High-level Synthesis Techniques Research Articles

Accelerating Weather Prediction Using Near-Memory Reconfigurable Fabric

Ongoing climate change calls for fast and accurate weather and climate modeling. However, when solving large-scale weather prediction simulations, state-of-the-art CPU and GPU implementations suffer from limited performance and high energy consumption. These implementations are dominated by complex irregular memory access patterns and low arithmetic intensity that pose fundamental challenges to acceleration. To overcome these challenges, we propose and evaluate the use of near-memory acceleration using a reconfigurable fabric with high-bandwidth memory (HBM). We focus on compound stencils that are fundamental kernels in weather prediction models. By using high-level synthesis techniques, we develop NERO, an field-programmable gate array+HBM-based accelerator connected through Open Coherent Accelerator Processor Interface to an IBM POWER9 host system. Our experimental results show that NERO outperforms a 16-core POWER9 system by \( 5.3\times \) and \( 12.7\times \) when running two different compound stencil kernels. NERO reduces the energy consumption by \( 12\times \) and \( 35\times \) for the same two kernels over the POWER9 system with an energy efficiency of 1.61 GFLOPS/W and 21.01 GFLOPS/W. We conclude that employing near-memory acceleration solutions for weather prediction modeling is promising as a means to achieve both high performance and high energy efficiency.

End-to-End Synthesis of Dynamically Controlled Machine Learning Accelerators

Edge systems are required to autonomously make real-time decisions based on large quantities of input data under strict power, performance, area, and other constraints. Meeting these constraints is only possible by specializing systems through hardware accelerators purposefully built for machine learning and data analysis algorithms. However, data science evolves at a quick pace, and manual design of custom accelerators has high non-recurrent engineering costs: general solutions are needed to automatically and rapidly transition from the formulation of a new algorithm to the deployment of a dedicated hardware implementation. Our solution is the SOftware Defined Architectures (SODA) Synthesizer, an end-to-end, multi-level, modular, extensible compiler toolchain providing a direct path from machine learning tools to hardware. The SODA Synthesizer frontend is based on the multilevel intermediate representation (MLIR) framework; it ingests pre-trained machine learning models, identifies kernels suited for acceleration, performs high-level optimizations, and prepares them for hardware synthesis. In the backend, SODA leverages state-of-the-art high-level synthesis techniques to generate highly efficient accelerators, targeting both field programmable devices (FPGAs) and application-specific circuits (ASICs). In this paper, we describe how the SODA Synthesizer can also assemble the generated accelerators (based on the finite state machine with datapath model) in a custom system driven by a distributed controller, building a coarse-grained dataflow architecture that does not require a host processor to orchestrate parallel execution of multiple accelerators. We show the effectiveness of our approach by automatically generating ASIC accelerators for layers of popular deep neural networks (DNNs). Our high-level optimizations result in up to 74x speedup on isolated accelerators for individual DNN layers, and our dynamically scheduled architecture yields an additional 3x performance improvement when combining accelerators to handle streaming inputs.

Sparse and dense matrix multiplication hardware for heterogeneous multi-precision neural networks

In this paper, we present hardware accelerators created with high-level synthesis techniques for sparse and dense matrix multiplication operations. The cores can operate with different precisions and are designed to be integrated in a heterogeneous CPU-FPGA system for Edge AI applications. The methodology involves quantization-sparsity aware training and it is applied to a case study consisting of human activity classification. We initially investigate the effects of quantization and sparsity on the accuracy of neural networks with convolution, dense and recurrent layers observing better tolerance to pruning when recurrent layers are present. Then, we propose the hardware accelerators that can switch precision at run-time and work with any matrix size up to a maximum configured at compile time. We compare the performance of these accelerators at different levels of precision and sparsity levels and create a performance model to enable workload balancing. The results show that the proposed sparse matrix multipliers can outperform dense multipliers when sparsity levels are higher than 70% and the improvements are more evident when higher precision arithmetic or structural pruning is used. Additionally, sparsity levels as high as 99% can maintain the level of accuracy required in the network especially when recurrent layers are deployed. Overall, the balance between sparse and dense performance depends on matrix shape, precision, structural pruning and sparsity levels and performance modelling can be used to balance concurrent execution in a heterogeneous configuration.

Lossy Hyperspectral Image Compression on a Reconfigurable and Fault-Tolerant FPGA-Based Adaptive Computing Platform

This paper describes a novel hardware implementation of a lossy multispectral and hyperspectral image compressor for on-board operation in space missions. The compression algorithm is a lossy extension of the Consultative Committee for Space Data Systems (CCSDS) 123.0-B-1 lossless standard that includes a bit-rate control stage, which in turn manages the losses the compressor may introduce to achieve higher compression ratios without compromising the recovered image quality. The algorithm has been implemented using High-Level Synthesis (HLS) techniques to increase design productivity by raising the abstraction level. The proposed lossy compression solution is deployed onto ARTICo3, a dynamically reconfigurable multi-accelerator architecture, obtaining a run-time adaptive solution that enables user-selectable performance (i.e., load more hardware accelerators to transparently increase throughput), power consumption, and fault tolerance (i.e., group hardware accelerators to transparently enable hardware redundancy). The whole compression solution is tested on a Xilinx Zynq UltraScale+ Field-Programmable Gate Array (FPGA)-based MPSoC using different input images, from multispectral to ultraspectral. For images acquired by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS), the proposed implementation renders an execution time of approximately 36 s when 8 accelerators are compressing concurrently at 100 MHz, which in turn uses around 20% of the LUTs and 17% of the dedicated memory blocks available in the target device. In this scenario, a speedup of 15.6× is obtained in comparison with a pure software version of the algorithm running in an ARM Cortex-A53 processor.

Special issue on SoC and AI processors

Special issue on SoC and AI processors

Development of Multiobjective High-Level Synthesis for FPGAs

Traditionally, the High-Level Synthesis (HLS) for Field Programmable Gate Array (FPGA) devices is a methodology that transforms a behavioral description, as the timing-independent specification, to an abstraction level that is synthesizable, like the Register Transfer Level. This process can be performed under a framework that is known as Design Space Exploration (DSE), which helps to determine the best design by addressing scheduling, allocation, and binding problems, all three of which are NP-hard problems. In this manner, and due to the increased complexity of modern digital circuit designs and concerns regarding the capacity of the FPGAs, designers are proposing novel HLS techniques capable of performing automatic optimization. HLS has several conflicting metrics or objective functions, such as delay, area, power, wire length, digital noise, reliability, and security. For this reason, it is suitable to apply Multiobjective Optimization Algorithms (MOAs), which can handle the different trade-offs among the objective functions. During the last two decades, several MOAs have been applied to solve this problem. This paper introduces a comprehensive analysis of different MOAs that are suitable to perform HLS for FPGA devices. We highlight significant aspects of MOAs, namely, optimization methods, intermediate structures where the optimizations are performed, HLS techniques that are addressed, and benchmarks and performance assessments employed for experimentation. In addition, we show the analysis of how multiple objectives are optimized currently in the algorithms and which are the objective functions that are optimized. Finally, we provide insights and suggestions to contribute to the solution of major research challenges in this area.

Verification of Scheduling of Conditional Behaviors in High-Level Synthesis

High-level synthesis (HLS) technique translates the behaviors written in high-level languages like C/C++ into register transfer level (RTL) design. Due to its complexity, proving the correctness of an HLS tool is prohibitively expensive. Translation validation is the process of proving that the target code is a correct translation of the source program being compiled. The path-based equivalence checking (PBEC) method is a widely used translation validation method for verification of the scheduling phase of HLS. The existing PBEC methods cannot handle significant control structure modification that occurs in the efficient scheduling of conditional behaviors. Hence, they produce a false-negative result. In this article, we identify some scenarios involving path merge/split where the state-of-the-art PBEC approaches fail to show the equivalence even though behaviors are equivalent. We propose a value propagation-based PBEC method along with a new cutpoint selection scheme to overcome this limitation. Our method can also handle the scenario where adjacent conditional blocks (CBs) having an equivalent conditional expression are combined into one CB. Experimental results demonstrate the usefulness of our method over the existing methods.

FPGA Based Real Time Simulations of the Face Milling Process

The article presents a successful implementation of the milling process simulation at the Field-Programmable Gate Array (FPGA). By using FPGA, very rigorous Real-Time (RT) simulation requirements can be met. The response time of the FPGA simulations is significantly reduced, and the time synchronization is better than in a typical RT system implemented in software. The FPGA-based approach is characterized by enormous flexibility when it comes to input and output operations that can be implemented deterministically in RT. Complex simulation software has been implemented using the High Level Synthesis technique, which is a relatively easy and fast approach for FPGA programming without using complex Hardware Description Languages. The hardware functions are based on procedures written in high-level C programming language. The mathematical descriptions of simulations, results of computer simulations, Hardware-in-the-Loop Simulation experiments, and real experiments are presented. The approach presented in this paper can be used to simulate the dynamics of various mechatronic systems.

A Design Flow Engine for the Support of Customized Dynamic High Level Synthesis Flows

High Level Synthesis is a set of methodologies aimed at generating hardware descriptions starting from specifications written in high-level languages. While these methodologies share different elements with traditional compilation flows, there are characteristics of the addressed problem which require ad hoc management. In particular, differently from most of the traditional compilation flows, the complexity and the execution time of the High Level Synthesis techniques are much less relevant than the quality of the produced results. For this reason, fixed-point analyses, as well as successive refinement optimizations, can be accepted, provided that they can improve the quality of the generated designs. This article presents a design flow engine for the description and the execution of complex and customized synthesis flows. It supports dynamic addition of passes and dependencies, cyclic dependencies, and selective pass invalidation. Experimental results show the benefits of such type of design flows with respect to static linear design flows when applied to High Level Synthesis.

Hybrid Quick Error Detection: Validation and Debug of SoCs Through High-Level Synthesis

Validation and debug challenges of system-on-chips (SoCs) are getting increasingly difficult. As we reach the limits of Dennard scaling, efforts to improve system performance and energy efficiency have resulted in the integration of a wide variety of complex hardware accelerators in SoCs. Hence, it is essential to address the validation and debug of hardware accelerators. High-level synthesis (HLS) is a promising technique to rapidly create customized hardware accelerators. In this paper, we present the hybrid quick error detection ( H -QED) approach that overcomes validation and debug challenges for hardware accelerators by leveraging HLS techniques in both the presilicon and post-silicon stages. H-QED improves error detection latencies (time elapsed from when a bug is activated to when it is detected) by 2–5 orders of magnitude with one cycle latencies in presilicon scenarios and bug coverage threefold higher compared to traditional validation techniques. H-QED also uncovered previously unknown bugs in the CHStone benchmark suite, which is widely used by the HLS community. H-QED incurs an 8% accelerator area overhead with negligible silicon performance impact for post-silicon stage, and we also introduce techniques to minimize any possible intrusiveness introduced by H-QED.

A Streaming Dataflow Engine for Sparse Matrix-Vector Multiplication Using High-Level Synthesis

Using high-level synthesis techniques, this paper proposes an adaptable high-performance streaming dataflow engine for sparse matrix dense vector multiplication (SpMV) suitable for embedded FPGAs. As the SpMV is a memory-bound algorithm, this engine combines the three concepts of loop pipelining , dataflow graph , and data streaming to utilize most of the memory bandwidth available to the FPGA. The main goal of this paper is to show that FPGAs can provide comparable performance for memory-bound applications to that of the corresponding CPUs and GPUs but with significantly less energy consumption. The experimental results indicate that the FPGA provides higher performance compared to that of embedded GPUs for small and medium-size matrices by an average factor of 3.25 whereas the embedded GPU is faster for larger size matrices by an average factor of 1.58. In addition, the FPGA implementation is more energy efficient for the range of considered matrices by an average factor of 8.9 compared to the embedded CPU and GPU. A case study based on adapting the proposed SpMV optimization to accelerate the support vector machine (SVM) algorithm, one of the successful classification techniques in the machine learning literature, justifies the benefits of utilizing the proposed FPGA-based SpMV compared to that of the embedded CPU and GPU. The experimental results show that the FPGA is faster by an average factor of 1.7 and consumes less energy by an average factor of 6.8 compared to the GPU.

Case study of an HEVC decoder application using high-level synthesis: intraprediction, dequantization, and inverse transform blocks

Due to the increasing need for testing solutions to complex hardware designs, several efforts were made in order to improve high-level synthesis (HLS) techniques. These solutions are conceived in such a way that they had to provide reasonable agreements in terms of design time, resources involved, and performance. Generally speaking, two main constraints should be satisfied for HLS applications. The first constraint consists in the ability to process complex systems at a reasonable cost, whereas the second revolves around considering some test constraints in the first tasks of the HLS flow. To fulfill these two constraints, we treated a case study using HLS for intraprediction, dequantization, and inverse transform decoding blocks of an high efficiency video coding (HEVC) decoder. For this experiment, version 10 was used of the HEVC test model (HM) reference software containing >200 functions and over 8000 lines of code. In addition, the suggested algorithm was implemented in an software/hardware (SW/HW) environment using Xilinx ZC 702-based platform. Finally, taking advantage of HLS optimization methods, the hardware design can process 6, 13, 71, and 285 video frames per second for 1600p, 1080p, 480p, and 240p video resolutions, respectively. Yet, the SW/HW designs can only decode 0.5, 1.5, 4, and 15.2 frames per second for the same video resolutions, i.e., with a gain of 3% in frame rate and 60% in power consumption compared to SW implementation.

Exact and Practical Modulo Scheduling for High-Level Synthesis

Loop pipelining is an essential technique in high-level synthesis to increase the throughput and resource utilisation of field-programmable gate array--based accelerators. It relies on modulo schedulers to compute an operator schedule that allows subsequent loop iterations to overlap partially when executed while still honouring all precedence and resource constraints. Modulo schedulers face a bi-criteria problem: minimise the initiation interval (II; i.e., the number of timesteps after which new iterations are started) and minimise the schedule length. We present Moovac, a novel exact formulation that models all aspects (including the II minimisation) of the modulo scheduling problem as a single integer linear program, and discuss simple measures to prevent excessive runtimes, to challenge the old preconception that exact modulo scheduling is impractical. We substantiate this claim by conducting an experimental study covering 188 loops from two established high-level synthesis benchmark suites, four different time limits, and three bounds for the schedule length, to compare our approach against a highly tuned exact formulation and a state-of-the-art heuristic algorithm. In the fastest configuration, an accumulated runtime of under 16 minutes is spent on scheduling all loops, and proven optimal IIs are found for 179 test instances.

Low-Power and High-Speed Deep FPGA Inference Engines for Weed Classification at the Edge

Deep neural networks (DNNs) have recently achieved remarkable performance in a myriad of applications, ranging from image recognition to language processing. Training such networks on graphics processing units (GPUs) currently offers unmatched levels of performance; however, GPUs are subject to large-power requirements. With recent advancements in high-level synthesis (HLS) techniques, new methods for accelerating deep networks using field programmable gate arrays (FPGAs) are emerging. FPGA-based DNNs present substantial advantages in energy efficiency over conventional CPU- and GPU-accelerated networks. Using the Intel FPGA software development kit (SDK) for OpenCL development environment, networks described using the high-level OpenCL framework can be accelerated targeting heterogeneous platforms including CPUs, GPUs, and FPGAs. These networks, if properly customized on GPUs and FPGAs, can be ideal candidates for learning and inference in resource-constrained portable devices such as robots and the Internet of Things (IoT) edge devices, where power is limited and performance is critical. Here, we introduce GPU- and FPGA-accelerated deterministically binarized DNNs, tailored toward weed species classification for robotic weed control. Our developed networks are trained and benchmarked using a publicly available weed species dataset, named DeepWeeds, which include close to 18 000 weed images. We demonstrate that our FPGA-accelerated binarized networks significantly outperform their GPU-accelerated counterparts, achieving a>7-fold decrease in power consumption, while performing inference on weed images 2.86 times faster compared to our best performing baseline full-precision GPU implementation. These significant benefits are gained whilst losing only 1.17% of validation accuracy. In this paper, this is a significant step toward enabling deep inference and learning on IoT edge devices, and smart portable machines such as agricultural robots, which is the target application.

Enabling Automated Bug Detection for IP-Based Designs Using High-Level Synthesis

Editor’s note: This article presents an automated approach for detecting system- level bugs in SoC designs that are composed of many IP blocks, without exposing sensitive information. The approach leverages high-level synthesis techniques. —Binoy Ravindran, Virginia Tech

Data Transfers Analysis in Computer Assisted Design Flow of FPGA Accelerators for Aerospace Systems

The integration of Field Programmable Gate Arrays (FPGAs) in an aerospace system improves its efficiency and its flexibility thanks to their programmability, but increases the design complexity. The design flows indeed have to be composed of several steps to fill the gap between the starting solution, which is usually a reference sequential implementation, and the final heterogeneous solution which includes custom hardware accelerators. Among these steps, there are the analysis of the application to identify the functionalities that gain advantages in execution on hardware and the generation of their implementations by means of Hardware Description Languages. Generating these descriptions for a software developer can be a very difficult task because of the different programming paradigms of software programs and hardware descriptions. To facilitate the developer in this activity, High Level Synthesis techniques have been developed aiming at (semi-)automatically generating hardware implementations of specifications written in high level languages (e.g., C). With respect to other embedded systems scenarios, the aerospace systems introduce further constraints that have to be taken into account during the design of these heterogeneous systems. In this type of systems explicit data transfers to and from FPGAs are preferred to the adoption of a shared memory architecture. The first approach indeed potentially improves the predictability of the produced solutions, but the sizes of all the data transferred to and from any devices must be known at design time. Identifying the sizes in presence of complex C applications which use pointers can be a not so easy task. In this paper, a semi-automatic design flow based on the integration of an aerospace design flow, an application analysis technique, and High Level Synthesis methodologies is presented. The initial reference application is analyzed to identify which are the sizes of the data exchanged among the different components of the application. Next, starting from the high level specification and from the results of this analysis, High Level Synthesis techniques are applied to automatically produce the hardware accelerators.

Functional Model-Based Design Methodology for Automotive Cyber-Physical Systems

The high complexity of cross-domain engineering in combination with the pressure for system innovation, higher quality, time to market, and budget constraints makes it imperative for automotive companies to use integrated engineering methods and tools. Computer engineering tools are mainly focused on a particular domain, and therefore, it is difficult to integrate different tools for system-level analysis. In this paper, a novel multidisciplinary systems engineering methodology and associated design automation algorithms for the complex automotive cyber-physical systems are presented. Rather than starting from the domain-specific architecture/simulation models where most resources are spent, we preemptively target the early design stage at the functional level that determines 75% of an automobile's cost. In our methodology, the marriage of systems engineering principles with high-level synthesis techniques from design automation area results in a functional modeling compiler capable of generating high-fidelity simulation models for the design space exploration and validation of multiple cyber-physical automotive architectures. Using real-world automotive use cases, we demonstrate how functional models capturing integrated cyber-physical aspects are synthesized into high-fidelity multidomain simulation models.

Lossless Compression Decoders for Bitstreams and Software Binaries Based on High-Level Synthesis

As the density of field-programmable gate arrays continues to increase, the size of configuration bitstreams grows accordingly. Compression techniques can reduce memory size and save external memory bandwidth. To accelerate the configuration process and reduce the software startup time, four open-source lossless compression decoders developed using high-level synthesis techniques are presented. Moreover, in order to balance the objectives of compression ratio, decompression throughput, and hardware resource overhead, various improvements and optimizations are proposed. Full bitstreams and software binaries have been collected as a benchmark, and 33 partial bitstreams have also been developed and integrated into the benchmark. Evaluations of the synthesizable compression decoders are demonstrated on a Xilinx ZC706 board, showing higher decompression throughput than those of the existing lossless compression decoders using our benchmark. The proposed decoders can reduce software startup time by up to 31.23% in embedded systems and 69.83% reduction of reconfiguration time for partial reconfigurable systems.

Custom Multicache Architectures for Heap Manipulating Programs

Memory-intensive implementations often require access to an external, off-chip memory which can substantially slow down an field-programmable gate array accelerator due to memory bandwidth limitations. Buffering frequently reused data on chip is a common approach to address this problem and the optimization of the cache architecture introduces yet another complex design space. This paper presents a high-level synthesis (HLS) design aid that automatically generates parallel multicache systems which are tailored to the specific requirements of the application. Our program analysis identifies nonoverlapping memory regions, supported by private caches, and regions which are shared by parallel units after parallelization, which are supported by coherent caches and synchronization primitives. It also decides whether the parallelization is legal with respect to data dependencies. The novelty of this paper is the focus on programs using dynamically allocated, pointer-based data structures which, while common in software engineering, remain difficult to analyze and are beyond the scope of the overwhelming majority of HLS techniques to date. Second, we devise a high-level cache performance estimation to find a heterogeneous configuration of cache sizes that maximizes the performance of the multicache system subject to an on-chip memory resource constraint. We demonstrate our technique with three case studies of applications using dynamic data structures and use Xilinx Vivado HLS as an exemplary HLS tool. We show up to $\boldsymbol {15}\boldsymbol {\times }$ speed-up after parallelization of the HLS implementations and the insertion of the application-specific distributed hybrid multicache architecture.

On supporting rapid prototyping of embedded systems with reconfigurable architectures

Reducing time-to-market while improving product quality is a big challenge. This paper proposes a software-supported framework for rapid prototyping that offers a concurrent fast hardware/software system-level design. The introduced framework enables the constant evaluation and verification of the prototype under development, while it provides automatic functionality mapping to hardware via High-Level Synthesis techniques. We evaluate our framework and its software instantiation with a computer vision algorithm. Based on our experimentation, we show that our approach reduces the development time by almost 64×, it prunes the hardware design space by 34×, while maintaining designs that trade-off high Quality-of-Report on the Pareto frontier.