TheRock/rocm-libraries/projects/hipdnn/README.md

hipDNN

Caution

hipDNN is in the early stages of development. There is currently limited functionality available to execute graphs. See Operation Support for reference.

Note

The published hipDNN documentation is available here in an organized, easy-to-read format, with search and a table of contents. The documentation source files reside in the docs folder of this repository. As with all ROCm projects, the documentation is open source. For more information on contributing to the documentation, see Contribute to ROCm documentation.

Overview

hipDNN is a graph-based deep learning library for AMD GPUs that leverages a flexible plugin architecture to provide optimized implementations and utilities for various routines.

---

Table of Contents

• Getting Started
• Documentation
  • User Guides
  • Developer Guides
  • Testing
• Project Structure
• API Documentation
• Contributing

---

Getting Started

The fastest way to get started with hipDNN is to follow the quick start steps in the build guide.

---

Documentation

User Guides

• Building - Prerequisites, build configurations, and platform-specific instructions
• Consumer Quick Start - Using an installed hipDNN in your CMake project
• How-To - Using hipDNN components and extending the framework
• Environment Configuration - Runtime configuration and logging setup
• Operation Support - Currently supported operations and their status
• Samples - Frontend usage examples
• API Reference - Doxygen-generated API documentation

Developer Guides

• Design Overview - Architecture and design descriptions and diagrams
• Extending hipDNN - How to extend hipDNN functionality
• Plugin Development - Creating and using custom plugins for hipDNN
• Roadmap - Feature priorities and development plans

Testing

• Testing - Synopsis of testing information
• Testing Strategy - Specific testing approach
• Test Plan - Detailed test planning
• Test Run Template - Guidelines for test execution

---

Project Structure

hipDNN is organized into several key components. For detailed architecture descriptions, see the Design Overview.

Component	Description
Backend	Core shared library providing C API for operation graphs and managing plugins
Data SDK	Header-only library with shared types, tensor utilities, logging, and the engine name registry
Flatbuffers SDK	Header-only library with FlatBuffers schemas, generated headers, graph wrappers, and optional JSON helpers
Frontend	Header-only C++ API wrapper around the backend
Plugin SDK	Header-only library for plugin development
Samples	Example implementations demonstrating hipDNN usage
Tests	Tests for the public API (incl. frontend integration tests)
Tools	Experimental utilities (e.g., benchmarking, engine listing) — subject to change

Note

Official hipDNN plugins can be found in the dnn-providers folder (e.g., MIOpen Plugin).

Docker Support

See Docker README for containerized development environments.

---

API Documentation

hipDNN includes Doxygen-generated API documentation for the public C++ frontend.

Building the Documentation

1. Install Doxygen:

   # Ubuntu/Debian
   sudo apt-get install doxygen
   

   For Windows, download the installer from doxygen.nl/download.

2. Generate the documentation:

   cd <project-root>
   doxygen Doxyfile
   

3. Open the generated documentation:

   # Linux
   xdg-open build/docs/html/index.html
   
   # Windows
   start build\docs\html\index.html
   

The documentation covers the frontend API including:

• Graph construction and execution
• Tensor and operation attributes
• Engine configuration and knobs
• Error handling

---

Contributing

For information about contributing to the hipDNN project, please see the Contributing Guide.