Summary

AKG (Auto Kernel Generator) is a deep learning compiler optimizing neural network operators through automatic fusion. It targets developers needing efficient operator execution across diverse hardware, aiming to boost performance and simplify custom kernel development.

How It Works

Upgraded to include AIKG and AKG-MLIR sub-projects, AKG offers dual approaches. AIKG is an LLM-driven generator using multi-agents for DSL creation (Triton-Ascend, SWFT), backed by an evaluation platform. AKG-MLIR is an MLIR-based compiler pipeline for CPU, GPU, and Ascend, extending dialects (Linalg, Affine, GPU) and integrating with AscendNPU IR for fused operator generation. This provides both AI-driven innovation and a robust compilation framework.

Quick Start & Requirements

Installation and usage are managed independently per sub-project (AIKG, AKG-MLIR), with details in their respective documentation. Supports CPU, NVIDIA (V100/A100), and Ascend (Atlas 800T A2/A3, 300I DUO) backends. Specific prerequisites are detailed in sub-project docs.

Highlighted Details

• AIKG: LLM-driven operator generation via multi-agent framework.
• AscendKernelBench evaluation platform and operator knowledge base.
• AKG-MLIR: MLIR-based compilation with Linalg, Affine, GPU dialect extensions.
• Complete compilation pipeline for CPU, GPU, Ascend.
• AscendNPU IR integration for fused operator generation.

Maintenance & Community

Contributions are welcome via the MindSpore Contributor Wiki. Version details are in RELEASE. Community interaction via an AKG SIG WeChat group.

Licensing & Compatibility

Licensed under Apache License 2.0, permissive for commercial use and closed-source linking.

Limitations & Caveats

Setup complexity due to separate sub-project documentation. Specific performance benchmarks are not detailed. Support may be more mature for Ascend hardware.