msOpGen Architecture Design Specifications¶

1 Project Overview¶

1.1 Background and Motivation¶

Operator development involves extensive framework code (Host-side prototype registration, Tiling strategies, Kernel-side operator implementation, build configuration, etc.). Manually setting up operator projects is tedious and error-prone. msOpGen automatically generates a complete operator project framework from JSON prototype definition files, allowing developers to focus on core algorithm logic.

1.2 Feature List¶

Type	Feature	Description
Core	Operator Project Generation	Generates complete Ascend C/TBE/AI CPU operator projects from JSON prototypes
Core	Multi-Framework Adaptation	Supports TensorFlow, PyTorch, MindSpore, ONNX frameworks and aclnn direct invocation
Core	Operator Append	Supports appending new operators to existing projects (`-m 1` mode)
Core	Simulation Pipeline Visualization	Parses performance simulation dump data to generate Chrome tracing views
Core	Compilation & Deployment	Generates build.sh compilation scripts and .run deployment packages
Support	ST Testing	msOpST tool auto-generates test cases and executes them on hardware
Support	On-Board Test Framework	msOpST ascendc_test generates kernel direct-invoke test framework

2 Design Goals¶

Design Goal	Description
Completeness	Generated projects can be compiled and deployed directly without manual framework code
Multi-Framework Coverage	Unified JSON interface adapts to multiple AI frameworks, reducing learning costs
Configurable Build	Flexible configuration of build options, chip models, and distribution modes via CMakePresets.json
CLI Usability	Clear and intuitive parameter design with sensible defaults

3 Architecture Overview¶

3.1 System Architecture¶

┌──────────────────────────────────────────────┐
│                CLI Layer                      │
│   msopgen gen    msopgen sim    msopst        │
└──────────────────┬───────────────────────────┘
                   │
┌──────────────────▼───────────────────────────┐
│              Core Engine Layer                 │
│  ┌──────────┐  ┌──────────┐  ┌────────────┐  │
│  │ JSON     │  │ Template │  │ Dump       │  │
│  │ Parser   │  │ Engine   │  │ Analyzer   │  │
│  └──────────┘  └──────────┘  └────────────┘  │
│  ┌──────────┐  ┌──────────────────────────┐  │
│  │ ST Test  │  │ Project Builder          │  │
│  │ Generator│  │ (CMake/Build Integration)│  │
│  └──────────┘  └──────────────────────────┘  │
└──────────────────┬───────────────────────────┘
                   │
┌──────────────────▼───────────────────────────┐
│                Output Layer                    │
│  Operator project / .run package /             │
│  trace.json / ST case.json / st_report.json   │
└──────────────────────────────────────────────┘

3.2 Module Division¶

Module	Responsibility	Input	Output
JSON Parser	Parse and validate operator prototype definition files	`*.json` prototype	Structured operator description
Template Engine	Generate project templates from operator description and chip model	Operator description + soc_version	Complete operator project directory
Dump Analyzer	Parse performance simulation dump data	Dump data files	trace.json pipeline visualization
Project Builder	Generate CMakeLists.txt, CMakePresets.json, build.sh	Operator description + build options	Buildable project
ST Test Generator	Parse Host-side source code to generate ST test cases	`op_host/*.cpp`	`*_case.json`
ST Test Runner	Execute hardware tests and generate reports	`*_case.json` + soc	`st_report.json`

3.3 Data Flow¶

Operator JSON ──→ [JSON Parser] ──→ 算子描述结构体
                                       │
                                [Template Engine] ──→ 算子工程目录
                                       │
                                [用户编写 Kernel 实现]
                                       │
                              [build.sh 编译] ──→ .run 部署包
                                       │                               
                                [msopst create]                       
                                       │                                     
                                ST 用例 .json   
                                       │                         
                                [msopst run]                       
                                       │                                 
                                st_report.json

4 Key Technical Points¶

4.1 Template Substitution Mechanism¶

msOpGen uses a template engine to automatically replace placeholders in C++ source templates based on the operator name, input/output parameter types, and formats defined in the JSON prototype, generating both Host-side (prototype registration, Shape inference, Tiling, info library) and Kernel-side (operator logic) framework code.

4.2 Naming Rules¶

Strict conversion rules between operator type (OpType) and file names / kernel function names: - PascalCase → snake_case - Example: AddCustom → add_custom.cpp / add_custom

4.3 Distribution Modes¶

Source Distribution: Retain kernel source .cpp files, supporting online compilation and ATC model conversion
Binary Distribution: Compile to .o and .json info files for direct operator binary invocation

5 Directory Structure¶

├── example/       // Tool examples
├── docs/          // Project documentation
├── msopgen/       // msopgen source code
├── tools/msopst/  // msopst code
├── test/
│   ├── msopgen/   // msopgen unit tests
│   └── msopst/    // msopst unit tests
├── output/        // WHL package output, test reports
├── setup.py       // msopgen WHL build script
└── build.py       // Build entry script