SophonSDK Development Guide

Catalogue

• 1. Declaration
• 2. Introduction
  • 2.1. Explanation of Terms
  • 2.2. Upgrade Description
  • 2.3. Precautions
  • 2.4. Authorisation
  • 2.5. Help and support
• 3. SDK Software Package
  • 3.1. SDK Introduction
    • 3.1.1. SophonSDK includes the Basic Toolkit and the Advanced Toolkit
      • 3.1.1.1. Basic Toolkit
      • 3.1.1.2. Advanced Toolkit
    • 3.1.2. SophonSDK Folder Directory
    • 3.1.3. SophonSDK Unpacked File Directory
    • 3.1.4. SDK Main Modules
  • 3.2. Factsheet
  • 3.3. Get SDK
  • 3.4. Install SDK
    • 3.4.1. Environment Configuration -Linux
      • 3.4.1.1. Decompress the SDK package
      • 3.4.1.2. tpu-nntc environment initialization
      • 3.4.1.3. libsophon environment construction
      • 3.4.1.4. sophon-mw environment construction
      • 3.4.1.5. Cross-compilation environment construction
      • 3.4.1.6. Verification
    • 3.4.2. Environment Configuration -SoC
      • 3.4.2.1. Development environment configuration
      • 3.4.2.2. Operating environment configuration
  • 3.5. Update SDK
    • 3.5.1. Updates in PCIe mode
    • 3.5.2. Updates in SoC mode
  • 3.6. SDK Update Record
  • 3.7. Release Note
  • 3.8. SDK known problems
• 4. Quick Start
  • 4.1. Chip Introduction
  • 4.2. Overview of Transplant Development
    • 4.2.1. Algorithm Transplant Process
    • 4.2.2. Typical Video AI Analysis Tasks
  • 4.3. Key Concepts
    • 4.3.1. Working Mode
    • 4.3.2. Hardware Memory
    • 4.3.3. BModel
    • 4.3.4. bm_image
  • 4.4. Commonly Used Examples
• 5. Network Transfer
  • 5.1. Model Migration Overview
  • 5.2. FP32 Model Generation
  • 5.3. INT8 Model Generation
    • 5.3.1. Prepare the lmdb data set
    • 5.3.2. Generate FP32 Umodel
    • 5.3.3. Generate INT8 Umodel:
      • 5.3.3.1. Data preprocessing
      • 5.3.3.2. Quantization network
      • 5.3.3.3. Optimized network
      • 5.3.3.4. Cascade network quantization
    • 5.3.4. Accuracy test
    • 5.3.5. Generate INT8 Bmodel
    • 5.3.6. auto_cali one-click quantization tool
• 6. Algorithm Transplant
  • 6.1. Algorithm Migration Overview
  • 6.2. C/C++ Programming Details
    • 6.2.1. Load bmodel
    • 6.2.2. Preprocessing
      • 6.2.2.1. Preprocessing initialization
      • 6.2.2.2. Open the video stream
      • 6.2.2.3. Decoded video frame
      • 6.2.2.4. Mat converts bm_image
      • 6.2.2.5. Preprocessing
    • 6.2.3. Inference
    • 6.2.4. Post-processing
    • 6.2.5. Summary of considerations for algorithm development
  • 6.3. Python Programming Details
    • 6.3.1. Loading model
    • 6.3.2. Preprocessing
    • 6.3.3. Inference
  • 6.4. Decoding Module
    • 6.4.1. OpenCV Decoding
    • 6.4.2. FFmpeg Decoding
  • 6.5. Graphics Operation Acceleration Module
    • 6.5.1. C Language Programming Interface
    • 6.5.2. Python Language Programming Interface
  • 6.6. Model Reasoning
    • 6.6.1. Introduction of C language Interface of BMLib Module
    • 6.6.2. BMRuntime This section describes the interface of module C
    • 6.6.3. Python interface
• 7. Performance Test
  • 7.1. Overview
  • 7.2. Video Codec Performance
    • 7.2.1. Decode capability performance test
    • 7.2.2. Encode capability performance test
    • 7.2.3. Transcode capability performance test
  • 7.3. TPU Computing Power
    • 7.3.1. BM1684 Performance test
    • 7.3.2. BM1684X Performance test
• 8. Appendix