UCLA

In this thesis, we present a novel simulation framework designed for flexible design space exploration over accelerator architectures. A conventional cycle-level simulator can bring accurate simulation results, but it requires enormous efforts to simulate the performance of architectures. Trace-based and datagraph-based simulation frameworks, which modify or argument the ISA, bring some flexibility for architecture exploration, but work at the ISA level and lose high-level information from the compiler. Our framework tries to achieve flexibility of datagraph-based simulator while maintaining high-level compiler information. The main contribution of this work can be divided into two parts. First, an LLVM-IR tracer and parser are developed to generate a datagraph at LLVM-IR level, which contains high-level compiler information and is flexible to be manipulated to reflect the desired architecture change. Second, we build an LLVM-IR datagraph simulator inside gem5 to leverage gem5's existing memory framework and provide accurate simulation results. We believe this framework would be useful for rapid design and verification of new architecture proposals.