ASFRM: An Array State Self-feedback-Based Self-reconfiguration Mechanism in a Reconfigurable Array Processor

With complex tasks in neural network computing, existing reconfiguration methods have a long reconfiguration time and it is not easy to achieve fast task switching. Therefore, this paper presents an array state self-feedback based self-reconfiguration mechanism (ASFRM) to quickly achieve autonomous task switching and dynamic scheduling based on task execution status. This mechanism is based on the array state information autonomously fed back by the reconfigurable array, and the host controller makes autonomous judgments and decisions on the current execution status of the array, achieving dynamic task rearrangement, dynamic scheduling of the array and fast task switching. The proposed ASFRM can switch configurations between different tasks within 6 clock cycles. To verify the correctness and efficiency of ASFRM, we modeled the hardware with synthesizable RTL encoding and implemented it on FPGA and chip. The experimental results show that compared with traditional reconfiguration methods and SIMD-like methods, ASFRM can effectively reduce reconfiguration overhead. The volume of bits in the configuration file has decreased by an average of 11.79% and 4.51%. The reconfiguration time has decreased by an average of 52.79% and 37.12%. Based on the UMC 55[Formula: see text]nm process, the operating frequency can reach 400[Formula: see text]MHz. Meanwhile, the chip area is 64[Formula: see text]mm 2 and the throughput is 115.2[Formula: see text]GOPS.