[Audio recognizer based on GD32F350RB] 2. Test whether GD32F3xx can support edge computing

北方

[Audio recognizer based on GD32F350RB] 2. Test whether GD32F3xx can support edge computing [Copy link]

1. In the design of the next generation of electronic products, edge computing is currently the most cutting-edge perspective. This is because artificial intelligence has been given the title of the fourth industrial revolution, but to make these smart calculations land, countless smart electronic devices are needed to realize them, rather than just doing some image recognition and video capture functions on the computer. Currently, the more commonly used artificial intelligence solutions all need to be connected to the Internet through wifi, upload the pictures and sound signals collected by the front end to the cloud, and let the massive servers in the cloud realize the calculation function and feedback the results to the terminal. A typical example is Amazon's ECHO speaker. The multi-dimensional sound field captures only high-precision sound signals. Specifically, these represent the language. Converting them into text and accurately identifying them in the massive text semantics is indeed not the terminal's competence. However, ARM has not given up in this regard and has taken the lead in realizing the concept of edge computing, and the threshold is as low as ARM-M0++. In other words, it is to let the data collected by the terminal directly calculate the results at the terminal, without relying on the communication channel. Completing large-scale calculations at the terminal is simply edge computing. In this way, these development boards no longer need to simply play the blink game of lighting up lights, but can be given heavier computing tasks, and more interesting applications can also be realized. 2. The way to realize edge computing is to call the CMSIS-NN library, which contains many powerful computing function libraries. In particular, it is necessary to enable the DSP coprocessor, that is, to set DSP=1. At present, there are not many MCU manufacturers trying edge computing. NXP has provided some examples and has some experience with NXP. So can it be realized in GD32F350RB? This post will test it. 3. First, create a project and select the GD32F350RB chip 4. Secondly, select the initial library to be installed. Be sure to select NN and DSP. Then compile and find that the compilation is successful, but the system resources are insufficient and the memory is insufficient. Modify to use microLib and compile again, and it will show that some programs have insufficient resources. 5. This preliminary verification shows that the program and code can be executed on GD32F350, but how to determine the resources occupied by the code and optimize it is a programming problem. Compared with reducing the number of layers and reducing the occupancy of the intermediate buff, you can also load external memory and swap the intermediate data to the external memory, so that the calculation can be completed. Specifically, it is the optimization process. This preliminary verification can verify that this sample code can be executed on this platform. Specific code analysis and instructions for use are explained step by step.