Research on Convolution Operator Generation Based on MLIR

In recent years, convolutional neural networks have become increasingly important in deep learning. Convolution operation is the core computing unit in convolutional neural networks, which directly affects the computational efficiency of the network. However, there are a large number of convolution operations with different parameters on convolutional neural networks, and manually optimizing the convolution operations for each set of parameters is too costly. In addition, with the rapid iteration of hardware platforms and the continuous evolution of deep learning frameworks, convolution operators need to be manually optimized multiple times. To solve the problem of handwritten convolution operators mentioned above, this paper designs and implements an MLIR convolution operator generator adapted to domestic accelerators, aiming to combine the advantages of MLIR and the characteristics of domestic accelerators to automatically generate convolution operators. The main work of this article is as follows: a hierarchical convolution operator generation framework was constructed to achieve optimization experiments of convolution at different granularities; Based on the Implicit GEMM algorithm, Conv Reflect was designed to implement the computation logic of convolution on domestic accelerators. The experimental results show that under FP32 accuracy, the performance of the generator can reach 85% of the domestic handwritten operator library, which basically meets the requirements of industrial applications. However, there is still a certain gap in performance under FP16 accuracy, and there is room for further optimization.