Open Source Summit Europe 2022

Artificial intelligence frameworks such as MindSpore, TensorFlow or PyTorch aim at providing high-level abstractions to enable rapid development of new operators and at supporting various architectures to enable computation onto the desired target platform. Translating programs written by AI scientists to efficient implementations on complex parallel hardware is an extreme compilation challenge. Existing frameworks heavily rely on high-performance implementations of well-known operators through optimized libraries. However, those libraries may miss global optimization opportunities and scientists may design new operators not covered by existing libraries. In such scenarios, a more general compiler solution is required. This solution must generate high-performance code that exploits the target architecture and also be fast enough to allow researchers to quickly evaluate their new design. In this session, we will present to the community AKG (Auto Kernel Generator), which is the automatic AI/DL fused operator generation solution of the Open Source framework MindSpore. AKG relies on the polyhedral abstraction of programs to enable automatic parallelization and optimization. We will discuss the software, its architecture, abilities and roadmap, as well as the current top challenges.