A Temporal and Spatial Code Compression and Decompression Technique for Coarse-Grained Reconfigurable Architectures

Abstract

In this thesis, we propose an efficient and lightweight configuration memory compression technique to reduce the space required to store the execution plan and enhance runtime energy efficiency for coarse-grained reconfigurable architectures. As the industry requires more flexibility to adapt the fast growth of applications, coarse-grained reconfigurable architectures have received more attention. However, the area and power overhead of the configuration memory that stores the execution plan of a loop kernel are significant and hinder a broader deployment of CGRA chips. To relieve this problem, we introduce a method to compress configuration memory by removing consecutive duplicated lines. Our compression technique uses two optimizations, spatial and temporal optimization, to generate more compression-friendly code. The temporal optimization sets control signals of the hardware entities in the CGRA to the same values of its neighboring lines and makes them as similar to them as possible without affecting the outcome of the computation. As a result, it minimizes the number of configuration changes for individual entities. The spatial optimization divides the configuration memory into sub-partitions to increase the potential of duplication of neighboring configuration lines. This thesis suggests two techniques, each in order to effectively apply spatial and temporal optimization. Decompression is performed by a simple hardware decoder logic that is able to decode lines with no additional latency and negligible area overhead. Experiments with 193 loop kernels extracted from thirty real-world applications show that the proposed compression technique achieves a memory reduction rate by almost 60% on average. The compression of unseen code accomplishes a memory reduction of over 45 % on average. As a result, our technique results in about a 35 to 70% reduction in energy consumption in the configuration memory in diverse types of experiments.