Efficient execution of deep learning workloads on dataflow architectures is crucial for overcoming memory bottlenecks and maximizing performance.While streaming intermediate results between computation kernels can significantly improve efficiency, existing approaches struggle with inter-kernel correlations, external memory access management, and buffer optimization.In this work, we propose StreamTensor, a compiler framework that automatically constructs and optimizes stream-based dataflow accelerators.StreamTensor introduces a novel iterative tensor type system to explicitly encode stream layouts, enabling seamless kernel fusion, buffer allocation, and memory optimization.By systematically exploring three hierarchical design spaces, including tensor tiling, kernel fusion, and resource allocation, StreamTensor balances computational intensity, memory efficiency, and data streaming to maximize performance.Based on FPGA evaluations on Large Language Models (LLM), StreamTensor achieves up to 0.76x and 0.64x lower latency compared to the state-of-the-art FPGA LLM accelerators and GPUs, and up to 1.99x higher energy efficiency compared to GPUs, making it a promising approach for scalable dataflow-based deep learning acceleration.

Heterogeneous systems with hardware accelerators are increasingly common, and various optimized implementations/algorithms exist for computation kernels. However, no single best combination ofcode version and device(C&D) can outper-form others across all input cases, demanding a method to select the best C&D pair based on input. We presentmachinelearning-basedcode version anddevice selection method, namedMLCD, that uses input data characteristics to select the best C&D pair dynamically. We also apply active learning to reduce the number of samples needed to construct the model. Demonstrated on two different CPU-GPU systems, MLCD achieves near-optimal speed-up regardless of which systems tested. Concretely, reporting results from system one with mid-end hardwares, it achieves 99.9%, 95.6%, 99.9%, and 98.6% of the optimal acceleration attainable through the ideal choice of C&D pairs in General Matrix Multiply, PageRank, N-body Simulation, and K-Motif Counting, respectively. MLCD achieves a speed-up of 2.57×, 1.58×, 2.68×, and 1.09× compared to baselines without MLCD. Additionally, MLCD handles end-to-end applications, achieving up to 10% and 46% speed-up over GPU-only and CPU-only solutions with Graph Neural Networks. Furthermore, it achieves 7.28× average speed-up in execution latency over the state-of-the-art approach and determines suitable code versions for unseen input 108− 1010× faster.

The widespread adoption of Large Language Models (LLMs) is impeded by their demanding compute and memory resources. The first task of this paper is to explore optimization strategies to expedite LLMs, including quantization, pruning, and operation-level optimizations. One unique direction is to optimize LLM inference through novel software/hardware co-design methods. Given the accelerated LLMs, the second task of this paper is to study LLMs’ performance in the usage scenario of circuit design and verification. Specifically, we place a particular emphasis on functional verification. Through automated prompt engineering, we harness the capabilities of the established LLM, GPT-4, to generate High-Level Synthesis (HLS) designs with predefined errors based on 11 open-source synthesizable HLS benchmark suites. This dataset is a comprehensive collection of over 1000 function-level designs, and each of which is afflicted with up to 45 distinct combinations of defects injected into the source code. This dataset, named Chrysalis, expands upon what’s available in current HLS error models, offering a rich resource for training to improve how LLMs debug code. The dataset can be accessed at: https://github.com/UIUC-ChenLab/Chrysalis-HLS.

Dataflow architectures are growing in popularity due to their potential to mitigate the challenges posed by the memory wall inherent to the Von Neumann architecture. At the same time, high-level synthesis (HLS) has demonstrated its efficacy as a design methodology for generating efficient dataflow architectures within a short development cycle. However, existing HLS tools rely on developers to explore the vast dataflow design space, ultimately leading to suboptimal designs. This phenomenon is especially concerning as the size of the HLS design grows. To tackle these challenges, we introduce HIDA1, a new scalable and hierarchical HLS framework that can systematically convert an algorithmic description into a dataflow implementation on hardware. We first propose a collection of efficient and versatile dataflow representations for modeling the hierarchical dataflow structure. Capitalizing on these representations, we develop an automated optimizer that decomposes the dataflow optimization problem into multiple levels based on the inherent dataflow hierarchy. Using FPGAs as an evaluation platform, working with a set of neural networks modeled in PyTorch, HIDA achieves up to 8.54× higher throughput compared to the state-of-the-art (SOTA) HLS optimization tool. Furthermore, despite being fully automated and able to handle various applications, HIDA achieves 1.29× higher throughput over the SOTA RTL-based neural network accelerators on an FPGA.

Large Language Models (LLMs) have revolutionized a wide range of applications with their strong human-like understanding and creativity. Due to the continuously growing model size and complexity, LLM training and deployment have shown significant challenges, which often results in extremely high computational and storage costs and energy consumption. In this paper, we discuss the recent advancements and research directions on (1) LLM algorithm-level acceleration, (2) LLM-hardware co-design for improved system efficiency, (3) LLM-to-accelerator compilation for customized LLM accelerators, and (4) LLM-aided design for HLS (High-Level Synthesis) functional verification. For each aspect, we present the background study, our proposed solutions, and future directions. An extended version of this work can be found at: https://arxiv.org/abs/2406.10903.

This paper proposes ISDC, a novel feedback-guided iterative system of difference constraints (SDC) scheduling algorithm for high-level synthesis (HLS). ISDC leverages subgraph extraction-based low-level feedback from downstream tools like logic synthesizers to iteratively refine HLS scheduling. Technical innovations include: (1) An enhanced SDC formulation that effectively integrates low-level feedback into the linear-programming (LP) problem; (2) A fanout and window-based subgraph extraction mechanism driving the feedback cycle; (3) A no-human-in-loop ISDC flow compatible with a wide range of downstream tools and process design kits (PDKs). Evaluation shows that ISDC reduces register usage by 28.5% against an industrial-strength open-source HLS tool.

Large Language Models (LLMs) have made remarkable progress in processing extensive contexts, with the Key-Value (KV) cache playing a vital role in enhancing their performance. However, the growth of the KV cache in response to increasing input length poses challenges to memory and time efficiency. To address this problem, this paper introduces SnapKV, an innovative and fine-tuning-free approach that efficiently minimizes KV cache size while still delivering comparable performance in real-world applications. We discover that each attention head in the model consistently focuses on specific prompt attention features during generation. Meanwhile, this robust pattern can be obtained from an `observation' window located at the end of the prompts. Drawing on this insight, SnapKV automatically compresses KV caches by selecting clustered important KV positions for each attention head. Our approach significantly reduces the growing computational overhead and memory footprint when processing long input sequences. Specifically, SnapKV achieves a consistent decoding speed with a 3.6x increase in generation speed and an 8.2x enhancement in memory efficiency compared to baseline when processing inputs of 16K tokens. At the same time, it maintains comparable performance to baseline models across 16 long sequence datasets. Moreover, SnapKV can process up to 380K context tokens on a single A100-80GB GPU using HuggingFace implementation with minor changes, exhibiting only a negligible accuracy drop in the Needle-in-a-Haystack test. Further comprehensive studies suggest SnapKV's potential for practical applications.

Dense matrix multiply (MM) serves as one of the most heavily used kernels in deep learning applications. To cope with the high computation demands of these applications, heterogeneous architectures featuring both FPGA and dedicated ASIC accelerators have emerged as promising platforms. For example, the AMD/Xilinx Versal ACAP architecture combines general-purpose CPU cores and programmable logic with AI Engine processors optimized for AI/ML. An array of 400 AI Engine processors executing at 1 GHz can provide up to 6.4 TFLOPS performance for 32-bit floating-point (FP32) data. However, machine learning models often contain both large and small MM operations. While large MM operations can be parallelized efficiently across many cores, small MM operations typically cannot. We observe that executing some small MM layers from the BERT natural language processing model on a large, monolithic MM accelerator in Versal ACAP achieved less than 5% of the theoretical peak performance. Therefore, one key question arises: How can we design accelerators to fully use the abundant computation resources under limited communication bandwidth for end-to-end applications with multiple MM layers of diverse sizes? We identify the biggest system throughput bottleneck resulting from the mismatch between the massive computation resources of one monolithic accelerator and the various MM layers of small sizes in the application. To resolve this problem, we propose the CHARM framework to compose multiple diverse MM accelerator architectures working concurrently on different layers within one application. CHARM includes analytical models that guide design space exploration to determine accelerator partitions and layer scheduling. To facilitate system designs, CHARM automatically generates code, enabling thorough onboard design verification. We deploy the CHARM framework on four different deep learning applications in FP32, INT16, and INT8 data types, including BERT, ViT, NCF, and MLP, on the AMD/Xilinx Versal ACAP VCK190 evaluation board. Our experiments show that we achieve 1.46 TFLOPS, 1.61 TFLOPS, 1.74 TFLOPS, and 2.94 TFLOPS inference throughput for BERT, ViT, NCF, and MLP in FP32 data type, respectively, which obtain 5.29 \(\times\) , 32.51 \(\times\) , 1.00 \(\times\) , and 1.00 \(\times\) throughput gains compared to one monolithic accelerator. CHARM achieves the maximum throughput of 1.91 TOPS, 1.18 TOPS, 4.06 TOPS, and 5.81 TOPS in the INT16 data type for the four applications. The maximum throughput achieved by CHARM in the INT8 data type is 3.65 TOPS, 1.28 TOPS, 10.19 TOPS, and 21.58 TOPS, respectively. We have open-sourced our tools, including detailed step-by-step guides to reproduce all the results presented in this article and to enable other users to learn and leverage CHARM framework and tools in their end-to-end systems: https://github.com/arc-research-lab/CHARM .

Dense matrix multiply (MM) serves as one of the most heavily used kernels in deep learning applications. To cope with the high computation demands of these applications, heterogeneous architectures featuring both FPGA and dedicated ASIC accelerators have emerged as promising platforms. For example, the AMD/Xilinx Versal ACAP architecture combines general-purpose CPU cores and programmable logic (PL) with AI Engine processors (AIE) optimized for AI/ML. An array of 400 AI Engine processors executing at 1 GHz can theoretically provide up to 6.4 TFLOPs performance for 32-bit floating-point (fp32) data. However, machine learning models often contain both large and small MM operations. While large MM operations can be parallelized efficiently across many cores, small MM operations typically cannot. In our investigation, we observe that executing some small MM layers from the BERT natural language processing model on a large, monolithic MM accelerator in Versal ACAP achieved less than 5% of the theoretical peak performance. Therefore, one key question arises: How can we design accelerators to fully use the abundant computation resources under limited communication bandwidth for end-to-end applications with multiple MM layers of diverse sizes?

This paper proposes a High-Level Synthesis (HLS) framework for domain-specific computing. The framework contains three key components: 1) ScaleHLS, a multi-level HLS compilation flow. Aimed to address the lack of expressiveness and hardware-dedicated representation of traditional software-oriented compilers. ScaleHLS introduces a hierarchical intermediate representation (IR) for the progressive optimization of HLS designs defined in various high-level languages. ScaleHLS consists of three levels of optimizations, including graph, loop, and directive levels, to realize an efficient compilation pipeline and generate highly-optimized domain-specific accelerators. 2) AutoScaleDSE is an automated design space exploration (DSE) engine. Real-world HLS designs often come with large design spaces that are difficult for designers to explore. Meanwhile, the connections between different components of an HLS design further complicate the design spaces. In order to address the DSE problem, AutoScaleDSE proposes a random forest classifier and a graph-driven approach to improve the accuracy of estimating the intermediate DSE results while reducing the time and computational cost. With this new approach, AutoScaleDSE can evaluate thousands of HLS design points and find the Pareto-dominating design points within a couple of hours. 3) PyTransform is a flexible pattern-driven design customization flow. Existing HLS flows demand manual code rewriting or intrusive compiler customization to conduct domain-specific optimizations, leading to unscalable or inflexible compiler solutions. PyTransform proposes a Python-based flow that enables users to define custom matching and rewriting patterns at a high level of abstraction, being able to be incorporated into the DSL compilation flow in an automatic and scalable manner. In summary, ScaleHLS, AutoScaleDSE, and PyTransform aim to address the challenges present in the compilation, DSE, and customization of existing HLS flows, respectively. With the three key components, our newly proposed HLS framework can deliver a scalable and extensible solution for designing domain-specific languages to automate and speed up the process of designing domain-specific accelerators.

High-Level Synthesis (HLS) has enabled users to rapidly develop designs targeted for FPGAs from the behavioral description of the design. However, to synthesize an optimal design capable of taking better advantage of the target FPGA, a considerable amount of effort is needed to transform the initial behavioral description into a form that can capture the desired level of parallelism. Thus, a design space exploration (DSE) engine capable of optimizing large complex designs is needed to achieve this goal. We present a new DSE engine capable of considering code transformation, compiler directives (pragmas), and the compatibility of these optimizations. To accomplish this, we initially express the structure of the input code as a graph to guide the exploration process. To appropriately transform the code, we take advantage of ScaleHLS based on the multi-level compiler infrastructure (MLIR). Finally, we identify problems that limit the scalability of existing DSEs, which we name the “design space merging problem.” We address this issue by employing a Random Forest classifier that can successfully decrease the number of invalid design points without invoking the HLS compiler as a validation tool. We evaluated our DSE engine against the ScaleHLS DSE, outperforming it by a maximum of 59×. We additionally demonstrate the scalability of our design by applying our DSE to large-scale HLS designs, achieving a maximum speedup of 12× for the benchmarks in the MachSuite and Rodinia set.

This paper presents an enhanced version of a scalable HLS (High-Level Synthesis) framework named ScaleHLS, which can compile HLS C/C++ programs and PyTorch models to highly-efficient and synthesizable C++ designs. The original version of ScaleHLS achieved significant speedup on both C/C++ kernels and PyTorch models [14]. In this paper, we first highlight the key features of ScaleHLS on tackling the challenges present in the representation, optimization, and exploration of large-scale HLS designs. To further improve the scalability of ScaleHLS, we then propose an enhanced HLS transform and analysis library supported in both C++ and Python, and a new design space exploration algorithm to handle HLS designs with hierarchical structures more effectively. Comparing to the original ScaleHLS, our enhanced version improves the speedup by up to 60.9× on FPGAs. ScaleHLS is fully open-sourced at https://github.com/hanchenye/scalehls.

High-level synthesis (HLS) has been widely adopted as it significantly improves the hardware design productivity and enables efficient design space exploration (DSE). Existing HLS tools are built using compiler infrastructures largely based on a single-level abstraction, such as LLVM. How-ever, as HLS designs typically come with intrinsic structural or functional hierarchies, different HLS optimization problems are often better solved with different levels of abstractions. This paper proposes ScaleHLS1, a new scalable and customizable HLS framework, on top of a multi-level compiler infrastructure called MLIR. ScaleHLS represents HLS designs at multiple representation levels and provides an HLS-dedicated analysis and transform library to solve the optimization problems at the suitable levels. Using this library, we provide a DSE engine to generate optimized HLS designs automatically. In addition, we develop an HLS C front-end and a C/C++ emission back-end to translate HLS designs into/from MLIR for enabling an end-to-end compilation flow. Experimental results show that, comparing to the baseline designs without manual directives insertion and code-rewriting, that are only optimized by Xilinx Vivado HLS, ScaleHLS improves the performances with amazing quality-of-results – up to 768.1× better on computation kernel level programs and up to 3825.0× better on neural network models.

To speedup Deep Neural Networks (DNN) accelerator design and enable effective implementation, we propose HybridDNN, a framework for building high-performance hybrid DNN accelerators and delivering FPGA-based hardware implementations. Novel techniques include a highly flexible and scalable architecture with a hybrid Spatial/Winograd convolution (CONV) Processing Engine (PE), a comprehensive design space exploration tool, and a complete design flow to fully support accelerator design and implementation. Experimental results show that the accelerators generated by HybridDNN can deliver 3375.7 and 83.3 GOPS on a high-end FPGA (VU9P) and an embedded FPGA (PYNQ-Z1), respectively, which achieve a 1.8x higher performance improvement compared to the state-of-art accelerator designs. This demonstrates that HybridDNN is flexible and scalable and can target both cloud and embedded hardware platforms with vastly different resource constraints.

Existing FPGA-based DNN accelerators typically fall into two design paradigms. Either they adopt a generic reusable architecture to support different DNN networks but leave some performance and efficiency on the table because of the sacrifice of design specificity. Or they apply a layer-wise tailor-made architecture to optimize layer-specific demands for computation and resources but loose the scalability of adaptation to a wide range of DNN networks. To overcome these drawbacks, this paper proposes a novel FPGA-based DNN accelerator design paradigm and its automation tool, called DNNExplorer, to enable fast exploration of various accelerator designs under the proposed paradigm and deliver optimized accelerator architectures for existing and emerging DNN networks. Three key techniques are essential for DNNExplorer's improved performance, better specificity, and scalability, including (1) a unique accelerator design paradigm with both high-dimensional design space support and fine-grained adjustability, (2) a dynamic design space to accommodate different combinations of DNN workloads and targeted FPGAs, and (3) a design space exploration (DSE) engine to generate optimized accelerator architectures following the proposed paradigm by simultaneously considering both FPGAs' computation and memory resources and DNN networks' layer-wise characteristics and overall complexity. Experimental results show that, for the same FPGAs, accelerators generated by DNNExplorer can deliver up to 4.2x higher performances (GOP/s) than the state-of-the-art layer-wise pipelined solutions generated by DNNBuilder [1] for VGG-like DNN with 38 CONV layers. Compared to accelerators with generic reusable computation units, DNNExplorer achieves up to 2.0x and 4.4x DSP efficiency improvement than a recently published accelerator design from academia (HybridDNN [2]) and a commercial DNN accelerator IP (Xilinx DPU [3]), respectively.