Abstract

Previous work introduced a dynamic compute model aimed at eliminating existing barriers to the widespread efficient exploitation of reconfigurable devices. Among other achievements this model decoupled application design-time decisions from the run-time physical resource bindings. The compute model uses graphs of compute pages and memory blocks connected by stream links to capture the definition of a computation abstracted from the detailed hardware size. An automatic run-time scheduler is a required component in this compute model in that it selects the temporal sequencing of virtual resources onto the physical device, allocates hardware resources, and configures the device. Although such a scheduler could be computationally expensive, this work describes a quasi-static scheduling strategy that dramatically reduces run-time overhead without restricting the full semantic power of the dynamic dataflow graphs. This work describes the quasi-static scheduling system, analyzes the trade-offs involved in selecting a scheduler implementation, and highlights critical algorithms. It pays particular attention to the temporal partitioning of compute graphs and the management of live computation state.

Report

• Full report, PDF (115 pages)
• Abstract only, PDF (1 page)