OAA: An Abstraction for Efficient Accelerator Adaptation in Deep Learning Frameworks

Deep learning frameworks rely on specific runtime and computation libraries to rewrite the backend for the adaptation of specialized accelerators, which is inefficient and hard to guarantee performance. This paper solves the issue by proposing an Operator Adaptation Abstraction (OAA) that lies between the framework and the libraries. In addition, this paper optimizes training performance based on the hardware characteristics of the accelerator. We designed experiments based on Ascend 910 and OneFlow to verify the effectiveness of OAA. Our adaptation includes more than 50 Ascend operators, achieving up to 2.0x throughput compared to the official adaptation version of PyTorch. The experiments validate that the adaptation method proposed in this paper can effectively retain the advantages of both the framework and the accelerator.