General convolution acceleration, such as Winograd and FFT, is a promising direction to address the computational complexity of current convolutional neural networks (CNNs). However, the flexibility of these CNNs makes this kind of scheme always introduce massive redundant computations, damaging the acceleration effect. In this article, a two-stage splitting method for arbitrarily sized tensors and filters and a unified hardware architecture using layer-adaptive allocated Winograd units are proposed, achieving effective redundance elimination and unified architecture. First, a tensor adaptive presplitting method is proposed to divide the original tensors to match the rule of Winograd. Furthermore, a Winograd-based extended splitting scheme is designed to reduce the redundant calculations; therefore, a substantial reduction in multiplication operations in convolutional layers achieved 30.6%–75% savings. Finally, a unified hardware architecture with a layer-adaptive allocation method is proposed to evaluate and select the optimal Winograd F(${m}$,${r}$) units and input/output parallelisms. This architecture is evaluated based on the Xilinx XCVU9P platform and achieves 1.97/1.23/1.60/1.25 GOPS/DSP for AlexNet, VGG16, modified VGG16, and ResNet18, respectively. It achieves up to$5.81\times $improvements in DSP efficiency compared with previous FPGA-based designs.

Neural network pruning, which can be divided into unstructured pruning and structured pruning strategies, has been proven to be an efficient method to substantially reduce the number of computations of convolutional neural networks (CNNs). However, it remains difficult to combine the advantages of these two pruning strategies. This article proposes a high-performance accelerator for unstructured sparse CNNs. First, a convolution-based filter selection and clustering method (FSCM) is proposed to reorder unstructured sparse filters into uniform-size dense filters, eliminating redundant computations in the convolutional layers while maintaining a regular structure. In addition, a convolution and pooling calculation method (CPCM) is presented to reduce interlayers’ computational redundancy. Third, a hardware accelerator with high digital signal processing (DSP) efficiency is designed to take advantage of FSCM and CPCM. The proposed accelerator is implemented on a Xilinx XCVU9P platform at 300 MHz. With different CNN configurations, the performance and DSP efficiency for LeNet, AlexNet, and VGG16 are 1490.53 GOPS and 3.882, 913.85 GOPS and 1.785, and 862.16 GOPS and 1.68, respectively. Compared to previous field-programmable gate array (FPGA) and application-specific integrated circuit (ASIC) designs, the accelerator achieves up to$15.43\times $and$7.73\times $speedup in terms of performance and DSP efficiency.