency and larger capacity for connected devices. New performance requirements and increased flexibility challenge conventional physical layer (L1) solutions and boost research of new algorithmic approaches. One proposed new approach is deep learning (DL), which in recent studies have shown to provide a feasible alternative for existing algorithms in terms of computational load and accuracy. So far, neural networks (NNs) based solutions are not available in any commercial 5G physical layer product. They employ highly specialized hardware, which poses challenges when efficient NN inference implementations are considered. Convolutional neural networks (CNNs) are an essential subtype of NNs and have recently been introduced to replace several L1 processing blocks. This work aims to develop a comprehensive framework that enables accelerating CNN inferences on CEVA-XC4500 DSP, a current state of the art in L1 processing. The basic idea of the framework is presented earlier for multilayer perceptrons (MLPs), and this work elaborates the idea for 1-D CNNs. Essential parts of CNNs are covered, including convolution layers, pooling layers and some activation functions. The DSP implementation is optimized with the single instruction, multiple data (SIMD) intrinsics that the DSP offers broad support. Since the DSP is designed for fixed-point arithmetic, the framework includes a procedure for quantizing the network parameters as a part of the framework's offline preprocessing part. Implementation's real-time performance is evaluated by recording the processor cycle counts of running inference with five different CNN models. The results are recorded in a software simulator that simulates the operation of the DSP in a cycle-accurate manner. The implementation's accuracy and quantization procedure are also evaluated. The results show that the target DSP can accelerate optimized CNN inferences effectively. Compared to the unoptimized reference implementation, speedups ranging from 6.5 to 26.8 were observed. Most performance gains originate from optimizing the convolution layers, typically computationally the heaviest part of CNNs. Clear benefits can also be observed from optimizing pooling layers and activation functions.
Source: Aaltodoc (Aalto University)
Author: Ollila, Uula
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112325786