TIDL, as an important software library for deploying AI algorithms for TI embedded devices, is iterated and updated in each software version. Each iterative update is mainly to fix bugs in the old version and support the new features of the current deep learning algorithm. TIDL is mainly divided into two parts. The first part is to read, quantify, and convert models such as Tensorflow and ONNX of the current mainstream algorithm framework into the model format that TIDL can execute on the PC side. This part depends on the external open source libraries Opencv, Flatbufffers, and Protobuf, and different versions of the SDK have different version requirements for the three libraries. Therefore, when upgrading, you need to confirm whether to upgrade according to the TIDL User Guide. The second part is to infer the converted model on the board side. The board side inference is mainly to accelerate the calculation of operators. Some operators are calculated using C7X DSP, which is implemented by TIDL, and the other part is calculated using MMA, such as volume Product calculation, Fourier transform are calculated using MMALIB. Therefore, when TIDL is upgraded, MMALIB needs to be upgraded. MMALIB usage reference MMALIB User Guide.

Figure 1-3 TIDL Block Diagram