Accelerating Speculative Execution in High-Level Synthesis with Cancel Tokens

摘要

We present an improved method for scheduling speculative data paths which relies on cancel tokens to undo computations in mis-speculated paths. Performancewise, this method is considerably faster than lenient execution, and faster than any other known approach applicable for general (including non-pipelined) computation structures. We present experimental evidence obtained by implementing our method as part of the high-level language hardware/software compiler COMRADE.