Erwei Wang

Jiayi Nie, Haoran Wu, Yao Lai et al. · cambridge, tsinghua

New AI accelerators with novel instruction set architectures (ISAs) often require developers to manually craft low-level kernels, a time-consuming and error-prone process that does not scale across hardware targets. This delays emerging hardware platforms from reaching the market. While prior LLM-based code generation has shown promise in mature GPU ecosystems, it remains unclear whether agentic LLM systems can quickly produce valid and efficient kernels for emerging hardware with new ISAs. We present KernelCraft: the first benchmark for evaluating an LLM agent's ability to generate and optimize low-level kernels for customized accelerators through a function-calling, feedback-driven workflow. We evaluate agent performance across three emerging accelerators on more than 20 machine-learning tasks, each with five diverse task configurations. Across four leading reasoning models, the strongest agents generate functionally correct kernels for unseen ISAs within a few refinement steps and produce optimized kernels that match or outperform compiler baselines. These results demonstrate KernelCraft's potential to accelerate the accelerator chip development cycle. KernelCraft is available at https://kernelcraft-cam.github.io/.

Emily Tucker, Louis-Noël Pouchet, Erika Hunhoff et al.

Optimizing compilers have become a cornerstone for high-performance program generation in research and industry. Optimizations, including those implemented manually by a user and those target-specific and non-target-specific, are used to transform programs to achieve good performance. Although these optimizations are necessary for performance, assessing their correctness has remained a major challenge; the risk of incorrect code being deployed increases with unproven optimization flows. In this work, we target the formal verification of correctness of a transformed program by computing whether a pair of programs are semantically equivalent, one being a transformed version of the other. We restrict the class of programs supported to enable a hybrid concrete-symbolic interpretation approach to equivalence, which in turn is mostly agnostic to how the programs are implemented (syntax, schedule, storage, etc.). This approach can show equivalence in linear time with respect to the operations executed by the programs. We develop a verifier for a meaningful subset of MLIR, and report on the verification of the AMD MLIR-AIR and MLIR-AIE toolchains, as well as the standard mlir-opt on hundreds of benchmarks variants.

Erwei Wang, Samuel Bayliss, Andra Bisca et al.

General-purpose compilers abstract away parallelism, locality, and synchronization, limiting their effectiveness on modern spatial architectures. As modern computing architectures increasingly rely on fine-grained control over data movement, execution order, and compute placement for performance, compiler infrastructure must provide explicit mechanisms for orchestrating compute and data to fully exploit such architectures. We introduce MLIR-AIR, a novel, open-source compiler stack built on MLIR that bridges the semantic gap between high-level workloads and fine-grained spatial architectures such as AMD's NPUs. MLIR-AIR defines the AIR dialect, which provides structured representations for asynchronous and hierarchical operations across compute and memory resources. AIR primitives allow the compiler to orchestrate spatial scheduling, distribute computation across hardware regions, and overlap communication with computation without relying on ad hoc runtime coordination or manual scheduling. We demonstrate MLIR-AIR's capabilities through two case studies: matrix multiplication and the multi-head attention block from the LLaMA 2 model. For matrix multiplication, MLIR-AIR achieves up to 78.7% compute efficiency and generates implementations with performance almost identical to state-of-the-art, hand-optimized matrix multiplication written using the lower-level, close-to-metal MLIR-AIE framework. For multi-head attention, we demonstrate that the AIR interface supports fused implementations using approximately 150 lines of code, enabling tractable expression of complex workloads with efficient mapping to spatial hardware. MLIR-AIR transforms high-level structured control flow into spatial programs that efficiently utilize the compute fabric and memory hierarchy of an NPU, leveraging asynchronous execution, tiling, and communication overlap through compiler-managed scheduling.

Erwei Wang, James J. Davis, Daniele Moro et al.

The ever-growing computational demands of increasingly complex machine learning models frequently necessitate the use of powerful cloud-based infrastructure for their training. Binary neural networks are known to be promising candidates for on-device inference due to their extreme compute and memory savings over higher-precision alternatives. However, their existing training methods require the concurrent storage of high-precision activations for all layers, generally making learning on memory-constrained devices infeasible. In this article, we demonstrate that the backward propagation operations needed for binary neural network training are strongly robust to quantization, thereby making on-the-edge learning with modern models a practical proposition. We introduce a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions while inducing little to no accuracy loss vs Courbariaux & Bengio's standard approach. These decreases are primarily enabled through the retention of activations exclusively in binary format. Against the latter algorithm, our drop-in replacement sees memory requirement reductions of 3--5$\times$, while reaching similar test accuracy in comparable time, across a range of small-scale models trained to classify popular datasets. We also demonstrate from-scratch ImageNet training of binarized ResNet-18, achieving a 3.78$\times$ memory reduction. Our work is open-source, and includes the Raspberry Pi-targeted prototype we used to verify our modeled memory decreases and capture the associated energy drops. Such savings will allow for unnecessary cloud offloading to be avoided, reducing latency, increasing energy efficiency, and safeguarding end-user privacy.

Erwei Wang, James J. Davis, Georgios-Ilias Stavrou et al.

FPGA-specific DNN architectures using the native LUTs as independently trainable inference operators have been shown to achieve favorable area-accuracy and energy-accuracy tradeoffs. The first work in this area, LUTNet, exhibited state-of-the-art performance for standard DNN benchmarks. In this paper, we propose the learned optimization of such LUT-based topologies, resulting in higher-efficiency designs than via the direct use of off-the-shelf, hand-designed networks. Existing implementations of this class of architecture require the manual specification of the number of inputs per LUT, K. Choosing appropriate K a priori is challenging, and doing so at even high granularity, e.g. per layer, is a time-consuming and error-prone process that leaves FPGAs' spatial flexibility underexploited. Furthermore, prior works see LUT inputs connected randomly, which does not guarantee a good choice of network topology. To address these issues, we propose logic shrinkage, a fine-grained netlist pruning methodology enabling K to be automatically learned for every LUT in a neural network targeted for FPGA inference. By removing LUT inputs determined to be of low importance, our method increases the efficiency of the resultant accelerators. Our GPU-friendly solution to LUT input removal is capable of processing large topologies during their training with negligible slowdown. With logic shrinkage, we better the area and energy efficiency of the best-performing LUTNet implementation of the CNV network classifying CIFAR-10 by 1.54x and 1.31x, respectively, while matching its accuracy. This implementation also reaches 2.71x the area efficiency of an equally accurate, heavily pruned BNN. On ImageNet with the Bi-Real Net architecture, employment of logic shrinkage results in a post-synthesis area reduction of 2.67x vs LUTNet, allowing for implementation that was previously impossible on today's largest FPGAs.

Zhiqiang Que, Erwei Wang, Umar Marikar et al.

This paper presents novel reconfigurable architectures for reducing the latency of recurrent neural networks (RNNs) that are used for detecting gravitational waves. Gravitational interferometers such as the LIGO detectors capture cosmic events such as black hole mergers which happen at unknown times and of varying durations, producing time-series data. We have developed a new architecture capable of accelerating RNN inference for analyzing time-series data from LIGO detectors. This architecture is based on optimizing the initiation intervals (II) in a multi-layer LSTM (Long Short-Term Memory) network, by identifying appropriate reuse factors for each layer. A customizable template for this architecture has been designed, which enables the generation of low-latency FPGA designs with efficient resource utilization using high-level synthesis tools. The proposed approach has been evaluated based on two LSTM models, targeting a ZYNQ 7045 FPGA and a U250 FPGA. Experimental results show that with balanced II, the number of DSPs can be reduced up to 42% while achieving the same IIs. When compared to other FPGA-based LSTM designs, our design can achieve about 4.92 to 12.4 times lower latency.

Erwei Wang, James J. Davis, Peter Y. K. Cheung et al.

Research has shown that deep neural networks contain significant redundancy, and thus that high classification accuracy can be achieved even when weights and activations are quantized down to binary values. Network binarization on FPGAs greatly increases area efficiency by replacing resource-hungry multipliers with lightweight XNOR gates. However, an FPGA's fundamental building block, the K-LUT, is capable of implementing far more than an XNOR: it can perform any K-input Boolean operation. Inspired by this observation, we propose LUTNet, an end-to-end hardware-software framework for the construction of area-efficient FPGA-based neural network accelerators using the native LUTs as inference operators. We describe the realization of both unrolled and tiled LUTNet architectures, with the latter facilitating smaller, less power-hungry deployment over the former while sacrificing area and energy efficiency along with throughput. For both varieties, we demonstrate that the exploitation of LUT flexibility allows for far heavier pruning than possible in prior works, resulting in significant area savings while achieving comparable accuracy. Against the state-of-the-art binarized neural network implementation, we achieve up to twice the area efficiency for several standard network models when inferencing popular datasets. We also demonstrate that even greater energy efficiency improvements are obtainable.

Yiren Zhao, Xitong Gao, Xuan Guo et al.

Modern deep Convolutional Neural Networks (CNNs) are computationally demanding, yet real applications often require high throughput and low latency. To help tackle these problems, we propose Tomato, a framework designed to automate the process of generating efficient CNN accelerators. The generated design is pipelined and each convolution layer uses different arithmetics at various precisions. Using Tomato, we showcase state-of-the-art multi-precision multi-arithmetic networks, including MobileNet-V1, running on FPGAs. To our knowledge, this is the first multi-precision multi-arithmetic auto-generation framework for CNNs. In software, Tomato fine-tunes pretrained networks to use a mixture of short powers-of-2 and fixed-point weights with a minimal loss in classification accuracy. The fine-tuned parameters are combined with the templated hardware designs to automatically produce efficient inference circuits in FPGAs. We demonstrate how our approach significantly reduces model sizes and computation complexities, and permits us to pack a complete ImageNet network onto a single FPGA without accessing off-chip memories for the first time. Furthermore, we show how Tomato produces implementations of networks with various sizes running on single or multiple FPGAs. To the best of our knowledge, our automatically generated accelerators outperform closest FPGA-based competitors by at least 2-4x for lantency and throughput; the generated accelerator runs ImageNet classification at a rate of more than 3000 frames per second.

Erwei Wang, James J. Davis, Peter Y. K. Cheung et al.

Research has shown that deep neural networks contain significant redundancy, and that high classification accuracies can be achieved even when weights and activations are quantised down to binary values. Network binarisation on FPGAs greatly increases area efficiency by replacing resource-hungry multipliers with lightweight XNOR gates. However, an FPGA's fundamental building block, the K-LUT, is capable of implementing far more than an XNOR: it can perform any K-input Boolean operation. Inspired by this observation, we propose LUTNet, an end-to-end hardware-software framework for the construction of area-efficient FPGA-based neural network accelerators using the native LUTs as inference operators. We demonstrate that the exploitation of LUT flexibility allows for far heavier pruning than possible in prior works, resulting in significant area savings while achieving comparable accuracy. Against the state-of-the-art binarised neural network implementation, we achieve twice the area efficiency for several standard network models when inferencing popular datasets. We also demonstrate that even greater energy efficiency improvements are obtainable.

Erwei Wang, James J. Davis, Ruizhe Zhao et al.

Deep neural networks have proven to be particularly effective in visual and audio recognition tasks. Existing models tend to be computationally expensive and memory intensive, however, and so methods for hardware-oriented approximation have become a hot topic. Research has shown that custom hardware-based neural network accelerators can surpass their general-purpose processor equivalents in terms of both throughput and energy efficiency. Application-tailored accelerators, when co-designed with approximation-based network training methods, transform large, dense and computationally expensive networks into small, sparse and hardware-efficient alternatives, increasing the feasibility of network deployment. In this article, we provide a comprehensive evaluation of approximation methods for high-performance network inference along with in-depth discussion of their effectiveness for custom hardware implementation. We also include proposals for future research based on a thorough analysis of current trends. This article represents the first survey providing detailed comparisons of custom hardware accelerators featuring approximation for both convolutional and recurrent neural networks, through which we hope to inspire exciting new developments in the field.