1 Introduction

Deep learning algorithms based on deep neural networks (DNNs) are used in many industries, such as robotics, industrial, and automotive. More and more DNN-based machine learning algorithms are used in ADAS products, such as lane line detection, traffic light recognition, pedestrian recognition and other basic ADAS functions are implemented using DNN algorithms. TI's latest generation automotive processor TDA4VM integrates high-performance computing unit C7x Digital Signal Processor (DSP) and Deep-learning Matrix Multiply Accelerator (MMA), which can efficiently perform some basic deep learning operators such as convolution calculations and matrix transformations. These DNN neural network algorithms usually require a lot of computation, and the C7x and MMA in TI TDA4 series processors can accelerate the computation of some DNN operators to achieve fast inference and recognition results. TI Deep Learning Library (TIDL) is a software ecosystem based on deep learning algorithms for TI platforms, which can quickly deploy some common deep learning algorithm models to TI embedded platforms.

TDA4xx is a multi-core heterogeneous SOC launched by TI. It has built-in TI's latest generation C7X DSP and Matrix Multiply Accelerator (MMA), which can efficiently process digital signals, especially good at processing audio, millimeter wave radar, lidar, and camera data. For the current hot AI technology, TI has also integrated related AI operators to facilitate customers to quickly deploy deep learning models. The software part of C7 mainly depends on the following parts in the whole SOC development process

• Compiler of C7X DSP: It mainly provides some inline library functions and functions such as compiling, linking, and forming executable files written by developers.
• Operating system: used for DSP's own task management, memory management, and so forth.
• Low-level related drivers: mainly the code that depends on the implementation of related functions, such as DRU, UDMA, timer and other drivers.
• TIOVX middleware: Provide a calling interface of multi-core heterogeneous accelerators for the entire SOC, and perform parallel computing of heterogeneous accelerators through pipeline settings, thereby accelerating data processing.
• TIDL: TIDL is a comprehensive software product for acceleration of Deep Neural Networks (DNNs) on TI's embedded devices.
• Memory segment management between multi-core heterogeneous: TDA4xx has multiple built-in cores suitable for different application scenarios, and each core can access the DDR space, so it is necessary to manage the memory space to avoid memory space conflicts.