Optimization of weight update for convolutional operators for domestic accelerators

Convolutional neural networks play an important role in deep learning, which includes three processes: forward propagation, backpropagation, and weight update. On domestic heterogeneous accelerators, the performance of the weight update operator is somewhat weak. In this paper, we handwritten and optimized the weight update operator on domestic heterogeneous accelerators, including both ordinary convolution and grouped convolution. Among them, ordinary convolution uses the implicit gemm algorithm, Grouped convolution uses a parallelized direct convolution algorithm, which reduces memory usage and has strong universality. Parallel direct convolution algorithms make better use of it.