Diffusion models have gained significant popularity in image generation tasks. However, generating high-quality content remains notably slow because it requires running model inference over many time steps. To accelerate these models, we propose to aggressively quantize both weights and activations, while simultaneously promoting significant activation sparsity. We further observe that the stated sparsity pattern varies among different channels and evolves across time steps. To support this quantization and sparsity scheme, we present a novel diffusion model accelerator featuring a heterogeneous mixed-precision dense-sparse architecture, channel-last address mapping, and a time-step-aware sparsity detector for efficient handling of the sparsity pattern. Our 4-bit quantization technique demonstrates superior generation quality compared to existing $\mathbf{4}$-bit methods. Our custom accelerator achieves $6.91 \times$ speed-up and 51.5% energy reduction compared to traditional dense accelerators.

With the ever-increasing hardware design complexity comes the realization that efforts required for hardware verification increase at an even faster rate. Driven by the push from the desired verification productivity boost and the pull from leap-ahead capabilities of machine learning (ML), recent years have witnessed the emergence of exploiting ML-based techniques to improve the efficiency of hardware verification. In this article, we present a panoramic view of how ML-based techniques are embraced in hardware design verification, from formal verification to simulation-based verification, from academia to industry, and from current progress to future prospects. We envision that the adoption of ML-based techniques will pave the road for more scalable, more intelligent, and more productive hardware verification.

Self-attention matrices of Transformers are often highly sparse because the relevant context of each token is typically limited to just a few other tokens in the sequence. To reduce the computational burden of self-attention on Transformer inference, we propose static, structured, sparse attention masks that split attention matrices into dense regions, skipping computations outside these regions while reducing computations inside these regions. To support the proposed mask structure, we design an entropy-aware finetuning algorithm to naturally encourage attention sparsity while maximizing task accuracy. Furthermore, we extend a typical dense deep learning accelerator to efficiently exploit our structured sparsity pattern. Compared to a dense baseline, we achieve 56.6% reduction in energy consumption, 58.9% performance improvement with <1% accuracy loss and 2.6% area overhead.

Reasoning high-level abstractions from bit-blasted Boolean networks (BNs) such as gate-level netlists can significantly benefit functional verification, logic minimization, datapath synthesis, malicious logic identification, etc. Mostly, conventional reasoning approaches leverage structural hashing and functional propagation, suffering from limited scalability and inefficient usage of modern computing power. In response, we propose a novel symbolic reasoning framework exploiting graph neural networks (GNNs) and GPU acceleration to reason high-level functional blocks from gate-level netlists, namely Gamora, which offers high reasoning performance w.r.t exact reasoning algorithms, strong scalability to BNs with over 33 million nodes, and generalization capability from simple to complex designs. To further demonstrate the capability of Gamora, we also evaluate its reasoning performance after various technology mapping options, since technology-dependent optimizations are known to make functional reasoning much more challenging. Experimental results show that (1) Gamora reaches almost 100% and over 97% reasoning accuracy for carry-save-array (CSA) and Booth-encoded multipliers, respectively, with up to six orders of magnitude speedups compared to the state-of-the-art implementation in the ABC framework; (2) Gamora maintains high reasoning accuracy (>92%) in finding functional modules after complex technology mapping, and we comprehensively analyze the impacts on Gamora reasoning from technology mapping. Gamora is available at https://github.com/Yu-Utah/Gamora.

Data clipping is crucial in reducing noise in quantization operations and improving the achievable accuracy of quantization-aware training (QAT). Current practices rely on heuristics to set clipping threshold scalars and cannot be shown to be optimal. We propose Optimally Clipped Tensors And Vectors (OCTAV), a recursive algorithm to determine MSE-optimal clipping scalars. Derived from the fast Newton-Raphson method, OCTAV finds optimal clipping scalars on the fly, for every tensor, at every iteration of the QAT routine. Thus, the QAT algorithm is formulated with provably minimum quantization noise at each step. In addition, we reveal limitations in common gradient estimation techniques in QAT and propose magnitude-aware differentiation as a remedy to further improve accuracy. Experimentally, OCTAV-enabled QAT achieves state-of-the-art accuracy on multiple tasks. These include training-from-scratch and retraining ResNets and MobileNets on ImageNet, and Squad fine-tuning using BERT models, where OCTAV-enabled QAT consistently preserves accuracy at low precision (4-to-6-bits). Our results require no modifications to the baseline training recipe, except for the insertion of quantization operations where appropriate.

Representing deep neural networks (DNNs) in low-precision is a promising approach to enable efficient acceleration and memory reduction. Previous methods that train DNNs in low-precision typically keep a copy of weights in high-precision during the weight updates. Directly training with low-precision weights leads to accuracy degradation due to complex interactions between the low-precision number systems and the learning algorithms. To address this issue, we develop a co-designed low-precision training framework, termed LNS-Madam, in which we jointly design a logarithmic number system (LNS) and a multiplicative weight update algorithm (Madam). We prove that LNS-Madam results in low quantization error during weight updates, leading to stable performance even if the precision is limited. We further propose a hardware design of LNS-Madam that resolves practical challenges in implementing an efficient datapath for LNS computations. Our implementation effectively reduces energy overhead incurred by LNS-to-integer conversion and partial sum accumulation. Experimental results show that LNS-Madam achieves comparable accuracy to full-precision counterparts with only 8 bits on popular computer vision and natural language tasks. Compared to FP32 and FP8, LNS-Madam reduces the energy consumption by over 90% and 55%, respectively.

Transformers have transformed the field of natural language processing. Their superior performance is largely attributed to the use of stacked “self-attention” layers, each of which consists of matrix multiplies as well as softmax operations. As a result, unlike other neural networks, the softmax operation accounts for a significant fraction of the total run-time of Transformers. To address this, we propose Softermax, a hardware-friendly softmax design. Softermax consists of base replacement, low-precision softmax computations, and an online normalization calculation. We show Softermax results in 2.35x the energy efficiency at 0.90x the size of a comparable baseline, with negligible impact on network accuracy.

Image processing applications can benefit tremendously from FPGA acceleration. However, hardware accelerators for these applications look very different from the programs that image processing algorithm designers are accustomed to writing. As a result, many image processing hardware compilers have been designed to generate hardware accelerators from high-level specifications of image processing algorithms. Unfortunately, all of these compilers either exclude crucial access patterns, do not scale to realistic size applications, or rely on a compilation process in which each stage of the application is an independently scheduled module that sends data to its consumers through FIFOs which adds resource and energy overhead while inhibiting synthesis optimizations. In this paper we present a new algorithm for compiling image processing applications, Clockwork, that uses a combination of techniques from polyhedral analysis and synchronous dataflow (SDF) to overcome these limitations. Clockwork compiles the entire application into one flat, statically scheduled module. As a result, accelerators produced by Clockwork have fixed latency, cannot deadlock, and have no resource overhead from inter-stage FIFOs. We show that designs generated by Clockwork achieve on average a 55% reduction in LUTs, a 30% reduction in flip-flops, and a 22% reduction in BRAMs compared to a state-of-the-art stencil compiler at the same throughput, while handling a wider range of access patterns. Clockwork scales to applications with more than 100,000 LUTs. For an application with dozens of stages, Clockwork achieves energy efficiency 260x that of an 8 thread Intel CPU, 17x that of an NVIDIA K80 GPU, and 2.4x that of an NVIDIA V100 GPU.

Image processing algorithms can benefit tremendously from hardware acceleration. However, hardware accelerators for image processing algorithms look very different from the programs that image processing algorithm designers are accustomed to writing. Many image processing hardware compilers have been proposed to close this gap. Unfortunately, all of them either exclude crucial access patterns, do not scale to realistic size applications, or rely on a compilation process in which each stage of the application is an independently scheduled module that sends data to its consumers through FIFOs, which adds resource and energy overhead while inhibiting synthesis optimizations. In this work we present a new algorithm for compiling image processing applications to hardware, Clockwork, that combines insights from polyhedral analysis and synchronous dataflow to overcome these limitations. Clockwork achieves an average of 43% reduction in LUTs, 22% reduction in flip-flops, and 17% reduction in BRAMs compared to a state-of-the-art stencil compiler at the same throughput while handling a wider range of access patterns. For an image processing application with dozens of stages Clockwork achieves energy efficiency 265x that of an 8 core CPU, 17x that of an NVIDIA K80 GPU, and 2.4x that of an NVIDIA V100 GPU.

Loop pipelining is an important optimization in high-level synthesis to enable high-throughput pipelined execution of loop iterations. However, current pipeline scheduling approach relies on fundamentally inexact heuristics based on ad hoc priority functions and lacks guarantee on achieving the best throughput. To address this shortcoming, we propose a scheduling algorithm based on system of integer difference constraints (SDC) and Boolean satisfiability (SAT) to exactly handle various pipeline scheduling constraints. Our techniques take advantage of conflict-driven learning and problem-specific specialization to optimally yet efficiently derive pipelining solutions. Experiments demonstrate that our approach achieves notable speedup in comparison to integer linear programming based techniques.

Deep neural networks have been adopted in a wide range of application domains, leading to high demand for inference accelerators. However, the high cost associated with ASIC hardware design makes it challenging to build custom accelerators for different targets. To lower design cost, we propose MAGNet, a modular accelerator generator for neural networks. MAGNet takes a target application consisting of one or more neural networks along with hardware constraints as input and produces synthesizable RTL for a neural network accelerator ASIC as well as valid mappings for running the target networks on the generated hardware. MAGNet consists of three key components: (i) MAGNet Designer, a highly configurable architectural template designed in C++ and synthesizable by high-level synthesis tools. MAGNet Designer supports a wide range of design-time parameters such as different data formats, diverse memory hierarchies, and dataflows. (ii) MAGNet Mapper, an automated framework for exploring different software mappings for executing a neural network on the generated hardware. (iii) MAGNet Tuner, a design space exploration framework encompassing the designer, the mapper, and a deep learning framework to enable fast design space exploration and co-optimization of architecture and application. We demonstrate the utility of MAGNet by designing an inference accelerator optimized for image classification application using three different neural networks-AlexNet, ResNet, and DriveNet. MAGNet-generated hardware is highly efficient and leverages a novel multi-level dataflow to achieve 40 fJ/op and 2.8 TOPS/mm2in a 16nm technology node for the ResNet-50 benchmark with <; 1% accuracy loss on the ImageNet dataset.

While high-level synthesis (HLS) offers sophisticated techniques to optimize designs for area and performance, HLS-estimated resource usage and timing often deviate significantly from actual quality of results (QoR) achieved by FPGA-targeted designs. Inaccurate HLS estimates prevent designers from performing meaningful design space exploration without resorting to the time-consuming downstream implementation process. To address this challenge, we first build a large collection of C-to-FPGA results from a diverse set of realistic HLS applications and identify relevant features from HLS reports for estimating post-implementation metrics. We then leverage these features and data to train and compare a number of promising machine learning models to effectively and efficiently bridge the accuracy gap. Experiments demonstrate that our proposed approach is able to dramatically reduce the estimation errors for different families of FPGA devices. By extracting domain-specific insights from our experiments, we explore the implications of our models and predictive influence of various features for enabling fast and accurate QoR estimation in HLS. We have released our dataset to springboard future efforts in this area.

Despite increasing adoption of high-level synthesis (HLS) for its design productivity advantage, success in achieving high quality-of-results out-of-the-box is often hindered by the inexactness of the common HLS optimizations. In particular, while scheduling forms the algorithmic core to HLS technology, current scheduling algorithms rely heavily on fundamentally inexact heuristics that make ad hoc local decisions and cannot accurately and globally optimize over a rich set of constraints. To tackle this challenge, we propose a scheduling formulation based on system of integer difference constraints (SDC) and Boolean satisfiability (SAT) to exactly handle a variety of scheduling constraints. We develop a specialized scheduler based on conflict-driven learning and problem-specific knowledge to optimally and efficiently solve the resource-constrained scheduling problem. By leveraging the efficiency of SDC algorithms and scalability of modern SAT solvers, our scheduling technique is able to achieve on average over 100x improvement in runtime over the integer linear programming (ILP) approach while attaining optimal latency. By integrating our scheduling formulation into a state-of-the-art open-source HLS tool, we further demonstrate the applicability of our scheduling technique with a suite of representative benchmarks targeting FPGAs.

Modern high-level synthesis (HLS) tools greatly reduce the turn-around time of designing and implementing complex FPGA-based accelerators. They also expose various optimization opportunities, which cannot be easily explored at the register-transfer level. With the increasing adoption of the HLS design methodology and continued advances of synthesis optimization, there is a growing need for realistic benchmarks to (1) facilitate comparisons between tools, (2) evaluate and stress-test new synthesis techniques, and (3) establish meaningful performance baselines to track progress of the HLS technology. While several HLS benchmark suites already exist, they are primarily comprised of small textbook-style function kernels, instead of complete and complex applications. To address this limitation, we introduce Rosetta, a realistic benchmark suite for software programmable FPGAs. Designs in Rosetta are fully-developed applications. They are associated with realistic performance constraints, and optimized with advanced features of modern HLS tools. We believe that Rosetta is not only useful for the HLS research community, but can also serve as a set of design tutorials for non-expert HLS users. In this paper we describe the characteristics of our benchmarks and the optimization techniques applied to them. We further report experimental results on an embedded FPGA device as well as a cloud FPGA platform.

Specialized hardware accelerators are being increasingly integrated into today's computer systems to achieve improved performance and energy efficiency. However, the resulting variety and complexity make it challenging to ensure the security of these accelerators. To mitigate complexity while guaranteeing security, we propose a high-level synthesis (HLS) infrastructure that incorporates static information flow analysis to enforce security policies on HLS-generated hardware accelerators. Our security-constrained HLS infrastructure is able to effectively identify both explicit and implicit information leakage. By detecting the security vulnerabilities at the behavioral level, our tool allows designers to address these vulnerabilities at an early stage of the design flow. We further propose a novel synthesis technique in HLS to eliminate timing channels in the generated accelerator. Our approach is able to remove timing channels in a verifiable manner while incurring lower performance overhead for high-security tasks on the accelerator.

Modern high-level synthesis (HLS) tools commonly employ pipelining to achieve efficient loop acceleration by overlapping the execution of successive loop iterations. While existing HLS pipelining techniques obtain good performance with low complexity for regular loop nests, they provide inadequate support for effectively synthesizing irregular loop nests. For loop nests with dynamic-bound inner loops, current pipelining techniques require unrolling of the inner loops, which is either very expensive in resource or even inapplicable due to dynamic loop bounds. To address this major limitation, this paper proposes ElasticFlow, a novel architecture capable of dynamically distributing inner loops to an array of processing units (LPUs) in an area-efficient manner. The proposed LPUs can be either specialized to execute an individual inner loop or shared among multiple inner loops to balance the tradeoff between performance and area. A customized banked memory architecture is proposed to coordinate memory accesses among different LPUs to maximize memory bandwidth without significantly increasing memory footprint. We evaluate ElasticFlow using a variety of real-life applications and demonstrate significant performance improvements over a state-of-the-art commercial HLS tool for Xilinx FPGAs.

Traditional techniques for pipeline scheduling in high-level synthesis for FPGAs assume an additive delay model where each operation incurs a pre-characterized delay. While a good approximation for some operation types, this fails to consider technology mapping, where a group of logic operations can be mapped to a single look-up table (LUT) and together incur one LUT worth of delay. We propose an exact formulation of the throughput-constrained, mapping-aware pipeline scheduling problem for FPGA-targeted high-level synthesis with area minimization being a primary objective. By taking this cross-layered approach, our technique is able to mitigate the pessimism inherent in static delay estimates and reduce the usage of LUTs and pipeline registers. Experimental results using our method demonstrate improved resource utilization for a number of logic-intensive, real-life benchmarks compared to a state-of-the-art commercial HLS tool for Xilinx FPGAs.

Scheduling plays a central role in high-level synthesis, as it inserts clock boundaries into the untimed behavioral model and greatly impacts the performance, power, and area of the synthesized circuits. While current scheduling techniques can make use of pre-characterized delay values of individual operations, it is difficult to obtain accurate timing estimation on a cluster of operations without considering technology mapping. This limitation is particularly pronounced for FPGAs where a large logic network can be mapped to only a few levels of look-up tables (LUT).

Modern high-level synthesis (HLS) tools commonly employ pipelining to achieve efficient loop acceleration by overlapping the execution of successive loop iterations. However, existing HLS techniques provide inadequate support for pipelining irregular loop nests that contain dynamic-bound inner loops, where unrolling is either very expensive or not even applicable. To overcome this major limitation, we propose ElasticFlow, a novel architectural synthesis approach capable of dynamically distributing inner loops to an array of loop processing units (LPUs) in a complexity-effective manner. These LPUs can be either specialized to execute an individual loop or shared amongst multiple inner loops for area reduction. We evaluate ElasticFlow using a variety of real-life applications and demonstrate significant performance improvements over a widely used commercial HLS tool for Xilinx FPGAs.

Loop pipelining is a widely-accepted technique in high-level synthesis to enable pipelined execution of successive loop iterations to achieve high performance. Existing loop pipelining methods provide inadequate support for pipeline flushing. In this paper, we study the problem of enabling flushing in pipeline synthesis and examine its implications in scheduling and binding. We propose novel techniques for synthesizing a conflict-aware flushing-enabled pipeline that is robust against potential resource collisions. Experiments with real-life benchmarks show that our methods significantly reduce the possibility of resource collisions compared to conventional approaches while conserving hardware resources and achieving near-optimal performance.

Pipelining is an important technique in high-level synthesis, which overlaps the execution of successive loop iterations or threads to achieve high throughput for loop/function kernels. Since existing pipelining techniques typically enforce in-order thread execution, a variable-latency operation in one thread would block all subsequent threads, resulting in considerable performance degradation. In this paper, we propose a multithreaded pipelining approach that enables context switching to allow out-of-order thread execution for data-parallel kernels. To ensure that the synthesized pipeline is complexity effective, we further propose efficient scheduling algorithms for minimizing the hardware overhead associated with context management. Experimental results show that our proposed techniques can significantly improve the effective pipeline throughput over conventional approaches while conserving hardware resources.