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.

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.

Due to the fact that FPGA on-chip memory capacity increases significantly, the feature maps and weights of convolutional layers can be stored on chip, which can reduce the data movement between on-chip memory and off-chip memory. Hence, the bottleneck can shift from the bandwidth to the computing resources in convolutional layers, which will improve the performance dramatically. Under this circumstance, this paper quantitatively analyzes how to design the hardware architecture based on the roofline model to optimize the performance under the constraints of available on-chip computing resources and propose an efficient architecture. Our accelerator is implemented on Xilinx UltraScale+ FPGA with the performance of 9.39 TOPS and 6.86 TOPS for 8-bit data width with 100MHz main frequency and 400MHz DSP frequency on ResNet-50 and AlexNet, which outperforms the existing FPGA-based CNN accelerator.

In this brief, an FPGA-based solution is proposed to show the computing efficiency on rotated object detection based on R ³ Det algorithm. The key idea of our approach is firstly to design reconfigurable neural processing units (NPU) for convolutional neural networks (CNN) and a specific architecture for spatial operations, and then to adopt a novel scheduling scheme to deal with the data dependency on these modules. When implemented on cost-effective Kintex Ultrascale+ XCKU15P FPGA, the proposed solution achieves nearly 3.20 times energy efficiency compared to NVIDIA V100 GPU. The proposed reconfigurable NPU achieves 73.20% DSP efficiency, 30.39% improvement compared to the 56.14% DSP efficiency of the state-of-the-art CNN accelerator on ResNet-50.

Fast regular expression matching is an essential task for deep packet inspection. In previous works, the regular expression matching engine on FPGA struggled to achieve an ideal balance between resource consumption and throughput. Speculation and enumerative computation exploits the statistical properties of deterministic finite automata, allowing for more efficient pattern matching. Existing related designs mostly revolve around vector instructions and multiple processors/cores or SIMD instruction sets, with a lack of implementation on FPGA platforms. We design a parallelized two-character matching engine on FPGA for efficiently fast filtering off fields with no pattern features. We transform the state transitions with sequential dependencies to the existing problem of elements in one set, enabling the proposed design to achieve high throughput with low resource consumption and support dynamic updates. Results show that compared with the traditional DFA matching, with a maximum resource consumption of 25% for on-chip FFs (74323/1045440) and LUTs (123902/522720), there is an improvement in throughput of 8.08–229.96× speedup and 87.61–99.56% speed-up(percentage improvement) for normal traffic, and 11.73–39.59× speedup and 91.47–97.47% speed-up(percentage improvement) for traffic with high-frequency match hits. Compared with the state-of-the-art similar implementation, our circuit on a single FPGA chip is superior to existing multi-core designs.

Fast sorting of large-scale data is an essential task for data centers. In previous works, the existing computational model of sorting kernel still results in lower bandwidth utilization on the external memory bus. And the execution of merge operations in merge sort circuit on FPGAs depends on control commands from the host CPU. In this case, the merge sort circuit is not fully offloaded to hardware layer for acceleration, resulting in a performance loss. We design an on-chip merge sort controller to efficiently command the merge sort process. The proposed controller has the ability to schedule multiple on-chip computing kernels simultaneously in a more efficient mode, thus ensuring that the circuit has a better bandwidth utilization. Meanwhile, fundamental factors affecting the performance of merge sort are studied and analyzed, and we propose a high-performance merge sort architecture. Results show that using the proposed controller-centered architecture, an overall improvement of 20%–30% in sorting throughput can be achieved. Compared with the state-of-the-art previous merge sorting implementation on FPGA, our circuit can achieve 1.22/1.46 \(\times\) speedup.