SkyCastle: A Resource-Aware Multi-Loop Scheduler for High-Level Synthesis

Abstract

A common optimisation problem in the high-level synthesis (HLS) of FPGA-based accelerators is to find a microarchitecture that maximises the performance while keeping the utilisation of the device s low-level resources below certain limits. We propose to tackle it directly as part of the HLS scheduler. To that end, we formalise a general, integrated scheduling and allocation problem for HLS kernels, and present SkyCastle, a novel resource-aware multi-loop scheduler using integer linear programming to solve it for a subclass of kernels composed of multiple, nested loops. In order to demonstrate the practical applicability of the approach, we model the scheduler in such a way as to be plug-in compatible with the Xilinx Vivado HLS engine, allowing the computed solutions to be fed back into its synthesis flow. We evaluate SkyCastle for three non-trivial kernels from the machine learning, signal processing, and physical simulation domains, on two FPGA devices. Additionally, we investigate the replication of slightly slower, but smaller accelerators as a means to further boost the overall performance. In contrast to Vivado HLS default settings, which aim at maximum performance but may fail in later synthesis steps, the solutions computed by our scheduler always result in synthesisable designs.