Reconfigurable architectures have been widely used as single core processor accelerators. In the multi-core era, however, it is necessary to review the way that reconfigurable arrays are integrated into multi-core processor. Generally, a set of reconfigurable functional units are employed in a similar way as they are used in single core processors. Unfortunately, a considerable increase in the area ensues from this practice. Besides, in applications with unbalanced workload in their threads this approach can lead to a inefficient use of the reconfigurable architecture in cores with a low or even idle workload. To cope with this issue, this work proposes and evaluates a partially shared thin reconfigurable array, which allows to share reconfigurable resources among the processor's cores. Sharing is performed dynamically by the configuration scheduler hardware. The results shows that the sharing mechanism provided 76% of energy savings, improving the performance 41% in average when compared with a version without the proposed reconfigurable array. A comparison with a version of the reconfigurable array without the sharing mechanism was performed and shows that the sharing mechanism improved up to 11.16% in the system performance.
UltraSynth: Insights of a CGRA Integration into a Control Engineering Environment
