Sophon SDK development Sophon SDK Sophon SDK is a customized deep learning SDK based on its self- developed AI chip, covering model optimization and efficient operation support required for the inference phase of neural networks, providing an easy-to-use and efficient full-stack solution for deep learning application development and deployment. SophonSDK is compatible with both the 3rd generation BM1684 chip and the 4th generation BM1684X chip. Components of Sophon SDK Basic Toolkit The basic toolkit includes: Tpu-nntc is responsible for the offline compilation and optimization of the neural network model trained under the third-party deep learning framework to generate the final run-time BModel. Caffe, Darknet, MXNet, ONNX, PyTorch, PaddlePaddle, TensorFlow and so on are supported. Libsophon provides BMCV, BMRuntime, BMLib and other libraries, which are used to drive VPP, TPU and other hardware to complete image processing, tensor operation, model inference and other operations for users to develop deep learning applications. Sophon-mw encapsulates BM-OpenCV, BM-FFmpeg and other libraries, which are used to drive VPU, JPU and other hardware, and support the parsing of RTSP stream and GB28181 stream, video image codec acceleration, etc., for users to carry out deep learning application development. Sophon-sail provides advanced interfaces supporting Python/C++, which encapsulates the underlying library interfaces such as BMRuntime, BMCV, BMDecoder, BMLib, etc., for users to carry out deep learning application development. Advanced Toolkit The Advanced Toolkit includes: Tpu-mlir provides a complete tool chain for the TPU compiler project, which can transform pre-trained neural networks in different frameworks into binary BModel that can run efficiently on the Sophon TPU. Frameworks that are currently directly supported include tflite, onnx, and Caffe. Tpu-perf provides a complete set of toolkits for model performance and accuracy verification. Tpu-kernel is the underlying development interface to the chip, allowing both dedicated instructions to accelerate deep learning business logic and generic instructions to accelerate various custom algorithms. SDK Directory Structure SophonSDK provides eleven folder modules, as shown in the following table: The structure of the unpacked SDK files is as follows: 1SophonSDK 2 3├── libsophon__ 4│   ├── BMCV_Technical_Reference_Manual.pdf 5│   ├── BMCV开发参考手册.pdf 6│   ├── BMLib_Technical_Reference_Manual.pdf 7│   ├── BMLIB开发参考手册.pdf 8│   ├── BMRuntime Technical Reference Manual.pdf 9│   ├── BMRUNTIME开发参考手册.pdf 10│   ├── libsophon__aarch64.tar.gz #Arm64 machine, libsophon installar for other LINUX systems 11│   ├── libsophon__x86_64.tar.gz #X86x64 machine, libsophon installar for other LINUX systems 12│   ├── libsophon_dockerfile 13│   ├── libsophon.MD5 14│   ├── LIBSOPHON_User_Guide.pdf 15│   ├── LIBSOPHON使用手册.pdf 16│   ├── release_version.txt 17│   ├── sophon-driver__amd64.deb #x86_64 machine，PCIe card driver installation file corresponding to Debian/Ubuntu system 18│   ├── sophon-driver__arm64.deb #arm64 machine，PCIe card driver installation file corresponding to Debian/Ubuntu system 19│   ├── sophon-libsophon__amd64.deb #x86_64 machine，libsophon runtime environment installation file corresponding to Debian/Ubuntu system 20│   ├── sophon-libsophon__arm64.deb #arm64 machine，libsophon runtime environment installation file corresponding to Debian/Ubuntu system 21│   ├── sophon-libsophon-dev__amd64.deb #x86_64 machine，libsophon runtime environment installation file corresponding to Debian/Ubuntu system 22│   └── sophon-libsophon-dev__arm64.deb #arm64 machine，libsophon runtime environment installation file corresponding to Debian/Ubuntu system 23├── sophon-demo__ 24│   ├── release_version.txt 25│   ├── sophon-demo.MD5 26│   └── sophon-demo_v_b909566_20221027.tar.gz #Comprehensive routines for a single model or scenario 27├── sophon-img__ 28│   ├── bsp-debs 29│   ├── bsp_update.tgz 30│   ├── libsophon_soc__aarch64.tar.gz #Files required for cross-compilation 31│   ├── release_version.txt 32│   ├── sdcard.tgz #SoC card brush package 33│   ├── SOPHON_BSP_Technical_Reference_Manual.pdf 34│   ├── SOPHON BSP开发参考手册.pdf 35│   ├── sophon-img.MD5 36│   ├── sophon-soc-libsophon__arm64.deb #Libsophon installation files corresponding to SoC platform 37│   ├── sophon-soc-libsophon-dev__arm64.deb #Libsophon installation files corresponding to SoC platform 38│   ├── system.tgz 39│   └── tftp.tgz 40├── sophon-mw__ 41│   ├── Multimedia FAQ.pdf 42│   ├── Multimedia Technical Reference Manual.pdf 43│   ├── Multimedia User Guide.pdf 44│   ├── MULTIMEDIA使用手册.pdf 45│   ├── MULTIMEDIA常见问题手册.pdf 46│   ├── MULTIMEDIA开发参考手册.pdf 47│   ├── release_version.txt 48│   ├── sophon-mw__aarch64.tar.gz #Arm64 machine, sophon-mw installar for other LINUX systems 49│   ├── sophon-mw__x86_64.tar.gz #Arm64 machine, sophon-mw installar for other LINUX systems 50│   ├── sophon-mw.MD5 51│   ├── sophon-mw-soc__aarch64.tar.gz 52│   ├── sophon-mw-soc-sophon-ffmpeg__arm64.deb #SoC platform, ffmpeg runtime environment installation files 53│   ├── sophon-mw-soc-sophon-ffmpeg-dev__arm64.deb #SoC platform, ffmpeg development environment installation files 54│   ├── sophon-mw-soc-sophon-opencv__arm64.deb #SoC platform, opencv runtime environment installation files 55│   ├── sophon-mw-soc-sophon-opencv-dev__arm64.deb #SoC platform, opencv development environment installation files 56│   ├── sophon-mw-soc-sophon-sample__arm64.deb #SoC platform, multimedia program sample file 57│   ├── sophon-mw-sophon-ffmpeg__amd64.deb #X86x64 machine, installation files for ffmpeg runtime environment corresponding to Debian/Ubuntu system 58│   ├── sophon-mw-sophon-ffmpeg__arm64.deb #arm64 machine, installation files for ffmpeg runtime environment corresponding to Debian/Ubuntu system 59│   ├── sophon-mw-sophon-ffmpeg-dev__amd64.deb #x86_64 machine, ffmpeg development environment installation file corresponding to Debian/Ubuntu system 60│   ├── sophon-mw-sophon-ffmpeg-dev__arm64.deb #arm64 machine, ffmpeg development environment installation file corresponding to Debian/Ubuntu system 61│   ├── sophon-mw-sophon-opencv__amd64.deb #x86_64 machine, opencv runtime environment installation file corresponding to Debian/Ubuntu system 62│   ├── sophon-mw-sophon-opencv__arm64.deb #arm64 machine, opencv runtime environment installation file corresponding to Debian/Ubuntu system 63│   ├── sophon-mw-sophon-opencv-abi0__amd64.deb #X86x64 machine, installation files for opencv runtime environment corresponding to CenterOS system 64│   ├── sophon-mw-sophon-opencv-abi0__arm64.deb #arm64 machine, opencv runtime environment installation file corresponding to CenterOS system 65│   ├── sophon-mw-sophon-opencv-abi0-dev__amd64.deb #x86x64 machine, opencv development environment installation file corresponding to CenterOS system 66│   ├── sophon-mw-sophon-opencv-abi0-dev__arm64.deb #arm64 machine, opencv development environment installation file corresponding to CenterOS system 67│   ├── sophon-mw-sophon-opencv-dev__amd64.deb #x86x64 machine, opencv development environment installation file corresponding to Debian/Ubuntu system 68│   ├── sophon-mw-sophon-opencv-dev__arm64.deb #arm64 machine, opencv development environment installation file corresponding to Debian/Ubuntu system 69│   ├── sophon-mw-sophon-sample__amd64.deb #x86x64 machine, sample files for multimedia programs 70│   └── sophon-mw-sophon-sample__arm64.deb #arm64 machine, sample files of multimedia programs 71├── sophon-pipeline__ 72│   ├── release_version.txt 73│   ├── sophon-pipeline.MD5 74│   └── sophon-pipeline_v_586366b_20221027.tar.gz #High performance inference framework based on pipeline 75├── sophon-rpc__ 76│   ├── release_version.txt 77│   ├── sophon-rpc__amd64.deb #x86x64 machine, sophon-rpc installation file corresponding to Debian/Ubuntu system 78│   ├── sophon-rpc__arm64.deb #arm64 machine, sophon-rpc installation file corresponding to Debian/Ubuntu system 79│   ├── sophon-rpc_.tar.gz #Sophon-rpc installation packages corresponding to other systems 80│   ├── sophon-rpc.MD5 81│   └── sophon-rpc使用指南.pdf 82├── sophon-sail__ 83│   ├── release_version.txt 84│   ├── Sophon_Inference_zh.pdf 85│   ├── sophon-sail_.tar.gz #Advanced interface library encapsulating BMLib, BMDecoder, BMCV, BMRuntime 86│   └── sophon-sail.MD5 87├── Sophon SDK文件清单.docx 88├── tpu-kernel__ 89│   ├── release_version.txt 90│   ├── tpu-kernel-1684x_v-0d0e513e-221027.tar.gz #Custom operator development tool 91│   └── tpu-kernel.MD5 92├── tpu-mlir__ 93│   ├── release_version.txt 94│   ├── tpu-mlir.MD5 95│   └── tpu-mlir_v--.tar.gz #Compiler tool chain 96├── tpu-nntc__ 97│   ├── release_version.txt 98│   ├── tpu-nntc.MD5 99│   └── tpu-nntc_v--.tar.gz #Model compilation quantization tool chain 100└── tpu-perf_v 101 ├── md5sum.txt 102 ├── tpu_perf--py3-none-manylinux2014_aarch64.whl 103 ├── tpu_perf--py3-none-manylinux2014_x86_64.whl 104 └── tpu-perf-.tar.gz #Tools for performance analysis and accuracy verification of the model SDK Main Modules Hardware Driver and Runtime Library LIBSOPHON : contains libraries such as BMCV, BMRuntime, BMLib, etc., which are used to drive hardware such as VPP, TPU, etc. toperform operations such as image processing, tensor operations, model inference, etc Multimedia library SOPHON-MW : supports BM-OpenCV and BM-FFmpeg for hardware accelerate of Sophon devices, parsing of RTSP streams, GB28181 streams, video and image encoding and decoding Model Compilation Quantization Tool Chain TPU-NNTC : supports model conversion of Caffe,Tensorflow,Pytorch,MXNet,Darknet, Paddle Paddle, ONNX and other framework models; supports model quantization: original model -> FP32 UModel -> INT8 UModel -> INT8 BModel, and also provides auto-cali automatic quantization tool. Tensor Operation and Image Processing Library BMCV : colour space conversion, scale transformation, affine transformation, projection transformation, linear transformation, frame, JPEG coding, BASE64 coding, NMS, sorting, feature matching Equipment Management BMLib : basic interfaces: management of device Handle, memory management, data handling, API sending and synchronisation, A53 enable, etc SAIL, the Arithmetic AI acceleration library : supports high-level interfaces to Python/C++ and is wrapper for underlying library interfaces such as BMRuntime、BMCV、BMDecoder Custom Operator Advanced Programming Library BMLang: A C++ based high- level programming library for Sophon TPUs, decoupled from hardware information, no knowledge of hardware architecture is required, code is written using tensor data (bmlang::Tensor) and computation operations (bm- lang::Operator), and finally bmlang::compile or bmlang::compile_with_check is used to generate a BModel that the TPU can run; it also supports the use of the arm cpu in BM168X to implement arithmetic that is not yet supported by the TPU. Algorithm Parallel Accelerated Programming Library TPUKernel: underlying programming interface based on the Sophon chip’s underlying atomic operations interface, requiring familiarity with hardware architecture and instruction set Model Performance and Accuracy Verification Tool TPUPerf: enables performance analysis and accuracy validation of models Details Get SDK（v23.05.01 version） Install SDK（v23.05.01 version） For additional resources, please visit here .