关键词: 通用矩阵乘法, 向量运算, 脉动阵列, 向量处理单元, 数据流调度, 编译器

Abstract: Systolic arrays, as energy-efficient accelerators for general matrix multiplication (GEMM) operators, have garnered widespread attention from both academia and industry. However, they often occupy a substantial amount of area and typically require collaboration with VPU (vector processing unit) components, a combination frequently seen in neural network accelerators. Additionally, they suffer from issues such as low temporal and spatial utilization rates and limited performance in end-to-end scenarios. To address these challenges, a hybrid vector systolic array (HVSA) is proposed by integrating systolic arrays with vector processors. By reusing the storage, broadcasting, and inter-channel communication units within the VPU, this architecture enables reconfigurable capabilities in terms of array shape and data flow, allowing for more efficient support of GEMM and vector operations within an acceptable hardware area overhead. Furthermore, an end-to-end compilation framework tailored for HVSA is introduced, encompassing an MLIR-based compilation frontend, data flow scheduling, and a programming model compatible with the RISC-V vector extension. Experimental data demonstrates that HVSA achieves a 30.30-fold increase in computational speed compared to a systolic array of equivalent area. In end-to-end applications, the average operating time of HVSA is reduced to around 4.7% of the original compared to the "VPU+SA" of the same area, and energy consumption is reduced by approximately 58.7%.

Key words: general matrix multiplication(GEMM), vector operation, systolic array, vector proces- sing unit(VPU), dataflow scheduling, compiler