high performance computing on graphics processing units: hgpu.org

Using high-level synthesis (HLS) tools for field-programmable gate array (FPGA) design is becoming an increasingly popular choice because HLS tools can generate a high-quality design in a short development time. However, current HLS tools still cannot adequately support users in understanding and fixing the performance issues of the current design. That is, current HLS tools lack in performance debugging capability. Previous work on performance debugging automates the process of inserting hardware monitors in low-level register-transfer level (RTL) languages which limits the comprehensibility of the obtained result. Instead, our HLS-based flows offer analysis on a function or loop level and provide more intuitive feedback that can be used to pinpoint the performance bottleneck of a design. In this dissertation, we present a collection of HLS-based debugging frameworks for various purposes and characteristics of the design. First, we address the problem in the HLS synthesis step, where an inaccurate cycle estimation is provided if the program has input-dependent behavior. We propose a new performance estimator that automatically instruments code that models the hardware execution behavior and interprets the information from the HLS software simulation. However, the performance estimation result of this flow may not be accurate for a type of designs that cannot be simulated correctly by existing HLS software simulators. To handle such cases, we propose a new software simulator that provides cycle-accurate result based on the HLS scheduling information. If the input dataset is not available for software simulation or high-level models do not exist for all components of the FPGA design, we also present an on-board monitoring flow for automated cycle extraction and stall analysis. Finally, we address the needs of HLS programmers to automatically find the best set of directives for FPGA designs. We propose a design space exploration (DSE) framework to optimize applications with variable loop bounds in Polybench benchmark. A quantitative comparison among the proposed frameworks is shown using the sparse matrix-vector multiplication benchmark.