Diastolic arrays : throughput-driven reconfigurable computing

Abstract

In this thesis, we propose a new reconfigurable computer substrate: diastolic arrays. Diastolic arrays are arrays of processing elements that communicate exclusively through First-In First-Out (FIFO) queues, and provide hardware support to guarantee bandwidth and buffer space for all data transfers. FIFO control implies that a module idles if its input FIFOs are empty, and stalls if its output FIFOs are full. The timing of data transfers between processing elements in diastolic arrays is therefore significantly more relaxed than in systolic arrays or pipelines. All specified data transfers are statically routed, and the routing problem to maximize average throughput can be optimally or near-optimally solved in polynomial time by formulating it as a maximum concurrent multicommodity flow problem and using linear programming. We show that the architecture of diastolic arrays enables efficient synthesis from high-level specifications of communicating finite state machines, providing a high-performance, off-the-shelf computer substrate that can be easily programmed.