Optimizing Half Precision Winograd Convolutions on Multi-Core Vector Processors

Abstract Winograd-based algorithm is currently one of the optimization methods in convo-lutional computation. It can effectively reduce the number of multiplications and computational complexity. This paper proposes a Winograd-based convolution accuracy compensation method for low precision vector processor based on the architecture of low precision multi-core vector processors. And The we implement a multi-level parallel optimization scheme for Winograd-based convolution on the multi-core vector processor. The proposed low precision compensation scheme for Winograd-based convolution can effectively improve the adverse effects of accuracy loss caused by Winograd-based convolution on low precision platforms. The implemented parallel optimization scheme develops parallelism from different dimensions, and effectively improves computational performance. The experimental results show that the proposed parallel optimization method achieves a performance acceleration of 26.81 ∼ 131.18 times compared with the fast con-volution implementation in NNPACK. Compared with direct convolution on the same platform, it can achieve an acceleration ratio of 1.18 ∼ 3.28 times.