The demand for scalable, high-performance computing has increased as the size of datasets has grown in recent years. However, the breakdown of Dennard's scaling has led to energy efficiency becoming an important concern in datacenters, and spawned exploration into using power-efficient processors such as GPUs (Graphic Processing Units) and FPGAs (Field-Programmable Gate Arrays) as accelerators in datacenters. In particular, the FPGA's low power consumption and the re-programmability allow datacenters to use FPGAs as highly energy-efficient accelerators for a variety of application. On the other hand, FPGA has poor programmability compared to instructions-based architectures like CPU and GPU. To facilitate the process of implementing and deploying FPGA accelerators, High-Level Synthesis (HLS) that generates functional-equivalent RTL from C-based programming languages attracts more and more attention since past decades. Nowadays, both FPGA vendors have their commercial HLS products -- Xilinx SDx and Intel FPGA SDK for OpenCL. However, modern HLS is still not friendly for software designers who have limited FPGA domain knowledge. Since the hardware architecture inferred from a syntactic C implementation could be ambiguous, current commercial HLS tools usually generate architecture structures according to specific HLS C code patterns. As a result, even though the authors have illustrated that the HLS tool is capable of generating FPGA designs with competitive performance as the one in RTL, designers must manually reconstruct the HLS C kernel with specific code patterns to achieve high performance. This problem becomes one of the main impediments to consolidating the FPGA community on cooperation and developments.
In this dissertation, we first present an automated framework that frees human efforts from code reconstruction and design space exploration (DSE). The framework creates a more comprehensive micro-architecture design space from user-written C-based kernel with the Merlin compiler, so the design point should cover the design point with better performance when compared to the HLS-pragma-based design space. To efficiently identify the best design configuration in the tremendous design space, we first propose efficient design space pruning processes that reduce the design space by 24.65x. Accordingly, we develop and evaluate several approaches, including multi-armed bandit hyper heuristic approach, gradient-based approach, and design bottleneck optimization approach. The evaluation result shows that our DSE framework is able to identify the design point that achieves on average (using geometric mean) 93.78% QoR compared to the corresponding manual design.
Based on the proposed DSE framework, we further support automated design optimization for high level domain specific languages (DSLs). Since DSLs might not explicitly provide interfaces for users to specify design configurations, automatic DSE becomes even more important when supporting DSLs for FPGAs. Specifically, we adopt Merlin C, an OpenMP-like C-based programming model, as the intermediate representation (IR) and implement DSL-to-Merlin front-end compilers while preserving the semantic and domain-specific information such as parallel patterns, systolic patterns, and scheduling functions. We first implement Spark-to-Merlin front-end compiler that translates Spark applications in Scala to Merlin C for FPGA acceleration. By leveraging parallel patterns as scheduling hints, the generated accelerators are able to achieve 50x speedup on geometric mean for a set of machine learning kernels. In addition, we also demonstrate that our DSE framework can be even more practical for the DSLs with plenty scheduling functions. Specifically, we implement HeteroCL-to-Merlin front-end that takes HeteroCL programming model embedded in Python. Our DSE framework is capable of exploring a subset of HeteroCL scheduling primitives and let users focus on the platform independent loop transformations. With the help from the DSE framework, we achieve 27.62x speedup on geometric mean over a CPU core for a variety of compute-intensive kernels (chapter 3).
On the other hand, a main challenge of performing design space exploration for a design with arbitrary functionality is the lack of the assumption to underlying micro-architectures. As we will illustrate in the dissertation, the cost of evaluating the quality of a design point is extremely expensive (15-60 minutes) so only a limited number of design points can be explored. In addition, due to the uncertainty of vendor tool behaviors, the development of performance and resource modeling is also unrealistic. As a result, we propose composable, parallel and pipeline (CPP) architecture template to limit the design space to a certain region that is more practical and has less exceptions (chapter 4). With the CPP architecture, we are able to derive an incremental analytical model, which only requires a few HLS run to be initialized, to facilitate the DSE process.
In the last part of this dissertation, we use convolutional neural network (CNN) to demonstrate that the HLS runtime cost can be totally saved with the use of a more domain specific architecture (chapter 5). Specifically, we leverage a systolic array architecture template for CNN accelerator generation. By mapping a CNN model to the pre-defined systolic array template, we can guarantee the model accuracy and DSE efficiency. The experimental result shows that our analytical model for the architecture template achieves 96% accuracy, and the mapped CNN model achieves up to 1.2 Tops throughput on Intel Arria 10 FPGA.
