Many convolutional neural network (CNN) accelerators are proposed to exploit the sparsity of the networks recently to enjoy the benefits of both computation and memory reduction. However, most accelerators cannot exploit the sparsity of both activations and weights. For those works that exploit both sparsity opportunities, they cannot achieve the stable load balance through a static scheduling (SS) strategy, which is vulnerable to the sparsity distribution. In this work, a balanced compressed sparse row format and a dynamic scheduling strategy are proposed to improve the load balance. A set-associate structure is also presented to tradeoff the load balance and hardware resource overhead. We propose SWM to accelerate the CNN inference, which supports both sparse convolution and sparse fully connected (FC) layers. SWM provides Winograd adaptability for large convolution kernels and supports both 16-bit and 8-bit quantized CNNs. Due to the activation sharing, 8-bit processing can achieve theoretically twice the performance of the 16-bit processing with the same sparsity. The architecture is evaluated with VGG16 and ResNet50, which achieves: at most 7.6 TOP/s for sparse-Winograd convolution and three TOP/s for sparse matrix multiplication with 16-bit quantization on Xilinx VCU1525 platform. SWM can process 310/725 images per second for VGG16/ResNet50 with 16-bit quantization. Compared with the state-of-the-art works, our design can achieve at least 1.53 × speedup and 1.8 × energy efficiency improvement.

Partially reconfigurable systems are promising computing platforms for streaming applications, which demand both hardware efficiency and reconfigurable flexibility. To realize the full potential of these systems, a streaming-based partially reconfigurable architecture and unified software/hardware multithreaded programming model (SPREAD) is presented in this paper. SPREAD is a reconfigurable architecture with a unified software/hardware thread interface and high throughput point-to-point streaming structure. It supports dynamic computing resource allocation, runtime software/hardware switching, and streaming-based multithreaded management at the operating system level. SPREAD is designed to provide programmers of streaming applications with a unified view of threads, allowing them to exploit thread, data, and pipeline parallelism; it enhances hardware efficiency while simplifying the development of streaming applications for partially reconfigurable systems. Experimental results targeting cryptography applications demonstrate the feasibility and superior performance of SPREAD. Moreover, the parallelized Advanced Encryption Standard (AES), Data Encryption Standard (DES), and Triple DES (3DES) hardware threads on field-programmable gate arrays show 1.61-4.59 times higher power efficiency than their implementations on state-of-the-art graphics processing units.

With the development of machine learning technology, the exploration of energy-efficient and flexible architectures for object inference algorithms is of growing interest in recent years. However, not many publications concentrate on a coarsegrained reconfigurable architecture (CGRA) for object inference algorithms. This paper provides a stream processing, dual-track programming CGRA-based approach to address the inherent computing characteristics of algorithms in object inference. Based on the proposed approach, an architecture called stream dual-track CGRA (SDT-CGRA) is presented as an implementation prototype. To evaluate the performance, the SDT-CGRA is realized in Verilog HDL and implemented in Semiconductor Manufacturing International Corporation 55-nm process, with the footprint of 5.19 mm ² at 450 MHz. Seven object inference algorithms, including convolutional neural network (CNN), k-means, principal component analysis (PCA), spatial pyramid matching (SPM), linear support vector machine (SVM), Softmax, and Joint Bayesian, are selected as benchmarks. The experimental results show that the SDT-CGRA can gain on average 343.8 times and 17.7 times higher energy efficiency for Softmax, PCA, and CNN, 621.0 times and 1261.8 times higher energy efficiency for k-means, SPM, linear-SVM, and Joint-Bayesian algorithms when compared with the Intel Xeon E5-2637 CPU and the Nvidia TitanX graphics processing unit. When compared with the state-of-the-art solutions of AlexNet on field-programmable gate array and CGRA, the proposed SDT-CGRA can achieve a 1.78 times increase in energy efficiency and a 13 times speedup, respectively.

Approximate computing is an efficient approach to reduce the design complexity for error-resilient applications. Multipliers are key arithmetic units in many applications, such as deep neural networks (DNNs) and digital signal processing (DSP) systems. In this article, an open-source adaptable approximate multiplier design driven by input distribution and polarity is proposed to generate optimized approximate multipliers to trade off between the application-level performance and the hardware cost. The proposed method minimizes the average square of the absolute error of an approximate multiplier according to the probability distributions of operands extracted from the target application with consideration of input polarity, achieving low hardware cost and negligible application-level performance loss. The proposed method can generate unsigned multipliers (or signed multipliers) based on the Braun multiplier (or Baugh–Wooley multiplier). To demonstrate the effectiveness of the method, three different-scale quantized DNNs, including LeNet, AlexNet, and VGG16 with $8$ $\times$ $8$ unsigned multiplication and an adaptive least mean square (LMS)-based finite impulse response (FIR) filter with $16$ $\times$ $16$ fixed-point signed multiplication, are evaluated. In the DNN training process, a noise training technique is adopted to reduce the accuracy loss due to the approximation. When compared to the state-of-the-art approximate multipliers, the generated multipliers can achieve up to 26.4% and 27.1% product of power, delay, and area gains with negligible application-level performance loss in VGG16 and FIR applications, respectively.

Logic synthesis is a crucial step in electronic design automation tools. The rapid developments of reinforcement learning (RL) have enabled the automated exploration of logic synthesis. Existing RL based methods may lead to data inefficiency, and the exploration approaches for FPGA and ASIC technology mapping in recent works lack the flexibility of the learning process. This work proposes ESE, a reinforcement learning based framework to efficiently learn the logic synthesis process. The framework supports the modeling of logic optimization and technology mapping for FPGA and ASIC. The optimization for the execution time of the synthesis script is also considered. For the modeling of FPGA mapping, the logic optimization and technology mapping are combined to be learned in a flexible way. For the modeling of ASIC mapping, the standard cell based optimization and LUT optimization operations are incorporated into the ASIC synthesis flow. To improve the utilization of samples, the Proximal Policy Optimization model is adopted. Furthermore, the framework is enhanced by supporting MIG based synthesis exploration. Experiments show that for FPGA technology mapping on the VTR benchmark, the average LUT-Level-Product and script runtime are improved by more than 18.3% and 12.4% respectively than previous works. For ASIC mapping on the EPFL benchmark, the average Area-Delay-Product is improved by 14.5%.