MIOpen documentation

MIOpen is the AMD open-source, deep-learning primitives library for GPUs. It implements fusion to optimize for memory bandwidth and GPU launch overheads, providing an auto-tuning infrastructure to overcome the large design space of problem configurations. It also implements different algorithms to optimize convolutions for different filter and input sizes.

MIOpen is one of the first libraries to publicly support the bfloat16 data type for convolutions, which allows for efficient training at lower precision without loss of accuracy.

The MIOpen public repository is located at ROCm/MIOpen.

Install

• MIOpen prerequisites

• Install MIOpen

• Build MIOpen from source

• Build MIOpen for embedded systems

• Build MIOpen using Docker

Conceptual

• Find database

• Kernel cache

• Performance database

• Manual tuning

• MI200 alternate implementation

• Porting to MIOpen

How to

• Use the fusion API

• Log and debug

• Use the find APIs and immediate mode

Reference

• API library

  • Modules

  • Datatypes

For information on contributing to the MIOpen code base, see Contributing to ROCm.

For licensing information for all ROCm components, see ROCm licensing.