Multi-level parallelism optimization for two-dimensional convolution vectorization method on multi-core vector accelerator

The widespread application of convolutional neural network across diverse domains has highlighted the growing significance of accelerating convolutional computations. In this work, we design a multi-level parallelism optimization method for direct convolution vectorization algorithm based on a channel-first data layout on a multi-core vector accelerator. This method calculates based on the input row and weight column in a single core, and achieves the simultaneous computation of more elements, thereby effectively hiding the latency of instructions and improving the degree of parallelism at instruction-level. This method can also substantially eliminates data overlap caused by convolutional windows sliding. Among multiple cores, the data flow is optimized with various data reuse methods for different situations. Experimental results show that the computational efficiency on multi-core can be improved greatly, up to 80.2%. For the typical network ResNet18, compared with existing method on the accelerator, a performance acceleration of 4.42-5.63 times can be achieved.