摘要：

近年来, 随着人工智能(AI)技术的快速发展, 神经网络以较强的建模性能和优秀的多场景适应能力被广泛应用于各人工智能领域。然而, 随着逻辑和算法复杂度的增加, 对平台的算力需求日益增加, 这一问题在资源受限的嵌入式异构平台尤为突出, 已成为研究热点。深入分析国内外并行计算引擎和软件定义架构的发展规律, 探索软件定义架构的发展特点, 提出一种基于软件定义的并行计算架构(SDPCA)方案。该方案采用"数据-控制-能力"模块化设计结构, 并综合运用软件定义资源、动态调度策略、虚拟化内存、算子优化等关键技术, 有效提升SDPCA运行的计算效率, 并具备轻量化和动态重构特征。在嵌入式异构平台典型硬件条件下, 使用神经网络卷积计算和模型推理对该计算架构进行测试, 实验结果表明, SDPCA功能稳定、运行准确, 卷积计算性能较OpenCL提升4倍以上, 在4种常见神经网络模型的图像推理测试中, SDPCA识别效率较OpenCL提升2.6倍以上, 在不同模型计算时抖动更小。在识别准确率方面, SDPCA对4种模型的识别准确率均在91.7%以上, 推理平均准确度较OpenCL提升3.9百分点, 功能和性能指标符合设计预期, 具有一定的参考价值。

关键词: 神经网络, 嵌入式异构平台, 软件定义架构, 分布式计算, 算子优化

Abstract:

In recent years, with the rapid development of Artificial Intelligence (AI), neural networks have been widely used in various artificial intelligence fields. However, the demand for computing power is increasing with the complexity of logic and algorithms. This issue is particularly prominent in resource-constrained embedded heterogeneous platforms. Therefore, it has become a research hotspot. This paper analyzes the development of parallel computing engines and software-defined architecture globally, explores the characteristics of software-defined architecture, and proposes a design scheme for a Parallel Computing Architecture based on Software Definition (SDPCA). This computing architecture adopts a modular structure of ″data-control-capability″ and comprehensively utilizes key technologies of parallel computing architecture such as software-defined resources, dynamic scheduling strategies, virtualized memory, and operator optimization. These technologies effectively improve the runtime computational efficiency of SDPCA, providing a lightweight framework with dynamic reconstruction features. Under the typical hardware conditions of an embedded heterogeneous platform, SDPCA is tested using neural network convolutional layer computation and model inference. The experimental results show that the SDPCA function is stable and runs accurately, and that the computational performance of the convolutional layer is more than four times higher than that of OpenCL. The image inference tests of the four common neural network models reveal that the recognition efficiency of SDPCA increases by more than 2.6 times compared to OpenCL and it is more stable in different model calculations. Additionally, SDPCA achieves a recognition accuracy of over 91.7% for all four models, with an average inference accuracy improvement of 3.9 percentage points compared with OpenCL. The functional and performance indicators meet the design expectations and have certain reference values.

Key words: neural network, embedded heterogeneous platform, software defined architecture, distributed computing, operator optimization