Optimization of Convolutional Operators for Domestic Accelerators

Convolutional operation is the core operation of convolutional neural networks, which is widely used in deep learning tasks such as image classification, object detection, and semantic segmentation, which occupies most of the computational time and resource consumption. The Winograd convolution algorithm is widely used for the accelerated calculation of small convolutional kernels because it significantly reduces the amount of multiplication operations. However, although the algorithm can significantly reduce the computational complexity in theory, the actual performance is still affected by the memory access mode, computational efficiency and chunking strategy. How to optimize the synergy between these three aspects on the domestic accelerator has become the key to improve the performance of Winograd convolution in practical applications. In this paper, a fixed block Winograd convolution optimization algorithm for domestic accelerators is proposed to improve the performance of the convolution operator. Firstly, the thread block is divided based on the number of hardware computing unit arrays, so as to reduce the scheduling overhead caused by the switching of different thread blocks between computing resources. Secondly, the multilevel storage system is used to cache and manage the data at each stage of the Winograd algorithm to improve the efficiency of memory access. Finally, combined with the domestic accelerator hardware TensorCore, the core multiplication stage is optimized to further shorten the calculation time. Experimental results show that the parameter performance of the optimized operator in this paper exceeds that of the same type of operator in the domestic acceleration library by 60%, and the overall performance is about 10% higher than that of the optimal operator in the database.