文档

Independent Building of Operator Deliverables

Overview

After obtaining the operator deliverables and storing them based on the directory structure requirements, you can build the deliverables independently to generate a custom operator installation package (*.run). The operations are as follows:

  • Compile the AI CPU operator code implementation files (.h and .cc) into libcust_aicpu_kernels.so.
  • Compile the operator information library definition file (.ini) into a .json file.
  • Build the operator prototype definition files (.h and .cc) into libcust_op_proto.so.
  • Compile the plugin implementation files (.h and .cc) of the TensorFlow/Caffe/ONNX operator into libcust_{tf/caffe/onnx}_parsers.so.
  • Build-related files, such as CMakeLists.txt, are automatically added.

Procedure

  1. Before performing independent building, ensure that the directory structure of the operator deliverables is as follows:
    • Directory structure of the TBE operator deliverables:
      ├── framework      // Directory of the operator plugin implementation files. Ignore this directory for a PyTorch operator.
│   ├── tf_plugin    // Directory of the generated operator plugin code when the source framework is TensorFlow
│       └── tensorflow_conv2_d_plugin.cc       // Implementation file of the operator plugin
│   └── onnx_plugin   // Directory of the generated operator plugin code when the source framework is ONNX
│       └── conv2_d_plugin.cc   // Implementation file of the operator plugin
├── op_proto     // Directory of the operator prototype definition file (mandatory deliverable)
│   ├── conv2_d.h
│   ├── conv2_d.cc
├── tbe     // Mandatory deliverable
│   ├── impl    // Directory of operator implementation files
│       └── conv2_d.py      // Operator implementation file
│   ├── op_info_cfg   // Directory of the operator information library files
│       └── ai_core
│                 ├── {Soc Version}        // Ascend AI Processor version
│                     ├── conv2_d.ini         // Operator information library definition file
├── op_tiling  // Directory of files related to operator tiling. Ignore this directory if operator tiling is not involved.
    • Directory structure of the AI CPU operator deliverables:
      ├── framework      // Directory of the operator plugin implementation files. Ignore this directory for a PyTorch operator.
│   ├── tf_plugin    // Directory of the generated operator plugin code when the source framework is TensorFlow
│       └── tensorflow_conv2_d_plugin.cc       // Implementation file of the operator plugin
│   └── onnx_plugin   // Directory of the generated operator plugin code when the source framework is ONNX
│       └── conv2_d_plugin.cc   // Implementation file of the operator plugin
├── op_proto     // Directory of the operator prototype definition file (mandatory deliverable)
│   ├── conv2_d.h
│   ├── conv2_d.cc
├── cpukernel    // Mandatory deliverable
│   ├── impl        // Directory of operator code implementation files
│   │   ├── conv2_d_kernels.cc
│   │   └── conv2_d_kernels.h
│   ├── op_info_cfg
│   │   └── aicpu_kernel
│   │       └── conv2_d.ini      // Operator information library definition file
├── op_tiling  // Directory of files related to operator tiling. Ignore this directory if operator tiling is not involved.
  2. Build the operator deliverables.

    Go to the directory where the msopgen tool is located (python/site-packages/bin under the CANN component directory) and run the following command. Table 1 describes the parameters.

    ./msopgen compile -i {operator deliverables directory} -c {CANN installation paths}

    Table 1 Command-line options

    Option

    Description

    Required

    compile

    Builds the operator deliverables.

    Yes

    -i,

    --input

    Path of the operator deliverables. The path can be either absolute or relative. The user who executes the tool must have the read permission on the path.

    Yes

    -c,

    --cann

    CANN installation directory

    No

    -h,

    --help

    Help information

    No
  3. After the project is built successfully, an OPP named custom_opp_<target os>_<target architecture>.run is generated in the build_out directory under the path of the operator deliverables.
  4. Check the path of the operator deliverables and add files required for independent building.
    • The following shows the directory structure of the TBE operator deliverables after building.
      ├── build_out    // Build output directory
├── build.sh     // Entry script for the build
├── cmake       // New build directory
│   ├── config.cmake
│   ├── util        
├── CMakeLists.txt    // New CMakeLists.txt
├── framework      // Directory of the operator plugin implementation files. Ignore this directory for a PyTorch operator.
│   ├── CMakeLists.txt    // New file
│   ├── tf_plugin    // Directory of the generated operator plugin code when the source framework is TensorFlow
│       └── tensorflow_conv2_d_plugin.cc       // Implementation file of the operator plugin
│       └── CMakeLists.txt   // New file
│   └── onnx_plugin   // Directory of the generated operator plugin code when the source framework is ONNX
│       ├── CMakeLists.txt    // New file
│       └── conv2_d_plugin.cc   // Implementation file of the operator plugin
├── op_proto     // Directory of the operator prototype definition files and the CMakeLists file
│   ├── conv2_d.h
│   ├── conv2_d.cc
│   ├── CMakeLists.txt     // New file
├── tbe 
│   ├── CMakeLists.txt     // New file
│   ├── impl    // Directory of operator implementation files
│       └── conv2_d.py      // Operator implementation file
│   ├── op_info_cfg   // Directory of the operator information library files
│       └── ai_core
│                 ├── {Soc Version}        // Ascend AI Processor version
│                     ├── conv2_d.ini         // Operator information library definition file
├── op_tiling  // Directory of files related to operator tiling. Ignore this directory if operator tiling is not involved.
│   ├── CMakeLists.txt     // New file
├── scripts     // Directory of scripts used for custom operator packing
    • The following shows the directory structure of the AI CPU operator deliverables after building.
      ├── build_out    // Build output directory
├── build.sh     // Entry script for the build
├── cmake    // New build directory
│   ├── config.cmake
│   ├── util        
├── CMakeLists.txt    // New file
├── framework      // Directory of the operator plugin implementation files. Ignore this directory for a PyTorch operator.
│   ├── CMakeLists.txt    // New file
│   ├── tf_plugin    // Directory of the generated operator plugin code when the source framework is TensorFlow
│       └── tensorflow_conv2_d_plugin.cc       // Implementation file of the operator plugin
│       └── CMakeLists.txt   // New file
│   └── onnx_plugin   // Directory of the generated operator plugin code when the source framework is ONNX
│       ├── CMakeLists.txt    // New file
│       └── conv2_d_plugin.cc   // Implementation file of the operator plugin
├── op_proto     // Directory of the operator prototype definition files and the CMakeLists file
│   ├── conv2_d.h
│   ├── conv2_d.cc
│   ├── CMakeLists.txt     // New file
├── cpukernel
│   ├── CMakeLists.txt     // New file
│   ├── impl        // Directory of operator code implementation files
│   │   ├── conv2_d_kernels.cc
│   │   └── conv2_d_kernels.h
│   ├── op_info_cfg
│   │   └── aicpu_kernel
│   │       └── conv2_d.ini      // Operator information library definition file
│   └── toolchain.cmake    // New build file
├── op_tiling    // Directory of files related to operator tiling. Ignore this directory if operator tiling is not involved.
│   ├── CMakeLists.txt     // New file
├── scripts     // Directory of scripts used for custom operator packing