Auto Kernel Generator (AKG)

Project Introduction

AKG (Auto Kernel Generator) is a deep learning compiler designed to optimize operators in deep neural networks and provide automatic operator fusion capabilities under specific patterns.
After version 2.3, we have completely upgraded AKG. The current AKG project includes four subprojects: AIKG, AKG-TVM, AKG-MLIR, and SWFT. It supports automatic operator generation for multiple hardware backends such as CPU, NVIDIA V100/A100, Atlas 800T A2/A3, and 300I DUO.

• AIKG: AIKG (AI Driven Kernel Generator) is our latest incubated tool for AI-driven operator generation. It provides a Multi-agent-based operator auto-generation framework, the AscendKernelBench evaluation platform, and an operator knowledge base. Currently, it supports generating Triton-Ascend (Atlas A2/A3) and SWFT (300I DUO) DSLs through LLM and compiling operator kernels.
• AKG-TVM: AKG-TVM retains the features and code from versions prior to 2.2. It offers operator auto-generation technology based on polyhedral scheduling, supporting capabilities like auto-vectorization, auto-tiling, and memory promotion. Combined with MindSpore's graph operator fusion, it enhances the performance of large model training and inference across different hardware backends.
• AKG-MLIR: A deep learning compiler evolved from the open-source MLIR project, providing a complete operator compilation pipeline for CPU/GPU/Ascend. It currently includes the MindSpore Dialect for graph compilation and extends dialects such as Linalg, Affine, and GPU, enhancing loop fusion scheduling capabilities. Additionally, AKG-MLIR integrates with AscendNPU IR, supporting fused operator generation for the Ascend backend.
• SWFT: An operator development tool for the 300I DUO, offering a Python-based, AscendC-compliant, dataflow-oriented macro-instruction programming model. By allowing users to define operator fusion boundaries, it supports extensive deep/shallow operator fusion.

Installation and Usage Instructions

Each subproject in the AKG repository supports independent installation and building. Please refer to the respective documentation for each subproject.

• AIKG: Refer to the AIKG documentation.
• AKG-TVM: Refer to the AKG-TVM documentation.
• AKG-MLIR: Refer to the AKG-MLIR documentation.
• SWFT: Refer to the SWFT documentation.

Contributions

We welcome your contributions. For details, please refer to the MindSpore Contributor Wiki.

Release Notes

For release notes, see RELEASE.

License

Apache License 2.0.

AKG SIG WeChat Group