Key Points

During application development, the Matrix module processes data in the following phases:

1. Calls an API to transfer data (images) from the host to the device.
2. Calls DVPP APIs to encode, decode, and resize data.
3. Calls the model manager APIs provided by Framework to infer data.
4. Calls an API to send the inference result from the device to the host.

Figure 1 Data processing

During Matrix module processing, data transfer, DVPP processing, and model inference are most time-consuming. This document mainly describes how to streamline these phases and boost the performance.

For details about each API, see Matrix API Reference and DVPP API Reference

• Memory Management
  
• Host-Device Data Transmission
  
• Usage of DVPP
  
• Model Conversion Pre-Processing Configuration
  
• Batch and Timeout
  
• Processing the Input and Output Data of Algorithm Inference
  
• Optimization of Data Backhaul