The use of multi-output look-up tables (LUTs) is a widely adopted approach in contemporary commercial field-programmable gate arrays (FPGAs). Larger LUT configurations (e.g., six-input LUTs) can be partitioned into smaller LUTs (e.g., two five-input LUTs, maintaining a total input count of less than six). This capability of generating a second output from a larger LUT is not only crucial for reducing logic cell count and enhancing the utilization efficiency of logic resources—thus conserving area—but also plays a key role in optimizing system-level delays and energy consumption. In this paper, we propose an efficient multi-output LUT mapping technique, incorporating several highly efficient technology mapping algorithms, which focus on optimizing the mapping from an interconnection perspective as alternatives to directly merging smaller LUTs. These algorithms include a side-fanout insertion algorithm, and a runtime multi-output cut generation algorithm. The proposed methods improve mapping efficiency and enhance performance. The benchmarking results demonstrate that the dual-output mapping algorithms achieve LUT area reductions of up to 35% and 6%, compared to the state-of-the-art ABC six-input, single-output LUT mapping technique and previous work focusing on dual-output LUT mapping techniques that optimize cut generation parameters. Moreover, FPGA system-level simulations also show that area, delay, and energy can all be optimized based on this multi-output mapping technique.
Vision transformers (ViTs) have emerged as one of the most popular computer vision models, achieving remarkable performance in image recognition. However, ViTs require large-scale, high-dimensional matrix computations, and traditional digital accelerators, such as graphics processing units (GPUs), have memory bandwidth limitations, leading to higher latency, increased energy consumption, and larger area. To address this challenge, this paper proposes a memristor-based analog accelerator that leverages memristor crossbar arrays for in-memory computing, reducing data movement and improving computational efficiency. Considering the non-ideal characteristics of memristor devices and the influence of analog circuitry, we incorporate Gaussian-distributed analog computation error at each step and memristor non-ideality modeling into the ViT inference to enable realistic evaluation under hardware-level conditions. Experimental evaluation on ImageNet-1k dataset with TIMM-pretrained ViT models shows that the proposed analog accelerator can achieve the same Top-1 accuracy as a custom-designed 5 nm digital baseline accelerator, even with ~35% analog computation error and ~10% memristor conductance variation injected at each step. Compared to the digital counterpart, the proposed design achieves an 11.9× reduction in energy-delay product (EDP) and a 137.2× reduction in energy-delay-area product (EDAP).