Combining dynamic and static scheduling in high-level synthesis

Abstract

Field Programmable Gate Arrays (FPGAs) are starting to become mainstream devices for custom computing, particularly deployed in data centres. However, using these FPGA devices requires familiarity with digital design at a low abstraction level. In order to enable software engineers without a hardware background to design custom hardware, high-level synthesis (HLS) tools automatically transform a high-level program, for example in C/C++, into a low-level hardware description.

A central task in HLS is scheduling: the allocation of operations to clock cycles. The classic approach to scheduling is static, in which each operation is mapped to a clock cycle at compile time, but recent years have seen the emergence of dynamic scheduling, in which an operation’s clock cycle is only determined at run-time. Both approaches have their merits: static scheduling can lead to simpler circuitry and more resource sharing, while dynamic scheduling can lead to faster hardware when the computation has a non-trivial control flow.

This thesis proposes a scheduling approach that combines the best of both worlds. My idea is to use existing program analysis techniques in software designs, such as probabilistic analysis and formal verification, to optimize the HLS hardware. First, this thesis proposes a tool named DASS that uses a heuristic-based approach to identify the code regions in the input program that are amenable to static scheduling and synthesises them into statically scheduled components, also known as static islands, leaving the top-level hardware dynamically scheduled. Second, this thesis addresses a problem of this approach: that the analysis of static islands and their dynamically scheduled surroundings are separate, where one treats the other as black boxes. We apply static analysis including dependence analysis between static islands and their dynamically scheduled surroundings to optimize the offsets of static islands for high performance. We also apply probabilistic analysis to estimate the performance of the dynamically scheduled part and use this information to optimize the static islands for high area efficiency. Finally, this thesis addresses the problem of conservatism in using sequential control flow designs which can limit the throughput of the hardware. We show this challenge can be solved by formally proving that certain control flows can be safely parallelised for high performance. This thesis demonstrates how to use automated formal verification to find out-of-order loop pipelining solutions and multi-threading solutions from a sequential program.